Datasets:

nickypro

/

minipile-split

like 0

--- abstract: | ${\ensuremath{\mathcal{PT}}}$-symmetric quantum mechanics began with a study of the Hamiltonian $H=p^2+ x^2(ix)^{\varepsilon}$. A surprising feature of this non-Hermitian Hamiltonian is that its eigenvalues are discrete, real, and positive when ${\varepsilon}\geq0$. This paper examines the corresponding quantum-field-theoretic Hamiltonian $H={\mbox{$\textstyle{\frac{1}{2}}$}}(\nabla \phi)^2+{\mbox{$\textstyle{\frac{1}{2}}$}}\phi^2(i\phi)^{\varepsilon}$ in $D$-dimensional spacetime, where $\phi$ is a pseudoscalar field. It is shown how to calculate the Green’s functions as series in powers of ${\varepsilon}$ directly from the Euclidean partition function. Exact finite expressions for the vacuum energy density, all of the connected $n$-point Green’s functions, and the renormalized mass to order ${\varepsilon}$ are derived for $0\leq D<2$. For $D\geq2$ the one-point Green’s function and the renormalized mass are divergent, but perturbative renormalization can be performed. The remarkable spectral properties of ${\ensuremath{\mathcal{PT}}}$-symmetric quantum mechanics appear to persist in ${\ensuremath{\mathcal{PT}}}$-symmetric quantum field theory. author: - 'Carl M. Bender$^{a,c}$' - Nima Hassanpour$^a$ - 'S. P. Klevansky$^b$' - Sarben Sarkar$^c$ title: '${\ensuremath{\mathcal{PT}}}$-symmetric quantum field theory in $D$ dimensions' --- =1 Introduction\[s1\] ================== The study of ${\ensuremath{\mathcal{PT}}}$-symmetric quantum theory may be traced back to a series of papers that proposed a new perturbative approach to scalar quantum field theory. Instead of a conventional perturbation expansion in powers of a coupling constant, it was proposed that a parameter $\delta$ that measures the nonlinearity of the theory could be used as a perturbation parameter [@r1; @r2]. Thus, to solve a $g\phi^4$ field theory we studied a $g\phi^2( \phi^2)^\delta$ theory and treated the parameter $\delta$ as small. The procedure was to obtain a perturbation expansion in powers of $\delta$ and then to set $\delta=1$ to obtain the results for the $g\phi^4$ theory. Detailed investigation showed that this perturbative calculation is numerically accurate and does not require the coupling constant $g$ to be small [@r1; @r2]. An important feature of this approach was that $\phi^2$ and not $\phi$ had to be raised to the power $\delta$ in order to avoid raising a negative quantity to a noninteger power, thereby generating complex numbers as an artifact of the procedure. This $\delta$ expansion was also used to solve nonlinear classical differential equations of physics [@r3]: the Thomas-Fermi equation (nuclear charge density) $y''(x)=[y(x)]^{3/2}/\sqrt{x}$ becomes $y''(x)=y(x)[y(x)/x]^\delta$; the Lane-Emdon equation (stellar structure) $y''(x)+2y'(x)/x+[y(x)]^n=0$ becomes $y''(x)+2y'(x)/x+[y(x)]^{1+\delta}$; the Blasius equation (fluid dynamics) $y''' (x)+y''(x)y(x)=0$ becomes $y'''(x)+y''(x)[y(x)]^\delta=0$; the Korteweg-de Vries equation (nonlinear waves) $u_t+uu_x+u_{xxx}=0$ becomes $u_t+u^\delta u_x+u_{xxx }=0$. In each case the quantity raised to the power $\delta$ is positive and when $\delta=0$ the equation becomes linear. Also, these $\delta$ expansions have a nonzero radius of convergence and are numerically accurate. ${\ensuremath{\mathcal{PT}}}$-symmetric quantum mechanics began with the surprising discovery that spurious complex numbers do not appear if the quantity raised to the power $\delta$ is ${\ensuremath{\mathcal{PT}}}$ symmetric (invariant under combined space and time reflection) [@r4; @r5]. This fact is highly nontrivial and was totally unexpected. Indeed, the eigenvalues of the non-Hermitian ${\ensuremath{\mathcal{PT}}}$-symmetric Hamiltonian $$H=p^2+x^2(ix)^{\varepsilon}\quad({\varepsilon}\geq0) \label{e1}$$ are entirely real, positive, and discrete when ${\varepsilon}\geq0$ because $ix$ is ${\ensuremath{\mathcal{PT}}}$ invariant. A proof that the spectrum is real when ${\varepsilon}>0$ was given by Dorey, Dunning, and Tateo [@r6; @r7]. Numerous ${\ensuremath{\mathcal{PT}}}$-symmetric model Hamiltonians have been studied at a theoretical level [@r8] and many laboratory experiments have been performed on ${\ensuremath{\mathcal{PT}}}$-symmetric physical systems [@r9; @r10; @r11; @r12; @r13; @r14; @r15; @r16; @r17; @r18; @r19; @r20; @r21]. The purpose of this paper is to introduce powerful new tools and techniques that can be used to investigate ${\ensuremath{\mathcal{PT}}}$-symmetric quantum field theories. We illustrate these tools by studying the quantum-field-theoretic analog of (\[e1\]) whose $D$-dimensional Euclidean-space Lagrangian density is $${\ensuremath{\mathcal{L}}}={\mbox{$\textstyle{\frac{1}{2}}$}}(\nabla\phi)^2+{\mbox{$\textstyle{\frac{1}{2}}$}}\phi^2(i\phi)^{\varepsilon}\quad({\varepsilon}\geq0), \label{e2}$$ where $\phi$ is a pseudoscalar field so that ${\ensuremath{\mathcal{L}}}$ is ${\ensuremath{\mathcal{PT}}}$ invariant. We treat ${\varepsilon}$ as small and show how to calculate the vacuum energy density $E_0$, the connected $n$-point Green’s functions $G_n$, and the renormalized mass $M_{\rm R}$ as series in powers of ${\varepsilon}$. In this paper we assume that $0\leq D<2$ to avoid the appearance of renormalization infinities and then we comment briefly on the perturbative renormalization procedure for the case $D\geq2$. To first order in ${\varepsilon}$ (${\varepsilon}<<1$), the unusual Lagrangian density ${\ensuremath{\mathcal{L}}}$ in (\[e1\]) has a [*logarithmic*]{} self-interaction term: $${\ensuremath{\mathcal{L}}}={\mbox{$\textstyle{\frac{1}{2}}$}}(\nabla\phi)^2+{\mbox{$\textstyle{\frac{1}{2}}$}}\phi^2+{\mbox{$\textstyle{\frac{1}{2}}$}}{\varepsilon}\phi^2\log(i\phi)+{\rm O}({\varepsilon}^2). \label{e3}$$ For a quantum field theory having a complex logarithmic interaction term it is not obvious whether one can find Feynman rules for performing perturbative diagramatic calculations. We will show how to construct such Feynman rules. We begin by replacing the complex logarithm with a real logarithm and we do so in such a way as to preserve ${\ensuremath{\mathcal{PT}}}$ symmetry; to wit, we define $$\log(i\phi)\equiv{\mbox{$\textstyle{\frac{1}{2}}$}}i\pi+\log(\phi)\quad({\rm if}~\phi>0)$$ and we define $$\log(i\phi)\equiv-{\mbox{$\textstyle{\frac{1}{2}}$}}i\pi+\log(-\phi)\quad({\rm if}~\phi<0).$$ Combining these two equations, we make the replacement $$\log(i\phi)={\mbox{$\textstyle{\frac{1}{2}}$}}i\pi\,|\phi|/\phi+{\mbox{$\textstyle{\frac{1}{2}}$}}\log(\phi^2). \label{e4}$$ Note that in (\[e4\]) the imaginary part is odd in $\phi$ and the real part is even in $\phi$. Thus, (\[e4\]) enforces the ${\ensuremath{\mathcal{PT}}}$ symmetry because the pseudoscalar field $\phi$ changes sign under parity ${\ensuremath{\mathcal{P}}}$ and $i$ changes sign under time reversal ${\ensuremath{\mathcal{T}}}$. \[To derive (\[e4\]) we must assume that $\phi$ is real. The reality of $\phi$ is explained in Sec. \[s2\].\] Graphical techniques were developed in Ref. [@r1] to handle real logarithmic interaction terms. These techniques are generalizations of the replica trick [@r22], which has been used in the study of spin glasses. The idea of the replica trick is that a logarithmic term $\log A$ can be reformulated as the limit $\log A=\lim_{N\to0}\textstyle{\frac{1}{N}}\big(A^N-1\big)$, or equivalently and slightly more simply as the limit $$\log A=\lim_{N\to0}\textstyle{\frac{d}{dN}}A^N. \label{e5}$$ One then regards $N$ as an integer and identifies $A^N$ as an $N$-point vertex in a graphical expansion. Of course, this procedure is not rigorous because it requires taking the [*continuous*]{} limit $N\to0$. The validity of this approach has not been proved, but when it is possible to compare with exactly known results in low-dimensional theories, the replica trick gives the correct answer. In this paper we verify our field-theoretic results by comparing them with the exact answers for $D=0$ (where the functional integral becomes an ordinary integral) and for $D=1$ (quantum mechanics). The graphical calculations in this paper are done in coordinate space. Once the vertices have been identified, all that one needs is the free propagator in $D$-dimensional Euclidean space $\Delta(x-y)$, which satisfies the differential equation $$\big(-\nabla_x^2+1\big)\Delta(x-y)=\delta^{(D)}(x-y). \label{e6}$$ Taking the Fourier transform of this equation gives the amplitude for the free propagator of a particle of mass $1$ in momentum space: $$\tilde\Delta(p)=1/(p^2+1).$$ The $D$-dimensional inverse Fourier transform of this expression then gives the $D$-dimensional coordinate-space propagator in terms of an associated Bessel function: $$\Delta(x_1-x_2)=(2\pi)^{-\frac{D}{2}}|x_1-x_2|^{1-\frac{D}{2}}{\rm K}_{1- \frac{D}{2}}(|x_1-x_2|). \label{e7}$$ If we let $x_1\to x_2$, we obtain the amplitude $\Delta(0)$ for a [*self loop*]{}, which is the amplitude for a line to originate from and return to the same point: $$\Delta(0)=(4\pi)^{-D/2}\Gamma(1-D/2). \label{e8}$$ This expression is finite and nonsingular for $0\leq D<2$. This paper is organized as follows. We calculate the ground-state energy density $E_0$ in Sec. \[s2\], the one-point Green’s function $G_1$ in Sec. \[s3\], the two-point Green’s function $G_2$ and the renormalized mass $M_{\rm R}$ in Sec. \[s4\], and the general connected $n$-point Green’s function $G_n$ in Sec. \[s5\], all to first order in ${\varepsilon}$. These quantities are finite when $0 \leq D<2$, but the three quantities $E_0$, $G_1$, and $M_{\rm R}$ diverge when $D\geq2$ so it is necessary to introduce a renormalization procedure. In Sec. \[s6\] we discuss the issues of renormalization. We show that a redefinition of the energy scale, an additive shift in the field, and a mass counterterm eliminate these infinities. In this section we also discuss our future calculational objectives, namely, calculating the Green’s functions to higher order in ${\varepsilon}$. First-order calculation of the ground-state energy density\[s2\] ================================================================ If we expand the partition function $$Z({\varepsilon})=\int\!{\ensuremath{\mathcal{D}}}\phi\,e^{-\int\! d^D\!x\,{\ensuremath{\mathcal{L}}}} \label{e9}$$ for the Lagrangian density ${\ensuremath{\mathcal{L}}}$ to first order in ${\varepsilon}$ and use (\[e4\]), the functional integral (\[e9\]) becomes $$\begin{aligned} Z({\varepsilon}) &=& \int\!{\ensuremath{\mathcal{D}}}\phi\,e^{-\int\! d^D\!x\,{\ensuremath{\mathcal{L}}}_0} \bigg(1-\frac{{\varepsilon}}{4}\int\! d^Dy\,\big\{i\pi\phi(y)|\phi(y)\nonumber\\ &&\qquad+\phi^2(y)\log\big[\phi^2(y)\big]\big\} +{\rm O}({\varepsilon}^2)\bigg), \label{e10}\end{aligned}$$ where the free Lagrangian density $${\ensuremath{\mathcal{L}}}_0={\mbox{$\textstyle{\frac{1}{2}}$}}(\nabla\phi)^2+{\mbox{$\textstyle{\frac{1}{2}}$}}\phi^2 \label{e11}$$ is obtained by setting ${\varepsilon}=0$ in ${\ensuremath{\mathcal{L}}}$. Note that the imaginary part of the functional integrand in (\[e10\]) is odd in $\phi$, so $Z({\varepsilon})$ is [*real*]{}. We emphasize that the functional integration in (\[e10\]) is performed along the [*real*]{}-$\phi$ axis and not in the complex-$\phi$ domain. This justifies the use of (\[e4\]). We are not concerned here with complex functional integration paths that terminate in complex Stokes sectors because the functional integral (\[e10\]) converges term-by-term in powers of ${\varepsilon}$. In ${\ensuremath{\mathcal{PT}}}$-symmetric quantum mechanics the boundary conditions on the Schrödinger equation associated with the Hamiltonian (\[e1\]) \[see (\[a2\])\] are imposed in complex Stokes sectors [@r5]. However, in the context of quantum field theory it would be hopelessly unwieldy to consider [*functional*]{} Stokes sectors. This is why we treat ${\varepsilon}$ as small. This paper is concerned with calculating the coefficients in the ${\varepsilon}$ series and we do not consider here the mathematical issues involved with the summation of such a series for large values of ${\varepsilon}$. In general, a partition function is the exponential of the ground-state energy density $E_0$ multiplied by the volume of spacetime $V$: $Z=e^{-E_0V}$. Thus, the [*shift*]{} in the ground-state energy density $\Delta E$ to order ${\varepsilon}$ is given by $$\Delta E=\frac{{\varepsilon}}{4Z(0)V}\int\!{\ensuremath{\mathcal{D}}}\phi\,e^{-\int d^D\!x\,{\ensuremath{\mathcal{L}}}_0} \!\int\!d^D\!y\,\phi^2(y)\log\big[\phi^2(y)\big].$$ Hence, from (\[e5\]) we obtain $$\Delta E=\lim_{N\to1}\frac{{\varepsilon}}{4Z(0)V}\,\frac{d}{dN} \int\!{\ensuremath{\mathcal{D}}}\phi\,e^{-\int d^D\!x\,{\ensuremath{\mathcal{L}}}_0}\!\int\!d^D y\,\phi^{2N}(y).$$ This expression has a graphical interpretation as the product of $N$ self loops from the spacetime point $y$ back to $y$. There are exactly $(2N-1)!!$ ways to construct these self loops, so the expression for $\Delta E$ simplifies to $$\Delta E=\lim_{N\to1}\frac{{\varepsilon}}{4V} \,\frac{d}{dN}\int\!d^D\!y\,[\Delta(0)]^N (2N-1)!!.$$ Next, we note that the integral $\int d^Dy$ is the volume of spacetime $V$, so this formula simplifies further: $$\Delta E=\lim_{N\to1}\frac{{\varepsilon}}{4}\,\frac{d}{dN}[\Delta(0)]^N(2N-1)!!.$$ Finally, we use the duplication formula for the gamma function [@r23] to write $$(2N-1)!!=2^N\Gamma\big(N+{\mbox{$\textstyle{\frac{1}{2}}$}}\big)/\sqrt{\pi}$$ and then take the derivative with respect to $N$ to get $$\Delta E={\mbox{$\textstyle{\frac{1}{4}}$}}{\varepsilon}\Delta(0)\left\{\log[2\Delta(0)]+ \Gamma'\big({\mbox{$\textstyle{\frac{3}{2}}$}}\big)\big/\Gamma\big({\mbox{$\textstyle{\frac{3}{2}}$}}\big)\right\}, \label{e12}$$ where $\Gamma'\big({\mbox{$\textstyle{\frac{3}{2}}$}}\big)\big/\Gamma\big({\mbox{$\textstyle{\frac{3}{2}}$}}\big)=2-\gamma-2 \log2$. The result in (\[e12\]) may be verified for the special cases of $D=0$ and $D=1$. [*Special case*]{} $D=0$: For $D=0$ the normalized partition function becomes an ordinary integral, which we can expand to first order in ${\varepsilon}$: $$\begin{aligned} Z&=& \frac{1}{Z_0}\int_{-\infty}^\infty\!d\phi\,e^{-{\mbox{$\textstyle{\frac{1}{2}}$}}\phi^2(i\phi)^{\varepsilon}} \nonumber\\ &\sim& \frac{1}{Z_0}\int_{-\infty}^\infty\!d\phi\,e^{-{\mbox{$\textstyle{\frac{1}{2}}$}}\phi^2}\big[1-{\mbox{$\textstyle{\frac{1}{2}}$}}{\varepsilon}\phi^2\log(i\phi)\big], \label{e13}\end{aligned}$$ where $Z_0=\sqrt{2\pi}$. If we take the negative logarithm of this result, we obtain the first-order shift in the ground-state energy density $$\Delta E=\frac{{\varepsilon}}{2\sqrt{2\pi}}\int_{-\infty}^\infty\!d\phi\, e^{-{\mbox{$\textstyle{\frac{1}{2}}$}}\phi^2}\phi^2\log(i\phi).$$ We then integrate separately from $-\infty$ to $0$ and from $0$ to $\infty$ and combine the two integrals to obtain a single real integral that we evaluate as follows: $$\begin{aligned} \Delta E&=&\frac{{\varepsilon}}{\sqrt{2\pi}}\int_0^\infty\!d\phi\,e^{-{\mbox{$\textstyle{\frac{1}{2}}$}}\phi^2}\phi^2 \log\phi\nonumber\\ &=&\textstyle{\frac{{\varepsilon}}{4}}\big[\Gamma'\big({\mbox{$\textstyle{\frac{3}{2}}$}}\big) \big/\Gamma\big({\mbox{$\textstyle{\frac{3}{2}}$}}\big)+\log2\big]. \label{e14}\end{aligned}$$ Taking $D=0$ in (\[e8\]) gives $\Delta(0)=1$, so (\[e12\]) reduces exactly to (\[e14\]) in zero-dimensional spacetime. [*Special case*]{} $D=1$: In quantum mechanics, $\Delta E$ to leading order in ${\varepsilon}$ is the expectation value of the interaction Hamiltonian $H_I={\mbox{$\textstyle{\frac{1}{2}}$}}{\varepsilon}x^2\log(ix)$ \[see (\[e3\])\] in the unperturbed ground-state eigenfunction $\psi_0(x)=\exp\big(-{\mbox{$\textstyle{\frac{1}{2}}$}}x^2\big)$: $$\begin{aligned} \Delta E&=&\frac{{\varepsilon}}{2}\int_{-\infty}^\infty\! dx\,e^{-x^2}x^2\log(ix)\Big/ \!\int_{-\infty}^\infty \!dx\,e^{-x^2}\nonumber\\ &=&{\mbox{$\textstyle{\frac{1}{8}}$}}{\varepsilon}\Gamma'\big({\mbox{$\textstyle{\frac{3}{2}}$}}\big)\big/\Gamma\big({\mbox{$\textstyle{\frac{3}{2}}$}}\big). \label{e15}\end{aligned}$$ Taking $D=1$ in (\[e8\]) gives $\Delta(0)={\mbox{$\textstyle{\frac{1}{2}}$}}$, so the general result in (\[e12\]) reduces exactly to (\[e15\]) in one-dimensional spacetime. Note that in ${\ensuremath{\mathcal{PT}}}$-symmetric quantum mechanics the calculation of expectation values requires the ${\ensuremath{\mathcal{C}}}$ operator [@r8]. However, the ${\ensuremath{\mathcal{C}}}$ operator is not needed for any of the calculations in this paper because a functional integral involves [*vacuum*]{} expectation values. The vacuum state is an eigenstate of the ${\ensuremath{\mathcal{C}}}$ operator with eigenvalue $1$, ${\ensuremath{\mathcal{C}}}|0\rangle=|0\rangle$, so all reference to ${\ensuremath{\mathcal{C}}}$ disappears. This simplification was first pointed out in Ref. [@r24]. First-order calculation of the one-point Green’s function\[s3\] =============================================================== The one-point Green’s function $G_1$ is [*nonperturbative*]{} in character but it can be calculated by following the approach used above to calculate $\Delta E$. Keeping terms that do not vanish under $\phi\to-\phi$, we evaluate directly the functional-integral representation $$\begin{aligned} G_1(a)&=&-\frac{{\varepsilon}}{4Z(0)}\int\!{\ensuremath{\mathcal{D}}}\phi\,\phi(a)e^{-\int d^D\!x\,{\ensuremath{\mathcal{L}}}_0} \nonumber\\ &&\quad\times\int\!d^D\!y\,i\pi\phi(y)|\phi(y)|, \label{e16}\end{aligned}$$ where ${\ensuremath{\mathcal{L}}}_0$ is the free Euclidean Lagrangian in (\[e11\]). We use the integral identity $$\phi|\phi|=\frac{2}{\pi}\phi^2\int_0^\infty\frac{dt}{t}\sin(t\phi) \label{e17}$$ to replace $\phi(y)|\phi(y)|$ in the functional integral (\[e16\]) and then we replace $\sin(t\phi)$ by its Taylor series: $$\sin(t\phi)=\sum_{n=0}^\infty\frac{(-1)^nt^{2n+1}}{(2n+1)!}\phi^{2n+1}. \label{e18}$$ This converts (\[e16\]) to the product of an infinite sum in $n$, a one-dimensional integral in $t$, a $D$-dimensional integral in $y$, and a functional integral in $\phi$: $$\begin{aligned} G_1(a)&=&-\frac{i{\varepsilon}}{2}\int_{t=0}^\infty dt\,\sum_{n=0}^\infty\frac{(-1)^n t^{2n}}{(2n+1)!} \int d^D\!y\nonumber\\ &&\!\!\!\!\!\!\!\!\!\!\!\!\times\frac{1}{Z(0)}\int{\ensuremath{\mathcal{D}}}\phi\,e^{-\int d^D\!x\,{\ensuremath{\mathcal{L}}}_0}\phi(a)[\phi(y)]^{2n+3}. \label{e19}\end{aligned}$$ The sum and multiple integrals in (\[e19\]) may appear to be difficult, but like the calculation of the ground-state energy density, the functional integral in the second line of (\[e19\]) also has a graphical interpretation; it is merely the product of the free propagator $\Delta(a-y)$ representing a line from $a$ to $y$ multiplied by $n+1$ self loops from $y$ to $y$, and this product is accompanied by the combinatorial factor $(2n+3)!!$. Thus, the second line in (\[e19\]) reduces to $(2n+3)!!\Delta(y-a)\Delta^{n+1}(0)$. This result simplifies further because, as we can see from (\[e6\]), the $D$-dimensional integral is trivial: $\int d^Dy\Delta(y-a)=1$. This establishes the translation invariance of $G_1$. The rest is straightforward: $$\begin{aligned} G_1&=&-\frac{i{\varepsilon}}{2}\int_{t=0}^\infty dt\,\sum_{n=0}^\infty \frac{(-1)^n t^{2n}\Delta^{n+1} (0)(2n+3)!!}{(2n+1)!}\nonumber\\ &=&-\frac{i{\varepsilon}}{2}\int_{t=0}^\infty dt\,\Delta(0)\big[3-\Delta(0)t^2\big] e^{-{\mbox{$\textstyle{\frac{1}{2}}$}}\Delta(0)t^2} \nonumber\\ &=&-i{\varepsilon}\sqrt{\pi\Delta(0)/2}. \label{e20}\end{aligned}$$ This expression for $G_1$ is exact to order ${\varepsilon}$. Observe that the expression for $G_1$ is a [*negative imaginary*]{} number. This is precisely what we would expect based on previous studies of classical ${\ensuremath{\mathcal{PT}}}$-symmetric systems. The classical trajectories in complex coordinate space of a particle described by the Hamiltonian (\[e1\]) are left-right symmetric but they lie mostly in the lower-half plane [@r8]. Thus, the average value of the classical orbits is a negative-imaginary number. [*Special case*]{} $D=0$: The one-point Green’s function in $D=0$ is given by the second line in (\[e13\]) with an additional extra factor of $\phi$: $$G_1\sim\frac{1}{Z(0)}\int_{-\infty}^\infty d\phi\,\phi e^{-{\mbox{$\textstyle{\frac{1}{2}}$}}\phi^2} \big[1-{\mbox{$\textstyle{\frac{1}{2}}$}}{\varepsilon}\phi^2\log(i\phi)\big],$$ where $Z(0)=\sqrt{2\pi}$. The integration over the first term in the square brackets vanishes by oddness. We evaluate the contribution of the second term by integrating first from $-\infty$ to $0$ and then from $0$ to $\infty$. Combining these two integrals, we obtain $$G_1= -i{\varepsilon}\sqrt{\pi/2}.$$ This result is in exact agreement with the general result in (\[e20\]) because $\Delta(0)=1$ when $D=0$. [*Special case*]{} $D=1$: When $D=1$, the expression for $G_1$ in (\[e20\]) reduces to $$G_1=-{\mbox{$\textstyle{\frac{1}{2}}$}}i{\varepsilon}\sqrt{\pi}. \label{e21}$$ In the Appendix we derive $G_1$ in quantum mechanics and verify (\[e21\]). First-order calculation of the two-point Green’s function\[s4\] =============================================================== To obtain the [*connected*]{} two-point Green’s function $G_2(a,b)$, one must subtract $G_1^2$ from the vacuum expectation value of $\phi(a)\phi(b)$. However, we have seen that $G_1$ is of order ${\varepsilon}$. Therefore, to first order in ${\varepsilon}$ we need only evaluate $Z({\varepsilon})$ in (\[e10\]) with $\phi(a)\phi(b)$ inserted after ${\ensuremath{\mathcal{D}}}\phi$ and then divide this integral by $Z({\varepsilon})$. We may neglect the imaginary terms in these integrals because they are odd under $\phi\to-\phi$. Thus, the expression that we must evaluate for $G_2(a,b)$ is $$\frac{\int{\ensuremath{\mathcal{D}}}\phi\,\phi(a)\phi(b)e^{-\int d^D\!x\,{\ensuremath{\mathcal{L}}}_0}\left\{1-\frac{{\varepsilon}} {4}\int d^D\!y\,\phi^2(y)\log\big[\phi^2(y)\big]\right\}}{\int{\ensuremath{\mathcal{D}}}\phi\,e^{-\int d^D\!x\,{\ensuremath{\mathcal{L}}}_0}\left\{1-\frac{{\varepsilon}}{4}\int d^D\!y\,\phi^2(y)\log\big[\phi^2(y) \big]\right\}}.$$ We expand this expression to first order in ${\varepsilon}$ as a sum of three functional integrals: $$G_2(a,b)=A+B+C, \label{e22}$$ where $$\begin{aligned} \!\!\!\!\!\!\!\!A&=&\frac{1}{Z(0)}\int\!{\ensuremath{\mathcal{D}}}\phi\,\phi(a)\phi(b)e^{-\int d^D\!x\, {\ensuremath{\mathcal{L}}}_0},\nonumber\\ \!\!\!\!\!\!\!\!B&=& -\frac{{\varepsilon}}{4Z(0)}\int\!{\ensuremath{\mathcal{D}}}\phi\,\phi(a)\phi(b)e^{-\int d^D\!x\,{\ensuremath{\mathcal{L}}}_0}\nonumber\\ &&\!\!\!\!\!\!\!\!\!\!\!\!\times\int\!d^D\!y\,\phi^2(y)\log\big[\phi^2(y) \big],\nonumber\\ \!\!\!\!\!\!\!\!C&=& \frac{{\varepsilon}}{4Z(0)}\int\!{\ensuremath{\mathcal{D}}}\phi\,\phi(a)\phi(b)e^{-\int d^D\! x\,{\ensuremath{\mathcal{L}}}_0}\nonumber\\ &&\!\!\!\!\!\!\!\!\!\!\!\!\times\frac{1}{Z(0)}\int\!{\ensuremath{\mathcal{D}}}\phi\,e^{-\int d^D \!x\,{\ensuremath{\mathcal{L}}}_0}\!\int\!d^D\!y\,\phi^2(y)\log\big[\phi^2(y)\big]. \label{e23}\end{aligned}$$ We must now evaluate the three contributions $A$, $B$, and $C$. The functional integral $A$ in (\[e23\]) is simply the free propagator $\Delta(a-b)$. This result verifies that if we set ${\varepsilon}=0$ in the Lagrangian (\[e2\]), we obtain a free field theory; the two-point Green’s function for such a field theory is $\Delta(a-b)$. The double functional integral $C$ presents a complication. The first line of $C$ is proportional to $A$ and evaluates to ${\mbox{$\textstyle{\frac{1}{4}}$}}{\varepsilon}\Delta(a-b)$. However, as we showed in the calculation of the ground-state energy, the next two lines of $C$ evaluate to $\frac{4}{{\varepsilon}}V\Delta E$, where $\Delta E$ is the first-order shift in the ground-state energy density (\[e12\]) and $V$ is the volume of Euclidean spacetime. Thus, $C=\Delta(a-b)V\Delta E$, and this quantity is [*divergent*]{} because $V$ is infinite. We resolve this divergence problem by calculating $B$. Using (\[e5\]) we express the $B$ integral as $$\begin{aligned} B&=& -\frac{{\varepsilon}}{4Z(0)}\lim_{N\to1}\frac{d}{dN}\int{\ensuremath{\mathcal{D}}}\phi\,e^{-\int d^D\!x\, {\ensuremath{\mathcal{L}}}_0} \nonumber\\ &&\qquad\times\int d^D\!y\,\phi(a)\phi(b)\phi^{2N}(y). \label{e24}\end{aligned}$$ This functional integral requires that we connect in a pairwise fashion the set of $2N+2$ points consisting of $a$, $b$, and the $2N$ points $y$ with the free propagator $\Delta$ in (\[e7\]). There are two cases to consider. In the first case $a$ is connected to $b$ and the remaining $2N$ points at $y$ are connected in pairs. Note that this reproduces the result above for $C$ except with the opposite sign. Thus, the volume divergences exactly cancel. In the second case $a$ is not connected to $b$. Instead, $a$ connects to a point $y$ (there are $2N$ ways to do this) and $b$ connects to one of the remaining $2N-1$ points $y$ (there are $2N-1$ ways to do this). The rest of the $2N-2$ points $y$ are joined in pairs \[there are $(2N-3)!!$ ways to do this\]. The amplitude for this case is $$-\frac{{\varepsilon}}{2}\int d^D\!y\,\Delta(a-y)\Delta(y-b)\lim_{N\to1}\frac{d}{dN}N (2N-1)!!\Delta^{N-1}(0),$$ which simplifies to $-{\varepsilon}K\int d^D\!y\,\Delta(a-y)\Delta(y-b)$, where $$\begin{aligned} K&=&{\mbox{$\textstyle{\frac{1}{2}}$}}+{\mbox{$\textstyle{\frac{1}{2}}$}}\Gamma'\big({\mbox{$\textstyle{\frac{3}{2}}$}}\big)/\Gamma\big({\mbox{$\textstyle{\frac{3}{2}}$}}\big) +{\mbox{$\textstyle{\frac{1}{2}}$}}\log[2\Delta(0)]\nonumber\\ &=&{\mbox{$\textstyle{\frac{3}{2}}$}}-{\mbox{$\textstyle{\frac{1}{2}}$}}\gamma+{\mbox{$\textstyle{\frac{1}{2}}$}}\log\big[{\mbox{$\textstyle{\frac{1}{2}}$}}\Delta(0)\big]. \label{e25}\end{aligned}$$ Thus, our final result for the coordinate-space two-point Green’s function to order ${\varepsilon}$ is $$G_2(a-b)=\Delta(a-b)-{\varepsilon}K\int d^D\!y\,\Delta(a-y)\Delta(y-b). \label{e26}$$ In momentum space this is $$\tilde{G}_2(p)=\frac{1}{p^2+1}-{\varepsilon}\frac{K}{(p^2+1)^2}+{\rm O}\big({\varepsilon}^2\big). \label{e27}$$ From (\[e27\]) we construct the $(0,1)$ Padé approximant, which is just the geometric sum of a chain of bubbles: $$\tilde{G}_2(p)=\frac{1}{p^2+1+{\varepsilon}K+{\rm O}\big({\varepsilon}^2\big)}. \label{e28}$$ We then read off the square of the renormalized mass to first order in ${\varepsilon}$: $$M_{\rm R}^2=1+K{\varepsilon}+{\rm O}\big({\varepsilon}^2\big). \label{e29}$$ [*Special case*]{} $D=0$: To verify (\[e28\]) and (\[e29\]) for the case $D=0$ we evaluate the ordinary one-dimensional integrals in $$\begin{aligned} G_2&=&\frac{\int dx\,x^2e^{-x^2(ix)^{\varepsilon}/2}}{\int dx\,e^{- x^2(ix)^{\varepsilon}/2}}\nonumber\\ &=&\frac{\int_0^\infty dx\,x^2e^{-x^2/2} \left[1-\frac{{\varepsilon}}{4} x^2\log(x^2)\right]} {\int_0^\infty dx\,e^{-x^2/2}\left[1-\frac{{\varepsilon}}{4} x^2\log(x^2)\right]}.\nonumber\end{aligned}$$ These integrals are not difficult and to order ${\varepsilon}$ we get $$G_2=\frac{1}{1+{\varepsilon}\big({\mbox{$\textstyle{\frac{3}{2}}$}}-{\mbox{$\textstyle{\frac{1}{2}}$}}\gamma-{\mbox{$\textstyle{\frac{1}{2}}$}}\log2\big)}. \label{e30}$$ This result agrees exactly with that in (\[e28\]) for $D=0$. [*Special case*]{} $D=1$: We verify (\[e29\]) for the case $D=1$ by calculating the energy level of the [*first*]{} excited state of the Hamiltonian $H$ in (\[e1\]). When ${\varepsilon}=0$, $H$ becomes the harmonic-oscillator Hamiltonian, so the first excited state eigenfunction is $xe^{-x^2/2}$ and the associated eigenvalue is ${\mbox{$\textstyle{\frac{3}{2}}$}}$. We solve the time-independent Schrödinger equation $H\psi=E\psi$ perturbatively by substituting $$\begin{aligned} \psi(x)&=&xe^{-x^2/2}+{\varepsilon}\psi_1(x)+{\rm O}\big({\varepsilon}^2\big),\nonumber\\ E&=&{\mbox{$\textstyle{\frac{3}{2}}$}}+{\varepsilon}E_1+{\rm O}\big({\varepsilon}^2\big), \label{e31}\end{aligned}$$ and collecting powers of ${\varepsilon}$. To first order in ${\varepsilon}$ the function $\psi_1(x)$ satisfies $$\begin{aligned} &&-{\mbox{$\textstyle{\frac{1}{2}}$}}\psi_1''(x)+{\mbox{$\textstyle{\frac{1}{2}}$}}x^2\psi_1(x)-{\mbox{$\textstyle{\frac{3}{2}}$}}\psi_1(x)\nonumber\\ && \qquad=\left[-{\mbox{$\textstyle{\frac{1}{2}}$}}x^2\log(ix)+E_1\right]xe^{-x^2/2}.\nonumber\end{aligned}$$ To solve this equation we use the technique of reduction of order and let $\psi_1(x)=xe^{-x^2/2}f(x)$. The equation for $f(x)$ is then $$xf''(x)+\left(2-2x^2\right)f(x)=-2E_1x+x^3\log(ix).$$ Multiplying this equation by the integrating factor $x\,\exp\big(-x^2\big)$ gives $$\frac{d}{dx}\left[x^2 f'(x)e^{-x^2}\right]=\left[-2E_1x^2+x^4\log (ix)\right]e^{-x^2}.$$ Therefore, if we integrate this equation from $-\infty$ to $\infty$, we obtain an equation for $E_1$: $$E_1=\frac{1}{4}\int_{-\infty}^\infty dx\,x^4\log\big(x^2\big)e^{-x^2} \Big/\int_{-\infty}^\infty dx\,x^2e^{-x^2},$$ where we have replaced $\log(ix)$ by ${\mbox{$\textstyle{\frac{1}{2}}$}}\log\big(x^2\big)$. These integrals are easy to evaluate and we get $$E_1=\textstyle{\frac{3}{8}}\left(\textstyle{\frac{8}{3}}-\gamma-2\log2 \right).$$ Thus, the first excited eigenvalue of $H$ to order ${\varepsilon}$ is $${\mbox{$\textstyle{\frac{3}{2}}$}}+\textstyle{\frac{3}{8}}{\varepsilon}\left(\textstyle{\frac{8}{3}}-\gamma -2\log2\right). \label{e32}$$ We have already calculated the ground-state energy in Sec. \[s2\]: $${\mbox{$\textstyle{\frac{1}{2}}$}}+\textstyle{\frac{1}{8}}{\varepsilon}(2-\gamma-2\log2). \label{e33}$$ The renormalized mass $M_{\rm R}$ is the first excitation above the ground state so $M_{\rm R}$ is the difference of these two energies: $$M_{\rm R}=1+{\mbox{$\textstyle{\frac{1}{4}}$}}{\varepsilon}(3-\gamma-2\log2). \label{e34}$$ If we square this result and keep terms of order ${\varepsilon}$, we get $$M_{\rm R}^2=1+{\varepsilon}\left({\mbox{$\textstyle{\frac{3}{2}}$}}-{\mbox{$\textstyle{\frac{1}{2}}$}}\gamma-\log2\right), \label{e35}$$ which exactly reproduces (\[e29\]) for the case $D=1$. Higher-order Green’s functions\[s5\] ==================================== The connected three-point Green’s function is given by the cumulant $$\begin{aligned} G_3(x,y,z)&=&\frac{\langle\phi(x)\phi(y)\phi(z)\rangle}{Z} -\frac{\langle\phi(x)\phi(y)\rangle\langle\phi(z)\rangle}{Z^2}\nonumber\\ &&-\frac{\langle\phi(x)\phi(z)\rangle\langle\phi(y)\rangle}{Z^2} -\frac{\langle\phi(y)\phi(z)\rangle\langle\phi(x)\rangle}{Z^2}\nonumber\\ &&+2\frac{\langle\phi(x)\rangle\langle\phi(y)\rangle\langle\phi(z)\rangle}{Z^3}. \nonumber\end{aligned}$$ However, to order ${\varepsilon}$ $$\begin{aligned} G_3(x,y,z)&=&\frac{\langle\phi(x)\phi(y)\phi(z)\rangle}{Z}-\Delta(x-y)G_1\nonumber\\ &&-\Delta(x-z)G_1-\Delta(y-z)G_1.\nonumber\end{aligned}$$ The calculation of $G_3$ is somewhat tedious, but the procedure follows exactly the calculation of the two-point Green’s function. The final result after the disconnected terms have canceled is $$\begin{aligned} G_3(x,y,z)&=&-i\frac{{\varepsilon}\sqrt{\pi}}{\sqrt{2\Delta(0)}}\nonumber\\ &&\hspace{-1.0cm}\times\int\!d^D\!u\,\Delta(x-u)\Delta(y-u)\Delta(z-u). \label{e36}\end{aligned}$$ The connected four-point function is defined by the cumulant $$\begin{aligned} G_4(x,y,z,w)&=&-\frac{\langle\phi(x)\phi(y)\phi(z)\phi(w)\rangle}{Z}\nonumber\\ &&\hspace{-2.5cm}-\frac{\langle\phi(x)\phi(y)\phi(z)\rangle\langle\phi(w) \rangle}{Z^2}-(\rm three~permutations)\nonumber\\ &&\hspace{-2.5cm}-\frac{\langle\phi(x)\phi(y)\rangle\langle\phi(z)\phi(w) \rangle}{Z^2}-(\rm two~permutations)\nonumber\\ &&\hspace{-2.5cm}+2\frac{\langle\phi(x)\phi(y)\rangle\langle\phi(z)\rangle \langle\phi(w)\rangle}{Z^3} +(\rm five~permutations)\nonumber\\ &&\hspace{-2.5cm}-6\frac{\langle\phi(x)\rangle\langle\phi(y)\rangle\langle\phi (z)\rangle\langle\phi(w)\rangle}{Z^4}. \nonumber\end{aligned}$$ Again, calculating $G_4$ is tedious but the result is simply $$\begin{aligned} G_4(x,y,z,w)&=&-\frac{{\varepsilon}}{\Delta(0)}\!\int\!d^D\!u\,\Delta(x-u) \nonumber\\ &&\hspace{-1.0cm}\times\Delta(y-u)\Delta(z-u)\Delta(w-u). \label{e37}\end{aligned}$$ The pattern is now evident and with some effort we can calculate the connected $n$-point Green’s function to order ${\varepsilon}$ and obtain a general formula valid for all $n\neq2$: $$\begin{aligned} G_n(x_1,x_2,\ldots x_n)&=&-{\mbox{$\textstyle{\frac{1}{2}}$}}{\varepsilon}(-i)^n\Gamma\big({\mbox{$\textstyle{\frac{1}{2}}$}}n-1\big) \nonumber\\ &&\hspace{-2.5cm}\times\big[{\mbox{$\textstyle{\frac{1}{2}}$}}\Delta(0)\big]^{1-n/2} \int d^D\!u\,\prod_{k=1}^n\Delta\big(x_k-u\big). \label{e38}\end{aligned}$$ Thus, the connected Green’s functions are all of order ${\varepsilon}$ except for $G_2$ in (\[e26\]), which is of order 1. Observe that (\[e38\]) reduces to (\[e20\]), (\[e36\]), and (\[e37\]) for $n=1$, 3, and 4. We emphasize that (\[e38\]) holds for both odd and even $n$. Our calculation of the odd-$n$ Green’s functions uses the [*first*]{} term on the right side of (\[e4\]) and our calculation proceeds by introducing the integral representation in (\[e17\]) followed by using the Taylor series in (\[e18\]). On the other hand, our calculation of the even-$n$ Green’s functions uses the [*second*]{} term on the right side of (\[e4\]) and the calculation proceeds by applying the derivative identity in (\[e5\]). It is satisfying that these two strikingly different techniques lead to the single universal formula in (\[e38\]) for $G_n$. Discussion and Future Work\[s6\] ================================ The principal advance reported in this paper is that we have developed all of the machinery necessary to calculate the Green’s functions of a ${\ensuremath{\mathcal{PT}}}$-symmetric quantum field theory in (\[e2\]) to any order in ${\varepsilon}$. Thus, this paper opens a vast area for future study and investigation; one can investigate the masses of the theory (the poles of the Green’s functions), scattering amplitudes, critical indices, and so on. The Green’s-function calculations done in Secs. \[s2\]-\[s5\] are exact to first order in ${\varepsilon}$. However, the procedures presented here immediately generalize to all orders in ${\varepsilon}$. Furthermore, as we show below, even in low orders the perturbation series in powers of ${\varepsilon}$ is highly accurate and it continues to be accurate for large ${\varepsilon}$. To illustrate, we calculate the one-point Green’s function $G_1$ in $D= 0$ to [*second*]{} order in ${\varepsilon}$. In $D=0$ this Green’s function is a ratio of two ordinary integrals: $$G_1=\frac{\int_{-\infty}^\infty dx\,x\,\exp\left[-{\mbox{$\textstyle{\frac{1}{2}}$}}x^2\left(1+{\varepsilon}L +{\mbox{$\textstyle{\frac{1}{2}}$}}{\varepsilon}^2L^2\right)\right]}{\int_{-\infty}^\infty dx\,\exp\left[-{\mbox{$\textstyle{\frac{1}{2}}$}}x^2 \left(1+{\varepsilon}L\right)\right]}, \label{e39}$$ where $L=\log(ix)={\mbox{$\textstyle{\frac{1}{2}}$}}i\pi|x|/x+{\mbox{$\textstyle{\frac{1}{2}}$}}\log\left(x^2\right)$. Evaluating these integrals is straightforward and the result is $$\begin{aligned} G_1&=&-i{\varepsilon}\sqrt{\textstyle{\frac{\pi}{2}}}\left[1+{\mbox{$\textstyle{\frac{1}{4}}$}}{\varepsilon}\left(\gamma-2-3 \log2\right)+{\rm O}\left({\varepsilon}^2\right)\right]\nonumber\\ &=& -i{\varepsilon}\sqrt{\textstyle{\frac{\pi}{2}}}\left[1-0.8756{\varepsilon}+{\rm O}\left( {\varepsilon}^2\right)\right]. \label{e40}\end{aligned}$$ To check of the accuracy of (\[e40\]) we calculate the one-point Green’s function for a cubic theory (${\varepsilon}=1$). We convert the expansion in (\[e40\]) to a $[0,1]$ Padé approximant, $$G_1=-i{\varepsilon}\sqrt{\textstyle{\frac{\pi}{2}}}\frac{1}{1+0.8756{\varepsilon}},$$ and then set ${\varepsilon}=1$ to obtain the approximate result $G_1=-0.6682i$. The exact value of $G_1$ for the zero-dimensional cubic theory ${\varepsilon}=1$ is given by the ratio of integrals $$G_1=\frac{\int_{-\infty}^\infty dx\,x\exp\left(-{\mbox{$\textstyle{\frac{1}{2}}$}}ix^3\right)} {\int_{-\infty}^\infty dx\,\exp\left(-{\mbox{$\textstyle{\frac{1}{2}}$}}ix^3\right)} =-i2^{\frac{1}{3}}\frac{\Gamma(2/3)}{\Gamma(1/3)}=-0.6369i.$$ Thus, the [*two*]{}-term ${\varepsilon}$ expansion (\[e40\]) has an accuracy of 5%, which is impressive for such a large value of ${\varepsilon}$. This good result is consistent with the results found in previous studies of the accuracy of the ${\varepsilon}$ expansion for various classical equations (see Ref. [@r3]). The third-order version of (\[e40\]) is $$\begin{aligned} G_1&=&-i{\varepsilon}\sqrt{\textstyle{\frac{\pi}{2}}}\left[1+{\mbox{$\textstyle{\frac{1}{4}}$}}{\varepsilon}(\gamma-2-3\log2) \right.\nonumber\\ &&~~+\textstyle{\frac{1}{192}}{\varepsilon}^2\left(54\log^22+144\log2-36\gamma\log2 \right.\nonumber\\ &&~~\left.\left.-\pi^2+6\gamma^2-48\gamma+48\right)+ {\rm O}\left({\varepsilon}^3\right)\right]\nonumber\\ &=&-i{\varepsilon}\sqrt{\textstyle{\frac{\pi}{2}}}\left[1-0.8756{\varepsilon}+0.6447{\varepsilon}^2+ {\rm O}\left({\varepsilon}^3\right)\right]. \nonumber\end{aligned}$$ Converting the expansion above to a $[1,1]$ Padé approximant and setting ${\varepsilon}=1$, we obtain the result that $G_1=-0.6213i$, which now differs from the exact result by only $-2$%. These numerical results strongly motivate us to extend our studies of the ${\varepsilon}$ expansion of ${\ensuremath{\mathcal{PT}}}$-symmetric quantum field theories to higher order in ${\varepsilon}$. We will publish the higher-order calculations in a future paper. A second issue that that needs to be examined in depth is that of renormalization. Because $\Delta(0)$ becomes singular when the dimension $D$ of Euclidean spacetime reaches $2$, the one-point Green’s function $G_1$ and the renormalized mass $M_{\rm R}$ become singular. (To first order in ${\varepsilon}$ the higher Green’s functions do not become infinite when $D=2$.) Thus, for $D\geq2$ we must undertake a perturbative renormalization procedure. For simplicity, in this paper we have worked entirely with dimensionless quantities. However, to carry out a perturbative renormalization one must work with the Lagrangian $${\ensuremath{\mathcal{L}}}={\mbox{$\textstyle{\frac{1}{2}}$}}(\nabla\phi)^2+{\mbox{$\textstyle{\frac{1}{2}}$}}\mu^2\phi^2+{\mbox{$\textstyle{\frac{1}{2}}$}}g\mu_0^2\phi^2 \big(i\mu_0^{1-D/2}\phi\big)^{\varepsilon}$$ for which the dimensional parameters are explicit: $\mu$ is the unrenormalized mass $\mu$, $\mu_0$ is a fixed parameter having dimensions of mass, and $g$ is a dimensionless unrenormalized coupling constant. The mass renormalization procedure consists of expressing the renormalized mass $M_{\rm R}$ in terms of these Lagrangian parameters and absorbing the divergence that arises when $D\geq 2$ into parameter $\mu$. The coupling-constant renormalization procedure is similar; we define the renormalized coupling constant $G_{\rm R}$ as the value of the three-point or four-point Green’s functions at particular values of the external momentum and again absorb the divergence that arises into the Lagrangian parameter $g$. One must then verify that all higher Green’s functions are finite when expressed in terms of $M_{\rm R}$ and $G_{\rm R}$. This program will be carried out explicitly in a future paper. CMB thanks the Heidelberg Graduate School of Fundamental Physics for its hospitality. Appendix ======== In this Appendix we calculate the one-point Green’s function $G_1$ for the case $D=1$ (quantum mechanics). In quantum mechanics $G_1$ is the expectation value of the operator $x$ in the ground state $|0\rangle$. Thus, if $\psi(x)$ is the (unnormalized) ground-state eigenfunction in coordinate space, we can express $G_1$ as the ratio of integrals $$G_1 \equiv\frac{\langle0|x|0\rangle}{\langle0|0\rangle} =\int dx\,x\psi^2(x)\Big/\int dx\,\psi^2(x). \label{a1}$$ For the quantum-mechanical Hamiltonian (\[e1\]) the ground-state eigenfunction obeys the time-independent Schrödinger equation $$-{\mbox{$\textstyle{\frac{1}{2}}$}}\psi''(x)+{\mbox{$\textstyle{\frac{1}{2}}$}}x^2(ix)^{\varepsilon}\psi(x)=E\psi(x), \label{a2}$$ where $E$ is the ground-state energy. To first-order in ${\varepsilon}$ this differential equation becomes $$-{\mbox{$\textstyle{\frac{1}{2}}$}}\psi''(x)+{\mbox{$\textstyle{\frac{1}{2}}$}}x^2\psi(x)+{\mbox{$\textstyle{\frac{1}{2}}$}}{\varepsilon}x^2\log(ix)\psi(x)=E\psi(x). \label{a3}$$ We solve this equation perturbatively by substituting $$\begin{aligned} \psi(x)&=&\psi_0(x)+{\varepsilon}\psi_1(x)+{\rm O}\big({\varepsilon}^2\big),\nonumber\\ E&=&E_0+{\varepsilon}E_1+{\rm O}\big({\varepsilon}^2\big), \label{a4}\end{aligned}$$ where $\psi_0(x)=\exp\big(-{\mbox{$\textstyle{\frac{1}{2}}$}}x^2\big)$, $E_0={\mbox{$\textstyle{\frac{1}{2}}$}}$, and as derived in (\[e15\]), $E_1={\mbox{$\textstyle{\frac{1}{8}}$}}\Gamma'({\mbox{$\textstyle{\frac{3}{2}}$}})\big/\Gamma\big({\mbox{$\textstyle{\frac{3}{2}}$}}\big)$. Collecting powers in ${\varepsilon}$, we see that $\psi_0(x)$ and $\psi_1(x)$ satisfy the differential equations $$\begin{aligned} -{\mbox{$\textstyle{\frac{1}{2}}$}}\psi_0''(x)+{\mbox{$\textstyle{\frac{1}{2}}$}}x^2\psi_0(x)-{\mbox{$\textstyle{\frac{1}{2}}$}}\psi_0(x)&=&0,\nonumber\\ -{\mbox{$\textstyle{\frac{1}{2}}$}}\psi_1''(x)+{\mbox{$\textstyle{\frac{1}{2}}$}}x^2\psi_1(x)-{\mbox{$\textstyle{\frac{1}{2}}$}}\psi_1(x)&&\nonumber\\ &&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!=-{\mbox{$\textstyle{\frac{1}{2}}$}}x^2\log(ix)\psi_0(x)+E_1\psi_0(x). \label{a5}\end{aligned}$$ The equation for $\psi_1(x)$ is an inhomogenous version of the equation for $\psi_0(x)$. Thus, we use reduction of order to solve the $\psi_1(x)$ equation; we substitute $$\psi_1(x)=\psi_0f(x)=e^{-{\mbox{$\textstyle{\frac{1}{2}}$}}x^2}f(x) \label{a6}$$ and obtain $$f''(x)-2xf'(x)=x^2\log(ix)-2E_1.$$ We then multiply this equation by the integrating factor $\exp(-x^2)$ and integrate from $-\infty$ to $x$: $$f'(x)e^{-x^2}=C+\int_{-\infty}^x \!ds\,\big[s^2\log(is)-2E_1\big]e^{-s^2}.$$ The integration constant $C$ vanishes because $$2E_1=\int_{-\infty}^\infty \!\!ds\,s^2\log(is)e^{-s^2}\Big/\int_{-\infty}^\infty \!\!ds\,e^{-s^2}.$$ Thus, we find that $$\psi_1(x)=e^{-{\mbox{$\textstyle{\frac{1}{2}}$}}x^2}\!\!\int_0^x\!\!dt\,e^{t^2}\!\!\int_{-\infty}^t \!\!\!ds\big[s^2\log(is)-2E_1\big]e^{-s^2}. \label{a7}$$ We can now evaluate the integrals in (\[a1\]) to first order in ${\varepsilon}$. This expression simplifies considerably because of the factor of $x$: $$\begin{aligned} G_1&=&-\frac{2{\varepsilon}}{\sqrt{\pi}}\int_{-\infty}^\infty dx\,xe^{-x^2}\!\int_0^x dt\,e^{t^2}\nonumber\\ &&\quad\times \int_t^\infty\!ds\big[s^2\log(is)-2E_1\big]e^{-s^2}. \label{a8}\end{aligned}$$ To evaluate this integral we integrate from $x=-\infty$ to $0$ and then from $x=0$ to $\infty$ and combine the two integrals. The resulting integral simplifies further because the logarithm in the integrand collapses to $\log(-1)=i\pi$: $$G_1=-2i{\varepsilon}\sqrt{\pi}\!\int_0^\infty \!\!dx\,xe^{-x^2}\!\!\int_0^x \!\!dt\,e^{t^2}\!\!\int_t^\infty \!\!ds\,s^2e^{-s^2}. \label{a9}$$ Evaluating this triple integral is not trivial, but by using some tricks it can be calculated exactly and in closed form. We begin by interchanging the order of the $s$ and $t$ integrals: $$G_1=-2i{\varepsilon}\sqrt{\pi}\!\int_0^\infty \!\!dx\,xe^{-x^2}\!\!\int_0^\infty \!\!ds\,s^2e^{-s^2}\!\!\int_0^{{\rm min}(s,x)}\!\!dt\,e^{t^2}. \label{a10}$$ Next, we introduce polar coordinates $x=r\cos\theta$ and $s=r\sin\theta$ and make the change of variable $t=rz$: $$\begin{aligned} G_1&=&-2i{\varepsilon}\sqrt{\pi}\int_0^\infty\!\!dr\,r^5\!\int_0^{\pi/2}\! d\theta\,\cos \theta\sin^2\theta\nonumber\\ && \quad\times\int_0^{{\rm min}(\sin\theta,\cos\theta)}\!dz\,e^{r^2z^2-r^2}. \label{a11}\end{aligned}$$ The $r$ integral can now be done, and we obtain a sum of two double integrals: $$\begin{aligned} G_1&=&-2i{\varepsilon}\sqrt{\pi}\left[\int_0^{\pi/4}\!d\theta\,\cos\theta\sin^2\theta\! \int_0^{\sin\theta}\!dz \frac{1}{(1-z^2)^3}\right.\nonumber\\ &&~~\left.+ \int_{\pi/4}^{\pi/2}\!d\theta\,\cos\theta\sin^2\theta\!\int_0^{\cos\theta}\!dz \frac{1}{(1-z^2)^3}\right]. \label{a12}\end{aligned}$$ These double integrals may be evaluated by using any algebraic manipulation code such as Mathematica. The final result is $$G_1=-{\mbox{$\textstyle{\frac{1}{2}}$}}i{\varepsilon}\sqrt{\pi}. \label{a13}$$ This verifies the general result in (\[e20\]) for the case $D=1$. [99]{} C. M. Bender, K. A. Milton, M. Moshe, S. S. Pinsky, and L. M. Simmons, Jr., Phys. Rev. Lett. [**58**]{}, 2615 (1987). C. M. Bender, K. A. Milton, M. Moshe, S. S. Pinsky, and L. M. Simmons, Jr., Phys. Rev. D [**37**]{}, 1472 (1988). C. M. Bender, K. A. Milton, S. S. Pinsky, and L. M. Simmons, Jr., J. Math. Phys. [**30**]{}, 1447 (1989). C. M. Bender and S. Boettcher, Phys. Rev. Lett. [**80**]{}, 5243 (1998). C. M. Bender, S. Boettcher, and P. N. Meisinger, J. Math. Phys. [**40**]{}, 2201 (1999). P. E. Dorey, C. Dunning, and R. Tateo, J. Phys. A: Math. Gen. [**34**]{}, 5679 (2001). P. E. Dorey, C. Dunning, and R. Tateo, J. Phys. A: Math. Theor. [**40**]{}, R205 (2007). For a review of some of the early work see C. M. Bender, Rep. Prog. Phys. [**70**]{}, 947 (2007). J. Rubinstein, P. Sternberg, and Q. Ma, Phys. Rev. Lett. [**99**]{}, 167003 (2007). A. Guo, G. J. Salamo, D. Duchesne, R. Morandotti, M. Volatier-Ravat, V. Aimez, G. A. Siviloglou, and D. N. Christodoulides, Phys. Rev. Lett. [**103**]{}, 093902 (2009). C. E. Rüter, K. G. Makris, R. El-Ganainy, D. N. Christodoulides, M. Segev, and D. Kip, Nat. Phys. [**6**]{}, 192-195 (2010). K. F. Zhao, M. Schaden, and Z. Wu, Phys. Rev. A [**81**]{}, 042903 (2010). Z. Lin, H. Ramezani, T. Eichelkraut, T. Kottos, H. Cao, and D. N. Christodoulides, Phys. Rev. Lett. [**106**]{}, 213901 (2011). L. Feng, M. Ayache, J. Huang, Y.-L. Xu, M. H. Lu, Y. F. Chen, Y. Fainman, and A. Scherer, Science [**333**]{}, 729 (2011). J. Schindler, A. Li, M. C. Zheng, F. M. Ellis, and T. Kottos, Phys. Rev. A [**84**]{}, 040101(R) (2011). S. Bittner, B. Dietz, U. Günther, H. L. Harney, M. Miski-Oglu, A. Richter, and F. Schäfer, Phys. Rev. Lett. [**108**]{}, 024101 (2012). N. Chtchelkatchev, A. Golubov, T. Baturina, and V. Vinokur, Phys. Rev. Lett. [**109**]{}, 150405 (2012). C. Zheng, L. Hao, and G. L. Long, Phil. Trans. R. Soc. A [**371**]{}, 20120053 (2013). C. M. Bender, B. Berntson, D. Parker, and E. Samuel, Am. J. Phys. [**81**]{}, 173 (2013). B. Peng, Ş. K. Özdemir, F. Lei, F. Monifi, M. Gianfreda, G. L. Long, S. Fan, F. Nori, C. M. Bender, and L. Yang, Nat. Phys. [**10**]{}, 394 (2014). S. Assawaworrarit, X. Yu, and S. Fan, Nature [**546**]{}, 387 (2017). M. Mezard, G. Parisi, and M. Virasoro, [*Spin Glass Theory and Beyond*]{} (World Scientific, Singapore, 1987). M. Abramowitz and I. A. Stegun, [*Handbook of Mathematical Functions*]{} (Dover, New York, 1972). H. F. Jones and R. J. Rivers, Phys. Rev. D [**75**]{}, 025023 (2007).

ArXiv

--- abstract: 'The emerging trend of edge computing has led several cloud providers to release their own platforms for performing computation at the ‘edge’ of the network. We compare two such platforms, Amazon AWS Greengrass and Microsoft Azure IoT Edge, using a new benchmark comprising a suite of performance metrics. We also compare the performance of the edge frameworks to cloud-only implementations available in their respective cloud ecosystems. Amazon AWS Greengrass and Azure IoT Edge use different underlying technologies, edge Lambda functions vs. containers, and so we also elaborate on platform features available to developers. Our study shows that both of these edge platforms provide comparable performance, which nevertheless differs in important ways for key types of workloads used in edge applications. Finally, we discuss several current issues and challenges we faced in deploying these platforms.' author: - 'Anirban Das, Stacy Patterson, Mike P. Wittie [^1][^2][^3]' bibliography: - 'biblio.bib' title: '**EdgeBench: Benchmarking Edge Computing Platforms**' --- INTRODUCTION {#sec.introduction} ============ As the Internet of Things (IoT) is becoming mainstream, the number of connected devices is growing at an exponential rate [@gartner21billion]. In this paradigm, IoT devices, which are often geographically distributed at the edge of the network, will generate a massive quantity of data. Transmitting, storing, and processing this huge amount of data in the cloud is expected to lead to high bandwidth usage and prohibitive costs [@edgevision]. Further, many applications that run at the edge of the network, such as autonomous vehicles and augmented reality, have real-time requirements that are difficult to meet with relatively distant cloud datacenters [@etsiedgecomputing]. Edge computing has the potential to mitigate these cost and performance bottlenecks. This computing paradigm enables applications to leverage compute nodes in close proximity to data sources to perform data processing, such as aggregation, filtering, and classification, before forwarding the results to other nodes and cloud servers [@edgevision; @satyanarayanan2017emergence]. For example, instead of sending an image to the cloud to perform facial recognition, and edge device may simply report whether the image contains the particular signature. This approach reduces bandwidth usage and can speed up application response. To make application development in the edge computing model easier, several cloud providers, have put forward their own edge computing platforms. These platforms provide the ability to deploy and orchestrate applications, such as machine learning models, on edge devices in the form of stateless serverless functions or user code in containers. The resource provisioning and runtime is provided by the edge platforms. Such serverless functions or user scripts in containers then act on raw data, and depending on the configuration, send results to the cloud and additionally also make them available in other cloud services. Since these platforms use different paradigms and technologies, it is important to compare them to understand their tradeoffs and to select the best platform for a given use case. Criteria of interest include the platform architecture, programmability, performance, and cost. To quantify these differences consistently requires benchmarks of common uses cases. Further, it requires standardized collection of metrics, such as end-to-end latency, device compute time, device resource utilization, bandwidth usage, and cost. To make an informed choice between edge and cloud platforms, the benchmarks and metrics must also allow fair comparison across different types of deployments. We present EdgeBench[^4] – an open-source benchmark suite for serverless edge computing platforms. EdgeBench features three key applications: a speech/audio-to-text decoder, an image recognition machine learning model, and a scalar value generator emulating a sensor. Each application processes a bank of input data on an edge device and sends results to cloud storage. We target EdgeBench for two of the most popular edge computing platforms currently available, AWS Greengrass [@aws_dev_guide] and Microsoft Azure IoT Edge [@azure_dev_guide]. EdgeBench also provides cloud-based workload implementations. Our aim is to quantify the differences between the different edge platforms, as well as the providers’ respective cloud-only alternatives. In future work, we plan to extend EdgeBench to other emerging edge platforms, such as Google’s Cloud IoT Edge [@googleiotedge] and IBM Watson IoT Platform Edge [@watsoniotedge], as these offerings mature. We report on initial experiments with EdgeBench using a Raspberry Pi 3B, a relatively resource-constrained device, to emulate the IoT device that sends traffic to the AWS and Azure cloud platforms. We provide a performance comparison across the different workloads and platforms. EdgeBench complements previous work on benchmarking serverless cloud computing platforms. Malawski et al. have developed two CPU-intensive benchmark suites and evaluated them on the different industry providers [@malawski2017benchmarking]. The recent work by McGrath and Brenner presents a comparison of the overhead of various platforms, measured using a custom-developed tool [@serverlessicdcs]. The work by Back and Andrikopoulos presents a performance study of industry serverless cloud platforms using compute/memory constrained workloads, with a focus on cost [@microbenchmark]. Finally, Deese presents a study of the performance of $K$-means clustering using AWS Lambda functions in the cloud [@Deese18]. To the best of our knowledge, our work is the first that benchmarks industry platforms that use the serverless paradigm for edge computing. The rest of the paper is organized as follows. Sec. \[sec.sysarchitecture\] provides details about the architectures of AWS Greengrass and Azure IoT Edge. Sec.\[sec.benchmarks\] describes the EdgeBench workloads and metrics. In Sec. \[sec.setup\_results\], we describe the experimental setup and results of the benchmark study and discuss some observations, and finally, we conclude in Sec. \[sec.conclusion\]. SYSTEM ARCHITECTURES {#sec.sysarchitecture} ==================== On an abstract level, both platforms, AWS Greengrass (henceforth, Greengrass) and Azure IoT Edge (henceforth, Azure Edge) share a common general architecture. There is an edge device that runs user code in the platform’s runtime system. This user code can access local volumes or local devices, such as sensors and cameras, performs computations, and sends messages to the cloud. In both platforms, the cloud has a high throughput message ingestion service, the IoT Hub. The cloud ingests the messages from the edge devices and sends them to configurable destinations, such as AWS S3 or Azure blob for storage using ‘Rule’ (for AWS) or a ‘Route’ (for Azure). Below, we highlight some of the details of each platform: AWS Greengrass -------------- {width="0.9\linewidth"} \[fig.Greengrass\_architecture\] The Greengrass pipeline is shown in Fig. \[fig.Greengrass\_architecture\]. Greengrass edge devices run the Greengrass core software. The core software allows users to run Lambda functions locally on the edge devices and manage, modify or update them through the AWS console website or deployment API. Developers can constrain the maximum memory usage of the local Lambda functions. The Greengrass core software also takes care of the authentication, authorization, and secure message routing, through the MQTT protocol [@mqtt], between the devices, Lambda functions, and the cloud. Greengrass core Lambda runtime currently supports code deployment in Python 2.7, Node.JS 6.10, Java 8, C, C++, and any language that supports importing C libraries. Code running inside Lambda functions can also access all other AWS services, such as Amazon S3 or DynamoDB, using the standard AWS SDKs. Once the AWS IoT Hub receives a message in the cloud, a ‘Rule’ can be defined to trigger one of 15 actions (as of now), including invoking Lambda functions that run in the cloud or saving data in S3 or DynamoDB. If the ‘Rule’ declares to save messages in S3 storage, the hub does so by creating one blob file for each message in the specified S3 bucket, as soon as the message is processed by the hub. Microsoft Azure IoT Edge ------------------------ {width="0.9\linewidth"} \[fig.azure\_iot\_architecture\] Azure Edge uses lightweight virtualization, specifically, Docker compatible containers, to deploy computation on edge devices. The Docker containers run as ‘edge modules’ in the Azure Edge platform, as shown in Fig. \[fig.azure\_iot\_architecture\]. The modules can contain Azure Functions, user code, and libraries. As of now, the platform supports five languages: C\#, C, Node.JS ver $>$ 0.4.x.x, Python (both 2.7 and 3.6), and Java 7+. It is also possible to deploy Streaming Analytics and Azure ML models directly in the containers. The former is a managed service from Azure for doing analytics on streaming data and the latter are the models developed in Azure’s machine learning service. Modules can be deployed, updated, and modified via the Azure IoT Edge Cloud web interface or the Azure command line interface. The runtime system on a single edge device consists of an `edgeAgent` module and an `edgeHub` module. The `edgeAgent` takes care of provisioning and monitoring user deployed modules. The `edgeHub` takes care of the connection between the modules and the cloud and also maintains security and authentication. The `edgeHub` supports edge-to-cloud connections using the MQTT and AMQP protocols The edgeAgent sends messages to the cloud-hosted Azure IoT Hub, as shown in Fig. \[fig.azure\_iot\_architecture\]. Messages are then routed from the IoT Hub to a user-specified IoT Hub Endpoint, such as Azure Blob Storage. For the Blob Storage Endpoint, the IoT Hub batches the incoming messages and writes multiple results in a single blob file. If user selects Blob Storage endpoint, batching is the only option. The batching window can be configured by either time window, the smallest being 60s, or by chunk size, the smallest being 10MB. EDGEBENCH {#sec.benchmarks} ========= In this section, we describe EdgeBench benchmark suite and the performance metrics. We also summarize the cloud-based implementations of the benchmark applications. Benchmark Applications and Workloads {#benchmark_apps.sec} ------------------------------------ We selected three canonical applications: a speech/audio-to-text application; an image recognition application; and a scalar value generator that emulates a sensor, e.g. a temperature sensor. With the popularity hike in use of a myriad of smart speakers, such as Amazon Echo and Google Home, has made audio and speech decoding and translation very relevant. Similarly, with the proliferation of smart cameras and autonomous vehicles, image processing and image classification has become very common. Currently, these applications are often executed in the cloud. Hence, it is interesting to investigate performance of such applications in an edge computing setting. The scalar benchmark, however, is an example of an extremely lightweight workload; it allows us to measure the performance of each framework when the computation and data volume at the edge are negligible. All benchmark codes are written in Python. In all three pipelines, the edge devices send data in messages to the IoT Hub We use a either a ‘Rule’ or ‘Route’ to push each message payload in the cloud to an AWS S3 bucket or an Azure Blob, respectively. - **Audio/Speech to Text Translation (Audio Pipeline)** Here, the edge application reads audio files from a local directory, decodes them to get the translated text, and then sends the text to the cloud. Each audio file is processed one at a time. For our experiments, we use a mobile version of Carnegie Mellon University’s Sphinx speech recognition system, called PocketShpinx [@pocketsphinx]. We use the default acoustic model provided with the package. For the audio workload, we use 104 samples contributed by user ‘**rhys\_mcg**’ in Tatoeba Corpus [@tatoeba], a free collaborative online database of example sentences. We have converted the audio files into 16khz, 16 bit, mono ‘wav’ file format to comply with the requirements of PocketSphinx. The average realtime length of files in the dataset is about 2.4s. - **Image Recognition (Image Pipeline):** The serverless function performs an image recognition task; specifically, given an image as input, the function recognize the objects present in the image and generates class labels for these objects. In both platforms, the edge application reads an image from a directory. It then uses OpenCV [@opencv] to resize the image to standard ($224\times224\times3$) size. Finally, the application uses the open-source, deep learning framework MXNet [@MXNet] to recognize and classify the objects in that image. The results are sent to the cloud. This is repeated for all images in the directory. For the classifier architecture, we chose a pretrained classifier, Squeezenet [@squeezenet] because its small model size ($\approx$ 5MB) and low compute footprint are suitable for resource-constrained edge devices. For the input workload, we select 500 images from the ILSVRC2012 image dataset [@imagenet2012]. The input is stored on a local directory on the device. - **Scalar Sensor (Scalar Pipeline):** The application is a simple sensor emulator. The serverless function generates random scalar values at a user-specified frequency. At a user-specified interval, e.g., 1s, the set of values generated in that interval is sent to the cloud in a single message. The cloud side of the pipeline then simply stores these values in the specified storage location. {width="0.9\linewidth"} \[fig.Cloud\_Architectures\] Metrics ------- In all pipelines of both platforms, each message receives three UTC timestamps during the pipeline execution, as shown in Fig. \[fig.Greengrass\_architecture\] and \[fig.azure\_iot\_architecture\], from which we calculate delay metrics. Here, for valid calculations on these metrics, we need the time of both the edge device and cloud to be independently synced with accurate clocks. The user code in the edge device adds the $T_1$ timestamp before sending the result message from the edge device. Timestamp $T_2$ is added automatically by the platform when the message is en-queued in the IoT Hub, and finally, $T_3$ is the creation timestamp of the blob file in which the message is stored after it is routed out of the IoT Hub. We capture the following metrics to study performance: - **Compute time:** This is the total time required for processing one image or audio or to generate the scalar values in the Raspberry Pi and is denoted by $C_{edge}$. - **Time-in-flight:** This is the time taken for a message to reach the IoT Hub after it is sent from the edge device. It is given by $T_2 - T_1$. - **End-to-end latency:** This is the difference between the time when the input is ingested at the edge device and the time when the final results are available in the storage. This value is given by $C_{edge} + (T_3 - T_1)$. - **Payload size**: This is the size of the message sent from the edge device without the framework overhead. - **CPU and memory utilization**: As the platforms use different architectures, i.e., Lambda vs Docker, it is interesting to look at the memory and CPU usage patterns. We measure the average CPU and memory utilization on the edge device over the execution of a given benchmark. For AWS, we use the `top` command in Linux and for Azure we use the `docker stats` command to obtain the resource utilization while the applications are running. Our applications log compute time, the payload size, and resource utilization locally on the edge device, while $T_1$ is added to the header. Therefore, $T_1$, $T_2$ and $T_3$ are retrieved from the message meta-data in the cloud. Cloud-Only Pipelines -------------------- We implement cloud-only versions of the three benchmark workloads described in Sec. \[benchmark\_apps.sec\] for both the Amazon and Microsoft Azure cloud platforms. The pipelines use the serverless architecture, as shown in Fig. \[fig.Cloud\_Architectures\], to process device data. All code is written in Python. For the image classification and speech-to-text benchmarks, we upload either image or audio file from the edge device to a S3 bucket using `boto3` library. For the scalar pipeline in Amazon AWS, we generate and upload the sensor values as JSON blob files in S3. Lambda functions are triggered by the creation of the new blob file in the bucket. The Lambda function reads the value from the blob file and simply stores it in another S3 bucket. The upload of a file triggers the Lambda function which, in turn, either performs the image recognition or the audio-to-text conversion. After the computation, the results are stored as blobs in another S3 bucket. In the Azure implementations, we use Azure Functions, which are similar to Lambda functions. The Azure functions are triggered by the upload of an audio or image file or a scalar value JSON file in Azure blob storage. After the computation, the results are stored in a different blob. The majority of the code in cloud and edge pipelines are the same, excepting changes for handling input/output and receiving and handling the events due to different API specifications. Note that in our benchmarks, Azure functions run on Windows while Lambda functions run on Amazon Linux. Linux is available on a preview basis for Azure Functions, but it supports only JavaScript and .Net runtime as of now. Moreover, even in Windows in Azure Functions, the support of Python is in the experimental phase. We manually installed MXNet, OpenCV, PocketSphinx and other necessary libraries from the KUDU console in Azure Python runtime to use the libraries in Azure Functions.In AWS, the dependencies are packaged along with the Lambda function. EXPERIMENTS {#sec.setup_results} =========== Experimental Setup ------------------ We run the set of benchmarks using a standard Raspberry Pi 3B model as the edge device. The edge device is connected to the internet via a wireless router using 2.4 Ghz spectrum. We have used a dedicated Stratum 1 NTP time server, TM2000A [@tm2000A] with accuracy $\approx$ 50$\mu$s to synchronize the time of the Raspberry Pi. Also, AWS and Azure are known to use highly precise clocks to accurately sync their services. We use AWS and Azure cloud services in the US East region, both in Virginia. We used `ping` to find the round trip latency from our institution’s server to virtual machines in both Azure and AWS. We are unable to measure this from the edge device, due to security restrictions on `ping`. The average delay for AWS is 9.5ms and for Azure is 11.36ms. Assuming the extra delay within the institution network is same for both Greengrass and Azure Edge, we observe latencies to both the cloud platforms are close. We use Greengrass core version 1.5.0 and Azure IoT Hub Device client 1.4.0. In the experiments with Greengrass, each Lambda function is provisioned with 256MB RAM and made ‘long lived’, i.e., it will run indefinitely. However, this option is absent in Azure Edge. In Greengrass, the image, audio and local statistics directories for storing metric values are mounted into the execution environment as ‘Local Resources’. In Azure Edge, the same directories are directly mounted as volumes in the Docker container. In Azure Edge, we use the geo-redundant storage option RA-GRS for blob storage in the cloud. AWS S3 replicates data automatically across at least 3 availability zones. We also measure performance of the cloud-only pipelines in AWS and Azure. For AWS, the memory of the Lambda functions is set to 3008MB. In AWS, CPU allocation is proportional to the memory, hence, this configuration has the highest memory and CPU performance. As Lambda CPU is not configurable, to keep the setups comparable, we select the Consumption Host Plan for Azure functions that auto-scales Azure functions based on system load. In all cloud pipelines, we wait for a period of 10 to 15s between uploading subsequent image/audio files to avoid congesting the system. Uploading too many image/audio files very quickly resulted in many functions being triggered concurrently and out of order. This results in reordering of results and some missed images, making it difficult to find the end-to-end latency. The input data sizes are shown in Table \[table.bandwidthusage\_total\]. Results ------- {width="0.75\linewidth"} \[fig.endtoend\_time\] ### End to End Latency {#subsec.endtoendlatency} We give the average end-to-end latency for each of the various benchmark configurations in Fig. \[fig.endtoend\_time\]. We observe that in all three applications, across both the cloud and edge pipelines, Azure Edge has the largest end-to-end latency. This is because the Azure IoT service batches the messages from the edge device in the IoT Hub in the cloud, and it writes the results from the multiple messages in a batch in a single blob file. We also found that when the batching interval is 60s, the average time spend in hub is $\approx$ 90s, while for a 90s batching interval, the average time in the hub is $\approx $ 93-94s. It appears that messages are held back in the IoT hub in Azure for some time interval before being written to the blob file, and this time does not coincide with the batching interval. If this were not the case, given that the messages are received by the IoT Hub approximately uniformly across a batching interval, the average time a message spends in hub should have been roughly equal to half the batching interval. This extra delay, adding to latency, exists irrespective of the blob storage type used. We observe that the end-to-end latency for Azure cloud pipelines is larger than both AWS edge and cloud pipelines across all applications. The majority of the latency is caused by the total time of execution of Azure function in the cloud. Though, the average time for audio to speech and image recognition in Azure cloud is 5.57s and 1.19s respectively, for each message, added to this is the time for loading libraries and trigger the function, which is very high for Azure python runtime. This may have been be caused by the fact that the Python runtime in Azure is still experimental and hence, not optimized. It may also be that importing the libraries takes a lot of time. We obtain the smallest end-to-end latency results using the AWS Cloud (1.79s for Audio, 0.87s for Image, and 0.936s for Scalar), followed by Greengrass (5.36s for Audio, 1.1s for Image and 0.66s for Scalar). It appears that image processing at the edge with Greengrass is highly feasible, as both cloud and edge end-to-end latencies are very close. In Fig. \[fig.flight\_time\] we observe, Azure takes on average from 1- 18ms longer to deliver the messages to the cloud for the edge pipelines compared to AWS. The flight times of AWS and Azure edge are very close, which suggests that as long as the user selects data centers with similar latencies, the flight time will not contribute much to the difference of end-to-end message latencies of AWS compared to Azure. Finally, Fig. \[fig.compute\_time\] shows that Azure, in general, has a higher compute time for all pipelines compared to Greengrass. The highest is the audio pipeline, with Azure Edge taking 6s, on average, and Greengrass taking 4.77s, on average. This difference may indicate a place where the different architectures (Lambda vs. Docker) may have made a difference. If this time is large, then it has a significant impact on end-to-end latency. We also observe that, for the audio pipeline in the edge, considering the average length of each clip is approximately 2.4s, it may not be possible to analyze the audio in real time using a Raspberry Pi without optimizing the code. However, using a more powerful edge device would help to reduce compute time. \[table.bandwidthusage\_total\] ### Bandwidth Utilization The average payload size in both platforms for audio pipeline is 162bytes, for image pipeline it is 752bytes and for scalar it is 234bytes. On comparing flight times for each edge pipeline in Fig. \[fig.flight\_time\] we observe that the transmission delay (flight time) of messages between the edge and cloud is roughly proportional to the message size. A cloud-only approach requires the upload of the raw image or audio file to the cloud service. In the edge pipelines, we only the send results of the applications as text to the cloud. Hence, we observe that there is drastic reduction in the per message size in edge pipelines compared to cloud. We also used `vnstat` [@vnstat] to obtain the total bandwidth usage of the applications in the edge and cloud pipelines, shown in Table \[table.bandwidthusage\_total\]. The results are measured with respect to 200 scalar values, 500 images, and 104 audio files, respectively. To avoid measuring the TLS handshakes and other module startup network overhead, we explicitly add a configurable delay (60s, in this case) before the application begins processing data. We see that framework data overhead itself is negligible and comparable in both platforms. Comparing the total data transmitted during pipeline executions, we see a massive reduction of data transmission while using the edge pipelines compared to cloud. AWS sends 36 times and 81 times more data when using the cloud pipelines compared to the edge, for the audio and image applications, respectively. Azure sends 36 times and 77 times more data using the cloud pipelines compared to the edge in audio and image applications, respectively. [.25]{} {width="\linewidth"} [.25]{} {width="\linewidth"} [.25]{} {width="\linewidth"} [.25]{} {width="\linewidth"} \[fig.aws\_azure\_timings\] ### Local Resource Utilization We study the average resource usage of the edge device (Raspberry Pi) across the edge pipelines over three separate runs of the experiments. In Azure Edge, we look at the average total CPU and memory percentage used by all of the containers running specific to Azure Edge. In Greengrass, we look at the total average CPU and memory percentage usage by all processes under the greengrass user, `ggc_user`. We observe in Fig. \[fig.CPU\_util\], \[fig.Mem\_util\] that the image recognition application is predominantly a CPU intensive job, with CPU utilization as high as 88-90% in both Azure and AWS. We also observe that audio-to-text is not very CPU-heavy, but it consumes more memory than the other applications. We further observe that within each application, the CPU % of Greengrass and Azure Edge are very similar, though the RAM consumed in Azure Edge is always higher. Azure Edge consumes on average about 29.5MB to 54.5MB more memory on the Raspberry Pi. We believe this difference would be less discernible if a more powerful edge device were used. Overall, we observe that the edge pipelines, on average, consume less than 200MB RAM and do not saturate the CPU usage. This strengthens the case for the feasibility of running some carefully chosen computations on resource-constrained devices. ### Infrastructure Cost We do a rough infrastructure cost estimate of running the applications in the edge pipelines versus the cloud-only pipelines. Cost of running pipelines in both vendors are comparable and so, for simplicity, here, we look only at the image pipeline in AWS. Assume there is one traffic camera, generating one image every 10s. Let us further assume the average image size is similar those used in our image benchmarks, i.e., 143.12 KB (In general, image size and rate would be larger in real world scenarios.) This amounts to $6\times60\times24\times30 = 259,200$ images per month. Also, on average, the duration for which AWS charges for Lambda function execution is 300ms in the image recognition cloud only pipeline in our study for image of that size. In the Greengrass image pipeline, the average size of a message to the cloud is 752bytes. We assume, with headers, it would be approximately 1KB per message. All prices are calculated in region US-East, Virginia. For Greengrass total cost is the expense of running Greengrass plus the cost of storing results in S3 plus put requests for results in S3, which equals $0.2627 + 0.0057 + 1.29=\;\$\;1.5584$/month. For the cloud pipeline, the cost is the expense of storing raw images and final results in S3 plus get and 2$\times$put requests cost in S3 plus the cost of running Lambda functions , which equals $0.814 + 0.0057 + 2.69 + 4.517 = \;\$\;8.027$ /month. Though this estimate appears cheap, if there are, for example, 50 road side cameras, the cost for the cloud-only pipeline escalates quickly. This rough cost estimate indicates that executing image recognition on the cloud setting is $\approx$ 5.2x more expensive than edge, with only an extra 230 ms in average end-to-end latency. It is possible to reduce the cost of the cloud-only pipeline by using less powerful Lambda functions, however, the storage cost alone is larger than the entire edge pipeline cost. Bandwidth usage-wise, the edge pipeline sends around 253.125MB data over the network per month, whereas uploading images to cloud requires sending 35.38GB data/month. Discussion {#sec.discussion} ---------- We observed that both platforms do not handle very high throughput messaging well yet. In this case either the messages are delayed or fail to reach the cloud from the edge device. In future we want to benchmark this throughput. Another important consideration is the relative ease of deploying dependencies and libraries. We can package all necessary libraries in the Docker container with a single Dockerfile in Azure Edge. On the contrary, in Greengrass Lambda functions, we need to compile external dependencies in required environment and add them in a zip file for deploying. We feel that the former is a cleaner choice for adding and managing a lot of external libraries. In the end, from our experience, we feel both these approaches are suitable for carrying out edge computation. However, the higher end-to-end latency of Azure may be a problem for latency sensitive applications. Although Azure has richer customization options, we found development and integration to be easier in the AWS Greengrass platform. CONCLUSION {#sec.conclusion} ========== We have presented EdgeBench, a benchmark suite for serverless edge computing plaforms. With this suite, we have studied two managed edge computing platforms, Greengrass and Azure Edge. Further, we have compared the performance of these platforms with that of cloud-only implementations of the same benchmarks. Our results show that the performance of Greengrass and Azure Edge are comparable, with the exception that Azure Edge exhibits higher end-to-end latency due to its batch-based processing approach. Further, our results show that for the image and scalar pipelines, the performance of Greengrass is comparable to that of the AWS cloud-only pipelines, while reducing the network bandwidth usage. These results indicate that edge computing is a promising alternative to cloud computing for CPU light workloads. In future work, we plan to extend EdgeBench to include additional applications and edge platforms. [^1]: Anirban Das and Stacy Patterson are with the Department of Computer Science, Rensselaer Polytechnic Institute, Troy, New York 12180, USA. [dasa2@rpi.edu, sep@cs.rpi.edu]{} [^2]: Mike P. Wittie is with the Gianforte School of Computing at Montana State University, Bozeman, Montana 59715, USA. [mwittie@cs.montana.edu]{} [^3]: This work was funded in part by NSF awards CNS-1527287, CNS-1553340, CNS-1555591, and CNS-1527097 [^4]: <https://github.com/akaanirban/edgebench>

ArXiv

--- abstract: 'We propose analytical models that allow to investigate the performance of Long Range Wide Area Network (LoRaWAN) uplink in terms of latency, collision rate, and throughput under the constraints of the regulatory duty cycling, when assuming exponential inter-arrival times. Our models take into account sub-band selection and the case of sub-band combining. Our numerical evaluations consider specifically the European ISM band, but the analysis is applicable to any coherent band. Protocol simulations are used to validate the proposed models. We find that sub-band selection and combining have a large effect on the QoS experienced in a LoRaWAN cell for a given load. The proposed models allow for optimizing resource allocation within a cell given a set of QoS requirements and a traffic model.' author: - '[^1] [^2]' bibliography: - 'lora.bib' title: 'Analysis of Latency and MAC-layer Performance for Class A LoRaWAN' --- LoRa; LoRaWAN; LPWA; IoT; QoS; latency; duty cycle; low power; long range. Introduction ============ Services utilizing communications between machines are expected to receive a lot of attention, such as health monitoring, security monitoring and smart grid services [@palattella2016internet]. These Internet of Things (IoT) services generate new demands for wireless networks. The spectrum of service scenarios in the IoT is wide and as a result the required quality of service (QoS) across IoT services is also wide. In some scenarios ultra high reliability is required, in others a low latency is required and supporting massive numbers of low-cost and low-complexity devices is still important issue. The devices can be served by the cellular networks and, specifically, by their M2M-evolved versions, such as Narrowband IoT (NB-IoT) [@wang2016primer]. However, there is a low-cost alternative for serving these devices using Low Power Wide Area (LPWA) networks that operate in unlicensed bands. The number of IoT devices connected by non-cellular technologies is expected to grow by 10 billions from 2015 to 2021 [@ericsson2016mobile]. It is therefore of interest to develop QoS models for the LPWA protocols in order to analyze which protocol is best suited for a given service. Long Range Wide-area Network (LoRaWAN) is an emerging protocol for low-complexity wireless communication in the unlicensed spectrum using Long Range (LoRa) modulation. The scalability and capacity of LoRaWAN is investigated in [@mikhaylov2016analysis] where it is implicitly assumed that the inter-arrival times are fixed. In [@scaleBor] the scalability is evaluated in terms of goodput and network energy consumption. One of the key elements of LoRaWAN is the use of duty cycling in order to comply with the requirements for unlicensed operation. Duty cycling is imposed per sub-band by regulation and optionally also aggregated for all bands. It is the central factor that sets limitation on the throughput and the latency of the network. The limits of duty-cycled LoRaWAN are pointed out in [@adelantado2016understanding], but only aggregated duty cycle and fixed inter-arrival arrivals are considered. The contribution of this paper is an analytical model of the LoRaWAN uplink (UL) that characterizes the performance, in terms of latency and collision rate, under the influence of regulatory and aggregated duty cycling, assuming exponential inter-arrival times. The obtained latency and collision rate results from the analysis are verified through simulation. We summarize the key features of LoRaWAN in Section \[sec:lora\]. A system model is presented in Section \[sec:model\] and analysed in Section \[sec:ana\]. Numerical results based on the analysis and simulation is shown in Section \[sec:eval\]. Concluding remarks are given in Section \[sec:conc\]. Long Range Wide Area Network {#sec:lora} ============================ LoRaWAN is a wireless communication protocol providing long range connectivity at a low bit rate. LoRaWAN is based on the LoRa modulation. LoRaWAN supports LoRa spreading factors 7 to 12. The overhead of a LoRaWAN message with a payload and no optional MAC command included is 13 bytes. LoRaWAN defines a MAC layer protocol to enable low power wide area networks (LPWAN) [@loraspec20015]. A gateway serves multiple devices in a star topology and relays messages to a central server. LoRaWAN implements an adaptive data rate (ADR) scheme, which allows a network server to select both the data rate and the channels to be used by each node. Three different classes (A, B and C) of nodes are defined in LoRaWAN. Class A has the lowest complexity and energy usage. All LoRaWAN devices must implement the class A capability. A class A device can receive downlink messages only in a receive window. There are two receive windows after a transmission in the uplink. The first window is scheduled to open 1 to 15 second(s) after the end of an uplink transmission with a negligible 20 ms margin of error. The second window opens 1 second after the end of the first. LoRaWAN utilizes the industrial, scientific and medical (ISM) radio bands, which are unlicensed and subject to regulations in terms of maximum transmit power, duty cycle and bandwidth. The end-device also obeys a duty cycling mechanism called the aggregated duty cycle, which limits the radio emission of the device. An aggregated duty cycle of 100 % corresponds to the device being allowed to transmit at any time, but still in accordance with the regulatory duty cycling. The lowest aggregated duty cycle of 0 % means that the particular device turns off the transmissions completely. System Model {#sec:model} ============ Consider $M$ devices connected to a single LoRaWAN gateway. Each device is assigned a spreading factor to use for transmission by a network server. We account for the interference through the collision model, where collision occurs when two or more devices try to transmit simultaneously in the same channel using the same spreading factor. We also consider a LoRa-only configuration, in this work, such that no interference from other technologies is present. Different spreading factors are considered to be entirely orthogonal. A fixed payload size is assumed. We further assume that all devices are class A and have successfully joined the network and transmit the messages without acknowledgement so that there are no downlink transmissions. Due to the absence of acknowledgements, retransmissions are not considered. Among all sub-bands, a device is given a subset of the sub-bands. Enumerate these sub-bands 1 through $c$. Let $n_i$, $i=1..c$ and $\delta_i$, $i=1..c$ be the number of channels and the duty-cycle[^3] in sub-band $i$, respectively. As described in the specifications [@loraspec20015] and in the source code of the reference implementation of a LoRa/LoRaWAN device[^4], the scheduling of a LoRaWAN transmission happens as follows: 1. A device waits until the end of any receive window. 2. A device waits for any off-period due to aggregated duty cycling. 3. A device checks for available sub-bands, i.e., ones that are not unavailable due to regulatory duty cycling: 1. A channel is selected uniformly randomly from the set of channels in all available sub-bands. 2. If there is no free sub-band, the transmission is queued in the first free sub-band. A random channel in that sub-band will be selected. A transmission, limited by the duty cycle $\delta$, with a transmission period $T_\mathrm{tx}$ infers a holding period, which, including the transmission itself is given by: $$\begin{aligned} \label{E:holdingtime} T_\mathrm{hold} = T_\mathrm{tx} + T_\mathrm{tx}\left(\frac{1}{\delta}-1\right) = T_\mathrm{tx}\frac{1}{\delta}.\end{aligned}$$ The *service rate* is the inverse of the holding time, $\mu=\delta/T_\mathrm{tx}$. Sub-bands can have different duty cycles and in turn different service rates. Let $\lambda$ be the generation rate of packets for a device. When several sub-bands are defined for the device the sub-band for the next transmission is selected according to the step 3-a) and 3-b). We define *service ratio* $r_i$ as the fraction of transmissions carried out in the $i-$th sub-band. Analytical Model {#sec:ana} ================ In this section the analytical models for latency and collision probability are presented. Single Device Model: Latency {#S:analat} ---------------------------- The latency of a transmission is the time spent on processing, queueing, transmission of symbols, and propagation. Assuming that the time for processing and propagation are negligible, we have: $$\begin{aligned} \label{E:totaldelay} T_\mathrm{total} = T_\mathrm{tx} + T_\mathrm{w}.\end{aligned}$$ We model the wait for reception windows and aggregated duty cycling ( steps 1) and 2) ) as a single traffic shaping $M/D/1$ queue. The service rate of this $M/D/1$ queue is the slowest mean rate of service in step 1) and 2). For step 3), we model the regulatory duty cycling as an $M/D/c$ queue with heterogeneous servers, where each server corresponds to a sub-band. The waiting time $ T_\mathrm{w}$ for a transmission and the service ratio of each sub-band can then be found from queue theory. The waiting time, $T_w$, due to regulatory duty-cycling can be calculated for asymmetric $M/D/c$ queue[^5] that models step 3) of the scheduling procedure, but as it is easier to model and compute on a $M/M/c$ queue relative to a $M/D/c$ queue, we use the rule of thumb that the waiting line of a symmetric $M/M/c$ queue is approximately twice that of an $M/D/c$ queue [@newoldmdc] to simplify our analysis. Our simulations show that this is a good approximation also for asymmetric queues. The waiting time in sub-band $i$ is then: $$\begin{aligned} \label{E:waitdelay} T_{w_i} = \frac{p_{busy,all}}{(\sum^c_{i=1}\mu_i+\lambda)\cdot 2},\end{aligned}$$ where $p_{busy,all}$ is the Erlang-C probability that all servers are busy. The transmission latency in each band can then be found from Eq. . The mean latency is given as a weighted sum of the transmission latencies in each sub-band, where the weights are given by the service rate of each sub-band. The fraction of transmissions in sub-band $i$, $\lambda_i$, is the product of the holding-efficiency of the sub-band (fraction of time it is held) and the service rate of the band throughout that period. Then the service ratio is: $$\begin{aligned} \label{E:serviceratio} r_i = \frac{\mu_i}{\lambda} \cdot({1-p_{i\mathrm{,idle}}}),\end{aligned}$$ where $p_{i\mathrm{,idle}}$ is the probability that the sub-band $i$ is idle. The service ratio can be expressed in short-hand forms for the two extreme cases of all sub-bands being available or busy all of the time. When all sub-bands are available at the time of a transmissionthe channel of transmission is selected uniformly from the set of all channels as per step 3-a): $$\begin{aligned} \label{E:rlimit1} \lim_{\lambda\rightarrow 0}(r_i) = \dfrac{n_i}{\sum_{j=1}^cn_j}.\end{aligned}$$ In the case that all sub-bands are unavailable at the time of a transmissionthe transmission is carried out in the next available sub-band as per step 3-b): $$\begin{aligned} \label{E:rlimit2} \lim_{\lambda\rightarrow \mu_c}(r_i) = \dfrac{\delta_i}{\sum_{j=1}^c\delta_j}.\end{aligned}$$ In order to describe $r_i$ between these extremes, we must find $p_{i\mathrm{,idle}}$. Hence, we wish to find the steady-state probabilities given a Markov model of the sub-band selection behaviour. For this purpose the model of a *jockeying*[^6] $M/M/c$ queue from [@generalJockey] has been adopted. The Markov model of the jockeying queue has a limited state space since, by definition, the difference in the number of queued transmissions in any two sub-bands may not be larger than one. This allows us to put up a matrix $\mathbf{A}$ containing all state transition probabilities, which can be used to evaluate the steady state probabilities, $\mathbf{P}$, by solving the linear system $\mathbf{A}\cdot\mathbf{P} = 0$. It also allows adoption of a Markov model for LoRaWAN device behaviour, which is step 3) in the sub-band selection, by introducing state transition probabilities based on the number of channels in each non-busy sub-band in $\mathbf{A}$. The jockeying queue does has a limited state space. As in [@generalJockey] we approximate the model by making it finite by limiting the queue sizes to 1000. The model now allows us to compute the steady state probabilities of all states; Amongst them $p_{busy,all}$ and $p_{i\mathrm{,idle}}$. Then $T_{w_i}$ and $r_i$ can be calculated from Eq.  and Eq. . Note that waiting times are lower for a jockeying queue than a regular queue. Hence applying the rule of thump for approximation of an $M/D/c$ queue from a $M/M/c$ queue on a $M/M/c:jockeying$ queue, will yield a *lower* latency approximation of the $M/D/c$ queue. Multiple Devices Model: Collisions ---------------------------------- In this work we assume that no devices are making use of the optional acknowledgement feature of LoRaWAN. Hence there is no DL in the model and as another consequence no retransmissions occur upon collision. It is empirically found in [@2017arXiv170404257M] that spreading factors are not orthogonal in practice and, due to capture effect, one transmission may be received successfully if the power of the wanted transmissions is sufficiently greater than the interfering one. Unfortunately, at present there is no model of capture effect in LoRaWAN and in this work, for simplicity, we assume that all channels and all SFs are orthogonal. When two or more transmissions happen in the same channel, using the same SF, at the same time, they collide. This means we can model the access scheme as multichannel ALOHA random access, as in [@mikhaylov2016analysis; @adelantado2016understanding; @augustin2016study]. Since there are 6 spreading factors defined for LoRaWAN, we have 6 sets of $n_i$ orthogonal Aloha-channels in sub-band $i$. The collision rate must be evaluated for each spreading factor. We found the service ratios of each sub-band in Section \[S:analat\]. Since the number of devices, the transmission time for the spreading factor being evaluated and the mean inter arrival time are known, we can calculate the load within a sub-band. The load within the sub-band is spread uniformly over the channels allocated to that band. Hence the traffic load of M devices, in sub-band i, given $SF_{i,j}$ is $$\begin{aligned} L(i,j) =\frac{\lambda\cdot r_i\cdot T_{\mathrm{tx,j}}\cdot M \cdot p_{\mathrm{SF}_{i,j}}}{n_i}\end{aligned}$$ where $p_{\mathrm{SF}_{i,j}}$ is the percentage of all devices $M$, which use the $j$’th spreading factor in sub-band $i$. The collision probability is then $$\begin{aligned} \label{E:colpr} p_{\mathrm{col},i,j} = \exp\left({-2\cdot L(i,j)}\right).\end{aligned}$$ In the paper, only unacknowledged UL transmissions are considered. So DL limitations and retransmissions are not considered in this work. Therefore the outage is caused by collisions can be quantified by our model. Performance Evaluation {#sec:eval} ====================== In this section the latency given by Eq. , the service ratios given by Eq.  and the collision probability given by Eq.  are evaluated numerically. The evaluation is done for SF 12 based on 125 kHz channels, 50 bytes payload, 13 bytes overhead, code rate 4 and preamble length $n_\mathrm{preamble}=8$. The latency including the transmission time and the waiting time due to regulatory duty cycling as a function of arrival rate are depicted in Fig. \[F:allSBlag\]. The latency is plotted for stand-alone usage of each sub-band (G to G4) and for two sub-band combinations (G+G1 and G+G2). The analytical approximation using Eq.  for a heterogeneous $M/M/c$ queue provides a tight upper bound of the cases for the multiple sub-bands (G+G1 and G+G2) and a tight approximation for the single band cases. The latency obtained by the jockeying $M/M/c$ queue provides a lower approximation. The results show that lower latencies and higher capacities can be achieved for sub-bands with higher duty-cycles and combinations of bands with high duty-cycles. ![Latencies on all sub-bands and combinations of sub-bands. Results denoted *Upper* and *Lower* are calculated using ordinary $M/M/c$ model and jockeying model, respectively.[]{data-label="F:allSBlag"}](allSBlag){width=".95\columnwidth"} The service ratios for the cases with combined sub-bands are plotted in Fig. \[F:SB1SB3col\]. We see that combining G with G1 and G2, respectively, leads to very different service ratios for the bands. G contains 15 channels and G1 contains just 3, but they have the same duty-cycle. The combination of G and G1 yields the service ratio limit $15/(15+3) = .834$ for G for low arrival rates, but since the duty-cycling is the same for the sub-bands we have the limit $.01/(.01+.01)$ for a high arrival rate. The consequence of the sub-band pairing becomes evident by the collision rates depicted in Fig. \[F:allSBcol\]. We see that the collision rate for G+G1 is larger than that of G alone or G+G2. This is due to the traffic not being spread equally on the channels for high arrival rates for G+G1. Since the limits of G+G2 are much closer, the load is spread more uniformly over the channels at high arrival rates and we see a drop in collision rate by adding the sub-band. Notice that the devices reach their capacities $\mu_c$ before the collision rate comes close to 1. In this way duty-cycling limits the collision rate for each band, allowing for more devices to share the band, but in practice arrivals beyond the capacity of each device would be dropped. ![Service ratios for G+G2 and G+G1.[]{data-label="F:SB1SB3col"}](SimAnaRatioSB1andSB3){width=".95\columnwidth"} ![Sub-band collision rates for 100 devices transmitting with SF$_{12}$.[]{data-label="F:allSBcol"}](SBcol100devs){width=".95\columnwidth"} From Fig. \[F:allSBlag\] it seems that the sub-band with the highest duty-cycle, G4, is attractive as it delivers low latency even at very high loads. However, when collisions are taken into account, we see that the sub-band has a very high collision rate as it only contains a single sub-channel. On the other hand, the lowest duty-cycle is found in sub-band G2, which has relatively high latency even at low loads, but with a lower collision rate than G4. ![Effect of aggregated duty cycle on service ratios.[]{data-label="F:aggDC"}](SimAnaAggregatedDutyCycle){width=".95\columnwidth"} In Fig. \[F:aggDC\] the service ratios for G+G4 with an aggregate duty cycle of 0.05 (equivalent to a service rate capacity of the $M/D/1$ queue is 0.0146) and an aggregate duty cycle of 0.075 (equivalent to 0.0219) are plotted. The introduction of the aggregated duty cycle ($M/D/1$ queue) was found to effect the regulatory duty cycle queue ($M/D/c$ queue) by the service capacity, which freezes the sub-band service ratios of the regulatory queue and limits the obtainable latency. Concluding Remarks {#sec:conc} ================== A model for evaluating the performance of LoRaWAN UL in terms of latency and collision probability was presented. The numerical evaluation was done for EU868 ISM band regulations, but the analysis is also valid for other bands utilizing duty cycling, such as the CN779-787 ISM band. Short-hand forms for the limits of $r_i$ were presented. Equalizing the limits keeps the collision rate of sub-band combining at a minimum. The trade-off for this is a higher latency. The traffic shaping effect of aggregated duty-cycling was shown and may be used as a built-in tool for collision-latency trade-off when combining sub-bands. The UL model presented in this work, can be combined with DL models for Class A, B and C LoRaWAN devices and more sophisticated collision models to give insight into the bi-directional performance in LoRaWAN. [^1]: This work has been supported by the European Research Council (ERC Consolidator Grant Nr. 648382 WILLOW) within the Horizon 2020 Program. [^2]: All authors are with the Department of Electronic Systems, Aalborg University, Denmark (Email: {rbs,dmk,jjn,petarp}@es.aau.dk). [^3]: $\delta_i$ is a normalized value between \[0, 1\]. [^4]: https://github.com/Lora-net [^5]: The term “asymmetric" captures the heterogeneous service rates of sub-bands. [^6]: Jockeying: A packet changes queue to a shorter queue if, upon the end of service of another packet, it is located in a longer queue.

ArXiv

--- abstract: | In this paper we describe a new method for detecting and counting a repeating object in an image. While the method relies on a fairly sophisticated deformable part model, unlike existing techniques it estimates the model parameters in an unsupervised fashion thus alleviating the need for a user-annotated training data and avoiding the associated specificity. This automatic fitting process is carried out by exploiting the recurrence of small image patches associated with the repeating object and analyzing their spatial correlation. The analysis allows us to reject outlier patches, recover the visual and shape parameters of the part model, and detect the object instances efficiently. In order to achieve a practical system which is able to cope with diverse images, we describe a simple and intuitive active-learning procedure that updates the object classification by querying the user on very few carefully chosen marginal classifications. Evaluation of the new method against the state-of-the-art techniques demonstrates its ability to achieve higher accuracy through a better user experience. author: - | Inbar Huberman            Raanan Fattal\ School of Computer Science and Engineering\ The Hebrew University of Jerusalem, Israel\ bibliography: - 'counting.bib' title: Detecting Repeating Objects using Patch Correlation Analysis --- counting\_intro counting\_prev counting\_method counting\_results counting\_disc counting\_acknowledgment

ArXiv

--- abstract: | Taking precautions before or during the start of a virus outbreak can heavily reduce the number of infected. The question which individuals should be immunized in order to mitigate the impact of the virus on the rest of population has received quite some attention in the literature. The dynamics of the of a virus spread through a population is often represented as information spread over a complex network. The strategies commonly proposed to determine which nodes are to be selected for immunization often involve only one centrality measure at a time, while often the topology of the network seems to suggest that a single metric is insufficient to capture the influence of a node entirely. In this work we present a generic method based on a genetic algorithm (GA) which does not rely explicitly on any centrality measures during its search but only exploits this type of information to narrow the search space. The fitness of an individual is defined as the estimated expected number of infections of a virus following SIR dynamics. The proposed method is evaluated on two contact networks: the Goodreau’s Faux Mesa high school and the US air transportation network. The GA method manages to outperform the most common strategies based on a single metric for the air transportation network and its performance is comparable with the best performing strategy for the high school network. **Keywords:** Complex networks; Genetic algorithm; Node centrality; SIR model; Damage control author: - 'V.V. Kashirin[^1]' - 'L.J. Dijkstra' nocite: - '[@Hethcote2000]' - '[@Brandes2001]' title: A heuristic optimization method for mitigating the impact of a virus attack --- Introduction {#sec:introduction} ============ Many viruses spread through human population by means of personal contact between infectious individuals (those who carry the virus) and susceptibles (those who are not ill at the moment but can catch the disease) [@Hethcote2000; @Newman2002]. A concept that proved to be very valuable in order to gain a better understanding of the virus spread process is the complex network [@Newman2010]. Nodes within this graph structure represent individuals and are associated with a certain state (e.g., susceptible or infectious). Edges between these nodes account for social interactions. The percolation of the virus through the population (network) is then predicated by a fixed set of rules [@Newman2002; @Boccaletti2006; @Pastor-Satorras2001; @Tao2006]. This approach has not only been employed in order to better understand the dynamics of such a disease spread, but also gives rise to a field of research with a more proactive attitude: which individuals in the population should be immunized to limit the spread of the virus most effectively? Or, to frame it a bit differently, which nodes in the network should be protected in order to limit the damage done as much as possible? The concept commonly used in order to find those nodes that are in need of protection (or removed in some cases) is *node centrality*[^2] [@Holme2002; @Memon2006; @Kitsak2010; @Bright2011; @Hou2012; @Chen2012]. Node centralities express to what extent the node facilitates the spread of information over the network (note that information spread is, to some extent, alike to a virus spread [@Daley1964]). In this article we consider three common traditional variants[^3] used in the literature, each of which formalizes the concept of centrality in a (slightly) different manner: 1) *degree centrality* which is equal to the node’s degree, 2) *betweenness centrality* which focusses rather on to what extent the node could influence the communication between other nodes in the network, and 3) *eigenvector centrality* which expresses the relative importance of a node in terms of importance of its direct neighbors (a having a few connections with important nodes is more important than the one having the same number of connections with less important ones). The definitions of these centrality measures are given in section \[sec:centrality\]. The usual approach to locate those nodes that influence the spread of information the most is to determine for all nodes in the network their score on one centrality measure only. A fraction of nodes that scored the highest are then proposed to be immunized. This approach has its drawbacks. Take, for example, the network in Fig. \[fig:example\]. It consists of two clusters connected by one single node. If one would only use a ranking based on degree centralities (as it rather common in the literature), one would start by immunizing several nodes in the clusters. The node in the center of the graph is left untouched, while immunizing it at an early phase, would prohibit the virus to spread from one cluster to the other. The importance of this node would be noted only when one would take betweenness centralities into account as well. Of course, this example is rather simple and one would be able to come up with such an analysis by simply examining the graph visually. When the network gets large, however, (e.g., 500 nodes as in one of the networks examined later – see section \[sec:networks\]) it is rather hard to make these kinds of assertions and a more structural approach is required. ![A rather simple hypothetical network. The values in each node represent the node’s degree and betweenness centrality, respectively.[]{data-label="fig:example"}](Figure1.pdf){width=".6\textwidth"} In this paper we propose a genetic algorithm [@SivanandamS.N.andDeepa2007] (GA) for finding those individuals that need to be protected in order to limit the spread of the virus through the rest of the network as much as possible. This heuristic search approach differs from the methods proposed in the literature, since the GA is allowed to search through a wide range of subsets of nodes and does not rely explicitly on any kind of centrality metrics during its search; centrality measures are only used to narrow the search space. Nodes that score low on all measures of centrality most likely do not play an important role in the spread of the virus and can, thus, be neglected. In order to emulate the spread of the virus over the network we define an adapted version of one of the most common models in epidemiology: the *SIR model* [@Hethcote2000; @Newman2002; @Newman2010; @Daley1964]. The model was first introduced in the twenties by Lowell Reed and Wade Hampton Frost. They proposed to divide the entire population into the following three distinct classes: - The *susceptibles*, $S$: the group of individuals who have not been infected but can catch the disease. - The *infectives*, $I$: the group consisting of all individuals that are currently infected by the disease and could infect others from the susceptible class. - The *removed*, $R$ (sometimes referred to in the literature as recovered): those who had the virus but either recovered and gained immunity or died. Note that an individual in the model can go through three strictly sequential phases: susceptible individuals can get infected by the virus and either recover or die; removed individuals never loose their gained immunity[^4]. In section \[sec:sir\] we discuss our adaptation of this model for simulating the spread on networks. The structure of the paper is organized as follows. In the following section we discuss the methods used. First we define the optimization task at hand formally in section \[sec:formal\]. Section \[sec:networks\] introduces the two networks that are used for validation of the methods. We proceed with defining the three different centrality measures discussed earlier in the text. In section \[sec:sir\] we formally define an adaption of the traditional SIR compartmental model for network structures. Section \[sec:ga\] contains a detailed description of the genetic algorithm for finding the optimal set of nodes that need to be protected using the various centrality rankings. The results are presented in section \[sec:results\]. We finish this paper with our conclusions, some discussion and a few pointers for future research. Methods {#sec:methods} ======= The optimization task {#sec:formal} --------------------- The optimization task presented in the introduction can be expressed a bit more formally in the following manner. Suppose we are dealing with a network $G = \left<V, E\right>$ where $V$ denotes the set of vertices in the network and $E$ is the set of (bidirectional) edges between pairs of nodes in $V$. The size of the network is given by $N = |V|$. The number of nodes that can be protected is limited due to time and resource constraints to a total of $k$ nodes. We are, thus, interested in finding a $k$-subset of nodes $I \subseteq V$ that, when protected, will most effectively limit the spread of the virus through the network $G$. This task is by no means easy. The number of possible $k$-subsets of nodes is $\binom{N}{k}$ which, for most real-world networks, is simply too large to brute-force. In addition, one is generally unaware where the virus starts which needs to be accounted for when assessing the successfulness of a solution. In the following sections we describe a method to find a (near-)optimal $k$-subset of nodes. The data sets used for validation {#sec:networks} --------------------------------- In this paper we consider two contact networks to validate our approach, one social and one transportation network: 1. The Goodreau’s Faux Mesa high school network which is the result of a simulation of a in-school friendship network based on data from a high school in the rural western United States [@Resnick1997]. The network consists of $147$ nodes and $202$ undirected edges (representing mutual friendship). See Fig. \[fig:topo\]a for a visual representation of the graph. 2. The US air transportation network [@Colizza2007], which consists of $500$ nodes (US airports with the highest amount of traffic) and $2980$ undirected edges (representing air travel connections). See Fig. \[fig:topo\]b for a visual representation of the graph. Both data sets are freely available and can be found, respectively, at the website of the CASOS group[^5] and the Cx-Nets website[^6]. Please, note that the models, methods and measures discussed in the following sections can be applied (or extended) to any network structure. The networks chosen here are only used as examples. Node centrality {#sec:centrality} --------------- Measures for centrality try to formalize the (relative) importance of a node in the network. In this paper we consider the following usual measures: 1) degree, 2) betweenness and 3) eigenvector centrality. Degree centrality, $C_D$, formalizes the importance of a node by setting it equal to the number of neighbors the node has; the more connections, the more important the node: $$C_D (v) = \mathrm{deg}(v). \label{eq:d}$$ Betweenness centrality, $C_B$, tries to capture a different kind of ‘importance’. The idea behind this formalism is that the importance of a node depends on to what extent it can influence the communication between the other nodes in the network [@Newman2010]. The measure is computed by determining the fraction of the shortest paths between each pair of nodes in the network that contains the node $v$, i.e.: $$C_B (v) = \displaystyle\sum_{s\neq v\neq t \in V} \frac{\sigma_{s,t} (v)}{\sigma_{s,t}}, \label{eq:b}$$ where $\sigma_{s,t}$ is the number of the shortest paths between the nodes $s$ and $t$ and $\sigma_{s,t}(v)$ returns the number of shortest paths that passes through the vertex of interest, $v$. As one can imagine, this measure for centrality is computationally rather hard to determine. In 2001 Ulrik Brandes presented a fast version of the algorithm; computation time is upper bounded by $O(|V||E|)$. We will use this algorithm for determining the betweenness centrality throughout this paper. Eigenvector centrality, $C_E$, distinguishes itself from the other measures discussed here, since it takes the importance of its neighbors explicitly into account: $$C_E (v) = \frac{1}{\lambda} \displaystyle\sum_{w \in V} a_{v,w} C_E (w), \label{eq:e1}$$ where $a_{v,w}$ is the $(v,w)$-entry of the adjacency matrix of the network in question. Rewriting this expression to matrix form yields $$\mathbf{A}\mathbf{x} = \lambda \mathbf{x}, \label{eq:e2}$$ where the $i$-th entry of $\mathbf{x}$ is equal to $C_E(i)$. In other words, the eigenvector centrality of a node $v$ is the $v$-th entry of the right eigenvector associated with the first eigenvalue $\lambda$ of the adjacency matrix $\mathbf{A}$ of the network. An adaption of the SIR model for networks {#sec:sir} ----------------------------------------- The SIR model is one of the most often used models for simulating the spread of a virus through a population. The original model in terms of a set of ordinary differential equations makes the assumptions that 1) the population size, $N$, is large and 2) the population is ‘fully mixed’, meaning that every individual has the same number of (randomly picked) connections and have (approximately) the same number of contacts at the same time [@Hethcote2000; @Newman2002]. Of course, these assumptions are rather unsatisfactory and can be overcome by extending the SIR model to networks [@Newman2002; @Newman2010]. Suppose we have the following network $G = \left<V,E\right>$ where the set $V$ denotes the vertices of the network and $E$ is the set of bidirectional edges between pairs of vertices in $V$. Each node $v$ in the set of vertices $V$ is associated with either the state $S$ (susceptible), $I$ (infectious) or $R$ (removed) at each moment in time, i.e., $\mathrm{St}\left( v, t\right) \in \left\{S, I, R\right\}$. The initial state of the network is given by $\{S_0 , I_0 , R_0 \}$, i.e., the sets of nodes who’s state at the start of the simulation, $t = 0$, are either $S, I$ or $R$. The dynamics of the virus are emulated by updating the state of all the nodes in the network simultaneously while stepping forward in time. When the state of node $v \in V$ at moment $t$ is susceptible, i.e., $\mathrm{St}(v, t) = S$, the state of this node at the next time step is given by $$\mathrm{St}\left(v, t + 1 \right) := \begin{cases} I & \text{with probability }1 - (1 - \beta)^{| \mathcal{N}_v(I) |}, \\ S & \text{otherwise,} \end{cases} \label{eq:s}$$ where $\beta$ is the chance of the susceptible node to be infected by a single infectious neighbor. The function $\mathcal{N}_v(I)$ returns the set of infectious neighbors in the direct vicinity of node $v$. The operator $|.|$ returns the cardinality of a set. Note that the chance of *not* becoming infected decreases exponentially with the number of infectious neighbors. When node $v$ is at time $t$ infective, its state becomes $$\mathrm{St}\left(v, t + 1\right) := \begin{cases} R & \text{with probability }\gamma, \\ I & \text{otherwise,} \end{cases} \label{eq:i}$$ where $\gamma$, thus, denotes the chance of recovering (or dying) from the virus during one time step. Since we assumed that once a node reached the removed state $R$, it is either indefinitely immune for the virus or dead, the state of that node will not change till the end of the simulation. The simulation ends when there are no more infective nodes in the networks; the spread of the virus grinded to a halt and the number of susceptible and removed nodes in the network will not change. The total number of casualties is equal to the number of nodes in state $R$. Genetic algorithm {#sec:ga} ----------------- In order to find a (near-optimal) set of nodes that need to be protected from the virus we employ a genetic algorithm [@SivanandamS.N.andDeepa2007]. First, we compute the degree, betweenness and eigenvector centrality for each node in the network (see section \[sec:centrality\]). By ranking the nodes according to their centrality scores, we obtain the following three ascending rankings: $$R_D = \{v_{(1)}^D, v_{(2)}^D, \dots, v_{(N)}^D\}, \quad R_B = \{v_{(1)}^B, v_{(2)}^B, \dots, v_{(N)}^B\} \quad \text{and} \quad R_E = \{v_{(1)}^E, v_{(2)}^E, \dots, v_{(N)}^E\}, \label{eq:rankings}$$ where $R_D$, $R_B$ and $R_E$ are the rankings based on, respectively, the degree (\[eq:d\]), betweenness (\[eq:b\]) and eigenvector (\[eq:e1\]) centrality. We limit the search space of the GA by taking into account that nodes with low centrality scores on all three centrality measure do most likely not play an important role in spreading the virus over the network. We, thus, keep only the first $l < N$ nodes in each ranking, i.e., $$R'_D = \{v_{(1)}^D, v_{(2)}^D, \dots, v_{(l)}^D\}, \quad R'_B = \{v_{(1)}^B, v_{(2)}^B, \dots, v_{(l)}^B\} \quad \text{and} \quad R'_E = \{v_{(1)}^E, v_{(2)}^E, \dots, v_{(l)}^E\}. \label{eq:r}$$ The genetic algorithm will only search for an optimal $k$-subset of nodes in the union of these reduced rankings: $$R' = R'_D \cup R'_B \cup R'_E. \label{eq:reduced}$$ Each individual in the GA is represented by a chromosome with $k$ genes: $$I = \{v_1, v_2, \ldots, v_k\} \qquad \text{where }v_i \in R'. \label{eq:individual}$$ The genes, thus, represent the nodes that need to be protected from the virus where we only take into account those nodes that score relatively high on at least one of the centrality measures. In order to determine the fitness of an individual, firstly we take the original network $G = \left<V, E\right>$ and remove the nodes present in the chromosome $I$. (Removing nodes corresponds here to immunizing the nodes from the virus). Secondly, in order to account for the fact that one is normally unaware where the virus starts, we select one node at random in the new graph to be infectious while keeping the other nodes susceptible. Thirdly, we apply the SIR model as described in section \[sec:sir\] and store the number of removed individuals at the end of the simulation which we denote with $N_{\text{casualties}}$. (Recall that the simulation ends when there are no more infectious nodes in the network). Since the selection of infectious individuals is a stochastic process we repeat the last two steps $m$ times. The fitness of an individual in the GA population is then defined as the expected number of casualties: $$\text{fitness}(I) = \mathrm{E}\left[N_{\text{casualties}} \right] \approx \frac{1}{m} \displaystyle\sum_{i = 1}^{m} N_{\text{casualties}}^{(i)} \label{eq:fitness}$$ where $N_\text{casualties}^{(i)}$ is the number of recovered nodes at the end of the $i$-th simulation of the SIR model applied to the updated graph $G$. \[tab:parameters\] In summary, the objective of the GA is find that subset $I^\ast$ that minimizes the expected number of casualties[^7]: $$I^\ast = \operatorname*{arg\,min}_{I \subseteq R', |I| = k} \text{fitness}(I).$$ The genetic algorithm simulates $100$ generations each with a total of $100$ individuals. Selection for mating is performed by applying tournament selection where the tournament size is set to $4$. As crossover operator we use an adaptation of uniform crossover. Suppose we selected two parents to mate, e.g., $P_1 = \{1,2,3,4\}$ and $P_2 = \{3,4,5,6\}$. We then concatenate the chromosomes of these parents and sort the resulting array. In our example, we find $\{1,2,3,3,4,4,5,6\}$. The chromosome of the first child, $C_1$, consists of the odd entries of the array; the chromosome of the second child, $C_2$, consists of the even entries, e.g., $C_1 = \{1,3,4,5\}$ and $C_2 = \{2,3,4,6\}$. This approach guarantees the absence of duplicates in the new chromosomes. The mutation rate per gene is set to $1/k$, i.e., each chromosome undergoes on average one mutation in one gene every generation. The mutation of a gene entails that its value is randomly set to a node in the set $R'$ that is not in the individual’s chromosome already. The initial population is randomly generated except for three individuals: their chromosomes are set to the first $k$ nodes in, respectively, the degree, betweenness and eigenvector centrality rankings as given in eq. (\[eq:r\]). In addition, we keep for every generation the $10$ best performing individuals from the previous generation in order not to loose good solutions that were found earlier. Results {#sec:results} ======= Table \[tab:parameters\] provides an overview of the parameters used in the SIR model and the GA introduced earlier. In addition, it presents the parameter settings that were used for producing the results presented in this section. Fig. \[fig:corr\] depicts the correlations between the degree, betweenness and eigenvector centralities for the nodes in both the networks. The reported $R^{2}$’s are determined by applying linear regression. Note that in both cases degree centrality seems to correlate with its betweenness and eigenvector counterparts while betweenness and eigenvector centralities do not seem to correspond. It is interesting to see that degree and eigenvector centralities seem to correlate more for the air transportation network (e) than for the high school network (b). This might be due to the higher number of edges in the air transportation network (2980 against 500 nodes) in contrast to the high school network (202 edges against 147 nodes). The main point of these figures is to show that there are discrepancies between the various centralities measures, which the genetic algorithm from section \[sec:ga\] is set out to exploit in order to find a more optimal set of nodes. So, to what extent does combining centralities measures help in finding a more optimal set of nodes that need to be protected from the virus? In order to answer this question, we compare four different strategies: 1. Provide protection for the nodes with the highest degree centrality, see eq. (\[eq:d\]); 2. Provide protection for the nodes with the highest betweenness centrality, see eq. (\[eq:b\]); 3. Provide protection for the nodes with the highest eigenvector centrality, see eq. (\[eq:e1\]); 4. Provide protection for the nodes found be the genetic algorithm, see section \[sec:ga\]. In order to keep the comparison fair, each strategy is allowed to protect exactly $k$ nodes. The parameters $k$ and $l$ are set to $10$ and $100$, respectively, for the high school network and to $50$ and $200$ for the air transportation network (see Table \[tab:parameters\]). ![The topology of the Goodreau’s Faux Mesa high school network (a) and the US air transportation network (b). The nodes depicted with  are the nodes that were suggested by both the genetic algorithm and the strategy based on degree centrality. The nodes denoted with $\blacksquare$ were only selected by the GA and not the degree strategy. Symbol $\square$ denotes the nodes that were only selected by the degree strategy. []{data-label="fig:topo"}](Figure4.pdf){width="\textwidth"} ![The degree distributions of the high school and air transportation networks. The bars marked black denote the number of nodes with that particular degree that were targeted by the genetic algorithm.[]{data-label="fig:degreeDistr"}](Figure5.pdf){width=".9\textwidth"} After applying a strategy on a network, we run $500$ SIR simulations and count the number of nodes that got infected by the virus. Before each SIR run we randomly select a single node to be infectious (the others are susceptible) in order to account for the fact that one is normally unaware where the virus starts. The results are presented in Fig. \[fig:boxplot\]. The ‘no protection’ case represents the situation when no preparations were undertaken. For both networks this would have disastrous consequences. In the air transportation network, almost all nodes will definitely get infected. The genetic algorithm and the degree centrality approach seem to be preferred for both networks. The GA seems to outperform the degree strategy, but only slightly. The methods based on betweenness and eigenvector centralities do improve the situation, but are clearly not optimal. Note that the performance of these two approaches switch places between the graphs. The reason why eigenvector centralities might outperform the betweenness approach when applied to the air transportation network might lie in the fact that degree and eigenvector centrality expose a higher correlation in this graph than in the high school network, see Fig. \[fig:corr\]b and \[fig:corr\]e. The genetic algorithm outperforms the strategy based on degree centrality only in the case of the US air transportation network. For the high school network their performances are comparable. Fig. \[fig:topo\] presents the topology of both networks and the differences between the sets of nodes proposed by the GA and the degree centrality strategy. The nodes depicted with  are the ones that were suggested by both strategies. The nodes depicted with $\blacksquare$ were only selected by the GA and not the degree strategy. Symbol $\square$ denotes the nodes that were only selected by the degree strategy. Note that the GA manages to locate nodes that when protected will help to ‘separate’ clusters, i.e., it becomes harder for the virus to move from one cluster to the next. Fig. \[fig:degreeDistr\] depicts the degree distributions of both the high school and the air transportation network. The bars marked black denote the number of nodes with that particular degree that were targeted by the genetic algorithm. Note that the nodes selected by the GA are not all in the tail but also appear at the beginning and in the middle of the degree distribution. Conclusions & Discussion {#sec:conclusions} ======================== The work presented in this paper was undertaken to design and evaluate the effectiveness of a heuristic optimization method for virus spread inhibition, which, in contrast to earlier research [@Holme2002; @Memon2006; @Bright2011; @Hou2012; @Chen2012; @Kitsak2010], does not rely explicitly on any centrality measures during its search. We found that the genetic algorithm proposed in this paper seems to outperform the standard approaches in the literature. This is a remarkable feat since the genetic algorithm searches through the immense set of possible subsets of nodes and does not base its decision directly on various centrality rankings. Fig. \[fig:topo\] shows that the genetic algorithm is able to find an adequate subset of nodes that need to be protected: it is able to construct a set of nodes that not only scores high on degree, but also on betweenness centrality. The virus is, thus, restricted in its movements within a cluster (degree) but also from one cluster to the other (betweenness). The degree distributions in Fig. \[fig:degreeDistr\] show the interesting result that the nodes with the biggest influence on the spread of the virus over the network are not necessarily to be found in the tail of the degree distribution; among them there are the nodes with no high degree at all, still, by being immunized, they seem to reduce the number of infected significantly. Future research is required to identify why exactly these nodes are of importance, and, most importantly, whether there are ways to simplify the identification of these individuals. Perhaps (non-topological) node attributes can provide important clues for the selection of individuals to be immunized? Being able to locate these individuals effectively without the need for the full topology of the network will aid tremendously in mitigating the impact of the virus on a population. The developed method could be applied rather easily to a wide variety of situations that require urgent computing since the method is not limited to a specific type of network: (near-)optimal solutions can be found for both regular and heterogenous networks. In addition, one does not have to restrict to one particular centrality measure. The search space of the genetic algorithm can be seamlessly extended for various fields of applications, regardless of the type of metrics used. The presented method is recommended for topologically heterogenous networks that are characterized by the absence of strong correlations between the various node centralities and the presence of modularity. The authors expect that for homogenous networks the traditional degree centrality approach (i.e., immunizing the nodes with the highest degree) is to be preferred. In addition, the GA requires more computation time. Especially when time is pressing, one is, therefore, recommended to resort to the traditional and fast-to-compute strategies. Further research might be done to explore different sets of metrics and heuristics in order to widen or narrow the search space. In addition, it might be interesting to take edge centralities into account as well, although the authors expect that this would not make a significant difference, since edge and node centralities are often highly correlated. The parameters for the genetic algorithm were set rather intuitively. Better results might be obtained when different parameter settings are applied. In this paper we only explored one particular compartmental model, the SIR model (see section \[sec:sir\]). The method can be extended to different kind of models as well [@Hethcote2000; @Newman2010]. It would be interesting to see how the method proposed here would perform when different virus dynamics are taking into account. The applicability and performance of other optimization methods such as simulated annealing might be a subject of future research as well. This paper has shown the promise of applying a genetic algorithm in order to mitigate the impact of a virus attack on a population. The authors would like to express their hope that continuing research along the lines set out in this paper will assist in gaining a better understanding of virus inhibition, and will ultimately provide better ways to avert future epidemics. Acknowledgements {#acknowledgements .unnumbered} ================ The authors thank Prof. dr. A.V. Boukhanovsky from NRU ITMO for his insightful comments and support. This work is supported by the *Leading Scientist Program* of the Russian Federation, contract 11.G34.31.0019. [^1]: Corresponding author. Tel.: +7-965-073-0861. E-mail address: kashirin.victor@gmail.com. [^2]: Measures for edge centrality exist as well, see [@Newman2010]. [^3]: More complicated centrality measures have been proposed as well, see, for example, [@Kitsak2010; @Hou2012; @Chen2012]. [^4]: Several extensions of the models exist, e.g., the SIRS model where recovered individuals can return to the susceptible state. See the paper by H.W. Hethcote (2000) for a splendid overview of these so-called compartmental models. [^5]: http://www.casos.cs.cmu.edu/computational\_tools/datasets/external/Goodreau/index11.php [^6]: http://sites.google.com/site/cxnets/usairtransportationnetwork [^7]: Although it is more common to maximize a fitness function rather than minimizing it, we felt minimization would be more appropriate, since we are dealing with the expected number of casualties.

ArXiv

--- abstract: | It is well known, as follows from the Banach-Steinhaus theorem, that if a sequence $\left\{ y_{n}\right\} _{n=1}^{\infty}$ of linear continuous functionals in a Fréchet space converges pointwise to a linear functional $Y,$ $Y\left( x\right) =\lim_{n\rightarrow\infty}\left\langle y_{n},x\right\rangle $ for all $x,$ then $Y$ is actually continuous. In this article we prove that in a Fréchet space the continuity of $Y$ still holds if $Y$ is the *finite part* of the limit of $\left\langle y_{n},x\right\rangle $ as $n\rightarrow\infty.$ We also show that the continuity of finite part limits holds for other classes of topological vector spaces, such as *LF*-spaces, *DFS*-spaces, and *DFS*$^{\ast}$-spaces, and give examples where it does not hold. address: - | R. Estrada, Department of Mathematics\ Louisiana State University\ Baton Rouge, LA 70803\ U.S.A. - | J. Vindas, Department of Mathematics\ Ghent University\ Krijgslaan 281 Gebouw S22\ B 9000 Gent\ Belgium author: - Ricardo Estrada - Jasson Vindas title: 'A generalization of the Banach-Steinhaus theorem for finite part limits' --- [^1] Introduction ============ Let $X$ be a topological vector space over $K$, $K$ being $\mathbb{R}$ or $\mathbb{C}.$ We denote as $X^{\prime}$ the dual space, that is, the space of continuous linear functionals on $X;$ the evaluation of $y\in X^{\prime}$ on $x\in X$ will be denoted as $\left\langle y,x\right\rangle $ or as $y\left( x\right) ;$ we shall denote as $X_{\mathrm{al}}^{\prime}$ the algebraic dual of $X,$ but if $z\in X_{\mathrm{al}}^{\prime}$ we denote evalutions as $z\left( x\right) $ only. Let $\left\{ y_{n}\right\} _{n=1}^{\infty}$ be a sequence of elements of $X^{\prime}$ and suppose that$$\lim_{n\rightarrow\infty}\left\langle y_{n},x\right\rangle =Y\left( x\right) \,, \label{I.1}$$ exists for each $x\in X,$ thus defining a function $Y:X\rightarrow K.$ It is clear that $Y$ is linear, an element of $X_{\mathrm{al}}^{\prime},$ and simple examples show that $Y$ does not have to be continuous, that is, maybe $Y\notin X^{\prime}.$ However, it is well known [@Horvath; @Treves] that if $X$ is barreled, in particular if $X$ is a Fréchet space or an *LF* space, then one must have that $Y\in X^{\prime};$ this result is quite important in the theory of distributions since the usual spaces of test functions are barreled and thus (\[I.1\]) provides a method, rather frequently employed, to construct new distributions as limits. Our aim is to consider the continuity of $Y$ in case the standard[^2] *finite part* of the limit$$\mathrm{F.p.}\lim_{n\rightarrow\infty}\left\langle y_{n},x\right\rangle =Y\left( x\right) \,, \label{I.2}$$ exists for each $x\in X.$ We will show that $Y\in X^{\prime}$ in case $X$ is a Fréchet space or in case it is an inductive limit of Fréchet spaces. Naturally several distributions are defined as finite parts, so such a result would be very useful. The meaning of (\[I.2\]) is that for each $x\in X$ there is $k=k_{x} \in\mathbb{N}_{0}=\mathbb{N}\cup\{0\},$ exponents $\alpha_{1}>\cdots>\alpha_{k}>0,$ scalars $R_{\alpha_{1}}\left( x\right) ,\ldots,R_{\alpha_{k}}\left( x\right) \in K\setminus\left\{ 0\right\} ,$ and $z_{n}\left( x\right) \in K$ for $n\geq0$ such that$$\left\langle y_{n},x\right\rangle =n^{\alpha_{1}}R_{\alpha_{1}}\left( x\right) +\cdots+n^{\alpha_{k}}R_{\alpha_{k}}\left( x\right) +z_{n}\left( x\right) \,, \label{I.3}$$ where$$\lim_{n\rightarrow\infty}z_{n}\left( x\right) =Y\left( x\right) \,. \label{I.4}$$ Observe that $k_{x}$ could be $0,$ meaning that $\left\langle y_{n},x\right\rangle =z_{n}\left( x\right) $ converges to $Y\left( x\right) . $ We call $n^{\alpha_{1}}R_{\alpha_{1}}\left( x\right) +\cdots+n^{\alpha_{k}}R_{\alpha_{k}}\left( x\right) $ the *infinite* part of $\left\langle y_{n},x\right\rangle $ as $n\rightarrow\infty$ and $z_{n}\left( x\right) $ the *finite* part. Clearly the infinite and finite part, if they exist, are *uniquely determined,* so that the finite part of the limit, if it exists, is likewise uniquely determined. It is important to point out that maybe $\sup\left\{ k_{x}:x\in X\right\} =\infty$ and that the set of exponents,$$\bigcup_{x\in X}\left\{ \alpha_{j}:1\leq j\leq k_{x}\right\} \,, \label{I.5}$$ does *not* have to be finite. We shall show that when $X$ is a Fréchet space then $\sup\left\{ k_{x}:x\in X\right\} <\infty$ and actually the set (\[I.5\]) is finite, but give examples in other types of spaces where these results do not hold. We shall also show that the $R_{\alpha}$ admit extensions as elements of the algebraic dual $X_{\mathrm{al}}^{\prime},$ and show that while in general they are not continuous, they must belong to $X^{\prime}$ when $X$ is a Fréchet space or an inductive limit of Fréchet spaces. The plan of the article is as follows. In Section \[Sect: General results\] we give some basic facts about finite parts that hold in any topological vector space. The central part of the article is Section \[Section: Finite parts in a Frechet space\], where we study finite parts in a Fréchet space. Extensions to more general finite parts and to more general topological vector spaces are considered in Sections \[Section: More General Finite Parts\] and \[Section: Other types of topological vector spaces\], respectively. Finally we present several illustrations in Section \[Sect: Examples\]. General results\[Sect: General results\] ======================================== We shall first consider several results that hold in any topological vector space. Thus let $\left\{ y_{n}\right\} _{n=1}^{\infty}$ be a sequence of elements of the dual space $X^{\prime}$ of the topological vector space $X,$ and suppose that for each $x\in X$ the finite part of the limit $$Y\left( x\right) =\mathrm{F.p.}\lim_{n\rightarrow\infty}\left\langle y_{n},x\right\rangle \,, \label{G.0}$$ exists, or, in other words, that the evaluation $\left\langle y_{n},x\right\rangle $ can be decomposed as $$\left\langle y_{n},x\right\rangle =u_{n}\left( x\right) +z_{n}\left( x\right) \,, \label{G.1}$$ with the infinite part of the form $$u_{n}\left( x\right) =n^{\alpha_{1}}R_{\alpha_{1}}\left( x\right) +\cdots+n^{\alpha_{k}}R_{\alpha_{k}}\left( x\right) \,, \label{G.2}$$ where $\alpha_{1}>\cdots>\alpha_{k}>0,$ and $R_{\alpha_{1}}\left( x\right) ,\ldots,R_{\alpha_{k}}\left( x\right) \in K\setminus\left\{ 0\right\} ,$ and with the finite part, $z_{n}\left( x\right) ,$ such that the limit$$Y\left( x\right) =\lim_{n\rightarrow\infty}z_{n}\left( x\right) \label{G.3}$$ exists. The following result is very easy to prove, but it is also very important. \[Lemma 1\]The decomposition (\[G.1\]) in finite and infinite parts is unique. It is convenient to define $R_{\alpha}\left( x\right) $ for all $\alpha>0$ and all $x\in X.$ We just put $R_{\alpha}\left( x\right) =0$ if $\alpha$ is not one of the exponents $\alpha_{1},\ldots,\alpha_{k}$ in the expression of the infinite part of $\left\langle y_{n},x\right\rangle .$ This allows us to rewrite (\[G.2\]) as$$u_{n}\left( x\right) =\sum_{\alpha>0}n^{\alpha}R_{\alpha}\left( x\right) \,, \label{G.4}$$ since only a finite number of terms of the uncountable sum do not vanish. \[Lemma 2\]If $Y\left( x\right) =\mathrm{F.p.}\lim_{n\rightarrow\infty }\left\langle y_{n},x\right\rangle $ exists for all $x\in X$ then $Y$ is linear: $Y\in X_{\mathrm{al}}^{\prime}.$ Indeed, if $x_{1},x_{2}\in X$ and $c\in K,$ then $\left\langle y_{n},x_{1}+cx_{2}\right\rangle $ admits the decomposition$$\begin{aligned} \left\langle y_{n},x_{1}+cx_{2}\right\rangle & =\left\langle y_{n},x_{1}\right\rangle +c\left\langle y_{n},x_{2}\right\rangle \\ & =\left( u_{n}\left( x_{1}\right) +cu_{n}\left( x_{2}\right) \right) +\left( z_{n}\left( x_{1}\right) +cz_{n}\left( x_{2}\right) \right) \,.\end{aligned}$$ Since$$u_{n}\left( x_{1}\right) +cu_{n}\left( x_{2}\right) =\sum_{\alpha >0}n^{\alpha}\left( R_{\alpha}\left( x_{1}\right) +cR_{\alpha}\left( x_{2}\right) \right) \,, \label{G.5}$$ has the form of an infinite part, the Lemma \[Lemma 1\] yields$$u_{n}\left( x_{1}+cx_{2}\right) =u_{n}\left( x_{1}\right) +cu_{n}\left( x_{2}\right) \,,\ \ \ \ \ \ z_{n}\left( x_{1}+cx_{2}\right) =z_{n}\left( x_{1}\right) +cz_{n}\left( x_{2}\right) \,,$$ and consequently $Y\left( x_{1}+cx_{2}\right) $ equals$$\lim_{n\rightarrow\infty}z_{n}\left( x_{1}+cx_{2}\right) =\lim _{n\rightarrow\infty}\left( z_{n}\left( x_{1}\right) +cz_{n}\left( x_{2}\right) \right) \,,$$ that is, $Y\left( x_{1}\right) +cY\left( x_{2}\right) .$ If we now use the fact that $R_{\alpha}\left( x\right) =\mathrm{F.p.}\lim_{n\rightarrow\infty}n^{-\alpha}\left\langle y_{n},x\right\rangle ,$ or employ (\[G.5\]), we immediately obtain the ensuing result. \[Lemma 3\]For any $\alpha>0$ the function $R_{\alpha}$ is linear, $R_{\alpha}\in X_{\mathrm{al}}^{\prime}.$ In general $Y$ nor all $R_{\alpha}$ will not be continuous, as the Example \[Example 4\] shows. Finite parts in a Fréchet space\[Section: Finite parts in a Frechet space\] =========================================================================== We shall now consider the continuity and structure of finite parts in a Fréchet space. We start with some useful preliminary results. \[Lemma Fr.1\]Let $X$ be a Fréchet space and let $\left\{ y_{n}\right\} _{n=0}^{\infty}$ be a sequence of non zero elements of $X^{\prime}.$ Then there exists $x\in X$ such that $\left\langle y_{n},x\right\rangle \neq0$ $\forall n\in\mathbb{N}_{0}.$ Indeed, since $y_{n}\neq0$ the kernel of $y_{n},$ $F_{n}=\left\{ x\in X:\left\langle y_{n},x\right\rangle =0\right\} $ is a closed proper subspace of $X$ and thus of first category. Hence $\cup_{n=0}^{\infty}F_{n}\neq X.$ Observe that this result fails in spaces that are not Fréchet. Consider, for example, the sequence $\left\{ \delta\left( t-n\right) \right\} _{n=0}^{\infty}$ in the space $\mathcal{D}^{\prime}\left( \mathbb{R}\right) .$ Recall that a function $f:W\rightarrow V,$ where $W$ and $V$ are topological spaces, is called a Baire function of the first class if there exists a sequence of continuous functions from $W$ to $V,$ $\left\{ f_{n}\right\} _{n=0}^{\infty},$ such that $f\left( w\right) =\lim_{n\rightarrow\infty }f_{n}\left( w\right) $ for *all* elements $w\in W.$ \[Lemma Fr.2\]Let $X$ be a Fréchet space and let $A:X\rightarrow \mathbb{R}$ be a function that satisfies the following three properties: 1\. $A$ is a Baire function of the first class; 2\. $A\left( x-y\right) \leq\max\left\{ A\left( x\right) ,A\left( y\right) \right\} ;$ 3\. $A\left( cx\right) =A\left( x\right) $ if $c\neq0.$ Then $A$ is bounded above in $X$ and it actually attains its maximum. We shall first show that $A$ is bounded above. If $F$ is a subset of $X,$ denote by $M_{F}=\sup\left\{ A\left( x\right) :x\in F\right\} .$ If $U$ is a neighborhood of $0,$ then *3* yields that $M_{X}=M_{U}.$ Let now $V$ be any set with non empty interior; then $V-V$ is a neighborhood of $0$ and thus *2* yields that $M_{X}=M_{V-V}\leq M_{V}\leq M_{X}$ so that $M_{X}=M_{V}.$ Let $\alpha_{n}$ be a sequence of continuous functions from $X$ to $\mathbb{R}$ that converges to $A$ everywhere. Then $X=\bigcup_{k=0}^{\infty }\left\{ x\in X:\alpha_{n}\left( x\right) \leq k\text{ },\forall n\right\} $ so that there exists $k\in\mathbb{N}$ such that the set $V_{k}=\left\{ x\in X:\alpha_{n}\left( x\right) \leq k\text{ }\forall n\right\} $ has non empty interior. This yields that $M_{X}=M_{V_{k}}\leq k,$ so that $A$ is bounded above by $k$ in the *whole* space $X.$ We should now show that there exists $\widetilde{x}\in X$ such that $A\left( \widetilde{x}\right) =M_{X}.$ If not, the function $B\left( x\right) =1/\left( M_{X}-A\left( x\right) \right) $ satisfies the same three conditions as $A,$ and from what we have already proved, $B$ must be bounded above by some constant $\lambda>0;$ but this means that $A\left( x\right) \leq M_{X}-1/\lambda,$ for all $x\in X,$ and consequently $M_{X}\leq M_{X}-1/\lambda,$ a contradiction. We now apply the Lemma \[Lemma Fr.2\] to the study of finite parts. Indeed, let $\left\{ y_{n}\right\} _{n=1}^{\infty}$ be a sequence of elements of the dual space $X^{\prime}$ of the Fréchet space $X,$ and suppose that for each $x\in X$ the finite part of the limit $Y\left( x\right) =\mathrm{F.p.}\lim_{n\rightarrow\infty}\left\langle y_{n},x\right\rangle $ exists. If the infinite part of $\left\langle y_{n},x\right\rangle $ has the expression as a finite sum, $$u_{n}\left( x\right) =n^{\alpha_{1}}R_{\alpha_{1}}\left( x\right) +\cdots+n^{\alpha_{k}}R_{\alpha_{k}}\left( x\right) =\sum_{\alpha >0}n^{\alpha}R_{\alpha}\left( x\right) \,, \label{Fr.0}$$ where $\alpha_{1}>\cdots>\alpha_{k}>0,$ and $R_{\alpha_{1}}\left( x\right) ,\ldots,R_{\alpha_{k}}\left( x\right) \in K\setminus\left\{ 0\right\} ,$ define $A\left( x\right) =0$ if $u_{n}\left( x\right) =0$ and as$$A\left( x\right) =\alpha_{1}=\max\left\{ \alpha>0:R_{\alpha}\left( x\right) \neq0\right\} \,, \label{Fr.1}$$ otherwise. \[Lemma Fr.3\]The function $A$ is bounded above and attains its maximum in $X.$ It is enough to prove that $A$ satisfies the three conditions of the Lemma \[Lemma Fr.2\]. However, condition *1* follows from the limit formula$$A\left( x\right) =\lim_{n\rightarrow\infty}\frac{\ln\left[ \left\vert \left\langle y_{n},x\right\rangle \right\vert +1\right] }{\ln n}\,, \label{Fr.2}$$ while *2* and *3* are obvious. The Lemma \[Lemma Fr.3\] not only means that if $\widetilde{\alpha}=\max\left\{ A\left( x\right) :x\in X\right\} $ then $R_{\alpha}\left( x\right) =0$ if $\alpha>\widetilde{\alpha},$ but it also means that $R_{\widetilde{\alpha}}\neq0.$ The linear form $R_{\widetilde{\alpha}}$ is actually *continuous* as follows from the Banach-Steinhaus theorem since$$R_{\widetilde{\alpha}}\left( x\right) =\lim_{n\rightarrow\infty }n^{-\widetilde{\alpha}_{1}}\left\langle y_{n},x\right\rangle \,, \label{Fr.3}$$ for each $x\in X.$ We can then replace $y_{n}$ by $y_{n}-n^{\widetilde{\alpha}}R_{\widetilde {\alpha}}$ and apply the same ideas as above. Therefore, for *some* integers $k$ we obtain exponents $\widetilde{\alpha}_{1}=\widetilde{\alpha }>\cdots>\widetilde{\alpha}_{k}>0$ such that $R_{\alpha}\left( x\right) =0$ if $\alpha>\widetilde{\alpha}_{k},$ $\alpha\neq\widetilde{\alpha}_{j},$ $1\leq j<k,$ while $R_{\widetilde{\alpha}_{j}}$ is continuous and $R_{\widetilde {\alpha}_{j}}\neq0$ for $1\leq j\leq k.$ In principle one could think that this is possible for each $k\geq0,$ but if it were then we would obtain an infinite sequence of non zero continuous functionals $\left\{ R_{\widetilde {\alpha}_{j}}\right\} _{j=0}^{\infty}$ and the Lemma \[Lemma Fr.1\] would give us the existence of $x^{\ast}\in X$ such that $R_{\widetilde{\alpha}_{j}}\left( x^{\ast}\right) \neq0$ for all $j,$ a contradiction, since for any $x\in X$ the set $\left\{ \alpha>0:R_{\alpha}\left( x\right) \neq0\right\} $ is finite. Summarizing, we have the following result. \[Theorem Fr.1\]Let $\left\{ y_{n}\right\} _{n=1}^{\infty}$ be a sequence of elements of the dual space $X^{\prime}$ of the Fréchet space $X,$ and suppose that for each $x\in X$ the finite part of the limit $Y\left( x\right) =\mathrm{F.p.}\lim_{n\rightarrow\infty}\left\langle y_{n},x\right\rangle $ exists. Then there exists $k\in\mathbb{N},$ exponents $\widetilde{\alpha}_{1}>\cdots>\widetilde{\alpha}_{k}>0,$ and continuous non zero linear functionals $\left\{ R_{\widetilde{\alpha}_{j}}\right\} _{j=1}^{k}$ such that for all $x\in X$$$\left\langle y_{n},x\right\rangle =n^{\widetilde{\alpha}_{1}}R_{\widetilde {\alpha}_{1}}\left( x\right) +\cdots+n^{\widetilde{\alpha}_{k}}R_{\widetilde{\alpha}_{k}}\left( x\right) +z_{n}\left( x\right) \,,\label{Fr.4}$$ where the finite part $z_{n}$ is continuous for all $n$ and where $$Y\left( x\right) =\lim_{n\rightarrow\infty}z_{n}\left( x\right) \,,\label{Fr.5}$$ is also a continuous linear functional on $X.$ The only thing left to prove is the continuity of the $z_{n}$’s and the continuity of $Y.$ But the continuity of the $R_{\widetilde{\alpha}_{j}}$’s yields the continuity of the $z_{n}$’s because of (\[Fr.4\]) while the continuity of $Y$ follows from (\[Fr.5\]) and the Banach-Steinhaus theorem. More general finite parts\[Section: More General Finite Parts\] =============================================================== One can consider a general finite part limit process as follows. Let $\Lambda\cup\left\{ \lambda_{0}\right\} $ be a topological space where $\lambda_{0}\in\overline{\Lambda}\setminus\Lambda,$ and let $\left( \mathfrak{E,}\prec\right) $ be a totally ordered set. Let $\mathsf{B}=\left\{ \rho_{\alpha}\right\} _{\alpha\in\mathfrak{E}}$ be the basic infinite functions, that is, a family of functions with the following properties: 1. For each $\alpha\in\mathfrak{E,}$ $\rho_{\alpha}:\Lambda\rightarrow \left( 0,\infty\right) ,$ and $\lim_{\lambda\rightarrow\lambda_{0}}\rho_{\alpha}\left( \lambda\right) =\infty;$ 2. If $\alpha\prec\beta$ then $\rho_{\alpha}\left( \lambda\right) =o\left( \rho_{\beta}\left( \lambda\right) \right) $ as $\lambda \rightarrow\lambda_{0}.$ Let $y_{\lambda}\in K,$ where $K$ is $\mathbb{R}$ or $\mathbb{C},$ for each $\lambda\in\Lambda.$ If we can write$$y_{\lambda}=u_{\lambda}+z_{\lambda}\,,\label{MGG.1}$$ where the infinite part has the form$$u_{\lambda}=\sum_{j=1}^{k}R_{\alpha_{j}}\rho_{\alpha_{j}}\left( \lambda\right) \,,\label{MGG.2}$$ where $\alpha_{1}\succ\cdots\succ\alpha_{k},$ and $R_{\alpha_{1}},\ldots,R_{\alpha_{k}}\in K\setminus\left\{ 0\right\} ,$ and where the finite part, $z_{\lambda},$ satisfies that the limit$$Y=\lim_{\lambda\rightarrow\lambda_{0}}z_{\lambda}\label{MGG.3}$$ exists, then we say that the finite part of the limit of $y_{\lambda}$ as $\lambda\rightarrow\lambda_{0}$ with respect to $\mathsf{B}$ exists and equals $Y,$ and write[^3] $$Y=\mathrm{F.p.}_{\mathsf{B}}\lim_{\lambda\rightarrow\lambda_{0}}y_{\lambda }\,.\label{MGG.4}$$ We have considered the *standard* system $\mathsf{B}=\left\{ \rho_{\alpha}\right\} _{\alpha>0}$ where $\Lambda=\mathbb{N},$ $\lambda _{0}=\infty,$ and $\rho_{\alpha}\left( \lambda\right) =\lambda^{\alpha}.$ Naturally one can consider the same standard system for functions defined in any unbounded set $\Lambda\subset(0,\infty),$ in particular for $\Lambda =(0,\infty).$ We can also consider *Hadamard finite part* limits[^4], where the infinite basic functions are products of powers and powers of logarithms. Explicitly, let $\mathfrak{E}=[0,\infty)^{2}\setminus\left\{ \left( 0,0\right) \right\} ,$ with the order given by $\left( \alpha_{1},\beta_{1}\right) \prec\left( \alpha_{2},\beta _{2}\right) $ if $\alpha_{1}<\alpha_{2}$ or if $\alpha_{1}=\alpha_{2}$ and $\beta_{1}<\beta_{2}.$ Here $\Lambda\subset(1,\infty),$ and the basic infinite functions $\mathsf{B}=\left\{ \rho_{\left( \alpha,\beta\right) }\right\} _{\left( \alpha,\beta\right) \in\mathfrak{E}}$ are given as $\rho_{\left( \alpha,\beta\right) }\left( \lambda\right) =\lambda^{\alpha}\ln^{\beta }\lambda.$ We can also take a set $\epsilon\subset\left( 0,1\right) $ and consider limits as $\varepsilon\in\epsilon$ tends to $0.$ For standard finite part limits $\mathfrak{E}=(0,\infty)$ and the basic infinite functions are $\rho_{\alpha}\left( \varepsilon\right) =\varepsilon^{-\alpha};$ for Hadamard finite part limits, $\mathfrak{E}=[0,\infty)^{2}\setminus\left\{ \left( 0,0\right) \right\} $ and $\rho_{\left( \alpha,\beta\right) }\left( \varepsilon\right) =\varepsilon^{-\alpha}\left\vert \ln \varepsilon\right\vert ^{\beta}.$ The continuity of the finite part of the limit in Fréchet spaces, Theorem \[Theorem Fr.1\], will also hold for these more general systems of basic infinite functions. This is of course the case for standard finite limits. For Hadamard finite parts the proof can be modified as follows. Indeed, let $X$ be a Fréchet space, $\Lambda\subset(1,\infty)$ is an unbounded set and $y_{\lambda}\in X^{\prime}$ for each $\lambda\in\Lambda.$ Suppose that for each $x\in X$ the evaluation $\left\langle y_{\lambda},x\right\rangle $ can be written as $$\left\langle y_{\lambda},x\right\rangle =u_{\lambda}\left( x\right) +z_{\lambda}\left( x\right) \,,\label{MG.1}$$ where the infinite part has the ensuing form for some $k=k_{x},$$$u_{\lambda}\left( x\right) =\lambda^{\alpha_{1}}\ln^{\beta_{1}}\lambda\,R_{\left( \alpha_{1},\beta_{1}\right) }\left( x\right) +\cdots+\lambda^{\alpha_{k}}\ln^{\beta_{k}}\lambda\,R_{\left( \alpha _{k},\beta_{k}\right) }\left( x\right) \,,\label{MG.2}$$ where the exponents $\left( \alpha_{j},\beta_{j}\right) \in\mathfrak{E}$ satisfy$\ \left( \alpha_{1},\beta_{1}\right) \succ\cdots\succ\left( \alpha_{k},\beta_{k}\right) ,$ where $R_{\left( \alpha_{1},\beta_{1}\right) }\left( x\right) ,\ldots,R_{\left( \alpha_{k},\beta_{k}\right) }\left( x\right) \in K\setminus\left\{ 0\right\} ,$ and where the finite part, $z_{\lambda}\left( x\right) ,$ satisfies that the limit$$Y\left( x\right) =\lim_{\lambda\rightarrow\infty,\lambda\in\Lambda }z_{\lambda}\left( x\right) \,,\label{MG.3}$$ exists. As before, we set $R_{\left( \alpha,\beta\right) }\left( x\right) =0$ if $\left( \alpha,\beta\right) \neq\left( \alpha_{j},\beta_{j}\right) $ for $1\leq j\leq k_{x}.$ Then we have the following generalization of the Lemma \[Lemma Fr.3\]. \[Lemma GM.1\]Let$$E\left( x\right) =\left( A\left( x\right) ,B\left( x\right) \right) =\max\left\{ \left( \alpha,\beta\right) \in\mathfrak{E}:R_{\left( \alpha,\beta\right) }\left( x\right) \neq0\right\} \,. \label{MG.5}$$ Then $E$ attains its maximum, $\left( \alpha^{\ast},\beta^{\ast}\right) ,$ and $R_{\left( \alpha^{\ast},\beta^{\ast}\right) }$ is continuous and not zero. The proof of the Lemma \[Lemma Fr.3\] applies to $A,$ so there exists $\alpha^{\ast}=\max_{x\in X}A\left( x\right) .$ For this exponent $\alpha^{\ast}$ we consider the function given by $B^{\ast}\left( x\right) =0$ if $R_{\left( \alpha^{\ast},\beta\right) }\left( x\right) =0$ for all $\beta$ and otherwise by$$B^{\ast}\left( x\right) =\max\left\{ \beta:R_{\left( \alpha^{\ast},\beta\right) }\left( x\right) \neq0\right\} \,. \label{MG.6}$$ The Lemma \[Lemma Fr.2\] yields the existence of $\beta^{\ast}=\max_{x\in X}B^{\ast}\left( x\right) $ because$$B^{\ast}\left( x\right) =\lim_{\lambda\rightarrow\infty}\frac{\ln\left[ \lambda^{-\alpha^{\ast}}\left\vert \left\langle y_{\lambda},x\right\rangle \right\vert +1\right] }{\ln\ln\lambda}\,. \label{MG.7}$$ Since$$R_{\left( \alpha^{\ast},\beta^{\ast}\right) }\left( x\right) =\lim_{\lambda\rightarrow\infty}(\lambda^{-\alpha^{\ast}}\ln^{-\beta^{\ast}}\lambda)\left\langle y_{\lambda},x\right\rangle \,, \label{MG.8}$$ the continuity of $R_{\left( \alpha^{\ast},\beta^{\ast}\right) }$ follows. Therefore we obtain that *the Theorem \[Theorem Fr.1\] holds for Hadamard finite parts.* These ideas can be further generalized. Let us take $m$ positive functions $F_{1},F_{2},F_{3},\dots,F_{m}$ defined on an unbounded set $\Lambda \subset(1,\infty),$ each of them tending to $\infty$ as $\lambda \rightarrow\infty$ and such that $F_{j+1}(\lambda)=o(F_{j}^{\alpha}(\lambda))$ as $\lambda\rightarrow\infty$ for all $\alpha>0.$ We now consider $\mathfrak{E}=[0,\infty)^{m}\setminus\left\{ (0,0,\dots,0)\right\} $ with the lexicographical order $\prec.$ Set $\mathbf{F}=(F_{1},F_{2},\dots,F_{m})$ and if $\vec{\alpha}\in\mathfrak{E}$ write $\mathbf{F}^{\vec{\alpha}}=F_{1}^{\alpha_{1}}F_{2}^{\alpha_{2}}\cdots F_{m}^{\alpha_{m}}$, where $\vec{\alpha}=(\alpha_{1},\alpha_{2},\dots,\alpha_{m})$. We now choose the basic infinite functions as $\mathsf{B}=\left\{ \mathbf{F}^{\vec{\alpha}}\right\} _{\vec{\alpha}\in\mathfrak{E}}$ . If $y_{\lambda}\in X^{\prime}$ for each $\lambda\in\Lambda$, where $X$ is again a Fréchet space, and for each $x\in X$ the evaluation $\left\langle y_{\lambda},x\right\rangle $ can be decomposed as in (\[MG.1\]) where the infinite part is now taken of the form (for some $k=k_{x},$) $$u_{\lambda}\left( x\right) =\mathbf{F}^{\vec{\alpha}_{1}}(\lambda )R_{\vec{\alpha}_{1}}\left( x\right) +\cdots+\mathbf{F}^{\vec{\alpha}_{k}}(\lambda)R_{\vec{\alpha}_{1}}\left( x\right) \,,\label{MG.12}$$ with $\vec{\alpha}_{1}\succ\cdots\succ\vec{\alpha}_{k},$ and $R_{\vec{\alpha }_{1}}\left( x\right) ,\ldots,R_{\vec{\alpha}_{k}}\left( x\right) \in K\setminus\left\{ 0\right\} ,$ and where $z_{\lambda}\left( x\right) $ satisfies (\[MG.3\]), we can then define the finite part limit of $y_{\lambda}(x)$ as $Y(x)$. Defining $R_{\vec{\alpha}}(x)=0$ if $\vec{\alpha}$ does not occur in (\[MG.12\]), the proof of Lemma \[Lemma GM.1\] can be readily adapted to show that the function $\mathbf{A}(x)=\max\left\{ \vec{\alpha}\in\mathfrak{E}:R_{\vec{\alpha}}\left( x\right) \neq0\right\} $ also attains its maximum, $\vec{\alpha}^{\ast},$ and that $R_{\vec{\alpha }^{\ast}}\in X^{\prime}\setminus\{0\}$. This leads to a general version of Theorem \[Theorem Fr.1\] for finite part limits with respect to the system of infinite functions $\mathsf{B}=\left\{ \mathbf{F}^{\vec{\alpha}}\right\} _{\vec{\alpha}\in\mathfrak{E}}$ . Naturally, the Hadamard finite part corresponds to the choices $m=2$, $F_{1}(\lambda)=\lambda$, and $F_{2}(\lambda)=\log\lambda$. Other types of topological vector spaces\[Section: Other types of topological vector spaces\] ============================================================================================= The continuity of finite part limits holds not only in Fréchet spaces, but in other types of spaces, those that carry a final locally convex topology given by a family of Fréchet spaces [@Horvath]. Indeed, let $(X_{i},u_{i})_{I}$ be a family of Fréchet spaces and linear mappings $u_{i}:X_{i}\rightarrow X$ for each $i\in I$. If $X$ is provided with the finest locally convex topology that makes all mappings $u_{i}$ continuous, the continuity of the finite part limits and the $R_{\vec{\alpha}}$’s follows at once from the fact that $y\in X^{\prime}$ if and only if $y\circ u_{i}\in X_{i}^{\prime}$, $\forall i\in I$. In particular, the result holds for *any* inductive limit of an inductive system of Fréchet spaces. Important instances of such inductive limits are those that can be written as countable inductive unions of Fréchet spaces, such as the *LF*-spaces, the *DFS*-spaces, and the *DFS*$^{\ast}$-spaces [@komatsu]. In these more general spaces, however, the set of exponents for which $R_{\vec{\alpha}}\left( x\right) \neq0$ does not have to be finite, not bounded, in general (see Examples \[Example 2\], \[Example 3\], and \[Example 3.5\] below). \[Theorem other 1\]Let $X$ be a locally convex space that is the inductive limit of a system of Fréchet spaces. Let $\Lambda$ be an unbounded subset of $\left( 1,\infty\right) $ and for each $\lambda\in\Lambda$ let $y_{\lambda}\in X^{\prime}.$ Suppose that for each $x\in X$ the finite part of the limit $Y\left( x\right) =\mathrm{F.p.}\lim_{\lambda\rightarrow\infty }\left\langle y_{\lambda},x\right\rangle $ exists. Then $Y\in X^{\prime}.$ Likewise, $R_{\vec{\alpha}}$ is continuous for each $\vec{\alpha},$ but while for each $x\in X$ the set $\left\{ \vec{\alpha}:R_{\vec{\alpha}}\left( x\right) \neq0\right\} $ is finite, the set $\left\{ \vec{\alpha}:R_{\vec{\alpha}}\neq0\right\} $ could be infinite and not bounded above. Examples\[Sect: Examples\] ========================== In order to better understand our results, it is useful to look at several examples. \[Example 1\]The best known example of finite parts are the distributions constructed as the finite part of divergent integrals [@GreenBook; @Kanwal]. Suppose $G$ is a homogenous continuous function in $\mathbb{R}^{d}\setminus\left\{ \mathbf{0}\right\} ,$ homogenous of degree $\lambda \in\mathbb{R}.$ Then $G$ gives a well defined distribution of the space $\mathcal{D}^{\prime}\left( \mathbb{R}^{d}\setminus\left\{ \mathbf{0}\right\} \right) ,$ which without loss of generality we can still denote by $G,$ as $\left\langle G,\phi\right\rangle =\int_{\mathbb{R}^{d}\setminus \left\{ \mathbf{0}\right\} }G\left( \mathbf{x}\right) \phi\left( \mathbf{x}\right) \,\mathrm{d}\mathbf{x},$ for $\phi\in\mathcal{D}\left( \mathbb{R}^{d}\setminus\left\{ \mathbf{0}\right\} \right) .$ When $\lambda<-d$ then the integral $\int_{\mathbb{R}^{d}}G\left( \mathbf{x}\right) \phi\left( \mathbf{x}\right) \,\mathrm{d}\mathbf{x},$ would be divergent in general if $\phi\in\mathcal{D}\left( \mathbb{R}^{d}\right) $ and thus there is no canonical distribution corresponding to $G$ in $\mathcal{D}^{\prime}\left( \mathbb{R}^{d}\right) .$ One can, however, define the distribution $\mathrm{F.p.}\left( G\right) ,$ the *radial* finite part[^5] of $G$ by setting for $\phi\in\mathcal{D}\left( \mathbb{R}^{d}\right) $$$\left\langle \mathrm{F.p.}\left( G\right) ,\phi\right\rangle =\mathrm{F.p.}\lim_{n\rightarrow\infty}\int_{\left\vert \mathbf{x}\right\vert \geq 1/n}G\left( \mathbf{x}\right) \phi\left( \mathbf{x}\right) \,\mathrm{d}\mathbf{x}\,, \label{Ex.1}$$ a standard finite part if $-\lambda\notin\mathbb{N},$ and a Hadamard finite part if $\lambda$ is an integer. Similar ideas are needed to construct *thick* distributions from locally integrable functions [@YE2]. \[Example 2\]Let $\mathbb{D}=\left\{ z\in\mathbb{C}:\left\vert z\right\vert <1\right\} $ be the unit disc in $\mathbb{C}.$ Let $H_{k}$ be the Banach space of functions continuous in $\overline{\mathbb{D}}\setminus\left\{ 0\right\} ,$ analytic in $\mathbb{D}\setminus\left\{ 0\right\} ,$ and that have a pole at $z=0$ of order $k;$ the norm being $\left\Vert f\right\Vert =\max_{\left\vert z\right\vert \leq1}\left\vert z\right\vert ^{k}\left\vert f\left( z\right) \right\vert .$ Let $H$ be the inductive limit of the $H_{k}$ as $k\rightarrow\infty.$ Consider the functionals $y_{n}\in H^{\prime}$ given as$$\left\langle y_{n},f\right\rangle =f\left( 1/n\right) \,,\label{Ex.2}$$ that is, $y_{n}=\delta\left( z-1/n\right) .$ For each $f\in H$ the finite part of the limit $\mathrm{F.p.}\lim_{n\rightarrow\infty}\left\langle y_{n},f\right\rangle =Y\left( f\right) $ exists, and equals the finite part of $f$ at $z=0;$ in fact, if $f\left( z\right) =\sum_{j=1}^{k}a_{j}z^{-j}+g\left( z\right) ,$ where $g$ is analytic at $0,$ then$$\left\langle y_{n},f\right\rangle =\sum_{j=1}^{k}a_{j}n^{j}+g\left( 1/n\right) \,.\label{Ex.3}$$ Observe that the infinite part is $\sum_{j=1}^{k}a_{j}n^{j},$ which has arbitrary large exponents; here the set (\[I.5\]) is infinite. Also $Y\left( f\right) =g\left( 0\right) ,$ the usual finite part of the analytic function at the pole. Our results will yield the continuity of $Y,$ but one can prove this directly, for example, by observing that $Y\left( f\right) =\left( 2\pi i\right) ^{-1}{\displaystyle\oint_{\left\vert z\right\vert =r}} z^{-1}f\left( z\right) \,\mathrm{d}z$ for any $r\in(0,1].$ Interestingly, if $X$ is the space of all analytic function in $\mathbb{D}\setminus\left\{ 0\right\} $ with its standard topology, then $H$ is dense in $X$ and the $y_{n}$’s and $Y$ admit continuous extensions to $X^{\prime}, $ but the extension of $Y$ is *not* the finite part of the limit of the extensions of the $y_{n}$’s. \[Example 3\]Let us consider the distributions $$f_{n}\left( x\right) =\sum_{k=1}^{\infty}\left( n^{1/k}+1\right) ^{k^{2}}\delta\left( x-k\right) \,, \label{Ex.4}$$ of the space $\mathcal{D}^{\prime}\left( \mathbb{R}\right) .$ If $\phi \in\mathcal{D}\left( \mathbb{R}\right) $ satisfies $\operatorname*{supp}\phi\subset(-\infty,k+1)$ then the infinite part of $\left\langle f_{n},\phi\right\rangle $ is the sum of $k^{2}-1$ terms, corresponding to the exponents $\alpha_{j}=j/k$ for $1\leq j\leq k^{2}.$ Hence the finite part of the limit is the Dirac comb $$\mathrm{F.p.}\lim_{n\rightarrow\infty}f_{n}\left( x\right) =\sum _{k=1}^{\infty}\delta\left( x-k\right) \,, \label{Ex.5}$$ while $R_{\alpha}\left( x\right) \neq0$ precisely when $\alpha$ is a positive rational number; and actually if $k$ is the smallest integer for which $\alpha=j/k$ and $j\leq k^{2},$ then$$R_{j/k}\left( x\right) =\sum_{q=1}^{\infty}\binom{q^{2}k^{2}}{q\,j}\delta\left( x-qk\right) \,. \label{Ex.6}$$ One could represent the infinite part of $f_{n}\left( x\right) $ as the infinite sum $\sum_{\alpha\in\mathbb{Q}_{+}}R_{\alpha}\left( x\right) ;$ the set of exponents $\alpha$ for which $R_{\alpha}\neq0$ is the infinite unbounded set $\mathbb{Q}_{+},$ but upon evaluation on a test function the sum becomes finite since $\left\langle R_{\alpha},\phi\right\rangle \neq0$ for only a finite set of exponents. In the previous example $X=\mathcal{D}\left( \mathbb{R}\right) $ is an *LF* space, and it is not hard to see that in an *LF* space the set of exponents $\alpha$ for which $R_{\alpha}\neq0$ is countable at the most. We can easily construct an example where this set of exponents is the whole $\left( 0,\infty\right) .$ \[Example 3.5\]Let $X$ be the space of functions $f:\left( 0,\infty \right) \rightarrow\mathbb{R}$ such that the set $\left\{ \alpha\in\left( 0,\infty\right) :f\left( \alpha\right) \neq0\right\} $ is finite. We give $X$ the inductive limit topology of the system $\left( \mathbb{R}^{F},i_{F}\right) ,$ where $F$ is a finite subset of $\left( 0,\infty\right) ,$ $F\nearrow,$ and if $f\in\mathbb{R}^{F},$ then $i_{F}\left( f\right) =f_{F}\in X$ is given by $f_{F}\left( \alpha\right) =f\left( \alpha\right) $ if $\alpha\in F$ and $f_{F}\left( \alpha\right) =0$ if $\alpha\notin F.$ Theorem \[Theorem other 1\] applies in $X.$ Let $y_{n}\in X^{\prime}$ be given as $$y_{n}\left( x\right) =\sum_{\alpha\in\left( 0,\infty\right) }n^{\alpha }\delta\left( x-\alpha\right) \,, \label{Ex.3.5.1}$$ that is $\left\langle y_{n},f\right\rangle $ is the *finite* sum $\sum_{\alpha\in\left( 0,\infty\right) }n^{\alpha}f\left( \alpha\right) .$ Then $R_{\alpha}\left( x\right) =\delta\left( x-\alpha\right) $ for all $\alpha\in\left( 0,\infty\right) ,$ so that the set (\[I.5\]) is the whole $\left( 0,\infty\right) .$ Notice also that $\mathrm{F.p.}\lim _{n\rightarrow\infty}y_{n}\left( x\right) =0.$ \[Example 4\]Consider the space $X$ whose elements are the continuous functions in $\left[ 0,1\right] ,$ with the topology of pointwise convergence on $\left[ 0,1\right] .$ If $0<\beta<1,$ let us consider the functional $f_{n}\in X^{\prime}$ given by$$f_{n}\left( x\right) =\sum_{k=0}^{n-1}\delta\left( x-\frac{k+\beta}{n}\right) \,, \label{Ex.7}$$ that is,$$\left\langle f_{n},\phi\right\rangle =\sum_{k=0}^{n-1}\phi\left( \frac{k+\beta}{n}\right) \,, \label{Ex.8}$$ for $\phi\in X.$ Then the Euler-Maclaurin formula [@Estrada97] yields$$\left\langle f_{n},\phi\right\rangle =n\int_{0}^{1}\phi\left( x\right) \,\mathrm{d}x+B_{1}\left( \beta\right) \left( \phi\left( 1\right) -\phi\left( 0\right) \right) \,, \label{Ex.9}$$ where $B_{1}\left( x\right) =x-1/2$ is the Bernoulli polynomial of order $1.$ The finite part is$$\mathrm{F.p.}\lim_{n\rightarrow\infty}f_{n}\left( x\right) =B_{1}\left( \beta\right) \left( \delta\left( x-1\right) -\delta\left( x\right) \right) \,, \label{Ex.10}$$ which is actually continuous in $X,$ but $R_{1},$ the coefficient of $n$ in the infinite part of $\left\langle f_{n},\phi\right\rangle $ is not continuous since it is given by $\phi\rightsquigarrow\int_{0}^{1}\phi\left( x\right) \,\mathrm{d}x,$ which belongs to $X_{\mathrm{al}}^{\prime}$ but not to $X^{\prime}.$ \[Example 5\]Let $\Omega$ be a complex region and let $\xi\in\Omega.$ Suppose $y_{\omega}\in X^{\prime}$ is weakly–$\ast$ analytic in $\omega \in\Omega\setminus\left\{ \xi\right\} ,$ that is, for each $x\in X$ the function $\left\langle y_{\omega},x\right\rangle $ is analytic in $\Omega\setminus\left\{ \xi\right\} .$ Suppose also that $\left\langle y_{\omega},x\right\rangle $ has a pole at $\omega=\xi$ for each $x.$ If $X$ is a Fréchet space then there exists a fixed number $N$ such that the order of the pole is $N$ at the most for all $x,$ and the finite part $$y_{\xi}^{\ast}\left( x\right) =\mathrm{F.p.}\lim_{n\rightarrow\infty }\left\langle y_{\xi+1/n},x\right\rangle \,, \label{Ex.11}$$ is an element of $X^{\prime}.$ If $X$ is an inductive limit of Fréchet spaces, then $y_{\xi}^{\ast}$ is still continuous, but the order of the pole of $\left\langle y_{\omega },x\right\rangle $ at $\xi$ does not have to be bounded, that is, maybe $R_{k}\left( x\right) =\mathrm{F.p.}\lim_{n\rightarrow\infty}n^{-k}\left\langle y_{\xi+1/n},x\right\rangle $ does not vanish in $X$ for an infinite number of values of $k.$ . [99]{} Estrada, R., On the Euler-Maclaurin asymptotic formula, *Bol. Soc. Mat. Mex.* **3** (1997), 117–133. Estrada, R. and Kanwal, R. P., Regularization and distributional derivatives of $\left( x_{1}^{2}+\cdots x_{p}^{2}\right) ^{n/2}$ in $\mathcal{D}^{\prime}\left( \mathbb{R}^{p}\right) ,$ *Proc. Roy. Soc. London A* **401** (1985), 281–297. Estrada, R. and Kanwal, R. P., Regularization, pseudofunction, and Hadamard finite part, *J. Math. Anal. Appl.* **141** (1989), 195–207. Estrada, R. and Kanwal, R.P., *A distributional approach to Asymptotics. Theory and Applications,* Second Edition, Birkhäuser, Boston, 2002. Farassat, F., *Introduction to generalized functions with applications in aerodynamics and aeroacoustics*, NASA Technical Paper 3248 (Hampton, VA: NASA Langley Research Center) (1996); http://ntrs.nasa.gov. Frahm, C.P., Some novel delta-function identities, *Am. J. Phys.* **51** (1983), 826–829. Hadamard, J., *Lectures on Cauchy’s Problem in Linear Partial Differential Equations,* Yale University Press, New Haven, 1923, (reprinted by Dover, New York, 1953). Hnizdo, V., Generalized second-order partial derivatives of $1/r,$ *Eur. J. Phys.* **32** (2011), 287–297. Horváth, J., *Topological Vector Spaces and Distributions*, vol. I., Addison-Wesley, Reading, Massachusetts, 1966. Kanwal, R.P., *Generalized Functions: Theory and Technique,* Third Edition, Birkhäuser, Boston, 2004. Komatsu, H., Projective and injective limits of weakly compact sequences of locally convex spaces, *J. Math. Soc. Japan* **19** (1967), 366–383. Trèves, F., *Topological Vector Spaces, Distributions, and Kernels,* Academic Press, New York, 1967. van der Corput, J. G., Neutrices, *J. Soc. Indust. Appl. Math.*, **7** (1959), 253–279. Yang, Y. and Estrada, R., Regularization using different surfaces and the second order derivatives of $1/r$, *Appl. Anal.* **92** (2013), 246–258. Yang, Y. and Estrada, R., Distributions in spaces with thick points, *J. Math. Anal. Appl.* **401** (2013), 821–835. [^1]: R. Estrada gratefully acknowledges support from NSF, through grant number 0968448. [^2]: We shall consider more complicated finite part limits later on. [^3]: Let $V$ be the vector space of all functions of the form $\sum_{j=1}^{k}c_{j}\rho_{\alpha_{j}}+\mu$ with $\lim_{\lambda \rightarrow\lambda_{0}}\mu(\lambda)=0$. The triple $N=(\Lambda,K,V)$ forms a *neutrix* in the sense of van der Corput [@Corput]. In his terminology, the finite part limit (\[MGG.4\]) coincides with the neutrix value $y_{N}$. [^4]: Hadamard was probably the first to use finite parts; in his 1923 work [@Hadamard], he employs them to find fundamental solutions of partial differential equations. [^5]: If instead of removing *balls* of small radius, solids of other shapes are removed one obtains a different finite part distribution [@Farassat; @VH1; @YE], an important fact in the numerical solution of integral equations [@Farassat]. The known formulas for the distributional derivatives of inverse power fields [@Frahm] and the corresponding finite parts [@EK85b; @EK88] hold for *radial* finite parts.

ArXiv

--- address: | Laboratoire de Physique Théorique,\ CEA/Saclay, Orme des Merisiers,\ 91191 Gif sur Yvette, FRANCE\ E-mail: reinosa@spht.saclay.cea.fr author: - 'U. Reinosa' title: Resummation in hot field theories --- Introduction ============ Self-consistent approximation schemes based on two particle irreducible functionals have proved to be successful for studying systems where the quasi-particle description gives a good understanding of the relevant degrees of freedom. This is for example the case for thermodynamic properties of the quark-gluon plasma \[1\]. These resummations schemes have also been recently applied to study the dynamics of quantum fields out of equilibrium \[2\]. In these schemes, any physical quantity is expressed in terms of the full propagator and the approximations used to compute the physical quantity and the propagator are “self-consistently” correlated with each other. For instance, for applications to thermodynamics, one starts by expressing the thermodynamic potential in terms of the full propagator \[3,4,5\]. A central quantity is then the sum of the 2 PI skeleton diagrams: $\Phi[D]$. The full self-energy $D$ is then given by: $$\label{eq:gap} \Pi=2\frac{\delta \Phi}{\delta D},$$ which, together with Dyson’s equation, defines the full propagator in a self-consistent manner. One of the main questions to be addressed before any numerical computation is that of renormalizability. Via the self-consistent equation for the self-energy, one is effectively resumming an infinite set of diagrams together with their UV singularities. One has then to devise a procedure where all the UV singularities disappear at once. The problem becomes even worse at finite temperature. In fact, since the UV singularities depend on the self-energy itself, when turning on temperature they become $T$-dependent. A major progress was recently achieved by H. van Hees and J. Knoll \[6,7\], who showed, in the real time formalism, how to renormalize in such a way that temperature dependent counterterms never appear. In a recent work \[8\], we have extended the analysis of van Hees and Knoll using the imaginary time formalism, and clarifying some aspects of renormalization of the vacuum sector giving an explicit construction of counterterms. In this short communication, we restrict ourselves to a simple model where only parts of the difficulties dealt with in \[6,7,8\] appear. One loop approximation in a scalar toy model ============================================ We work in a scalar theory with $\phi^3$ and $\phi^4$ interactions in $d=4$ space-time dimensions. To two loop order, the $\Phi$ functional is given in figure \[fig:approx\]. The gap equation is obtained by cutting one line on each diagram. We then obtain a one loop approximation for the self-energy. The diagrams entering this equation are given by the following sum-integrals: $$\Pi_{tad}=\frac{\lambda_0}{2}\int_{\omega_n,p}\hspace{-0.8cm}\Sigma D(i\omega_n,p),$$ for the tadpole diagram, and $$\Pi_{eye}(i\omega,k)=-\frac{g^2}{2}\int_{\omega_n,p}\hspace{-0.8cm}\Sigma D(i\omega_n,p)D(i\omega-i\omega_n,k-p),$$ for the eye diagram. In four dimensions, these integrals are divergent. One expects the effects of the $\phi^3$ interactions (here $\Pi_{eye}$) to be super-renormalizable, while those of the $\phi^4$ interactions ($\Pi_{tad}$) to be just renormalizable. The ultraviolet singularities should be absorbed in the usual counterterms $\delta Z$, $\delta m^2$ and $\delta \lambda$. Those have to be defined in the vacuum and be compatible with the underlying perturbative diagrams that are resummed by the equation. Large momentum behaviour of $\Pi(k)$ ------------------------------------ To understand the UV structure of the gap equation, one first studies the large momentum behavior of the self energy: - the contribution from the tadpole is a T-dependent mass: $m_{\infty}^2$; - the eye contribution consists in a vacuum piece which is logarithmic and a finite temperature piece which behaves like $\frac{T^2}{k^2}$. This last piece plays no role in the analysis of UV singularities and can be ignored in what follows; $$\Pi(k) \sim m_{\infty}^2(T)+m^2\log k^2 \mbox{ , for large k.}$$ In particular, there are no corrections in $k^2$ to $\Pi$, so that $\delta Z=0$. Analysis of divergences ----------------------- We first split each integral into a vacuum piece and a thermal piece: $$\begin{aligned} \Pi_{tad} & = & \frac{\lambda_0}{2}\int_{T=0}(dp)\frac{1}{p^2+m^2+\Pi}+\Pi^T_{tad},\\ \Pi_{eye} & = & -\frac{g^2}{2}\int_{T=0}(dp)\frac{1}{(p^2+m^2+\Pi)((k-p)^2+m^2+\Pi)}+\Pi^T_{eye}\nonumber.\end{aligned}$$ $\Pi^T_{tad}$ and $\Pi^T_{eye}$ are thermal 1 loop integrals and so are finite in the UV. The explicit integrals are divergent and depend on the temperature via the self-energy. The eye diagram is logarithmically divergent, but the dominant behavior ot the integrand does not depend on $\Pi$. Therefore, the divergence is independent of T, and can be removed by the standard one-loop counter-term of the vacuum perturbation theory: $$\delta m_{eye}^2=-\frac{g^2}{32\pi^2}\frac{1}{\varepsilon}.$$ The tadpole diagram is quadratically divergent and thus sensitive to the subleading behavior of the propagator, that is, to the self-energy $\Pi$. This is the origin of the $T$-dependent singularities. We can in fact be more precise: $T$-dependent singularities only arise from the insertion of the tadpole itself. This is because the insertion ot the dominant piece of the eye diagram ($\Pi_{eye}^{(0)}$) (see figure \[fig:eye\]) =0.5in leads to a $T$-independent singularity, i.e., a new contribution to the mass counterterm. The equation for the tadpole can then be written as follows: $$\begin{aligned} \label{eq:gap} \Pi_{tad} & = & \lambda_0\int_{T=0}(dp)\left(\frac{1}{p^2+m^2}-\frac{\Pi_{tad}+\Pi_{eye}^{(0)}}{(p^2+m^2)^2}\right)\nonumber\\\nonumber\\ & + & \lambda_0 F(\Pi)+\delta m^2_{tad},\end{aligned}$$ where $F(\Pi)$ is a finite functional of $\Pi$. $\delta m_{tad}^2$ absorbs usual mass singularities and the new one arising from the eye insertion: $$\label{eq:ct} \delta m^2_{tad}=\lambda_0\int_{T=0}(dp)\left(\frac{1}{p^2+m^2}-\frac{1}{p^2+m^2}\Pi_{eye}^{(0)}\frac{1}{p^2+m^2}\right).$$ The remaining terms need much more attention as they carry the $T$-dependent singularities. We know from the $g=0$ case that the divergence proportionnal to $\Pi_{tad}$ is absorbed in the bare coupling constant. In order to renormalize the ensuing equation, namely (cf. eqs (\[eq:gap\]) and (\[eq:ct\])): $$\Pi_{tad}\left\{\frac{1}{\lambda_0}+\int_{T=0}(dp)\frac{1}{\left(p^2+m^2\right)^2}\right\}=F(\Pi),$$ we set $$\frac{1}{\lambda}=\frac{1}{\lambda_0}+\int_{T=0}(dp)\frac{1}{\left(p^2+m^2\right)^2},$$ which is $T$-independent. Thus we have managed to renormalize the gap equation by avoiding $T$-dependent counterterms. The renormalized gap equations read: $$\begin{aligned} \tilde{\Pi}_{eye} & = & \Pi_{eye}+\delta m^2_{eye},\\ \Pi_{tad} & = & g^2F(\Pi_{tad}+\tilde{\Pi}_{eye}).\end{aligned}$$ First numerical computations ============================ With this finite set of equations, we can evaluate the full self-energy in this particular approximation and compare it to perturbative results. This is shown in figure \[fig:iter\]. Note that, for this particular model, the difference between the perturbative and the self-consistent results remains rather small because the well known infrared instability induced by $\phi^3$ interactions forces us to restrict ourselves to a very limited range of parameters. But the results in figure \[fig:iter\] demonstrate that, with our method the non local numerical problem is fully under control, and encourage us to use this method to address similar problems which are more directly relevant for physic, like QCD. New questions will arise, related to gauge invariance (see for example \[9\]). =1.6in =1.6in \[fig:iter\] Conclusion ========== We have shown in a simple model how to perform renormalization of the self-consistent approximation in the presence of non-local effects. [0]{} J. P. Blaizot, E. Iancu and A. Rebhan, Phys. Rev. D [**63**]{}, 065003 (2001) G. Aarts, D. Ahrensmeier, R. Baier, J. Berges\ and J. Serreau, Phys. Rev. D [**66**]{}, 045008 (2002) J.m. Luttinger and J.C. Ward, [*Phys. Rev.*]{} [**118**]{}, 1417 (1960) C. de Dominicis and P.C. Martin, J., [*J. Math. Phys.*]{} [**5**]{}, 14, 31 (1964) G.Baym, J., [*Phys. Rev.*]{} [**127**]{}, 1391 (1962) H. van Hees and J. Knoll, Phys. Rev. D [**65**]{}, 025010 (2002) H. Van Hees and J. Knoll, Phys. Rev. D [**65**]{}, 105005 (2002) J.-P. Blaizot, E. Iancu and U. Reinosa arXiv:hep-ph/0301201 A. Arrizabalaga and J. Smit, Phys. Rev. D [**66**]{}, 065014 (2002)

ArXiv

--- address: | Enric Nart\ Departament de Matemàtiques, Universitat Autònoma de Barcelona, E-08193 Bellaterra, Cerdanyola del Vallès, Barcelona\ Catalonia, Spain author: - Enric Nart title: On the equivalence of types --- [^1] Un type sur un corps de valuation discrète $(K,v)$ est un objet computationnel qui paramètrise une famille de polynômes unitaires irréductibles sur $K_v[x]$, où $K_v$ est le complété de $K$. Deux types sont équivalents s’ils determinent la même famille de polynômes irréductibles sur $K_v[x]$. Dans ce travail, nous donnons différentes caractérisations de la notion d’équivalence de types par rapport à certaines données et des opérateurs qui leur sont associés. Types over a discrete valued field $(K,v)$ are computational objects that parameterize certain families of monic irreducible polynomials in $K_v[x]$, where $K_v$ is the completion of $K$ at $v$. Two types are considered to be equivalent if they encode the same family of prime polynomials in $K_v[x]$. In this paper, we find diferent characterizations of the equivalence of types in terms of certain data and operators associated with them. Introduction ============ In the 1920’s, Ø. Ore developed a method to construct the prime ideals of a number field dividing a given prime number $p$, in terms of a defining polynomial $f\in{\mathbb Z}[x]$ satisfying a certain *$p$-regularity* condition [@ore1; @ore2]. The idea was to detect a $p$-adic factorization of $f$ from the factorization of certain residual polynomials over finite fields, attached to the sides of a Newton polygon of $f$. He raised then the question of the existence of a procedure to compute the prime ideals in the general case, based on the consideration of similar Newton polygons and residual polynomials “of higher order". S. MacLane solved this problem in 1936 in a more general context [@mcla; @mclb]. For any discrete valuation $v$ on an arbitrary field $K$, he described all valuations extending $v$ to the rational function field $K(x)$. Starting from the Gauss valuation $\mu_0$, MacLane constructed *inductive* valuations $\mu$ on $K(x)$ extending $v$, by the concatenation of augmentation steps $$\label{depth0} \mu_0\ \stackrel{(\phi_1,\nu_1)}\lra\ \mu_1\ \stackrel{(\phi_2,\nu_2)}\lra\ \cdots\ \lra\ \mu_{r-1} \ \stackrel{(\phi_{r},\nu_{r})}\lra\ \mu_{r}=\mu,$$ based on the choice of certain *key polynomials* $\phi_i\in K[x]$ and positive rational numbers $\nu_i$. Then, given an irreducible polynomial $f\in K[x]$, he characterized all extensions of $v$ to the field $L:=K[x]/(f)$ as limits of sequences of inductive valuations on $K(x)$ whose value at $f$ grows to infinity. In the case $K={\mathbb Q}$, Ore’s $p$-regularity condition is satisfied when all valuations on the number field $L$ extending the $p$-adic valuation are sufficiently close to an inductive valuation on $K(x)$ which may be obtained from $\mu_0$ by a single augmentation step. In 1999, J. Montes carried out Ore’s program in its original formulation [@HN; @Mo]. He introduced *types* as computational objects which are able to construct MacLane’s valuations and the higher residual polynomial operators foreseen by Ore. These ideas made the whole theory constructive and well-suited to computational applications, and led to the design of several fast algorithms to perform arithmetic tasks in global fields [@algorithm; @newapp; @bases; @GNP; @Ndiff]. In 2007, M. Vaquié reviewed and generalized MacLane’s work to non-discrete valuations. The introduction of the graded algebra $\ggm$ of a valuation $\mu$ led him to a more elegant presentation of the theory [@Vaq]. In the papers [@Rid] and [@gen], which deal only with discrete valuations, the ideas of Montes were used to develop a constructive treatment of Vaquié’s approach, which included the computation of generators of the graded algebras and a thorough revision and simplification of the algorithmic applications. In this paper we fill a gap concerning the notion of *equivalence of types*. Let $\oo\subset K$ be the valuation ring of $v$ and ${\mathbb F}$ its residue class field. A *type over* $(K,v)$ is an object carrying certain data distributed into several levels: $$\ty=(\psi_0;(\phi_1,\nu_1,\psi_1);\dots;(\phi_r,\nu_r,\psi_r)).$$ The pairs $\phi_i,\nu_i$ determine an inductive valuation $\mu_\ty:=\mu$ as in (\[depth0\]), and $\psi_i\in{\mathbb F}_i[y]$ are monic irreducible polynomials building a tower of finite extensions of ${\mathbb F}$: $${\mathbb F}={\mathbb F}_0\lra {\mathbb F}_1\lra \cdots\lra {\mathbb F}_r,\quad\ {\mathbb F}_{i+1}:={\mathbb F}_i[y]/(\psi_i),\ 0\le i<r.$$ These data facilitate a recurrent procedure to construct *residual polynomial operators*: $$R_i\colon K[x]\lra {\mathbb F}_i[y],\qquad 0\le i\le r,$$ having a key role in the theory. The last polynomial $\psi_r$ determines a maximal ideal $\ll_\ty$ of the piece of degree zero of the graded algebra $\gg(\mu_\ty)$. Two types are said to be equivalent when they yield the same pair $(\mu_\ty,\ll_\ty)$. This defines an equivalence relation $\equiv$ in the set ${\mathbb T}$ of all types over $(K,v)$. Any polynomial $g\in K[x]$ has an *order of divisibility* by the type $\ty$, defined as $\ord_\ty(g):=\ord_{\psi_r}(R_r(g))$ in ${\mathbb F}_r[y]$. Let $\rep(\ty)$ be the set of all *representatives* of $\ty$; that is, monic polynomials $\phi\in\oo[x]$ with minimal degree satisfying $\ord_\ty(\phi)=1$. We have $\rep(\ty)\subset\P$, where $\P$ is the set of monic irreducible polynomials with coefficients in $\oo_v$. The main result of the paper states that two types $\ty$, $\ty^*$ are equivalent if and only if $\rep(\ty)=\rep(\ty^*)$ (cf. Theorem \[finalchar\]). The outline of the paper is as follows. In section \[secML\] we recall some essential facts on MacLane valuations. In section \[secComparison\] we analyze to what extent different chains of augmentation steps as in (\[depth0\]) may build the same valuation $\mu$. In section \[secTypes\] we find a concrete procedure to decide whether two given types are equivalent, in terms of the data supported by them, and we describe then the relationship between their residual polynomial operators (Lemma \[optstep\] and Proposition \[charequiv\]). Finally, we prove Theorem \[finalchar\], which yields two more conceptual characterizations of the equivalence of types. Finally, let us add some remarks on the incidence of these results in the algorithmic applications of types and MacLane’s valuations. On the set $\P$ we may consider the following equivalence relation: two prime polynomials $F,G\in\P$ are *Okutsu equivalent*, and we write $F\approx G$, if $v(\res(F,G))$ is greater than certain *Okutsu bound* [@okutsu Sec. 4], [@Ok]. Equivalence of types had been considered in [@gen] only for *strongly optimal* types, which form a very special subset $\tst\subset {\mathbb T}$ (cf. section \[secTypes\]). In [@gen Thm. 3.9] it is shown that the assignment $\ty\mapsto\rep(\ty)$ induces a canonical bijection $$\label{main} \tst/\equiv \ \longrightarrow \ \P/\approx,$$ and the levels of $\ty\in\tst$ contain intrinsic data of the prime polynomials in the Okutsu class of any representative of $\ty$. Given a monic squarefree $f\in \oo[x]$, the *Montes algorithm*, known also as the *OM factorization algorithm*, computes a family of pairs $(\ty,\phi)$ parameterizing the prime factors of $f$ in $\oo_v[x]$. If a prime factor $F\in\oo_v[x]$ of $f$ is associated with a pair $(\ty,\phi)$, then $\ty$ is a strongly optimal type whose equivalence class is canonically attached to the Okutsu class of $F$ through the mapping of (\[main\]), and $\phi\approx F$ is a concrete choice in $\oo[x]$ of a polynomial in the Okutsu class of $F$. However, the algorithm is based on the construction of certain non-optimal types, which must then be converted into optimal types in the same equivalence class. In the original presentation of the algorithm in [@algorithm; @Mo], the discussion of these optimization steps was based on some excruciating arguments, due to the absence of the concept of equivalence of types. Thus, the results of this paper contribute to a great simplification of the analysis of this optimization procedure. This is illustrated in section \[secExample\], where we present a concrete example of OM factorization. MacLane chains of inductive valuations {#secML} ====================================== Let $K$ be a field equipped with a discrete valuation $v\colon K^*\to {\mathbb Z}$, normalized so that $v(K^*)={\mathbb Z}$. Let $\oo$ be the valuation ring of $K$, $\m$ the maximal ideal, $\pi\in\m$ a generator of $\m$ and ${\mathbb F}=\oo/\m$ the residue class field. Let $K_v$ be the completion of $K$ at $v$, with valuation ring $\oo_v\subset K_v$. Let $v\colon \kb^*\to {\mathbb Q}$ still denote the canonical extension of $v$ to a fixed algebraic closure of $K_v$. Graded algebra of a valuation {#subsecGradedAlg} ----------------------------- Let ${\mathbb V}$ be the set of all discrete valuations $\mu\colon K(x)^*\to {\mathbb Q}$ such that $\mu_{\mid K}=v$ and $\mu(x)\ge0$. In the set ${\mathbb V}$ there is a natural partial ordering: $$\mu\le \mu' \quad\mbox{ if }\quad\mu(g) \le \mu'(g), \ \forall\,g\in K[x].$$ Consider the Gauss valuation $\mu_0\in {\mathbb V}$ acting on polynomials as follows: $$\mu_0\left(\sum\nolimits_{0\le s}a_sx^s\right)=\mn_{0\le s}\left\{v(a_s)\right\}.$$ Clearly, $\mu_0\le \mu$ for all $\mu\in{\mathbb V}$. Let $\mu\in{\mathbb V}$ be a valuation. We denote by $\Gamma(\mu)=\mu\left(K(x)^*\right)\subset {\mathbb Q}$ the cyclic group of finite values of $\mu$. The *ramification index* of $\mu$ is the positive integer $e(\mu)$ such that $e(\mu)\Gamma(\mu)={\mathbb Z}$. For any $\alpha\in\Gamma(\mu)$ we consider the following $\oo$-submodules in $K[x]$: $$\ppa=\{g\in K[x]\mid \mu(g)\ge \alpha\}\supset \ppa^+=\{g\in K[x]\mid \mu(g)> \alpha\}.$$ The *graded algebra of $\mu$* is the integral domain: $$\ggm:=\bigoplus\nolimits_{\alpha\in\Gamma(\mu)}\ppa/\ppa^+.$$ Let $\Delta(\mu)=\pset_0/\pset_0^+$ be the subring determined by the piece of degree zero of this algebra. Clearly, $\oo\subset\pset_0$ and $\m=\pset_0^+\cap \oo$; thus, there is a canonical homomorphism ${\mathbb F}\to\Delta(\mu)$, equipping $\Delta(\mu)$ (and $\ggm$) with a canonical structure of ${\mathbb F}$-algebra. There is a natural map $\hm\colon K[x]\lra \ggm$, given by $\hm(0)=0$, and $$\hm(g)= g+\pset_{\mu(g)}^+\in\pset_{\mu(g)}/\pset_{\mu(g)}^+,$$ for $g\ne0$. Note that $\hm(g)\ne0$ if $g\ne0$. For all $g,h\in K[x]$ we have: $$\begin{array}{l} \hm(gh)=\hm(g)\hm(h), \\ \hm(g+h)=\hm(g)+\hm(h), \mbox{ if }\mu(g)=\mu(h)=\mu(g+h). \end{array}$$ If $\mu\le \mu'$ for some $\mu'\in{\mathbb V}$, we have a canonical homomorphism of graded algebras $$\ggm\to\gg(\mu'),\qquad g+\ppa^+(\mu)\mapsto g+\ppa^+(\mu').$$ The image of $\hm(g)$ is $\hmp(g)$ if $\mu(g)=\mu'(g)$, and zero otherwise. \[mu\] Consider $g,\phi\in K[x]$. We say that $g,\phi$ are *$\mu$-equivalent*, and we write $g\smu \phi$, if $\hm(g)=\hm(\phi)$. We write $\phi\mmu g$, if $\hm(g)$ is divisible by $\hm(\phi)$ in $\ggm$. We say that $\phi$ is $\mu$-irreducible if $\hm(\phi)\ggm$ is a non-zero prime ideal. We say that $\phi$ is $\mu$-minimal if $\phi\nmid_\mu h$ for all non-zero $h\in K[x]$ with $\deg h<\deg \phi$. Augmentation of valuations -------------------------- A *key polynomial* for the valuation $\mu$ is a monic polynomial in $K[x]$ which is $\mu$-minimal and $\mu$-irreducible. Let us denote by $\kpm$ the set of key polynomials for $\mu$. Every key polynomial has coefficients in $\oo$ and is irreducible in $\oo_v[x]$ [@Rid Lem. 1.8, Cor. 1.10]. [@Rid Lem. 1.4]\[mid=sim\] Consider $\phi\in\kpm$ and $g\in K[x]$ a monic polynomial such that $\phi\mmu g$ and $\deg g=\deg\phi$. Then, $\phi\smu g$ and $g$ is a key polynomial for $\mu$ too. \[muprima\] Take $\phi\in \kpm$. For $g\in K[x]$ let $g=\sum_{0\le s}a_s\phi^s$ be its canonical $\phi$-expansion in $K[x]$, uniquely determined by the condition $\deg a_s<\deg\phi$ for all $s\ge 0$. Take $\nu\in {\mathbb Q}_{>0}$. The augmented valuation $\mu'=[\mu;\phi,\nu]$ with respect to the pair $\phi,\nu$ is the valuation $\mu'$ determined by the following action on $K[x]$: $$\mu'(g):=\mn_{0\le s}\{\mu(a_s\phi^s)+s\nu\}=\mn_{0\le s}\{\mu'(a_s\phi^s)\}.$$ [@Rid Prop. 1.7]\[extension\] 1. The natural extension of $\mu'$ to $K(x)$ is a valuation and $\mu\le\mu'$. 2. For a non-zero $g\in K[x]$, $\mu(g)=\mu'(g)$ if and only if $\phi\nmid_{\mu}g$. 3. The polynomial $\phi$ is a key polynomial for $\mu'$ too. [@Rid Lem. 3.5]\[unicity\] Let $\mu''=[\mu;\phi^*,\nu^*]$ be another augmentation of $\mu$. We have $\mu'=\mu''$ if and only if $\deg\phi^*=\deg\phi$, $\mu'(\phi^*-\phi)\ge \mu'(\phi)$, and $\nu^*=\nu$. In this case, $\phi^*\smu\phi$. Denote $\Delta=\Delta(\mu)$, and let $I(\Delta)$ be the set of ideals in $\Delta$. Consider the following *residual ideal operator*: $$\rr=\rrm\colon K[x]\lra I(\Delta),\qquad g\mapsto \Delta\cap \hm(g)\ggm.$$ Let $\phi$ be a key polynomial for $\mu$. Choose a root $\t \in\kb$ of $\phi$ and denote by $K_\phi=K_v(\t)$ the finite extension of $K_v$ generated by $\t$. Also, let $\oo_\phi\subset K_\phi$ be the valuation ring of $K_\phi$, $\m_\phi$ the maximal ideal and ${\mathbb F}_\phi=\oo_\phi/\m_\phi$ the residue class field. [@Rid Prop. 1.12]\[sameideal\] If $\phi$ is a key polynomial for $\mu$, then 1. $\rr(\phi)$ is the kernel of the onto homomorphism $\Delta\twoheadrightarrow {\mathbb F}_\phi$ determined by $g+\pset^+_0\ \mapsto\ g(\t)+\m_\phi$. Hence, $\rr(\phi)$ is a maximal ideal of $\Delta$. 2. $\rr(\phi)=\op{Ker}(\Delta\to \Delta(\mu'))$ for any augmented valuation $\mu'=[\mu;\phi,\nu]$. Thus, the image of $\Delta\to\Delta(\mu')$ is a field canonically isomorphic to ${\mathbb F}_\phi$. The map $\rr\colon\kpm\to\mx(\Delta)$ is onto and its fibers coincide with the $\mu$-equivalence classes of key polynomials [@Rid Thm. 5.7]: $$\label{repr} \rr(\phi)=\rr(\phi^*) \sii \phi\smu\phi^*\sii \phi\mmu\phi^*.$$ MacLane chains {#subsecML} -------------- Let $\mu\in {\mathbb V}$ be an *inductive valuation*; that is, $\mu$ may be obtained from the Gauss valuation $\mu_0$ by a finite number of augmentation steps: $$\label{depth} \mu_0\ \stackrel{\phi_1,\nu_1}\lra\ \mu_1\ \stackrel{\phi_2,\nu_2}\lra\ \cdots \ \stackrel{\phi_{r-1},\nu_{r-1}}\lra\ \mu_{r-1} \ \stackrel{\phi_{r},\nu_{r}}\lra\ \mu_{r}=\mu$$ satisfying $\phi_{i+1}\nmid _{\mu_i}\phi_i$ for all $1\le i<r$. Such a chain of augmentations is called a *MacLane chain* of $\mu$. In a MacLane chain, the value group $\Gamma(\mu_{i})$ is the subgroup of ${\mathbb Q}$ generated by $\Gamma(\mu_{i-1})$ and $\nu_{i}$, for any $1\le i\le r$. In particular, $${\mathbb Z}=\Gamma(\mu_0)\subset \Gamma(\mu_1)\subset\cdots\subset\Gamma(\mu_{r-1})\subset \Gamma(\mu_r)=\Gamma(\mu).$$ A MacLane chain of $\mu$ supports several data and operators containing relevant information about $\mu$. Among them, the following deserve special mention: [**(1) A sequence of finite field extensions of the residue class field:**]{} $$\begin{array}{rcccccl} \Delta_0&\lra&\Delta_1&\lra&\cdots&\lra&\Delta_r=\Delta(\mu)\\ \cup&&\cup&&\cdots&&\cup\\ {\mathbb F}={\mathbb F}_0&\lra&{\mathbb F}_1&\lra&\cdots&\lra&{\mathbb F}_r \end{array}$$ where $\Delta_i=\Delta(\mu_i)$, the maps $\Delta_i\to\Delta_{i+1}$ are the canonical homomorphisms induced from the inequality $\mu_i\le\mu_{i+1}$, and ${\mathbb F}_i:=\op{Im}(\Delta_{i-1}\to\Delta_i)$. [**(2) Numerical data.**]{} Set $\phi_0=x$, $\nu_0=0$, $\mu_{-1}=\mu_0$ and ${\mathbb F}_{-1}={\mathbb F}_0$. For all $0\le i\le r$, we define integers: $$\begin{array}{lll} e_i:=e(\mu_i)/e(\mu_{i-1}),&\quad f_{i-1}:=[{\mathbb F}_{i}\colon {\mathbb F}_{i-1}], &\quad h_i:=e(\mu_i)\nu_i,\\ m_i:=\deg\phi_i,&\quad V_i:=e(\mu_{i-1})\mu_{i-1}(\phi_i). & \end{array}$$ which satisfy the following relations for $1\le i\le r$: $$\label{recurrence} \begin{array}{l} \gcd(e_i,h_i)=1,\\ e(\phi_i)=e(\mu_{i-1})=e_0\cdots e_{i-1},\\ f(\phi_i)=\left[{\mathbb F}_i\colon {\mathbb F}_0\right]=f_0\cdots f_{i-1},\\ m_i=e_{i-1}f_{i-1}m_{i-1}=(e_0\cdots e_{i-1})(f_0\cdots f_{i-1}),\\ V_i=e_{i-1}f_{i-1}(e_{i-1}V_{i-1}+h_{i-1}),\\ \end{array}$$ where $e(\phi_i)$, $f(\phi_i)$ denote the ramification index and residual degree of the finite extension $K_{\phi_i}/K_v$, respectively. [**(3) Generators of the graded algebras**]{}: $$p_i\in\gg(\mu_i)^*,\quad x_i\in\gg(\mu_i),\quad y_i\in\Delta_i,\quad 0\le i\le r,$$ such that $\Delta_i={\mathbb F}_i[y_i]$ and $\gg(\mu_i)=\Delta_i[p_i,p_i^{-1}][x_i]$. The elements $p_i,y_i$ are algebraically independent over ${\mathbb F}_i$ and $x_i$ satisfies the algebraic relation $x_i^{e_i}=y_ip_i^{h_i}$. In particular, we have a family of ${\mathbb F}_i$-isomorphisms: $$\label{ji} j_i\colon {\mathbb F}_i[y]\lra \Delta_i,\quad y\mapsto y_i,\quad 0\le i\le r.$$ Starting with $p_0=\op{H}_{\mu_0}(\pi)$, the generators are defined by the following recurrent relations: $$x_i=\op{H}_{\mu_i}(\phi_i)p_i^{-V_i},\quad y_i=x_i^{e_i}p_i^{-h_i},\quad p_{i+1}=x_i^{\ell_i}p_i^{\ell'_i},$$ where $\ell_i$, $\ell'_i\in{\mathbb Z}$ are uniquely determined by $\ell_ih_i+\ell'_ie_i=1$ and $0\le \ell_i<e_i$. In the relation concerning $p_{i+1}$ we identify the elements $x_i$, $p_i$ with their images under the canonical homomorphism $\gg(\mu_i)\to\gg(\mu_{i+1})$. [**(4) Newton polygon operators**]{}: $$N_i:=N_{\mu_{i-1},\phi_i}\colon K[x]\lra 2^{{\mathbb R}^2},\quad 1\le i\le r.$$ For any nonzero $g\in K[x]$ consider its canonical $\phi$-expansion $g=\sum_{0\le s}a_s\phi^s$, where $a_s\in K[x]$ have $\deg a_s<\deg\phi$. Then, $N_i(g)$ is the lower convex hull of the set of points $\left\{(s,\mu_{i-1}(a_s\phi_i^s))\mid s\ge0\right\}$ in the Euclidean plane. [**(5) Residual polynomial operators**]{}: $$R_i:=R_{\mu_{i-1},\phi_i,\nu_i}\colon K[x]\lra {\mathbb F}_i[y], \quad 0\le i\le r,$$ uniquely determined by the condition: $$\label{mainR} \op{H}_{\mu_i}(g)=x_i^{s_i(g)}p_i^{u_i(g)}R_i(g)(y_i),$$ for all nonzero $g\in K[x]$. For $i=0$ we define $s_0(g)=0$, $u_0(g)=\mu_0(g)$. For $i>0$, the point $(s_i(g),u_i(g)/e(\mu_{i-1}))$ is the left end point of the *$\nu_i$-component* $S_{\nu_i}(g)$ of the Newton polygon $N_i(g)$, which is defined as the intersection of $N_i(g)$ with the line of slope $-\nu_i$ first touching the polygon from below (see Figure \[figComponent\]). Let $s_i(g)\le s'_i(g)$ be the abscissas of the left end points of $S_{\nu_i}(g)$. The polynomial $R_i(g)$ has degree $(s'_i(g)-s_i(g))/e_i$, nonzero constant term, and it determines a generator of the residual ideal $\rr_{\mu_i}(g)$ as follows: $\rr_{\mu_i}(g)=y_i^{\lceil s_i(g)/e_i\rceil}R_i(g)(y_i)\Delta_i$. (14,10) (2.8,5.8)[$\bullet$]{}(7.8,3.3)[$\bullet$]{} (-1,1)[(1,0)[15]{}]{}(0,0)[(0,1)[10]{}]{} (-1,8)[(2,-1)[15]{}]{} (3,6)[(-1,2)[1.5]{}]{}(3,6.04)[(-1,2)[1.5]{}]{} (3,6)[(2,-1)[5]{}]{}(3,6.04)[(2,-1)[5]{}]{} (8,3.5)[(4,-1)[4]{}]{}(8,3.54)[(4,-1)[4]{}]{} (3,.9)(0,.25)[21]{}[height2pt]{} (8,.9)(0,.25)[11]{}[height2pt]{} (-.1,6.05)(.25,0)[13]{}[to 2pt]{} (7.3,.1) $s'_i(g)$ (2.3,.1) $s_i(g)$ (-5,5.8) $u_i(g)/e(\mu_{i-1})$ (12.3,1.5) line of slope $-\nu_i$ (-.6,.2) $0$ (3,8.2) $N_i(g)=N_{\mu_{i-1},\phi_i}(g)$ (5.5,5) $S_{\nu_i}(g)$ (-.15,7.5)[(1,0)[.3]{}]{} (-2.4,7.2) $\mu_i(g)$ [**(6) A family of maximal ideals**]{} $\ll_i\in\mx(\Delta_i)$, for $0\le i<r$. The ideals $\ll_i$ are determined by Proposition \[sameideal\] as: $$\ll_i:=\op{Ker}(\Delta_i\lra \Delta_{i+1})=\rr_{\mu_i}(\phi_{i+1}),\quad 0\le i<r.$$ Through the isomorphisms $j_i$ of (\[ji\]), these ideals yield monic irreducible polynomials $\psi_i\in{\mathbb F}_i[y]$ uniquely determined by the condition $j_i(\psi_i{\mathbb F}_i[y])=\psi_i(y_i)\Delta_i=\ll_i$, or alternatively, by the condition $\psi_i=R_i(\phi_{i+1})$. We have a commutative diagram with vertical isomorphisms: $$\as{1.2}\begin{array}{ccl}{\mathbb F}_i[y]&\twoheadrightarrow&{\mathbb F}_i[y]/(\psi_i)\\j_i\downarrow\hphantom{m}&&\quad\downarrow\\\Delta_i&\twoheadrightarrow&\Delta_i/\ll_i\iso{\mathbb F}_{i+1}\subset \Delta_{i+1}\end{array}$$ Hence, $\deg\psi_i=[{\mathbb F}_{i+1}\colon{\mathbb F}_i]=f_i$, for $0\le i<r$. Data comparison between MacLane chains {#secComparison} ====================================== Consider a MacLane chain of an inductive valuation $\mu$ as in (\[depth\]), supporting the data and operators described above. In this section, we analyze the variation of these data and operators when a different MacLane chain of the same valuation is chosen. Note that ${\mathbb F}_r$ is the algebraic closure of $ {\mathbb F}$ in $\Delta:=\Delta(\mu)$, through the canonical map ${\mathbb F}\to\Delta$. Thus, this field does not depend on the choice of the MacLane chain. We may denote it by ${\mathbb F}_\mu:={\mathbb F}_r$. It must not be confused with the residue class field $\kappa(\mu)$ of the valuation $\mu$. Actually, $\kappa(\mu)$ is isomorphic to the field of fractions of $\Delta$ [@Rid Prop. 3.9], so that ${\mathbb F}_\mu$ is isomorphic to the algebraic closure of ${\mathbb F}$ in $\kappa(\mu)$ too. \[proper\] A key polynomial $\phi \in\kpm$ is said to be *proper* if $\mu$ admits a MacLane chain such that $\phi\nmid_\mu\phi_r$, where $r$ is the length of the chain and $\phi_r$ is the key polynomial of the last augmentation step. For any MacLane chain of length $r$ of $\mu$, we have [@Rid Sec. 5.3]: $$\label{em} \begin{array}{l} m_r=\mn\left\{\deg\phi\mid\phi \in\kpm\right\}, \\ e_rm_r=\mn\left\{\deg\phi\mid\phi \in\kpm,\ \phi\mbox{ proper}\right\},\\ \phi\in\kpm \mbox{ proper }\sii \deg\phi\ge e_rm_r. \end{array}$$ Thus, the positive integers $m_\mu:=m_r$, $e_\mu:=e_r$ do not depend on the choice of the MacLane chain either. Independence of the lower levels -------------------------------- Our first aim is to prove the following result. \[laststep\] Let $\phi$ be a proper key polynomial for the inductive valuation $\mu$ and consider a MacLane chain of $\mu$ as in (\[depth\]) with $\phi\nmid_{\mu}\phi_r$. For any $\nu\in{\mathbb Q}_{>0}$ consider the MacLane chain of the augmented valuation $\mu'=[\mu;\phi,\nu]$ obtained by adding one augmentation step: $$\label{extchain} \mu_0\ \stackrel{\phi_1,\nu_1}\lra\ \mu_1\ \stackrel{\phi_2,\nu_2}\lra\ \cdots \ \lra\ \mu_{r-1} \ \stackrel{\phi_{r},\nu_{r}}\lra\ \mu_{r}=\mu\ \stackrel{\phi,\nu}\lra\ \mu_{r+1}=\mu'$$ Then, the elements $$p_{r+1}\in\ggmp^*,\quad x_{r+1}\in\ggmp,\quad y_{r+1}\in\Delta(\mu')$$ and the operators $N_{r+1}$, $R_{r+1}$ attached to this extended MacLane chain do not depend on the initial MacLane chain. In other words, the generators of $\ggmp$ and the operators $N_{r+1}$, $R_{r+1}$ depend on $\mu,\phi,\nu$, but not on the choice of a MacLane chain of $\mu$. In particular, we obtain a residual polynomial operator $$R_{\mu,\phi,\nu}\colon K[x]\lra {\mathbb F}_{\mu'},$$ defined as $R_{\mu,\phi,\nu}:=R_{r+1}$, which depends only on $\mu$, $\phi$ and $\nu$. The proof of Theorem \[laststep\] requires some previous work. \[augmentation\] Consider a MacLane chain of augmented valuations $$\mu^*\stackrel{\phi^*,\nu^*}\lra \mu\stackrel{\phi,\nu}\lra \mu'$$ with $\deg \phi=\deg\phi^*$. Then, $\phi\in\op{KP}(\mu^*)$ and $\mu'=[\mu^*;\phi,\nu^*+\nu]$. Further, consider the affine transformation $$\hh\colon {\mathbb R}^2\lra {\mathbb R}^2,\quad (x,y)\mapsto (x,y-\nu^*x).$$ Then, $N_{\mu^*,\phi}=\hh\circ N_{\mu,\phi}$. The first statement is just [@Rid Lem. 3.4]. For the comparison between $N_{\mu^*,\phi}$ and $N_{\mu,\phi}$, consider the $\phi$-expansion $g=\sum_{0\le s}a_s\phi^s$ of a nonzero $g\in K[x]$. By the definition of the augmented valuation $\mu=[\mu^*;\phi^*,\nu^*]$, $$\deg a_s<\deg \phi=\deg\phi^*\imp \mu(a_s)=\mu^*(a_s).$$ On the other hand, $\phi^*=\phi+a$ for some $a\in K[x]$ with $\deg a<\deg \phi$. By hypothesis, $\phi\nmid_{\mu}\phi^*$, and this implies $\phi^*\nmid_{\mu}\phi$ by Lemma \[mid=sim\]. Since $\phi$ and $ \phi^*$ are both $\mu$-minimal, [@Rid Lem. 1.3] shows that $$\mu(\phi)=\mu(a)=\mu(\phi^*)=\mu^*(\phi^*)+\nu^*.$$ Since $\mu^*(a)=\mu(a)>\mu^*(\phi^*)$, we deduce that $\mu^*(\phi)=\mu^*(\phi^*)=\mu(\phi)-\nu^*$. Thus, for each $s\ge0$ we have $\mu^*(a_s\phi^s)=\mu(a_s\phi^s)-s\nu^*$, or equivalently, $\hh(s,\mu(a_s\phi^s))=(s,\mu^*(a_s\phi^s))$. (14,13) (-.25,10.8)[$\bullet$]{}(1.8,6.75)[$\bullet$]{}(1.8,8.75)[$\bullet$]{} (4.8,2.75)[$\bullet$]{}(4.8,7.75)[$\bullet$]{} (6.8,.75)[$\bullet$]{}(6.8,7.75)[$\bullet$]{} (-1,0)[(1,0)[11]{}]{}(0,-1)[(0,1)[13]{}]{} (0,11)[(1,-1)[2]{}]{}(0,11)[(1,-2)[2]{}]{} (0,11.02)[(1,-1)[2]{}]{}(0,11.02)[(1,-2)[2]{}]{} (2,9.02)[(3,-1)[3]{}]{}(2,7.02)[(3,-4)[3]{}]{} (2,9)[(3,-1)[3]{}]{}(2,7)[(3,-4)[3]{}]{} (5,8)[(1,0)[2]{}]{}(5,3)[(1,-1)[2]{}]{} (5,8.02)[(1,0)[2]{}]{}(5,3.02)[(1,-1)[2]{}]{} (5,7.5)[(0,1)[.45]{}]{}(5,3.5)[(0,-1)[.45]{}]{} (2,-.1)(0,.25)[36]{}[height2pt]{} (5,-.1)(0,.25)[33]{}[height2pt]{} (7,-.1)(0,.25)[33]{}[height2pt]{} (4.8,-.7) $s$ (5.2,5.2) $s\nu^*$ (8,7.8) $N_{\mu,\phi}(g)$ (8,.8) $N_{\mu^*,\phi}(g)$ (-.5,-.7) $0$ The affinity $\hh$ acts as a translation on every vertical line and it keeps the vertical axis pointwise invariant. Thus, a side $S$ of slope $\rho$ of $N_{\mu,\phi}(g)$ is mapped to a side of slope $\rho-\nu^*$ of $N_{\mu^*,\phi}(g)$, whose end points have the same abscissas as those of $S$ (see Figure \[figNcomparison\]). Let us now consider a very particular instance of Theorem \[laststep\]. With the notation of that theorem, suppose that $r\ge 2$ and $\deg \phi_{r-1}=\deg\phi_r$, or equivalently, $e_{r-1}=f_{r-1}=1$. In this case, Lemma \[augmentation\] shows that $\phi_r$ is a key polynomial for $\mu_{r-2}$ and $\mu=\mu_r=[\mu_{r-2};\phi_r,\nu_{r-1}+\nu_r]$ can be obtained as a simple augmentation of $\mu_{r-2}$. Thus, we may consider two different MacLane chains of $\mu$: $$\begin{array}{l} \mu_0\ \stackrel{\phi_1,\nu_1}\lra\ \mu_1\ \stackrel{\phi_2,\nu_2}\lra\ \cdots\ \lra \ \mu_{r-2} \ \stackrel{\phi_{r-1},\nu_{r-1}}\lra\ \mu_{r-1} \ \stackrel{\phi_{r},\nu_{r}}\lra\ \mu_{r}=\mu\\ \mu^*_0\ \stackrel{\phi^*_1,\nu^*_1}\lra\ \mu^*_1\ \stackrel{\phi^*_2,\nu^*_2}\lra\ \cdots\ \lra \ \mu^*_{r-2} \ \stackrel{\phi^*_{r-1},\nu^*_{r-1}}\lra\ \mu^*_{r-1}=\mu \end{array}$$ where $\phi^*_{r-1}=\phi_r$, $\nu^*_{r-1}=\nu_{r-1}+\nu_r$, and $$\mu^*_i=\mu_i,\quad \phi^*_i=\phi_i,\quad \nu^*_i=\nu_i,\quad 0\le i\le r-2.$$ We use the standard notation for all data and operators attached to the upper MacLane chain and we mark with a superscript $(\ )^*$ all data and operators attached to the lower one. \[firstcase\] With the above notation, let $\hh(x,y)=(x,y-\nu_{r-1}x)$. 1. $p^*_{r-1}=p_r$, $x^*_{r-1}=x_rp_r^{h_{r-1}}$, $y^*_{r-1}=y_r$. 2. $N^*_{r-1}=\hh\circ N_r$, $R^*_{r-1}=R_r$. The generators of the graded algebra of $\mu$ were defined as follows: $$\as{1.4} \begin{array}{ll} p_r=x_{r-1}^{\ell_{r-1}}p_{r-1}^{\ell'_{r-1}},&p^*_{r-1}=p_{r-1}=x_{r-2}^{\ell_{r-2}}p_{r-2}^{\ell'_{r-2}},\\ x_r=\hm(\phi_r)p_r^{-V_r},&x^*_{r-1}=\hm(\phi_r)(p^*_{r-1})^{-V^*_{r-1}},\\ y_r=x_r^{e_r}p_r^{-h_r},&y^*_{r-1}=(x^*_{r-1})^{e^*_{r-1}}(p^*_{r-1})^{-h^*_{r-1}}. \end{array}$$ By hypothesis, $e_{r-1}=1$, so that $\ell_{r-1}=0$, $\ell'_{r-1}=1$; hence, $p_r=p_{r-1}=p^*_{r-1}$. On the other hand, the recurrences (\[recurrence\]) show that $$\begin{array}{lclcl} V_r&=&e_{r-1}f_{r-1}(e_{r-1}V_{r-1}+h_{r-1})&=&V_{r-1}+h_{r-1},\\ V^*_{r-1}&=&e^*_{r-2}f^*_{r-2}(e^*_{r-2}V^*_{r-2}+h^*_{r-2})&=&V_{r-1}, \end{array}$$ because for levels $i<r-1$ the data of the two MacLane chains coincide. Hence, $$x_r=\hm(\phi_r)p_r^{-V_r}=\hm(\phi_r)p_r^{-V_{r-1}-h_{r-1}}=x^*_{r-1}p_r^{-h_{r-1}}.$$ As mentioned at the beginning of the section, $e^*_{r-1}=e_\mu=e_r$. Hence, from the equalities: $$h^*_{r-1}/(e^*_1\cdots e^*_{r-1})=\nu^*_{r-1}=\nu_{r-1}+\nu_r=h_r/(e_1\cdots e_r)+h_{r-1}/(e_1\cdots e_{r-1}),$$we deduce $h^*_{r-1}=h_r+e_rh_{r-1}$. Therefore, $$y^*_{r-1}=(x^*_{r-1})^{e_r}(p^*_{r-1})^{-h^*_{r-1}}=x_r^{e_r}p_r^{e_rh_{r-1}}p_r^{-h_r-e_rh_{r-1}}=x_r^{e_r}p_r^{-h_r}=y_r.$$ This ends the proof of (1). By Lemma \[augmentation\], we have $$N^*_{r-1}=N_{\mu^*_{r-2},\phi^*_{r-1}}=N_{\mu_{r-2},\phi_r}=\hh\circ N_{\mu_{r-1},\phi_r}=\hh\circ N_r.$$ For any nonzero $g\in K[x]$, the affinity $\hh$ sends the $\nu_r$-component of $N_r(g)$ to the $\nu^*_{r-1}$-component of $N^*_{r-1}(g)$; hence, $$\hh(s_r(g),u_r(g)/e(\mu_{r-1}))=(s^*_{r-1}(g),u^*_{r-1}(g)/e(\mu^*_{r-2})).$$ Having in mind that $e(\mu_{r-1})=e(\mu_{r-2})=e(\mu^*_{r-2})$, this shows that $$\label{claim} s_r(g)=s^*_{r-1}(g),\qquad u_r(g)=u^*_{r-1}(g)+s_r(g)h_{r-1}.$$ Now, (\[mainR\]) shows that $$x_r^{s_r(g)}p_r^{u_r(g)}R_r(g)(y_r)=H_{\mu}(g)=(x^*_{r-1})^{s^*_{r-1}(g)}(p^*_{r-1})^{u^*_{r-1}(g)}R^*_{r-1}(g)(y^*_{r-1}).$$ From the identities in (\[claim\]) and $x_r=x^*_{r-1}p_r^{-h_{r-1}}$ we deduce: $$\begin{split} x_r^{s_r(g)}p_r^{u_r(g)}=&(x^*_{r-1})^{s_r(g)}p_r^{-s_r(g)h_{r-1}}p_r^{u^*_{r-1}(g)+s_r(g)h_{r-1}}\\=& (x^*_{r-1})^{s^*_{r-1}(g)}(p^*_{r-1})^{u^*_{r-1}(g)}. \end{split}$$ Therefore, $R_r(g)(y_r)=R^*_{r-1}(g)(y^*_{r-1})=R^*_{r-1}(g)(y_r)$ and this implies $R_r(g)=R^*_{r-1}(g)$ because $y_r$ is transcendental over ${\mathbb F}_r$ [@Rid Thm. 4.3]. This ends the proof of (2). These computations prove Theorem \[laststep\] in this particular situation. \[firstproof\] With the above notation, let $\phi$ be a proper key polynomial for $\mu$ such that $\phi\nmid_\mu \phi_r$ and consider the augmented valuation $\mu'=[\mu;\phi,\nu]$. Then, the generators of $\ggmp$ and the operators $N_{\mu,\phi}$, $R_{\mu,\phi,\nu}$ attached to the following MacLane chains coincide. $$\begin{array}{l} \mu_0\ \stackrel{\phi_1,\nu_1}\lra\ \mu_1\ \stackrel{\phi_2,\nu_2}\lra\ \cdots\ \ \stackrel{\phi_{r},\nu_{r}}\lra\ \mu_{r}=\mu\ \stackrel{\phi,\nu}\lra\ \mu_{r+1}=\mu'\\ \mu^*_0\ \stackrel{\phi^*_1,\nu^*_1}\lra\ \mu^*_1\ \stackrel{\phi^*_2,\nu^*_2}\lra\ \cdots\ \ \stackrel{\phi^*_{r-1},\nu^*_{r-1}}\lra\ \mu^*_{r-1}=\mu\ \stackrel{\phi,\nu}\lra\ \mu^*_r=\mu' \end{array}$$ Let us compare the Bézout identities: $$\begin{array}{cl} \ell_rh_r+\ell'_re_r=1,&\quad 0\le \ell_r<e_r,\\ \ell^*_{r-1}h^*_{r-1}+(\ell')^*_{r-1}e^*_{r-1}=1,&\quad 0\le \ell^*_{r-1}<e^*_{r-1}. \end{array}$$ From the identities $e^*_{r-1}=e_r$, $h^*_{r-1}=h_r+e_rh_{r-1}$, obtained during the proof of Lemma \[firstcase\], one deduces easily: $$\ell^*_{r-1}=\ell_r,\qquad (\ell')^*_{r-1}=\ell'_r-\ell_rh_{r-1}.$$ Let us denote $\phi=\phi_{r+1}=\phi^*_r$, $\nu=\nu_{r+1}=\nu^*_r$. Note that $e^*_r=e_{r+1}$ and $h^*_r=h_{r+1}$. Hence, the identities of Lemma \[firstcase\] show that $$p^*_r=(x^*_{r-1})^{\ell^*_{r-1}}(p^*_{r-1})^{(\ell')^*_{r-1}}=x_r^{\ell_r}p_r^{\ell_rh_{r-1}}p_r^{\ell'_r-\ell_rh_{r-1}}=p_{r+1}.$$ Also, from $V_{r+1}=e(\mu)\mu(\phi)=V^*_r$, we deduce $$\begin{array}{l} x^*_r=\op{H}_{\mu'}(\phi)(p^*_r)^{-V^*_r}=\op{H}_{\mu'}(\phi)p_{r+1}^{-V_{r+1}}=x_{r+1},\\ y^*_r=(x^*_r)^{e^*_r}(p^*_r)^{-h^*_r}=x_{r+1}^{e_{r+1}}p_{r+1}^{-h_{r+1}}=y_{r+1}. \end{array}$$ Thus, the generators of $\ggmp$ are the same for both MacLane chains of $\mu$. On the other hand, $N^*_r=N_{\mu,\phi}=N_{r+1}$ depends only on $\mu,\phi$ by definition. In particular, for any nonzero $g\in K[x]$ we have $$s_{r+1}(g)=s^*_r(g),\qquad u_{r+1}(g)=u^*_r(g).$$ This implies $R_{r+1}(g)=R^*_r(g)$ as in the proof of Lemma \[firstcase\]. In fact, (\[mainR\]) shows that $$x_{r+1}^{s_{r+1}(g)}p_{r+1}^{u_{r+1}(g)}R_{r+1}(g)(y_{r+1})=\op{H}_{\mu'}(g)=(x^*_r)^{s^*_r(g)}(p^*_r)^{u^*_r(g)}R^*_r(g)(y^*_r),$$ so that $R^*_r(g)(y^*_r)=R_{r+1}(g)(y_{r+1})=R_{r+1}(g)(y^*_r)$, which implies $R^*_r(g)=R_{r+1}(g)$ by the transcendence of $y^*_r$. \[optimal\] A MacLane chain of length $r$ is *optimal* if $\deg\phi_1<\cdots<\deg\phi_r$. By an iterative application of Lemma \[augmentation\], we may convert any MacLane chain of $\mu$ into an optimal MacLane chain. In fact, whenever we find an augmentation step with $\deg\phi_{i-1}=\deg\phi_i$, we may collapse this step to get a shorter MacLane chain. Let us call this “shrinking" procedure an *optimization step*. In an optimization step, all data of levels $0,\,1,\,\dots,\,i-1$ of the MacLane chain remain unchanged; the data of level $i-1$ are lost and the data of the $i$-th level change as indicated in Lemma \[firstcase\]. By Corollary \[firstproof\], the data of levels $i+1,\dots,r$ remain unchanged too. Let us now go back to the general situation of Theorem \[laststep\]. We have a MacLane chain of length $r$ of $\mu$ such that $\phi\nmid_\mu\phi_r$ and we extend it to a MacLane chain (\[extchain\]) of the augmented valuation $\mu'=[\mu;\phi,\nu]$. By applying a finite number of optimization steps to the MacLane chain of $\mu$, we may convert it into an optimal MacLane chain $$\mu^*_0\ \stackrel{\phi^*_1,\nu^*_1}\lra\ \mu^*_1\ \stackrel{\phi^*_2,\nu^*_2}\lra\ \cdots\ \ \stackrel{\phi^*_{r^*},\nu^*_{r^*}}\lra\ \mu^*_{r^*}=\mu$$ Since the polynomial $\phi^*_{r^*}=\phi_r$ remains unchanged, we may extend this chain as well to a MacLane chain of $\mu'$: $$\mu^*_0\ \stackrel{\phi^*_1,\nu^*_1}\lra\ \mu^*_1\ \stackrel{\phi^*_2,\nu^*_2}\lra\ \cdots\ \ \stackrel{\phi^*_{r^*},\nu^*_{r^*}}\lra\ \mu^*_{r^*}=\mu\ \stackrel{\phi,\nu}\lra\ \mu_{r^*+1}=\mu'$$ By an iterative application of Corollary \[firstproof\], all data and operators attached to $\mu'$ by this extension of an optimal chain coincide with the data and operators attached to $\mu'$ through the original extended chain (\[extchain\]). Therefore, in order to prove Theorem \[laststep\], we need only to compare the data attached to $\mu'$ through the MacLane chains obtained by extending two different optimal MacLane chains of $\mu$. Now, two optimal MacLane chains of the same valuation $\mu$ have the same length $r$, the same intermediate valuations $\mu_1,\dots,\mu_{r-1}$ and the same slopes $\nu_1,\dots,\nu_r$ [@Rid Prop. 3.6]. Also, by Lemma \[unicity\], two families $\phi_1,\dots,\phi_r$ and $\phi^*_1,\dots,\phi^*_r$ are the key polynomials of two optimal MacLane chains of $\mu$ if and only if $$\label{optimicity} \deg \phi_i=\deg \phi^*_i,\quad\ \mu_i(\phi_i-\phi^*_i)\ge\mu_i(\phi_i),\quad 1\le i\le r.$$ These polynomials satisfy $\phi^*_i\sim_{\mu_{i-1}}\phi_i$, but not necessarily $\phi^*_i\sim_{\mu_i}\phi_i$. [**Proof of Theorem \[laststep\]** ]{} As mentioned above, we may assume that we deal with two MacLane chains of $\mu'$ which have been obtained by adding the augmentation step $\mu'=[\mu;\phi,\nu]$ to two optimal MacLane chains of $\mu$: $$\as{.6} \mu_0\ \begin{array}{c}\phi_1,\nu_1\\\lra\\\lra\\\phi^*_1,\nu_1\end{array}\ \mu_1\ \begin{array}{c}\phi_2,\nu_2\\\lra\\\lra\\\phi^*_2,\nu_2\end{array}\ \cdots \ \begin{array}{c}\lra\\\lra\end{array}\ \mu_{r-1} \ \begin{array}{c}\phi_r,\nu_r\\\lra\\\lra\\\phi^*_r,\nu_r\end{array}\ \mu_r=\mu\ \stackrel{\phi,\nu}\lra \ \mu_{r+1}=\mu'$$ The key polynomials of both MacLane chains satisfy (\[optimicity\]). By hypothesis, $\phi\nmid_\mu\phi_r$ and $\phi\nmid_\mu\phi^*_r$. As usual, we mark with a superscript $(\ )^*$ all data and operators attached to the lower MacLane chain. Note that $V_{r+1}=e(\mu)\mu(\phi)=V^*_{r+1}$. Hence, the numerical data $$h_i,\ e_i,\ \nu_i,\ V_i,\ \ell_i,\ \ell'_i,\ \quad 0\le i\le r+1$$ coincide for both chains. By [@Rid Lem. 4.13], we have $$\begin{array}{ll} p^*_i=p_i,&\mbox{ for all } 1\le i\le r,\\ x^*_i=x_i,&\mbox{ for all } 1\le i\le r \mbox{ such that }e_i>1. \end{array}$$ Now, if $e_r>1$ we have $x^*_r=x_r$ and $$p^*_{r+1}=(x^*_r)^{\ell_r}(p^*_r)^{\ell'_r}=x_r^{\ell_r}p_r^{\ell'_r}=p_{r+1}.$$ If $e_r=1$ we have $\ell_r=0$, $\ell'_r=1$ and this leads to the same conclusion: $$p^*_{r+1}=(x^*_r)^{\ell_r}(p^*_r)^{\ell'_r}=p^*_r=p_r=x_r^{\ell_r}p_r^{\ell'_r}=p_{r+1}.$$ As a consequence, $$\begin{array}{l} x^*_{r+1}=\op{H}_{\mu'}(\phi)(p^*_{r+1})^{-V^*_{r+1}}=\op{H}_{\mu'}(\phi)(p_{r+1})^{-V_{r+1}}=x_{r+1},\\ y^*_{r+1}=(x^*_{r+1})^{e^*_{r+1}}(p^*_{r+1})^{-h^*_{r+1}}=(x_{r+1})^{e_{r+1}}(p_{r+1})^{-h_{r+1}}=y_{r+1}. \end{array}$$ By the very definition, $N_{r+1}=N_{\mu,\phi}=N^*_{r+1}$ depends only on $\mu$ and $\phi$. In particular, $s_{r+1}(g)=s^*_{r+1}(g)$, $u_{r+1}(g)=u^*_{r+1}(g)$, for any nonzero $g\in K[x]$. This leads to $R_{r+1}=R^*_{r+1}$ by the usual argument using (\[mainR\]) and the transcendence of $y_r$ over ${\mathbb F}_r$. Variation of the data attached to one level ------------------------------------------- Consider a fixed MacLane chain of $\mu$ of length $r$, as in (\[depth\]). Once we know that the data and operators attached to the $r$-th level do not depend on the previous levels, our second aim is to analyze the variation of these data and operators when the key polynomial $\phi_r$ of that level changes. By Lemma \[unicity\], the only way to obtain $\mu$ as an augmentation of $\mu_{r-1}$ is by taking $\mu=[\mu_{r-1};\phi^*_r,\nu_r]$, with $\phi^*_r=\phi_r+a$ such that $\deg a<\deg \phi_r$ and $\mu(a)\ge \mu(\phi_r)$. Since $\phi^*_r\sim_{\mu_{r-1}}\phi_r$, we have $\phi^*_r\nmid_{\mu_{r-1}}\phi_{r-1}$ too, so that it makes sense to consider another MacLane chain of $\mu$ as in (\[depth\]), just by replacing $\phi_r$ by $\phi^*_r$. As mentioned in section \[subsecML\], $\Gamma(\mu)$ is the subgroup of ${\mathbb Q}$ generated by $\Gamma(\mu_{r-1})$ and $\nu_r$; on the other hand, $e_r=e(\mu)/e( \mu_{r-1})$ is the least positive integer such that $e_r\Gamma(\mu)\subset\Gamma(\mu_{r-1})$. Hence, $\nu_r$ belongs to $\Gamma(\mu_{r-1})$ if and only if $e_r=1$. Hence, if $e_r>1$, then $\mu(\phi_r)=\mu_{r-1}(\phi_r)+\nu_r$ does not belong to $\Gamma(\mu_{r-1})$, and the equality $\mu(a)=\mu(\phi_r)$ cannot occur, because $\mu(a)=\mu_{r-1}(a)$ belongs to $\Gamma(\mu_{r-1})$. In other words, $$\label{e>1} e_r>1\imp \mu(a)>\mu(\phi_r)\imp \phi^*_r\smu\phi_r.$$ \[lastlevel\] Consider two MacLane chains of an inductive valuation $\mu$, which differ only in the last augmentation step: $$\as{.6} \mu_0\ \stackrel{\phi_1,\nu_1}\lra\ \mu_1\ \stackrel{\phi_2,\nu_2}\lra\ \cdots \ \lra\ \mu_{r-2} \ \stackrel{\phi_{r-1},\nu_{r-1}}\lra\ \mu_{r-1}\ \begin{array}{c}\phi_r,\nu_r\\\lra\\\lra\\\phi^*_r,\nu_r\end{array}\;\mu_r=\mu$$ Let us mark with a superscript $(\ )^*$ all data and operators attached to the lower MacLane chain. If $\phi^*_r\smu\phi_r$, we have $$p^*_r=p_r,\quad x^*_r=x_r,\quad y^*_r=y_r,\quad S^*_{\nu_r}=S_{\nu_r},\quad R^*_r=R_r.$$ Assume that $\phi^*_r\not\smu\phi_r$ and let $\eta:=R_r(\phi^*_r-\phi_r)\in{\mathbb F}_{\mu}^*$. Then, $$p^*_r=p_r,\quad x^*_r=x_r+p_r^{h_r}\eta,\quad y^*_r=y_r+\eta.$$Further, for any nonzero $g\in K[x]$ let $s:=\ord_{y+\eta}R_r(g)$ and denote $P(g):=R_r(g)/(y+\eta)^s$. Then, $$s^*_r(g)=s,\qquad R_r^*(g)(y)=(y-\eta)^{s_r(g)}P(g)(y-\eta).$$ By Lemma \[unicity\], $\phi_r^*=\phi_r+a$ with $\deg a<\deg\phi$ and $\mu(a)\ge \mu(\phi)$. All data attached to levels $i<r$ coincide for the both chains. Therefore, $$\begin{array}{l} V_r=e_{r-1}f_{r-1}(e_{r-1}V_{r-1}+h_{r-1})=V^*_r,\\ p_r=x_{r-1}^{\ell_{r-1}}p_{r-1}^{\ell'_{r-1}}=p^*_r. \end{array}$$ Also, since $\nu_r=\nu^*_r$, we have $h^*_r=h_r$ and $e^*_r=e_r$. Suppose $\phi^*_r\smu\phi_r$. Then, $$\begin{array}{l} x^*_r=\hm(\phi^*_r)(p^*_r)^{-V^*_r}=\hm(\phi_r)p_r^{-V_r}=x_r,\\ y^*_r=(x^*_r)^{e^*_r}(p^*_r)^{-h^*_r}=x_r^{e_r}p_r^{-h_r}=y_r. \end{array}$$ Now, consider a nonzero $g\in K[x]$, and let $S_{\nu_r}(g)$, $S^*_{\nu_r}(g)$ be the $\nu_r$-components of $N_r(g)$, $N^*_r(g)$, respectively. Both segments lie on the line of slope $-\nu_r$ cutting the vertical axis at the point $(0,\mu(g))$ (see Figure \[figComponent\]). Hence, in order to check that $S_{\nu_r}(g)=S^*_{\nu_r}(g)$ it suffices to show that the end points of both segments have the same abscissas. Let $\ord_{\mu,\phi_r}(g)$ be the largest integer $k$ such that $\phi_r^k\mid_\mu g$, namely the order with which the prime $\hm(\phi_r)$ divides $\hm(g)$ in $\ggm$. By [@Rid Lem. 2.6], the abscissas of the end points of $S_{\nu_r}(g)$ are: $$s_r(g)=\ord_{\mu,\phi_r}(g),\quad s'_r(g)=\ord_{\mu',\phi_r}(g),$$ where $\mu'=[\mu_{r-1};\phi_r,\nu_r-\epsilon]$ for a suficiently small positive rational number $\epsilon$. Since $\phi^*_r\smu\phi_r$, we have $\mu(a)>\mu(\phi_r)$, so that $\mu'(a)=\mu_{r-1}(a)=\mu(a)>\mu(\phi_r)>\mu'(\phi_r)$, and we have $\phi^*_r\sim_{\mu'}\phi_r$ as well. Hence, $$\as{1.2} \begin{array}{c} s_r(g)=\ord_{\mu,\phi_r}(g)=\ord_{\mu,\phi^*_r}(g)=s^*_r(g),\\ s'_r(g)=\ord_{\mu',\phi_r}(g)=\ord_{\mu',\phi^*_r}(g)=(s')^*_r(g). \end{array}$$ This implies $S_{\nu_r}(g)=S^*_{\nu_r}(g)$. In particular, $u_r(g)=u^*_r(g)$. We may now deduce $R_r(g)=R^*_r(g)$ by the usual argument using (\[mainR\]) and the transcendence of $y_r$ over ${\mathbb F}_r$. This ends the proof of the theorem in the case $\phi^*_r\smu\phi_r$. Suppose now $\phi^*_r\not\smu\phi_r$, or equivalently $\mu(a)=\mu(\phi)$, which implies $e_r=1$ by (\[e&gt;1\]). Both Newton polygons $N_r(a)=N^*_r(a)$ coincide with the point $(0,\mu(a))=(0,\mu(\phi_r))=(0,(V_r+h_r)/e(\mu_{r-1}))$. Hence, $$s_r(a)=0=s^*_r(a),\quad u_r(a)=V_r+h_r=u^*_r(a).$$ By (\[mainR\]), we have $$(p^*_r)^{V_r+h_r}R^*_r(a)=\hm(a)=(p_r)^{V_r+h_r}R_r(a),$$ which implies $\eta:=R_r(a)=R^*_r(a)=\hm(a)p_r^{-V_r-h_r}$, since $p^*_r=p_r$. Thus, $$x^*_r=\hm(\phi^*_r)p_r^{-V_r}=\left(\hm(\phi_r)+\hm(a)\right)p_r^{-V_r}=x_r+p^{h_r}\eta,$$ leading to $y^*_r=x^*_r(p^*_r)^{-h^*_r}=y_r+\eta$. Now, for a nonzero $g\in K[x]$, denote $\alpha=\mu(g)$ and $u_r(\alpha)=e(\mu)\alpha\in{\mathbb Z}$. Consider the polynomials $$R_{r,\alpha}(g)=y^{s_r(g)}R_r(g),\quad R^*_{r,\alpha}(g)=y^{s^*_r(g)}R^*_r(g).$$ By [@Rid Thm. 4.1], we have identities: $$(p_r^*)^{u_r(\alpha)}R^*_{r,\alpha}(g)(y^*_r)=\hm(g)= (p_r)^{u_r(\alpha)}R_{r,\alpha}(g)(y_r).$$ Since $p^*_r=p_r$, we deduce: $$R_{r,\alpha}(g)(y_r)= R^*_{r,\alpha}(g)(y^*_r)=R^*_{r,\alpha}(g)(y_r+\eta),$$ which implies $R_{r,\alpha}(g)(y)= R^*_{r,\alpha}(g)(y+\eta)$, by the transcendence of $y_r$ over ${\mathbb F}_r$. Let us rewrite this equality in terms of the original residual polynomials: $$\label{Ralpha} y^{s_r(g)}R_r(g)(y)=(y+\eta)^{s^*_r(g)}R_r^*(g)(y+\eta).$$ Since $r>0$, we have $y\nmid R_r(g)$, $y\nmid R^*_r(g)$ (cf. section \[subsecML\]). Hence, $(y+\eta)\nmid R^*_r(g)(y+\eta)$, and the equality (\[Ralpha\]) shows that $s^*_r(g)=\ord_{y+\eta}R_r(g)$ and $(y-\eta)^{s_r(g)}P(g)(y-\eta)=R_r^*(g)(y)$. Equivalence of types {#secTypes} ==================== Types are computational representations of certain mathematical objects. It is natural to consider two types to be equivalent when they represent the same objects. In sections \[subsecNormVal\], \[subsecTypes\], we recall the objects parameterized by types and in section \[subsecTypesEquiv\] we characterize the equivalence of types in terms of checkable conditions on the data supported by them (Lemma \[optstep\] and Proposition \[charequiv\]) and in terms of other invariants (Theorem \[finalchar\]). Normalized inductive valuations {#subsecNormVal} ------------------------------- In a computational context, it is natural to normalize inductive valuations in order to get groups of values equal to ${\mathbb Z}$. Given a MacLane chain of an inductive valuation $\mu$: $$\mu_0\ \stackrel{\phi_1,\nu_1}\lra\ \mu_1\ \stackrel{\phi_2,\nu_2}\lra\ \cdots \ \stackrel{\phi_{r-1},\nu_{r-1}}\lra\ \mu_{r-1} \ \stackrel{\phi_{r},\nu_{r}}\lra\ \mu_{r}=\mu$$ we consider the normalized valuations: $$v_i:=e(\mu_i)\mu_i=e_1\cdots e_i\,\mu_i,\quad 0\le i\le r,$$ with group of values $\Gamma(v_i)=v_i(K(x)^*)=e(\mu_i)\Gamma(\mu_i)={\mathbb Z}$. The property ${\mu_i}_{\mid K}=v$ translates into ${v_i}_{\mid K}=e_1\cdots e_i\, v$. The graded algebras $\gg(\mu_i)$ and $\gg(v_i)$ coincide up to the change of graduation given by the group isomorphism $$\Gamma(\mu_i)\iso \Gamma(v_i)={\mathbb Z},\quad\ \alpha\mapsto e_1\cdots e_i\,\alpha.$$ The piece of degree zero $\Delta_i:=\Delta(\mu_i)=\Delta(v_i)$ is the same for both valuations. Further, for any $g,h\in K[x]$ we obviously have $$g\mid_{\mu_i}h \sii g\mid_{v_i}h,\qquad g\sim_{\mu_i}h \sii g\sim_{v_i}h.$$ Also, consider the normalized slopes $$\la_i:=e(\mu_{i-1})\nu_i=h_i/e_i,\quad 0\le i\le r.$$ The augmentation step $\mu_i=[\mu_{i-1};\phi_i,\nu_i]$ translates into $v_i=[e_iv_{i-1};\phi_i,\la_i]$. If $g=\sum_{0\le s}a_s\phi_i^s$ is the $\phi_i$-expansion of a nonzero $g\in K[x]$, we have $$v_i(g)=\mn\left\{e_iv_{i-1}(a_s\phi_i^s)+s\la_i\mid 0\le s\right\}=\mn\left\{v_i(a_s\phi_i^s)\mid 0\le s\right\}.$$ The property $\mu_{i-1}<\mu_i$ translates into $e_iv_{i-1}<v_i$. With the obvious definition, we get a MacLane chain of $v_r$: $$\mu_0=v_0\ \stackrel{\phi_1,\la_1}\lra\ v_1\ \stackrel{\phi_2,\la_2}\lra\ \cdots \ \stackrel{\phi_{r-1},\la_{r-1}}\lra\ v_{r-1} \ \stackrel{\phi_{r},\la_r}\lra\ v_r$$ with attached data and operators as described in section \[subsecML\]. This approach has the advantage that the Newton polygons $N_{v_{i-1},\phi_i}(g)$ are derived from clouds of points in ${\mathbb R}^2$ with integer coordinates. The affinity $\hh(x,y)=(x,e_1\cdots e_{i-1}\,y)$ maps $N_{\mu_{i-1},\phi_i}(g)$ to $N_{v_{i-1},\phi_i}(g)$. This affinity maps a side of slope $\rho$ to a side of slope $e_1\cdots e_{i-1}\,\rho$ with the same abscissas of the end points. Thus, the role of the $\nu_i$-component is undertaken by the $\la_i$-component in the normalized context. Note that the left end point of the $\nu_i$-component of $N_{\mu_{i-1},\phi_i}(g)$ is $(s_i(g),u_i(g)/e_1\cdots e_{i-1})$, while the left end point of the $\la_i$-component of $N_{v_{i-1},\phi_i}(g)$ is simply $(s_i(g),u_i(g))$. The rest of data and operators attached to both MacLane chains coincide. Specially, for $0\le i<r$, we have the same residual polynomial operators: $$R_{\mu_{i-1},\phi_i,\nu_i}=R_i=R_{v_{i-1},\phi_i,\la_i}\colon K[x]\lra {\mathbb F}_i[y],$$ and the same family of prime polynomials $\psi_i=R_i(\phi_{i+1})\in{\mathbb F}_i[y]$. Types over $(K,v)$ {#subsecTypes} ------------------ A type of order $r$ is a collection of objects, distributed into levels: $$\ty=(\varphi_0;(\phi_1,\la_1,\varphi_1);\dots;(\phi_r,\la_r,\varphi_r)),$$ such that the pairs $\phi_i,\la_i$ determine a McLane chain of a normalized inductive valuation $v_\ty$: $$\label{depth2} v_0\ \stackrel{\phi_1,\la_1}\lra\ v_1\ \stackrel{\phi_2,\la_2}\lra\ \cdots \ \stackrel{\phi_{r-1},\la_{r-1}}\lra\ v_{r-1}\ \stackrel{\phi_r,\la_r}\lra\ v_r=v_\ty$$ and the data $\varphi_0,\dots,\varphi_{r}$ build a tower of finite field extensions of ${\mathbb F}$: $${\mathbb F}_{0,\ty}:={\mathbb F}\,\lra\, {\mathbb F}_{1,\ty}\,\lra\,\cdots\,\lra\, {\mathbb F}_{r,\ty}\,\lra\, {\mathbb F}_{r+1,\ty}$$ constructed as follows. Each $\varphi_i\in{\mathbb F}_{i,\ty}[y]$ is a monic irreducible polynomial, such that $\varphi_i\ne y$ for $i>0$. The field ${\mathbb F}_{i+1,\ty}$ is defined to be ${\mathbb F}_{i,\ty}[y]/(\varphi_i)$. Also, there is an specific procedure to compute certain residual polynomial operators $$R_{i,\ty}\colon K[x]\lra {\mathbb F}_{i,\ty}[y],\quad 0\le i\le r,$$ such that $\varphi_i=R_{i,\ty}(\phi_{i+1})$ for $0\le i<r$. The essential fact is that these objects reproduce the tower ${\mathbb F}_0\,\to\,\cdots \,\to\,{\mathbb F}_r$ and the residual polynomial operators $R_i$ attached to the MacLane chain of $v_\ty$. More precisely, there is a commutative diagram of vertical isomorphisms $$\as{1.2} \begin{array}{ccccccc} {\mathbb F}={\mathbb F}_{0,\ty}\ &\subset& {\mathbb F}_{1,\ty}\ &\subset&\cdots &\subset &{\mathbb F}_{r,\ty}\ \\ \quad\ \,\|\iota_0 &&\ \downarrow\iota_1&&\cdots&&\ \downarrow\iota_r\\ {\mathbb F}={\mathbb F}_0\quad &\subset& {\mathbb F}_1\ &\subset&\cdots &\subset &{\mathbb F}_r\ \end{array}$$ such that $R_i=\iota_{i}[y]\circ R_{i,\ty}$ for all $0\le i\le r$. In particular, $$\psi_i=R_i(\phi_{i+1})=\iota_i[y]\left(R_{i,\ty}(\phi_{i+1})\right)=\iota_i[y]\left(\varphi_i\right),\quad 0\le i<r.$$ The isomorphisms $\iota_0,\dots,\iota_r$ are uniquely determined by the isomorphisms $j_0,\dots,j_r$ defined in (\[ji\]). In fact, the isomorphism $\iota_0$ is the identity map on ${\mathbb F}_{0,\ty}={\mathbb F}={\mathbb F}_0$, while $\iota_{i+1}$ is determined by the following commutative diagram of vertical isomorphisms: (18,8) (0,3.5)[$\as{1.2} \begin{array}{ccl} {\mathbb F}_{i,\ty}[y]&\twoheadrightarrow&{\mathbb F}_{i,\ty}[y]/(\varphi_i)\ =\ {\mathbb F}_{i+1,\ty}\\ \!\!\!\!\!\iota_i[y]\downarrow\hphantom{m}&&\qquad\downarrow\\ {\mathbb F}_i[y]&\twoheadrightarrow&\ {\mathbb F}_i[y]/(\psi_i)\\ j_i\downarrow\hphantom{m}&&\qquad\downarrow\\ \Delta_i&\twoheadrightarrow&\quad \Delta_i/\ll_i\,\iso{\mathbb F}_{i+1}\subset \Delta_{i+1} \end{array} $]{} (13.2,5.5)[(0,-1)[3.4]{}]{} (13.8,3.6)[$\iota_{i+1}$]{} Therefore, for the theoretical considerations of this paper it will be harmless to consider the isomorphisms $\iota_0,\dots,\iota_r$ as identities. That is, we shall identify all data and operators supported by $\ty$ with the analogous data and operators attached to $v_\ty$: $${\mathbb F}_{i}={\mathbb F}_{i,\ty},\quad R_i=R_{i,\ty},\quad 0\le i\le r.$$ In particular, $\psi_i=\varphi_i$ for $0\le i<r$. According to this convention, from now on a type will be a collection of objects: $$\ty=(\psi_0;(\phi_1,\la_1,\psi_1);\dots;(\phi_r,\la_r,\psi_r)),$$ such that the pairs $\phi_i,\la_i$ determine a McLane chain of a normalized inductive valuation $v_\ty$ as in (\[depth2\]), and $\psi_i\in{\mathbb F}_i[y]$ are the monic irreducible polynomials determined by the MacLane chain too, for $0\le i<r$. What is the role of the prime polynomial $\psi_r\in{\mathbb F}_r[y]$? Let us denote by $$\mu_\ty:=(e_1\cdots e_r)^{-1}v_\ty,\qquad f_r:=\deg\psi_r,$$ the corresponding non-normalized inductive valuation attached to $\ty$ and the degree of $\psi_r$, respectively. Thanks to the isomorphism $j_r$, the polynomial $\psi_r$ determines a maximal ideal of $\Delta_r=\Delta(\mu_\ty)$: $$\ll_\ty:=j_r\left(\psi_r{\mathbb F}_r[y]\right)=\psi_r(y_r)\Delta(\mu_\ty)\in\mx(\Delta(\mu_\ty)).$$ The pair $(\mu_\ty,\ll_\ty)$, or equivalently $(v_\ty,\ll_\ty)$, is the “raison d’être" of $\ty$. Representatives of types {#subsecRepr} ------------------------ Denote $\mu:=\mu_\ty$, $\Delta:=\Delta(\mu)$, $\ll:=\ll_\ty$. The maximal ideal $\ll$ determines a certain subset of key polynomials for $\mu$, which are called *representatives* of the type $\ty$. By definition, the set $\rep(\ty)$ of all representatives of $\ty$ is: $$\rep(\ty)=\left\{\phi\in \kpm\mid \rr_\mu(\phi)=\ll_\ty \right\}\subset\kpm.$$ Since the residual ideal map $\rr_\mu\colon \kpm\to\mx(\Delta)$ is onto [@Rid Thm. 5.7], the set $\rep(\ty)$ is always non-empty. By (\[repr\]), the representatives of $\ty$ constitute one of the $\mu$-equivalence classes of the set $\kpm$. For any monic $\phi\in K[x]$, the property of being a representative of the type $\ty$ is characterized too by the following properties [@gen Lem. 3.1]: $$\label{charrep} \phi\in\rep(\ty)\sii\phi\in\oo[x],\ \deg\phi=e_rf_r m_r, \ R_r(\phi)=\psi_r.$$ By (\[charrep\]) and (\[em\]), the representatives of $\ty$ are proper key polynomials for $\mu$. Let $\phi$ be any representative of a type $\ty$ of order $r>0$. By [@Rid Cor. 5.3], $R_\mu(\phi_r)=y_r\Delta\ne\psi_r(y_r)\Delta=\ll_\ty$, because $\psi_r\ne y$. By (\[repr\]), $\phi\nmid_\mu\phi_r$, and we may extend the MacLane chain of $ \mu$ to a MacLane chain of length $r+1$ of the augmented valuation $\mu'=[\mu;\phi,\nu]$, where $\nu$ is an arbitray positive rational number. By choosing an arbitrary monic irreducible polynomial $\psi\in{\mathbb F}_{r+1}[y]={\mathbb F}_{\mu'}[y]$, we construct a type of order $r+1$ extending $\ty$: $$\ty'=(\ty;(\phi,\nu,\psi)):=(\psi_0;(\phi_1,\la_1,\psi_1);\dots;(\phi_r,\la_r,\psi_r);(\phi,\la,\psi)),$$ where $\la=e(\mu)\nu$. Let $\ty$ be a type of order $r\ge0$. For any $g\in K[x]$ we define $\ord_\ty(g):=\ord_{\psi_{r}}R_{r}(g)$; that is, the greatest integer $a$ such that $\psi_r^a$ divides $R_{r}(g)$ in ${\mathbb F}_r[y]$. Since the operators $R_i$ are multiplicative [@Rid Cor. 4.11], the identity $\ord_\ty(gh)=\ord_\ty(g)+\ord_\ty(h)$ holds for all $g,h\in K[x]$. Equivalence of types {#subsecTypesEquiv} -------------------- Let $\ty$ be a type of order $r$ with representative $\phi$, and let $\la\in{\mathbb Q}_{>0}$. We denote $$\begin{array}{l} N_i:=N_{v_{i-1},\phi_i},\quad 1\le i\le r;\qquad N_{\ty,\phi}:=N_{v_\ty,\phi},\\ R_{\ty,\phi,\la}:=R_{v_\ty,\phi,\la}=R_{\mu_\ty,\phi,\la/e(\mu_\ty)}. \end{array}$$ Note that $R_{\ty,\phi,\la}$ is well-defined by Theorem \[laststep\]. \[defequiv\] Two types $\ty$, $\ty^*$ are equivalent if $v_\ty=v_{\ty^*}$ and $\ll_\ty=\ll_{\ty^*}$. In this case we write $\ty\equiv\ty^*$. The next result is an immediate consequence of the definitions. \[samereps\] Let $\ty$, $\ty^*$ be two equivalent types. Then, 1. $\rep(\ty^*)=\rep(\ty)$. 2. For any $\phi\in\rep(\ty)$ and any $\la\in{\mathbb Q}_{>0}$, we have $N_{\ty^*,\phi}=N_{\ty,\phi}$ and $R_{\ty^*,\phi,\la}=R_{\ty,\phi,\la}$. The order of a type is not preserved by equivalence. In order to find a characterization of the equivalence of types in terms of the data supported by them, we consider *optimization steps* derived from the optimization steps for MacLane chains. \[stationary\] Let $\ty$ be a type of order $r$. We say that a level $(\phi_i,\la_i,\psi_i)$ of $\ty$ is *stationary* if $e_i=f_i=1$, or equivalently, if $\la_i\in{\mathbb Z}$ and $\deg\psi_i=1$. We say that $\ty$ is *optimal* if $\deg\phi_1<\cdots <\deg\phi_r$, or equivalently, if all levels $i<r$ are non-stationary. We say that $\ty$ is *strongly optimal* if all levels $i\le r$ are non-stationary. \[optstep\] For $r\ge 2$, let $\ty_0$ be a type of order $r-2$. Consider a type $$\ty=(\ty_0;(\phi_{r-1},\la_{r-1},\psi_{r-1});(\phi_r,\la_r,\psi_r))$$ of order $r$ whose $(r-1)$-th level is stationary. Then, $\phi_r$ is a representative of $\ty_0$ and the type $\ty^*=(\ty_0;(\phi_r,\la_{r-1}+\la_r,\psi_r))$ is equivalent to $\ty$. Moreover, $$\label{NR} N_{r-1}^*=\hh\circ N_r,\qquad R_{r-1}^*=R_r,$$ where $\hh$ is the affinity $\hh(x,y)=(x,y-\la_{r-1}x)$. Thus, $\ord_\ty=\ord_{\ty^*}$ as functions on $K[x]$. By Lemma \[augmentation\], $\phi_r$ is a key polynomial for $\mu_{\ty_0}$ and $$\mu_\ty=[\mu_{\ty_0};\phi_r,\nu_{r-1}+\nu_r]=\mu_{\ty^*},$$ where $\nu_{r-1}=\la_{r-1}/e_1\cdots e_{r-2}$ and $\nu_r=\la_r/e_1\cdots e_{r-1}=\la_r/e_1\cdots e_{r-2}$. By [@Rid Lem. 5.2], $N_{r-1}(\phi_r)$ is one-sided of negative slope $-\nu_{r-1}$; hence, [@Rid Lem. 2.1] shows that $\phi_{r-1}\mid_{\mu_{\ty_0}}\phi_r$. By (\[repr\]), we have $$\ll_{\ty_0}=\rr_{\mu_{\ty_0}}(\phi_{r-1})=\rr_{\mu_{\ty_0}}(\phi_r),$$ so that $\phi_r$ is a representative of $\ty_0$. The identities (\[NR\]) are a consequence of Lemma \[firstcase\]. Finally, let $\phi$ be a representative of $\ty$, so that $\ll_\ty=\rr_{\mu_\ty}(\phi)$. Since $R_r(\phi)=\psi_r$, we deduce that $R^*_{r-1}(\phi)=R_r(\phi)=\psi_r$. Hence, $\phi$ is a representative of $\ty^*$ too, because it satisfies the conditions of (\[charrep\]), characterizing the representatives of a type. Therefore, $$\ll_{\ty^*}=\rr_{\mu_{\ty^*}}(\phi)=\rr_{\mu_{\ty}}(\phi)=\ll_\ty,$$ and the types $\ty$, $\ty^*$ are equivalent. After a finite number of these optimization steps we may convert any type into an optimal type in the same equivalence class. Thus, in order to check if two types are equivalent we need only to characterize the equivalence of optimal types. The characterization we obtain is an immediate consequence of the characterization of MacLane optimal chains [@Rid Prop. 3.6] and Lemma \[lastlevel\]. \[charequiv\] Two optimal types $$\as{1.2} \begin{array}{l} \ty=(\psi_0;(\phi_1,\la_1,\psi_1);\dots;(\phi_r,\la_r,\psi_r)),\\ \ty^*=(\psi^*_0;(\phi^*_1,\la^*_1,\psi^*_1);\dots;(\phi^*_{r^*},\la^*_{r^*},\psi^*_{r^*})). \end{array}$$ are equivalent if and only if they satisfy the following conditions: - $r=r^*$. - $\la_i=\la^*_i$ for all $1\le i\le r$. - $\deg\phi_i=\deg\phi^*_i$ and $\mu_i(a_i)\ge \mu_i(\phi_i)$ for all $1\le i\le r$, where $a_i:=\phi^*_i-\phi_i$. - $\psi_r^*(y)=\psi_r(y-\eta_r)$, where $\eta_0:=0$ and for all $1\le i\le r$ we take $$\eta_i:= \begin{cases} 0,& \mbox{ if }\mu_i(a_i)>\mu_i(\phi_i) \quad\mbox{ (i.e. }\phi^*_i\sim_{\mu_i}\phi_i),\\ R_i(a_i)\in{\mathbb F}_i^*,& \mbox{ if }\mu_i(a_i)=\mu_i(\phi_i) \quad\mbox{ (i.e. }\phi^*_i\not\sim_{\mu_i}\phi_i). \end{cases}$$ In this case, $\psi_i^*(y)=\psi_i(y-\eta_i)$ for all $0\le i\le r$, and for any nonzero $g\in K[x]$ we have: $$s^*_i(g)=\ord_{y+\eta_i}R_i(g),\quad R^*_i(g)(y)=(y-\eta_i)^{s_i(g)}P_i(g)(y-\eta_i),\quad 1\le i\le r,$$ where $P_i(g)(y):=R_i(g)(y)/(y+\eta_i)^{s^*_i(g)}$. We may derive from this “practical" characterization of the equivalence of types some more conceptual characterizations. \[finalchar\] For any pair of types $\ty$, $\ty^*$, the following conditions are equivalent. 1. $\ty\equiv\ty^*$ 2. $\ord_\ty=\ord_{\ty^*}$ 3. $\rep(\ty)=\rep(\ty^*)$ Let us prove that (1) implies (2). By Lemma \[optstep\], the function $\ord_\ty$ is preserved by the optimization steps. Hence, we may assume that the types are optimal. Take $g\in K[x]$ a nonzero polynomial. For two equivalent types of order $r=0$ we have $R_0=R^*_0$ and $\psi_0=\psi^*_0$; thus, $$\ord_\ty(g)=\ord_{\psi_0}(R_0(g))=\ord_{\psi^*_0}(R^*_0(g))=\ord_{\ty^*}(g).$$ If $r>0$, we have $\psi_r\ne y$ and $\psi^*_r\ne y$. By Proposition \[charequiv\], $\psi^*_r(y)=\psi_r(y-\eta_r)$, and this implies $\psi_r\ne y+\eta_r$, $\psi^*_r\ne y-\eta_r$. Hence, Proposition \[charequiv\] shows that $$\begin{split} \ord_\ty(g)=&\ord_{\psi_r}R_r(g)=\ord_{\psi_r}P_r(g)=\ord_{\psi^*_r}P_r(g)(y-\eta_r)\\=&\ord_{\psi^*_r}R^*_r(g)=\ord_{\ty^*}(g). \end{split}$$ On the other hand, (\[charrep\]) characterizes the representatives of a type $\ty$ as monic polynomials $\phi\in\oo[x]$ with minimal degree satisfying $\ord_\ty(\phi)=1$; thus, (2) implies (3). Finally, let us prove that (3) implies (1). Let us denote $\mu=\mu_\ty$, $\mu^*=\mu_{\ty^*}$. It suffices to show that $\mu=\mu^*$, because then any common representative $\phi\in\rep(\ty)\cap\rep(\ty^*)$ leads to $\ll_\ty=\rr_{\mu}(\phi)=\rr_{\mu^*}(\phi)=\ll_{\ty^*}$, so that $\ty$ and $\ty^*$ are equivalent. Take $\phi\in\rep(\ty)\cap\rep(\ty^*)$ a common representative of $\ty$ and $\ty^*$. Let $\mu_{\infty}$ be the pseudo-valuation on $K[x]$ obtained as the composition: $$\mu_{\infty}\colon K[x] \hookrightarrow K_v[x]\lra K_\phi\stackrel{v}\lra {\mathbb Q}\cup\{\infty\},$$the second mapping being determined by $x\mapsto \t$, a root of $\phi$ in $\overline{K}_v$. By [@Rid Prop. 1.9], we have $\mu<\mu_\infty$, $\mu^*<\mu_\infty$, and for any nonzero $g\in K[x]$: $$\label{a} \mu(g)<\mu_\infty(g)\sii \phi\mmu g,\quad \mu^*(g)<\mu_\infty(g)\sii \phi\mid_{\mu^*} g.$$ Since the interval $[\mu_0,\mu_\infty]$ is totally ordered [@Rid Thm. 7.5], after exchanging the role of $\mu$ and $\mu^*$ if necessary, we must have $$\mu\le \mu^*<\mu_\infty.$$ The proof will be complete if we show that the conditions $\mu<\mu^*<\mu_\infty$ and $\rep(\ty^*)=\rep(\ty)$ lead to a contradiction. Let $\Phi_{\mu,\mu_\infty}$ be the set of all monic polynomials $\varphi\in K[x]$ of minimal degree satisfying $\mu(\varphi)<\mu_\infty(\varphi)$. Let $\deg \Phi_{\mu,\mu_\infty}$ be the common degree of all polynomials in this set. We claim that $\phi$ belongs to $\Phi_{\mu,\mu_\infty}$. In fact, the inequality $\mu(\phi)<\mu_\infty(\phi)=\infty$ is obvious. On the other hand, for any $a\in K[x]$ of degree less than $\deg\phi$, the $\mu$-minimality of $\phi$ implies that $\phi\nmid_\mu a$; by (\[a\]), we deduce that $\mu(a)=\mu_\infty(a)$. By Lemma \[inequality\] below, there is a unique maximal ideal $\ll\in\mx(\Delta(\mu))$ such that $$\Phi_{\mu,\mu_\infty}=\left\{\varphi \in\kpm\mid \rr_{\mu}(\varphi)=\ll\right\}.$$ Since $\phi\in\Phi_{\mu,\mu_\infty}$ and $\rr_\mu(\phi)=\ll_\ty$, we see that $$\Phi_{\mu,\mu_\infty}=\left\{\varphi \in\kpm\mid \rr_{\mu}(\varphi)=\ll_\ty\right\}=\rep(\ty)=\left\{\varphi \in \kpm\mid \varphi\smu\phi\right\}.$$ An analogous argument shows that $$\Phi_{\mu^*,\mu_\infty}=\rep(\ty^*)=\left\{\varphi \in \op{KP}(\mu^*)\mid \varphi\sim_{\mu^*}\phi\right\}.$$ Also, Lemma \[inequality\] shows that $\Phi_{\mu,\mu^*}$ is one of the $\mu$-equivalence classes in $\kpm$. Hence, if we show that $\Phi_{\mu,\mu^*}\subset \Phi_{\mu,\mu_\infty}$, these two sets must coincide. In fact, a polynomial $\varphi\in\Phi_{\mu,\mu^*}$ is a key polynomial for $\mu$ with $\mu(\varphi)<\mu^*(\varphi)\le \mu_\infty(\varphi)$. By (\[a\]), we have $\phi\mmu\varphi$, which implies $\rr_\mu(\phi)\supset\rr_\mu(\varphi)$; since $\rr_\mu(\phi)$, $\rr_\mu(\varphi)$ are maximal ideals of $\Delta(\mu)$, they coincide. Thus, $\phi\smu\varphi$, so that $\varphi$ belongs to $\Phi_{\mu,\mu_\infty}$. In particular, $\phi$ belongs to $\Phi_{\mu,\mu^*}=\Phi_{\mu,\mu_\infty}$. Consider the positive rational number $\nu=\mu^*(\phi)-\mu(\phi)$. By [@Vaq Thm. 1.15], the augmented valuation $\mu'=[\mu;\phi,\nu]$ satisfies $\mu<\mu'\le\mu^*$ and $\mu'(\phi)=\mu(\phi)+\nu=\mu^*(\phi)$. We claim that $\mu'=\mu^*$. In fact, if $ \mu'<\mu^*$, then we could replace $\mu$ by $\mu'$ in the above arguments to deduce $\Phi_{\mu',\mu^*}=\Phi_{\mu',\mu_\infty}$. Therefore, $$\deg\phi=\deg\Phi_{\mu,\mu^*}\le\deg\Phi_{\mu',\mu^*}=\deg\Phi_{\mu',\mu_\infty}\le\deg \Phi_{\mu^*,\mu_\infty}=\deg\phi.$$ We deduce $\deg\Phi_{\mu,\mu^*}=\deg\Phi_{\mu',\mu^*}$, and this leads to $\Phi_{\mu,\mu^*}\supset\Phi_{\mu',\mu^*}$, because $\mu'(\varphi)<\mu^*(\varphi)$ implies obviously $\mu(\varphi)<\mu^*(\varphi)$. Similarly, $\deg\Phi_{\mu',\mu_\infty}=\deg \Phi_{\mu^*,\mu_\infty}$, leading to $\Phi_{\mu',\mu_\infty}\supset\Phi_{\mu^*,\mu_\infty}$. Hence, $$\rep(\ty^*)=\Phi_{\mu^*,\mu_\infty}\subset\Phi_{\mu',\mu_\infty}=\Phi_{\mu',\mu^*}\subset \Phi_{\mu,\mu^*}=\Phi_{\mu,\mu_\infty}=\rep(\ty).$$ The hypothesis $\rep(\ty)=\rep(\ty^*)$ implies $\Phi_{\mu',\mu^*}=\Phi_{\mu,\mu^*}$, which is impossible, because $\phi$ does not belong to $\Phi_{\mu',\mu^*}$. Therefore, $\mu^*=\mu'=[\mu;\phi,\nu]$. Since $\phi$ is a proper key polynomial for $\mu$, there exists a MacLane chain of $\mu^*$ such that $\phi,\nu$ are the augmentation data of the last level. Hence, $m_{\mu^*}= \deg \phi$ and $e_{\mu^*}$ is the least positive integer such that $e_{\mu^*} \nu\in\Gamma(\mu)$. Since $\phi$ is a proper key polynomial for $\mu^*$, (\[em\]) shows that $\deg\phi\ge e_{\mu^*}m_{\mu^*}=e_{\mu^*}\deg\phi$. Thus, $e_{\mu^*}=1$, or equivalently, $\nu \in\Gamma(\mu)$. By Lemma \[values\] below, there exists $a\in K[x]$ of degree less than $\deg\phi$, such that $\mu(a)=\mu(\phi)+\nu$. Take $\varphi=\phi+a$. Since $\varphi\smu \phi$ and $\deg\varphi=\deg\phi$, Lemma \[mid=sim\] shows that $\varphi$ is a key polynomial for $\mu$, and $\varphi\in\rep(\ty)$ by (\[repr\]). However, $\varphi\not\sim_{\mu^*}\phi$, because $\mu^*(a)=\mu(a)=\mu^*(\phi)$ is not greater than $\mu^*(\phi)$. Hence, $\varphi \not\in\rep(\ty^*)$, and this contradicts our hypothesis. We recall that a pseudo-valuation on $K[x]$ is a map $K[x]\to {\mathbb Q}\cup\{\infty\}$ having the same properties as a valuation, except for the fact that the pre-image of $\infty$ is a prime ideal which is not necessarily zero. \[inequality\] Let $\mu_\infty$ be a pseudovaluation on $K[x]$, and let $\mu$ be an inductive valuation such that $\mu<\mu_\infty$. Let $\Phi_{\mu,\mu_\infty}$ be the set of all monic polynomials $\phi\in K[x]$ of minimal degree satisfying $\mu(\phi)<\mu_\infty(\phi)$. Then, there is a unique $\ll\in\mx(\Delta(\mu))$ such that $$\Phi_{\mu,\mu_\infty}=\left\{\phi \in\kpm\mid \rr_{\mu}(\phi)=\ll\right\}.$$ By [@Vaq Thm. 1.15], any $\phi\in\Phi_{\mu,\mu_\infty}$ is a key polynomial for $\mu$ such that $$\phi\mmu g \sii \mu(g)<\mu_\infty(g),$$for any nonzero $g\in K[x]$. For any fixed $\phi\in\Phi_{\mu,\mu_\infty}$, Lemma \[mid=sim\] shows that $$\Phi_{\mu,\mu_\infty}=\left\{\varphi \in \kpm\mid \varphi\smu\phi\right\}$$is the $\mu$-equivalence class of $\phi$ inside $\kpm$. This ends the proof because, as seen in (\[repr\]), the fibers of the map $\rr_\mu\colon \kpm\to \mx(\Delta(\mu))$ are the $\mu$-equivalence classes in $\kpm$. \[values\] The group of values $\Gamma(\mu)$ of an inductive valuation $\mu$ satisfies $$\Gamma(\mu)=\{\mu(a)\mid a\in K[x],\ \deg a<e_\mu m_\mu\}.$$ By (\[em\]), $\mu$ admits a proper key polynomial $\phi$ of degree $e_\mu m_\mu$. Consider a MacLane chain of $\mu$ as in (\[depth\]) such that $\phi\nmid_\mu\phi_r$. Let $\mu'=[\mu;\phi,\nu]$ be any augmentation of $\mu$ determined by the choice of an arbitrary positive rational number $\nu$. The MacLane chain may be extended to a MacLane chain of length $r+1$ of $\mu'$ with last step $\mu\ \stackrel{\phi,\nu}\lra\ \mu'$. Now, the claimed identity on $\Gamma(\mu)$ is proved in [@Rid Lem. 3.2]. An example {#secExample} ========== Let $p$ be an odd prime number. Denote by $v$ the $p$-adic valuation on ${\mathbb Q}_p$ and let ${\mathbb F}={\mathbb Z}/p{\mathbb Z}$. Consider the polynomial: $$f=x^4-2(p+p^2-p^3)\,x^2+p^2+2p^3-p^4-2p^5+p^6+p^8\in{\mathbb Z}[x].$$ Let us apply the OM factorization method to compute the prime factors of $f$ in ${\mathbb Z}_p[x]$. Clearly, $R_0(f)=\overline{f(y)}=y^4$. Thus, the type of order zero, $\ty_0=(y)$, divides all prime factors of $f$, and we have $\ord_{\ty_0}(f)=4$. We choose $\phi_1=x$ as a representative of $\ty_0$. Let $\mu_0$ be the Gauss valuation extending $v$ to ${\mathbb Q}_p[x]$, introduced in section \[subsecGradedAlg\]. The Newton polygon $N_{\mu_0,x}(f)$ is one-sided of length $4$ and slope $-1/2$. For the computation of residual polynomials we use the explicit recurrent method described in [@gen Sec. 3.1]. we have: $$R_{\mu_0,x,1/2}(f)=y^2-2y+1=(y-1)^2.$$ Thus, we get a unique type of order one dividing all prime factors of $f$: $$\ty_1=(y;(x,1/2,y-1)),$$but we now have $\ord_{\ty_1}(f)=2$. Hence, either $f$ is irreducible over ${\mathbb Z}_p[x]$, or it is the product $f=FG$ of two quadratic polynomials with $\ord_{\ty_1}(F)=\ord_{\ty_1}(G)=1$. Take $\phi_2=x^2-p$ as a representative of $\ty_1$. The $\phi_2$-expansion of $f$ is: $$\label{exp2} f=\phi_2^2-2(p^2-p^3)\,\phi_2+p^4-2p^5+p^6+p^8.$$ The augmented valuation $\mu_1=[\mu_0;x,1/2]$ on ${\mathbb Q}_p[x]$ attached to $\ty_1$ acts on ${\mathbb Q}_p[x]$ as follows: $$\mu_1\left(\sum_{0\le s}a_sx^s\right)=\mn\left\{\mu_0(a_s)+s/2\right\}=\mn\left\{v(a_s)+s/2\right\}.$$ Since $\mu_1(\phi_2)=1$, the points $(s,\mu_1(a_s\phi_2^s))\in{\mathbb R}^2$ associated with the $ \phi_2$-expansion (\[exp2\]) are $(2,2)$, $(1,3)$, $(0,4)$. Thus, $N_{\mu_1,\phi_2}(f)$ is one-sided of length $2$ and slope $-1$. The corresponding residual polynomial is: $$R_{\mu_1,\phi_2,1}(f)=y^2-2y+1=(y-1)^2.$$ Again, we get only one type of order two dividing all prime factors of $f$: $$\ty_2=(y;(x,1/2,y-1);(\phi_2,1,y-1)),$$still satisfying $\ord_{\ty_2}(f)=2$. Let us take $\phi_3=\phi_2-p^2=x^2-p-p^2$ as the simplest representative of $\ty_2$. The $\phi_3$-expansion of $f$ is: $$\label{exp3} f=\phi_3^2+2p^3\,\phi_3+p^6+p^8.$$ The non-normalized valuation $\mu_2=[\mu_1;\phi_2,1]$ attached to $\ty_2$ acts on ${\mathbb Q}_p[x]$ as follows: $$\mu_2\left(\sum_{0\le s}a_s\phi_2^s\right)=\mn\left\{\mu_1(a_s)+2s\right\}.$$ Since $\mu_2(\phi_3)=2$, the points in ${\mathbb R}^2$ associated with the $ \phi_3$-expansion (\[exp3\]) are $(2,4)$, $(1,5)$, $(0,6)$. Thus, $N_{\mu_2,\phi_3}(f)$ is again one-sided of length $2$ and slope $-1$. The corresponding residual polynomial is: $$R_{\mu_2,\phi_3,1}(f)=y^2+2y+1=(y+1)^2.$$ Again, we get only one type of order three dividing all prime factors of $f$: $$\ty_3=(y;(x,1/2,y-1);(\phi_2,1,y-1);(\phi_3,1,y+1)),$$still satisfying $\ord_{\ty_3}(f)=2$. Let us take $\phi_4=\phi_3+p^3=x^2-p-p^2+p^3$ as a representative of $\ty_3$. The $\phi_4$-expansion of $f$ is: $$\label{exp4} f=\phi_4^2+p^8.$$ The valuation $\mu_3=[\mu_2;\phi_3,1]$ attached to $\ty_3$ acts on ${\mathbb Q}_p[x]$ as follows: $$\mu_3\left(\sum_{0\le s}a_s\phi_3^s\right)=\mn\left\{\mu_2(a_s)+3s\right\}.$$ Since $\mu_3(\phi_4)=3$, the points in ${\mathbb R}^2$ associated with the $ \phi_4$-expansion (\[exp4\]) are $(2,6),\, (0,8)$. Thus, $N_{\mu_3,\phi_4}(f)$ is again one-sided of length $2$ and slope $-1$. The corresponding residual polynomial is: $$R_{\mu_3,\phi_4,1}(f)=y^2+1.$$ The factorization of this polynomial in ${\mathbb F}[y]$ depends on the class of $p$ modulo $4$. The method proceeds in a different way according to this class. [**Case $p\equiv -1\md{4}$.** ]{} The polynomial $y^2+1$ is irreducible in ${\mathbb F}[y]$ and we get a unique type of order four dividing all prime factors of $f$: $$\ty_4=(y;(x,1/2,y-1);(\phi_2,1,y-1);(\phi_3,1,y+1);(\phi_4,1,y^2+1)),$$for which $\ord_{\ty_4}(f)=1$. This implies that $f$ is irreducible in ${\mathbb Z}_p[x]$. Also, if $L/{\mathbb Q}_p$ is the finite extension of ${\mathbb Q}_p$ determined by $f$, we have $$e(L/{\mathbb Q}_p)=e_1e_2e_3e_4=2,\quad f(L/{\mathbb Q}_p)=f_0f_1f_2f_3f_4=2,$$ where $e_i$ are the lowest term denominators of the slopes of $\ty_4$ and $f_i$ are the degrees of the $\psi$-polynomials of all levels of $\ty_4$. However, the information about $f$ we have been collecting in the type $\ty_4$ is not intrinsic. It depends on the choices of representatives for the types $\ty_0$, $\ty_1$, $\ty_2$, $\ty_3$. Let us consider the following optimal type equivalent to $\ty_4$: $$\ty=(y;(x,1/2,y-1);(\phi_4,3,y^2+1)),$$ obtained by an iterative application of Lemma \[optstep\]. By Theorem \[finalchar\], $\ord_\ty(f)=1$ and $f$ is a representative of $\ty$. Moreover, since the type $\ty$ is strongly optimal, the equivalence class of $\ty$ is the canonical class attached to the Okutsu class of $f$ through the mapping of (\[main\]). Therefore, the data supported by $\ty$ are intrinsic data of $f$. For instance, the *Okutsu depth* of $f$ is two and $[x,\,\phi_4]$ is an *Okutsu frame* of $f$ [@okutsu]. This means that $$\begin{array}{rl} \frac 12 = v(\t)\ge v(h(\t)),&\mbox{ for all monic }h\in{\mathbb Z}_p[x]\mbox{ with }\deg h<2,\\ 3 = v(\phi_4(\t))\ge v(h(\t)),&\mbox{ for all monic }h\in{\mathbb Z}_p[x]\mbox{ with }\deg h<4, \end{array}$$ where $\t$ is a root of $f$ in $\overline{{\mathbb Q}}_p$. In particular, the slopes $1/2$ and $3$ are intrinsic data of $f$. [**Case $p\equiv 1\md{4}$.** ]{} The polynomial $y^2+1$ splits as $(y-i)(y+i)$ in ${\mathbb F}[y]$, where $i\in{\mathbb F}$ satisfies $i^2=-1$. We get then two inequivalent types dividing $f$: $$\begin{array}{c} \ty_4=(y;(x,1/2,y-1);(\phi_2,1,y-1);(\phi_3,1,y+1);(\phi_4,1,y-i)),\\ \ty'_4=(y;(x,1/2,y-1);(\phi_2,1,y-1);(\phi_3,1,y+1);(\phi_4,1,y+i)), \end{array}$$ with $\ord_{\ty_4}(f)=\ord_{\ty'_4}(f)=1$. This implies that $f=FF'$ splits in ${\mathbb Z}_p[x]$ into the product of two monic quadratic irreducible polynomials $F$, $F'$ such that $$\ord_{\ty_4}(F)=1,\ \ord_{\ty'_4}(F)=0;\quad \ord_{\ty_4}(F')=0,\ \ord_{\ty'_4}(F')=1.$$ If $L/{\mathbb Q}_p$, $L'/{\mathbb Q}_p$ are the quadratic extensions of ${\mathbb Q}_p$ determined by these prime factors, we have $$e(L/{\mathbb Q}_p)=e(L'/{\mathbb Q}_p)=2,\quad f(L/{\mathbb Q}_p)=f(L'/{\mathbb Q}_p)=1.$$ Also, by taking representatives of these types we obtain concrete Okutsu approximations to the unknown factors $F$, $F'$: $$\label{gg} \as{1.2} \begin{array}{l} G:=\phi_4-ip^4=x^2-p-p^2+p^3-ip^4\approx F,\\ G':=\phi_4+ip^4=x^2-p-p^2+p^3+ip^4\approx F', \end{array}$$ where now $i\in{\mathbb Z}$ is an arbitrary lifting of $i\in{\mathbb F}$. Again, the information about $F$, $F'$ contained in the types $\ty_4$, $\ty'_4$, respectively, is not intrinsic. Consider the optimal types equivalent to $\ty_4$, $\ty'_4$, respectively: $$\ty=(y;(x,1/2,y-1);(\phi_4,3,y-i)),\qquad \ty'=(y;(x,1/2,y-1);(\phi_4,3,y+i)),$$ obtained by an iterative application of Lemma \[optstep\]. By Theorem \[finalchar\], these types satisfiy $$\ord_{\ty}(F)=1,\ \ord_{\ty'}(F)=0;\quad \ord_{\ty}(F')=0,\ \ord_{\ty'}(F')=1,$$ and the polynomials $G$, $G'$ of (\[gg\]) are representatives of $\ty$, $\ty'$, respectively. Caution! The types $\ty$ and $\ty'$ are optimal, but not strongly optimal. Hence, the information contained in the last level of $\ty$, $\ty'$ is not intrinsic either. In this case, the equivalence class of strongly optimal types associated with the Okutsu class of $F$ is the class of the type $\ty_1$. In fact, by Lemma \[optstep\], $F$ and $F'$ are representatives of $\ty_1$. This means that the prime polynomials $F$, $F'$ both correspond to the same strongly optimal type $\ty_1$ by the mapping of (\[main\]); hence, these polynomials are Okutsu equivalent. Actually, if we denote by $[g]$ the Okutsu class of a prime polynomial $g\in\oo_v[x]$, we have: $$[x^2-p]=[\phi_4]=[F]=[F']=[G]=[G'],$$ and all these polynomials determine the same quadratic extension of ${\mathbb Q}_p$. In general, the extensions determined by two Okutsu equivalent prime polynomials in ${\mathbb Z}_p[x]$ have isomorphic maximal tamely ramified subextensions [@okutsu]. The type $\ty_1$ contains intrinsic information about all these Okutsu equivalent prime polynomials in ${\mathbb Z}_p[x]$. They all have Okutsu depth one, the family $[x]$ is an Okutsu frame and the slope $1/2$ has the following intrinsic meaning: $$\frac 12 = v(\t)\ge v(h(\t)),\mbox{ for all monic }h\in{\mathbb Z}_p[x]\mbox{ with }\deg h<2.$$ This situation enlightens an important feature of the OM factorization algorithm. When some prime factors of the input polynomial are in the same Okutsu class, the algorithm computes first the common strongly optimal (equivalence class of the) type attached to them, but then it must work further to enlarge this type with an adequate last level which enables one to distinguish the different prime factors. , *Residual ideals of MacLane valuations*, Journal of Algebra [**427**]{} (2015), 30–75. , *Okutsu invariants and Newton polygons*, Acta Arithmetica [**145**]{} (2010), 83–108. , *Higher Newton polygons in the computation of discriminants and prime ideal decomposition in number fields*, Journal de Théorie des Nombres de Bordeaux [**23**]{} (2011), no. 3, 667–696. , *Newton polygons of higher order in algebraic number theory*, Transactions of the American Mathematical Society [**364**]{} (2012), no. 1, 361–416. , *A new computational approach to ideal theory in number fields*, Foundations of Computational Mathematics [**13**]{} (2013), 729–762. , *Higher Newton polygons and integral bases*, Journal of Number Theory [**147**]{} (2015), 549–589. , *Genetics of polynomials over local fields*, Proceedings of AGCT14, Contemporary Mathematics [**637**]{} (2015), 207–241. , *Single-factor lifting and factorization of polynomials over local fields*, Journal of Symbolic Computation [**47**]{} (2012), 1318–1346. , *A construction for absolute values in polynomial rings*, Transactions of the American Mathematical Society, [**40**]{} (1936), pp. 363–395. , *A construction for prime ideals as absolute values of an algebraic field*, Duke Mathematical Journal [**2**]{} (1936), pp. 492–510. , *Polígonos de Newton de orden superior y aplicaciones aritméticas*, PhD Thesis, Universitat de Barcelona, 1999. , *Local computation of differents and discriminants*, Mathematics of Computation 83 (2014), no 287, 1513–1534. , *Construction of integral basis, I, II*, Proceedings of the Japan Academy [**58**]{}, Ser. A (1982), 47–49, 87–89. , *Zur Theorie der algebraischen Körper*, Acta Mathematica [**44**]{} (1923), pp. 219–314. , *Newtonsche Polygone in der Theorie der algebraischen Körper*, Mathematische Annalen [**99**]{} (1928), pp. 84–117. , *Extension d’une valuation*, Transactions of the American Mathematical Society [**359**]{} (2007), no. 7, 3439–3481. [^1]: Partially supported by MTM2013-40680-P from the Spanish MEC

ArXiv

--- abstract: 'Let $\mathcal{Z}(X)$ be the number of degree-d extensions of $\mathbb{F}_q(t)$ with bounded discriminant and some specified Galois group. The problem of computing $\mathcal{Z}(X)$ can be related to a problem of counting $\mathbb{F}_q$-rational points on certain Hurwitz spaces. Ellenberg and Venkatesh used this idea to develop a heuristic for the asymptotic behavior of $\mathcal{Z}''(X)$, the number of -geometrically connected- extensions, and showed that this agrees with the conjectures of Malle for function fields. We extend Ellenberg-Venkatesh’s argument to handle the more complicated case of covers of $\mathbb{P}^1$ which may not be geometrically connected, and show that the resulting heuristic suggests a natural modification to Malle’s conjecture which avoids the counterexamples, due to Klüners, to the original conjecture.' author: - Seyfi Türkelli title: 'Connected Components of Hurwitz Schemes and Malle’s Conjecture' --- Introduction ============ Let $k$ be a number field and let $N\leq S_n$ be a transitive group with one-point stabilizer $H\leq N$. By an $N$*-extension*, we mean a Galois extension $K/k$ with Galois group $G(K/k)\cong N$. We will denote the discriminant of a finite extension $K/k$ by $D(K/k)$. It is well known that the number of extensions $K/k$ with $\textbf{N}^k_{\mathbb{Q}}D(K/k)<X$ is finite. In [@ma2], Malle conjectures an asymptotic formula for the number of $N-$extensions $K/k$ of a fixed number field $k$ with $\textbf{N}^k_{\mathbb{Q}}D(K^H/k)<X$ where $K^H/k$ is the intermediate extension corresponding to $H$. In order to state Malle’s conjecture precisely, we need to introduce some invariants of the group $N$. Let $g\in N$. We define the *index* of $g$ to be the number $ind(g)=n-r$ where $r$ is the number of orbits of $g$ on set $\{1,2,...,n\}$. We define the *index* of the group $N$, $ind(N)$, to be the minimum of all $\{ind(g)\mid g\in N^{\#} \}$ where $N^{\#}=N-\{1\}$. Finally, we define our first invariant $a(N)=1/ind(N)$. In order to define our second invariant, we let $\mathcal{C}(N)$ be the set of conjugacy classes of $N^{\#}$ whose index is equal to the index of $N$ (the ones with the minimal index). We define a $G(\bar{k}/k)$ action on set $\mathcal{C}(N)$ via the cyclotomic character $\chi$ as $g.c:=c^{\chi(g)}$ for $g\in N$ and $c\in \mathcal{C}(N)$. Now, we define our second invariant to be the positive integer $b(N,k)=|\mathcal{C}(N)/G(\overline{k}/k)|$. Fix $k$ and denote the number of $N$-extensions $K/k$ with $\textbf{N}^k_{\mathbb{Q}}D(K^H/k)<X$ by $\mathcal{Z}_N(k,X)$. *Malle’s conjecture* [@ma2] is stated as follows: (Malle)\[malle\_con\] Let $k$ be a number field and $N$ be a transitive subgroup of $S_n$. Then, $$\mathcal{Z}_N(k,X) \asymp {X^{a(N)}}(logX)^{b(N,k)-1}.$$ Note that this conjecture is known for abelian groups and $k=\mathbb{Q}$ by the work of Wright [@wr], for $N=S_3$ by the work of Davenport-Heilbronn [@da], for $N=D_4,S_4,S_5$ and $k=\mathbb{Q}$ by the work of Bhargava, see [@bh1] and [@bh2]. In [@kl1], Klüners and Malle proved that for each positive $\epsilon >0$ there are positive constants $c_{\epsilon},C_{\epsilon}$ such that $c_{\epsilon} X^{a(N)} < \mathcal{Z}_N(\mathbb{Q},X) < C_{\epsilon} X^{a(N)+\epsilon}$ for any Nilpotent group $N$ given with its regular representation $N \hookrightarrow S_{|N|}$. Recently, Klüners gave a counterexample to this conjecture, see [@kl2]. Indeed, take $k=\mathbb{Q}$ and $N=(\langle(123)\rangle\oplus \langle(456)\rangle)\rtimes \langle(14)(25)(36)\rangle\leq S_6$ which is isomorphic to $C_3\wr C_2$ where $C_r$ is the cyclic group with $r$ elements. Then, one can easily see that the conjecture predicts that $\mathcal{Z}_N(\mathbb{Q},X) \asymp X^{1/2}$. Klüners shows that $\mathcal{Z}_N(\mathbb{Q},X) \asymp X^{1/2}logX$. He also points out that if one counts only the [*regular*]{} extensions [*i.e.*]{} extensions without an intermediate cyclotomic field, then one gets the asymptotic in the conjecture. For the rest of the paper, fix a $q$ which is coprime to $|N|$. One can state the conjecture for $k=\mathbb{F}_q (t)$ with evident modifications. In this setting, constant intermediate fields correspond to intermediate cyclotomic extensions. Subject to some heuristics on Hurwitz schemes, Ellenberg and Venkatesh [@el2] compute the size of the main term in the asymptotic for the number of $N$-extensions without constant subextensions, and obtain the analogue of Malle’s conjecture. This suggests that Malle’s conjecture may correctly predict the asymptotics for extensions without cyclotomic (constant) subextensions. Using the idea of Ellenberg-Venkatesh [@el2], we will count $N$-extensions $K/\mathbb{F}_q (t)$ with a maximal constant subextension corresponding to a fixed normal subgroup $G$; we will call such an extension $N_G$*-extension*. Then, we will take the maximum of the asymptotics of $N_G$-extensions as $G$ varies to get the asymptotic for $\mathcal{Z}_N(\mathbb{F}_q(t),X)$. We will propose a conjecture eliminating the Klüners’ counterexamples and compatible with the existing results. The idea is as follows: the category of $N_G$-extensions of $\mathbb{F}_q (t)$ is equivalent to the category of (connected but not necessarily geometrically connected) $N_G$-covers of $\mathbb{P}^1_{\mathbb{F}_q}$. Therefore, counting extensions is more or less equivalent to counting $\mathbb{F}_q$-rational points on the moduli space of certain covers of $\mathbb{P}^1$, namely Hurwitz schemes. Assuming the [*heuristic*]{}:  \ \[heu\](A) Let $\mathcal{H}$ be a geometrically connected scheme of dimension $d$ defined over $\mathbb{F}_q$; then one has $|\mathcal{H}(\mathbb{F}_q)|=q^d$,  \ we reduce the problem to one of counting irreducible components of Hurwitz spaces, and computing their dimension. More precisely, let $\mathcal{Z}_{G,N}(\mathbb{F}_q,X)$ be the number of $N_G$-extensions $K/\mathbb{F}_q(t)$ with $Norm D(K^H/\mathbb{F}_q(t))<X$. We will prove: \[main\_thm\] Assume (A). If $G$ splits in $N$, then $$\mathcal{Z}_{G,N}(\mathbb{F}_q,X)\asymp X^{a(G)}(logX)^{b(G,N,\mathbb{F}_q)-1}$$ where $b(G,N,\mathbb{F}_q)$ is an explicitly computable positive integer. As an immediate consequence, we have the result of Ellenberg-Venkatesh: \[ell\_ven\] Assume (A). Then, $\mathcal{Z}_{N,N}(\mathbb{F}_q,X)\asymp X^{a(N)}(logX)^{b(N,\mathbb{F}_q)-1}.$ In section 4, we will conjecture an asymptotic for the number of $N$-extensions of a global field. As evidence in the favor of our modification of Malle’s conjecture, we show that our version is not contradicted by Klüners’ counterexample from [@kl2]. \[counterexample\] Assume (A). Let $N=(\langle(123)\rangle\oplus \langle(456)\rangle)\rtimes \langle(14)(25)(36)\rangle\leq S_6$ and $q=2$ $(mod3)$. Then, we have $\mathcal{Z}_{N}(\mathbb{F}_q,X)\asymp X^{1/2}logX.$ We want to count branch covers of $\mathbb{P}^1_{\mathbb{F}_q}$ corresponding to $N_G$-extensions of $\mathbb{F}_q(t)$, namely $N_G$*-covers*. These covers are parameterized by certain $\mathbb{F}_q$-rational points of Hurwitz schemes of $(G,N)$-covers of $\mathbb{P}^1$. In section 2, we will state well known facts about these moduli schemes. In particular, we will introduce discrete invariants, so called Nielsen tuples, parameterizing “almost” geometrically connected components of Hurwitz schemes and we will determine the components parameterizing $N_G$-covers. In section 3, using the heuristic, we will reduce the problem of counting covers to counting Nielsen tuples. In section 4, we prove Theorem \[main\_thm\] and we conjecture an asymptotic for $\mathcal{Z}_N(k,X)$ where $k$ is a global field. Finally, we give some examples, as corollaries of Theorem \[main\_thm\], supporting our conjecture.  \ [*Acknowledgments*]{}. The author is very grateful to Jordan Ellenberg for his excellent guidance in preparation of this paper, and thanks to Jürgen Klüners for his very useful comments and pointing out an important mistake in an earlier version of the paper. Hurwitz Schemes of $(G,N)$-covers ================================= For the rest of the paper, fix a transitive subgroup $N\leq S_n$ and a normal subgroup $G$ of $N$ of cardinality $m$ with cyclic quotient $N/G$. Also, fix a one-point stabilizer $H\leq N$. Let $q$ be such that $(q,|N|)=1$. In this paper, we will prove the asymptotic formula under the assumption that $G$ splits in $N$, that is $N=G\rtimes T$ for some cyclic group $T$. We hope to treat the general case in a future paper. So, once and for all, fix a cyclic complement $T$ of $G$ in $N$ and an element $\tau\in T$ generating $T$. Let $C_T$ be the kernel of the projection $T\tilde{\rightarrow} N/G\twoheadrightarrow N/GCen_N(G)$, $T':=T/C_T\cong N/GCen_N(G)$ and $\tau_1\in T'$ be the image of $\tau$. Set $|T|=d$ and $|T'|=d'$. In this section, first, we will introduce $(G,N)$-*covers* of $\mathbb{P}^1/\mathbb{F}_q$ and discuss basic facts about their moduli spaces, namely Hurwitz schemes. Secondly, we will discuss the decomposition of Hurwitz schemes into their almost-geometrically-connected components. Finally, we will determine the components which parameterize the covers we want to count, namely $N_G$-*covers*. Most of the work we summarize in this section is due to Fried [@fr] and Wewers [@we]. \[defn\_gn\_extn\] An $N_G$-*extension* of $\mathbb{F}_q(t)$ is an $N$-extension $K/\mathbb{F}_q(t)$ such that $K^G/\mathbb{F}_q(t)$ is the maximal constant subextension of $K/\mathbb{F}_q(t)$ where $K^G$ is the subfield of $K$ corresponding to group $G$. Let $\rho: N\rightarrow S_m$ be the representation given by the action of $N$ on the left coset space $N/T$. Note that, for $g\in G$, $ind(g)$ will denote the index of $g\in S_n$, not the index of $\rho(g)$ in $S_m$. Let $Y/\mathbb{F}_q$ be a geometrically connected smooth curve and $f:Y\rightarrow \mathbb{P}^1/\mathbb{F}_q$ be a finite branched cover of degree $m=|G|$. Note that the field of moduli of the cover $(Y,f)$ is also the field of definition, see [@de1 Corollary 3.3]. Let $K/\mathbb{F}_q$ be the field of definition of the cover $(Y,f)$ and let $f_K:Y_K\rightarrow \mathbb{P}^1_K$ be one of its models over $K$ with branch locus $B$ where $B$ is a proper closed reduced subscheme of $\mathbb{P}^1_K$ of degree $k$. Consider pairs $(b,\sigma)$ where $b$ is a geometric point on $\mathbb{P}^1/\mathbb{F}_q$ and $\sigma:\{1,...,m\} \rightarrow f^{-1}(b)$ is a bijection such that 1. The image of the homomorphism $\tilde{\phi}:\Pi_1(\mathbb{P}^1_K\backslash B,b)\rightarrow S_m$ induced by $f$ and $\sigma$ is included in $\rho(N)$; 2. The image of $\phi:\Pi_1(\mathbb{P}^1_{\bar{\mathbb{F}}_{q}}\backslash B,b)\rightarrow S_m$ is $\rho(G)$ where $f^{-1}(b)$ denotes the geometric fiber. Two pairs $(b,\sigma)$ and $(b',\sigma ')$ are $N$-*equivalent* if there is a path $\gamma\in \Pi_1(\mathbb{P}^1_K\backslash B;b,b')$ such that $\sigma^{-1}\tilde{\gamma}\sigma '\in \rho(N)$ where $\tilde{\gamma}:f^{-1}(b')\rightarrow f^{-1}(b)$ is the bijection obtained from lifting $\gamma$ to $Y$. A *$(G,N)$-structure on $f$* is the $N$-equivalence class $\mathcal{S}=[(b,\sigma)]_N$ of a pair $(b,\sigma)$. A $(G,N)-cover$ is a pair $(f,\mathcal{S})$ where $f:Y\rightarrow \mathbb{P}^1/\mathbb{F}_q$ is a cover as above and $\mathcal{S}$ is a $(G,N)-$structure on $f$. We remark that $f:Y\rightarrow\mathbb{P}^1/\mathbb{F}_q$ is not a Galois cover in general but $\bar{f}=f\otimes \bar{\mathbb{F}}_q:\bar{Y}\rightarrow \mathbb{P}^1_{\bar{\mathbb{F}}_q}$ is Galois, see [@we Remark 3.4.4]. Abusing the notation, by $Im(\tilde{\phi})$ ( or $Im(\phi)$ ) we mean homomorphic image of $\tilde{\phi}$ ( or $Im(\phi)$ ) in $S_n$. In other words, a $(G,N)$-cover is a cover of $\mathbb{P}^1/\mathbb{F}_q$ with *geometric fundamental group*, $Im(\phi)$, $G$ and with *arithmetic fundamental group*, $Im(\tilde{\phi})$, at most $N$. Note that we want to count covers with arithmetic fundamental group exactly $N$. An $N_G-cover$ is a $(G,N)$-cover $(f,\mathcal{S})$ with $Im(\tilde{\phi})=N$. Note that given a geometric point $b$ on $\mathbb{P}^1/{\mathbb{F}_q}$ one can represent any $(G,N)$-structure $[(b',\sigma')]_N$ by $[(b,\sigma)]_N$ for some $\sigma$ because $\mathbb{P}^1/\mathbb{F}_q$ is path connected. A morphism between two $(G,N)$-covers $(f,[(b,\sigma)]_N)$ and $(f',[(b,\sigma')]_N)$ is a morphism $\psi:Y\rightarrow Y'$ of schemes over $\mathbb{F}_q$ such that $f=f'\psi$ and $[(b,\sigma)]_N=[(b,\psi^{-1}\sigma')]_N$. \[thm\_modspace\] There exists a smooth scheme $\mathcal{H}_{G,N}$ over $\mathbb{Z}[1/|G|]$ which is a coarse moduli scheme for $(G,N)$-covers of $\mathbb{P}^1$. Moreover, the fibers of the natural map $$\xymatrix{\Delta : \{ (G,N)\text{-covers of } \mathbb{P}^1 \text{ defined over } \mathbb{F}_q \}/ \cong \ar[r] & \mathcal{H}_{G,N}(\mathbb{F}_q)}$$ has size at most $|Cen_N(G)|$. Wewers carries out a detailed construction of $\mathcal{H}_{G,N}$ in [@we Section 4]. Let $P\in \mathcal{H}_{G,N}(\mathbb{F}_q)$. By section $4.1.3$ of [@we] and [@de1], the obstruction to $P$ arising from a cover lies in the second cohomology group $H^2(\mathbb{F}_q,Cen_N(G))$. Since $\mathbb{F}_q$ has cohomological dimension one, there is no obstruction. By Lemma \[lemma\_par\] below, $\Delta^{-1}(P)$ is parameterized by $H^1(\mathbb{F}_q, Cen_N(G))$. This cohomology group has size at most $|Cen_N(G)|$. \[lemma\_par\] The isomorphism classes of $(G,N)$-covers of $\mathbb{P}^1$ parameterized by the point $P\in \mathcal{H}_{G,N}(\mathbb{F}_q)$ is in one-to-one correspondence with the cohomology group $H^1(\mathbb{F}_q,Cen_N(G))$. Let $P\in \mathcal{H}_{G,N}(\mathbb{F}_q)$ and let $(f,[b,\sigma]_N)$ be a $(G,N)$-cover defined over $\mathbb{F}_q$ corresponding to the point $P$. Let $(f_1,[b,\sigma_1])$ be another one. Then, there is an isomorphism of $(G,N)$-covers $\psi:\bar{Y}\rightarrow \bar{Y}_1$ with $\bar{f}=\bar{f}_1\psi$ and $[b,\sigma]_N=[b,\psi^{-1}\sigma_1]_N$ over $\bar{\mathbb{F}}_q$ where $\bar{f}_i=f_i\times _{\mathbb{F}_q}\bar{\mathbb{F}}_q$ for $i=1,2$. This induces a map $$\delta _{\psi}: G_{\mathbb{F}_q}\rightarrow N \text{\indent defined by \indent } x\mapsto (\psi^{-1}\psi^x)^\sigma .$$ It is easy to see that $\psi^{-1}\psi^x\in Aut((\bar{f},[b,\sigma]_N))$ for all $x\in G_{\mathbb{F}_q}$. Since $Aut((\bar{f},[b,\sigma]_N))\cong Cen_N(G)$, Im($\delta_\psi$)$\subseteq Cen_N(G)$ [@we Lemma 3.4.2]. One checks that $\delta _{\psi}$ is a cocycle and $\delta$ defines the desired correspondence. Later, we will make use of the important fact that $\mathcal{H}_{G,G}\cong \mathcal{H}_G$ as schemes defined over $\mathbb{F}_q$ where $\mathcal{H}_G$ is the coarse moduli scheme of geometrically connected Galois $G$-covers of $\mathbb{P}^1$, see [@we proposition 4.2.2]. Let $Ni_k(G)$ be the set of $k$-tuples $\bar{g}=(g_1,...,g_k)$ generating $G$ and satisfying $g_1...g_k=1$. Note that $N$ acts on the set of such $k$-tuples by component-wise-conjugation; denote the orbit of $\bar{g}$ in this action by $[\bar{g}]_N$. Denote the set of such $N$-orbits of $k$-tuples by $Ni_k(G,N)$ and define the set of $G-N-$*Nielsen classes* to be the set $Ni(G,N):=\bigcup_k Ni_k(G,N)$. Given a $G-N-$Nielsen class $[\bar{g}]_N\in Ni_k(G,N)$ and a proper closed reduced subscheme $B$ of $\mathbb{P}^1_{\bar{\mathbb{F}}_{q}}$ of degree $k$, one has surjective homomorphisms $\pi_1(\mathbb{P}_{\bar{\mathbb{F}}_q}^1\backslash B,b_0)\rightarrow G$ modulo conjugation with elements of $N$. After fixing a set of generators of $\pi_1(\mathbb{P}_{\bar{\mathbb{F}}_q}^1\backslash B,b_0)$, one gets a one-to-one correspondence between these homomorphisms and $(G,N)$-covers$/\bar{\mathbb{F}}_q$ with branch locus $B$. Thus, such a orbit $[\bar{g}]_N\in Ni(G,N)$ and branch locus $B$ induces a $(G,N)$-cover and vice versa. Let $\mathcal{H}_{G,N,k}$ be the moduli space of $(G,N)$-covers with degree-$k$ branch locus and $U_k=\mathbb{P}^k\backslash \delta_k$ where $\delta_k$ is the discriminant locus. Then, the natural map $\pi: \mathcal{H}_{G,N,k}\rightarrow U_k$ taking a cover to its branch locus is étale. On the other hand, if $B$ is a geometric point of $U_k/\mathbb{F}_q$, then the braid group $\pi_1(\bar{U}_k,B)$, the geometric fundamental group of $U_k$, can be written in standard representation with generators $Q_1,...,Q_{k-1}$. In summary, there is a bijection between $Ni_k(G,N)$ and the geometric fiber $\pi^{-1}(B)$ [@we Proposition 4.3.1]. This bijection induces a well-known action of the braid group $\pi_1(\bar{U}_k,B)$ on $Ni_k(G,N)$ which is given by $$Q_i[(g_1,...,g_k)]_N:=[(g_1,...,g_i g_{i+1} g_i^{-1},g_i,...g_k)]_N$$ and the geometrically connected components of $\mathcal{H}_{G,N,k}$ correspond to the braid group orbits on $Ni_k(G,N)$. Given a $k$-tuple $\bar{C}=(C_1,...,C_k)$ of $G$-conjugacy classes, one can consider its orbit $[\bar{C}]_N$ under conjugation by $N$. Define $Ni(\bar{C})$ to be the set of $\bar{g}\in Ni_k(G)$ such that, after some permutation of the entries of $\bar{g}$, $g_i\in C_i$ for all $i$. Now, define $Ni([\bar{C}]_N)$ to be the set of $N-$orbits of elements of $Ni(\bar{C})$. Clearly, $Ni([\bar{C}]_N)$ is closed under the braid group action. Therefore, there is a closed subscheme (possibly empty) of $\mathcal{H}_{G,N}$ corresponding to the $N$-orbit $[\bar{C}]_N$, denoted by $\mathcal{H}_{[\bar{C}]_N}$, which parameterizes $(G,N)$-covers with ramification data $[\bar{C}]_N$. Thus, we have the decomposition $\mathcal{H}_{(G,N)}=\bigsqcup_{[\bar{C}]_N}\mathcal{H}_{[\bar{C}]_N}$ over $\bar{\mathbb{F}}_q$. For more details of the discussion above, we refer to [@we Section 4.3]. In summary, we have: \[thm\_decomp\] [@we Section 4] For each $N$-orbit $[\bar{C}]_N$ of a tuple of conjugacy classes $\bar{C}$, there is a Hurwitz scheme $\mathcal{H}_{[\bar{C}]_N}/\bar{\mathbb{F}}_q$ which is a coarse moduli scheme for $(G,N)$-covers $Y\rightarrow \mathbb{P}^1_{\bar{\mathbb{F}}_q}$ with ramification data $[\bar{C}]_N$. A Galois transformation $\sigma\in G_{\mathbb{F}_q}$ acts on $\mathcal{H}_{[\bar{C}]_N}$ by $\mathcal{H}^\sigma_{[\bar{C}]_N}=\mathcal{H}_{[\bar{C}]_N^\sigma}$. Moreover, the map $\pi: \mathcal{H}_{[\bar{C}]_N}\rightarrow U_k$ sending a cover to its branch locus is étale and geometric points of the fiber over $B\in U_k(\bar{\mathbb{F}}_q)$ can be identified with $Ni([\bar{C}]_N)$. The induced action of the monodromy group $\Pi_1(U_k)$ on $Ni([\bar{C}]_N)$ is given by $$Q_i(g_1,...,g_k)=(g_1,..,g_i g_{i+1} g_i^{-1},g_i,..,g_k)$$ and, thus geometrically connected components of $\mathcal{H}_{[\bar{C}]_N}$ correspond to $\Pi_1(U_k)$-orbits of $Ni([\bar{C}]_N)$. Note that the arithmetic fundamental group of a $(G,N)$-cover is a subgroup of $N$ containing $G$ and we want to count the ones defined over $\mathbb{F}_q$ whose arithmetic fundamental group is exactly $N$, namely $N_G$ covers defined over $\mathbb{F}_q$. Note also that an $N_G$-cover defined over $\mathbb{F}_q$ corresponds to a $N_G$-extension of $\mathbb{F}_q(t)$ and vice versa. Let $\mathcal{H}_{G,N}(\mathbb{F}_q)^N$ be the subset of $\mathcal{H}_{G,N}(\mathbb{F}_q)$ consisting of all the points parameterizing $N_G$-covers defined over $\mathbb{F}_q$; that is $$\mathcal{H}_{G,N}(\mathbb{F}_q)^N:=\Delta(\{N_G\text{-covers of } \mathbb{P}^1 \text{ defined over }\mathbb{F}_q \}/\cong)$$ where $\Delta$ is the natural map in Theorem \[thm\_modspace\]. Below, we determine the components whose $\mathbb{F}_q$-rational points parameterize $N_G$-covers defined over $\mathbb{F}_q$. For any given subgroup $N'\subseteq N$ containing $G$, the natural map $$\Lambda_N^{N'}:\mathcal{H}_{G,N'}\rightarrow \mathcal{H}_{G,N} \text{\indent induced by \indent} (f,[(b,\sigma)]_{N'})\mapsto (f,[(b,\sigma)]_{N})$$ is an étale cover of degree $|Out_N(N')|=| N/N'Cen_N(N')|$ and the map $\pi': \mathcal{H}_{G,N'}\rightarrow U_k$ factors through $\pi: \mathcal{H}_{G,N}\rightarrow U_k$, see [@fr Section 6.1] and [@we Section 4.2]. Note that the image of $\mathcal{H}_{[\bar{C}]_{N'}}$ under $\pi$ is $\mathcal{H}_{[\bar{C}]_N}$. In fact, if $\mathcal{H}_{[\bar{C}]_{N'}}$ is connected and defined over $\mathbb{F}_q$, then $\Lambda_N^{N'}$ is a Galois cover with automorphism group isomorphic to $Out_N(N')$ where $Out_N(N')$ denotes the image of $N$ in $Out(N')$. In particular, by taking $N'=G$ we obtain an étale cover $\Lambda:\mathcal{H}_{G}\rightarrow \mathcal{H}_{G,N}/\mathbb{F}_q$ of degree $d'=|T'|=|T/C_T|$ (recall that there is a canonical isomorphism $\mathcal{H}_{G}\cong\mathcal{H}_{G,G}/\mathbb{F}_q$). Given a geometric point $P$ in $\mathcal{H}_{G,N}/\mathbb{F}_q$, $C_T$ acts on the geometric fiber $\Lambda^{-1}(P)$ trivially and this induces a sharply transitive action of $T'$ on $\Lambda^{-1}(P)$. Let $P\in \mathcal{H}_{G,N}(\mathbb{F}_q)$ and $Q\in \Lambda^{-1}(P)$. Then, the geometric fiber $\Lambda^{-1}(P)$ is defined over $\mathbb{F}_q$ and for all $\sigma\in G_{\mathbb{F}_q}$ there exists a unique $\zeta(\sigma)\in T'$ such that $Q^{\sigma}=Q^{\zeta(\sigma)}\in \Lambda^{-1}(P)$. One can easily see that $\zeta\in H^1(\mathbb{F}_q,T')$ (here the action of the absolute Galois group $G_{\mathbb{F}_q}$ on $T'$ is trivial so indeed $\zeta\in Hom(G_{\mathbb{F}_q},T')$) and it is independent of the choice of $Q\in \Lambda^{-1}(P)$. Therefore, $\Lambda^{-1}(P)\subseteq \mathcal{H}^{\zeta}_G(\mathbb{F}_q)$ where $\mathcal{H}^{\zeta}_G$ is the twist of $\mathcal{H}_G$ via the composition of $\zeta$ with the embedding of $T'$ in $Aut(\mathcal{H}_G/ \mathcal{H}_{G,N})$. On the other hand, given a point $Q\in \mathcal{H}^{\zeta}_G(\mathbb{F}_q)$ for some $\zeta\in H^1(\mathbb{F}_q,T')$, $P:=\Lambda(Q)\in \mathcal{H}_{G,N}(\mathbb{F}_q)$. Thus, we have $$\Lambda^{-1}(\mathcal{H}_{G,N}(\mathbb{F}_q))=\bigsqcup_{\zeta\in H^1(\mathbb{F}_q,T')}\mathcal{H}^{\zeta}_G(\mathbb{F}_q).$$ Now, we want to determine the cocycles $\zeta\in H^1(\mathbb{F}_q,T')$ for which the twists $\mathcal{H}^{\zeta}_G(\mathbb{F}_q)$ induce $N_G$-covers via $\Lambda$. So, let $|T'|=d'$ and let $e$ be a positive integer with $1\leq e\leq d'$. We denote the $1$-cocycle sending $Frob_q$ to $\tau_1^e$ by $\zeta_e\in H^1(\mathbb{F}_q,T')$ where $\tau_1$ is the image of $\tau\in T$ under the projection $T\twoheadrightarrow T'$. The following proposition tells us the schemes we should consider. \[prop\_dec\] We have the following decomposition $$\Lambda^{-1}(\mathcal{H}_{G,N}(\mathbb{F}_q)^N)= \bigsqcup_{\substack{1\leq e \leq d' \\ (e,d')=1}}\mathcal{H}^{\zeta_e}_G(\mathbb{F}_q).$$ Giving a $(G,N)$-cover defined over $\mathbb{F}_q$ with branch locus $B$ is equivalent to giving the following diagram of fundamental groups: $$\xymatrix{ 1\ar[r] & \Pi_1(\mathbb{P}^1_{\bar{\mathbb{F}}_{q}}\backslash B,b)\ar[r]\ar@{>>}[d]_{\phi} & \Pi_1(\mathbb{P}^1_{\mathbb{F}_{q}}\backslash B,b)\ar[r]\ar[d]_{\tilde{\phi}} & G(\bar{\mathbb{F}}_{q}/\mathbb{F}_{q})\ar[r]\ar[d]_{\bar{\phi}} & 1 \\ 1 \ar[r] & G \ar[r] & N \ar[r] & T }.$$ The cover corresponding to the above diagram is actually an $N_G$-cover if and only if $\bar{\phi}$ is surjective *i.e.* $\bar{\phi}(Frob_q)=\tau^e$ for some positive integer $e$ with $1\leq e \leq d$ and $(e,d)=1$. Let $P\in \mathcal{H}_{G,N}(\mathbb{F}_q)^N$ and let $Q\in \Lambda^{-1}(P)$. Then, by the definition of $\Lambda$, $$Q^{Frob_q}=Q^{\tau_1^e}$$ for some unique integer $e$ with $$1\leq e \leq d' \text{ \indent and\indent } (e,d')=1.$$ Therefore, $Q\in \mathcal{H}^{\zeta_e}_G(\mathbb{F}_q)$ and $\Lambda^{-1}(\mathcal{H}_{G,N}(\mathbb{F}_q)^N)\subseteq \bigsqcup_{\substack{1\leq e \leq d' \\ (e,d')=1}}\mathcal{H}^{\zeta_e}_G(\mathbb{F}_q).$ The difficult part is to prove the inequality “$\supseteq$”; that’s what we do below. Let $e$ be a positive integer with $(e,d')=1$ and $1\leq e\leq d'$ and let $Q\in \mathcal{H}^{\zeta_e}_G(\mathbb{F}_q)$. We want to show that $Q\in \Lambda^{-1}(\mathcal{H}_{G,N}(\mathbb{F}_q)^N).$ In other words, we want to show that there exists an $N_G$-cover $(f,\mathcal{S})$ such that $\Delta((f,\mathcal{S}))=P$ where $P=\Lambda(Q)$ and $$\xymatrix{\Delta : \{ (G,N)\text{-covers of } \mathbb{P}^1 /\mathbb{F}_q \}/ \cong \ar[r] & \mathcal{H}_{G,N}(\mathbb{F}_q)}$$ is the natural map defined in Theorem \[thm\_modspace\]. Let $(f_0,\mathcal{S}_0)\in \Delta^{-1}(P)$. Then $(f_0,\mathcal{S}_0)$ is an $M_G$-cover for some subgroup $M\subseteq N$ containing $G$. Let $T_0\subseteq T$ be the complement of $G$ in $M$, that is $M=G\rtimes T_0$. $(f_0,\mathcal{S}_0)$ induces a diagram as above, which is induced by a surjective morphism of groups $\tilde{\phi}_0: \Pi_1(\mathbb{P}^1_{\mathbb{F}_{q}}\backslash B,b)\twoheadrightarrow M$. Then, $\bar{\phi}_0(Frob_q)=\tau^{e'}$ for some $e'$ with $1\leq e'\leq d$. This implies that $Q^{Frob_q}=Q^{\tau^{e'}}$. On the other hand, $Q^{Frob_q}=Q^{\tau^e}$ since $Q\in \mathcal{H}^{\zeta_e}_G(\mathbb{F}_q)$. Therefore, $Q^{\tau^{e-e'}}=Q$ and $\tau^{e-e'}\in C_T$, that is $\tau^e=\tau^{e'}$ in $T/C_T$. Since $(e,d')=1$, $\tau^e$ generates $T/C_T\cong N/GCen_N(G)$ and so does $\tau^{e'}$. Thus, we have natural isomorphisms: $$N/Cen_N(G)\cong M/Cen_{M}(G) \indent \text{and} \indent N/GCen_N{G}\cong M/GCen_{M}(G).$$ Recall that $|T/C_T|=d'$ and $|C_T|=d''$. One can easily see that $(e+ad',d)=1$ for some $a$ with $0\leq a \leq d''-1$. So, let $e''=e+ad'<d$ be such that $(e'',d)=1$. Note that $\tau^e=\tau^{e'}=\tau^{e''}$ in $N/GCen_N(G)$. Let $\phi:=\phi': \Pi_1(\mathbb{P}^1_{\bar{\mathbb{F}}_{q}}\backslash B,b)\twoheadrightarrow G$ be a morphism induced by the $M_G$-cover. Define $\bar{\phi}:G_{\mathbb{F}_{q}}\twoheadrightarrow N/G$ by $Frob_q\mapsto \tau^{e''}G$. Now, we have the following commutative diagram: $$\xymatrix{ 1\ar[r] & \Pi_1(\mathbb{P}^1_{\bar{\mathbb{F}}_{q}}\backslash B,b)\ar[r]\ar@{>>}[d]_{[\phi']}\ar@{>>}[ddr]^{\phi} & \Pi_1(\mathbb{P}^1_{\mathbb{F}_{q}}\backslash B,b)\ar[r]\ar@{>>}[d]_{[\tilde{\phi}']}\ar@{-->>}[ddr]^{\tilde{\phi}} & G_{\mathbb{F}_{q}}\ar[r]\ar@{>>}[d]_{[\bar{\phi}']}\ar@{>>}[ddr]^{\bar{\phi}} & 1 & \\ 1\ar[r] & G/Cen(G)\ar[r] & N/Cen_N(G)\ar[r] & N/GCen_N(G)\ar[r] & 1 & \\ & 1\ar[r] & G\ar[r]\ar@{>>}[ul] & N\ar[r]\ar@{>>}[ul] & N/G\ar[r]\ar@{>>}[ul] & 1 }$$ where $[\tilde{\phi}']$ is the map induced by $\tilde{\phi}'$ and the canonical isomorphism $M/Cen_{M}(G)\cong N/Cen_N(G)$. The obstruction to the existence of $\tilde{\phi}$ lies in $H^2(\mathbb{F}_q,Cen_N(G))$ [@de1 Theorem 4.3]. Since $\mathbb{F}_q$ has cohomological dimension $1$, there exists such a lift $\tilde{\phi}:\Pi_1(\mathbb{P}^1_{\mathbb{F}_{q}}\backslash B,b)\rightarrow N$. Since $(e'',d)=1$, $\bar{\phi}$ is surjective and so is $\tilde{\phi}$. Hence, the diagram above corresponds to an $N_G$-cover, say $(f,\mathcal{S})$. By the construction, $(f,\mathcal{S})\in \Delta^{-1}(P)$ and this completes the proof. We call a tuple $\bar{C}$ of conjugacy classes of $G$ *Nielsen tuple* and its $N$-orbit $[\bar{C}]_N$ $N-Nielsen$ $tuple$. We will also refer to $N$-Nielsen tuples simply as Nielsen tuples when the meaning is clear from the context. Given two Nielsen tuples $\bar{C}$ and $\bar{C}'$, we write $\bar{C}=\bar{C}'$ if they differ only by a permutation of the entries. We denote their concatenation by $\bar{C}+\bar{C'}$. Given an integer $1\leq e\leq d'$ with $(e,d')=1$ and Nielsen tuple $\bar{C}$, we say $\bar{C}$ is $\mathbb{F}_q-rational$ *of type* $e$ if $\bar{C}^{q\tau^{-e}}=\bar{C}$. Our motivation for this definition is the following proposition. \[prop\_nil\] For every $e$ with $1\leq e \leq d'$ and $(e,d')=1$, we have $$\mathcal{H}^{\zeta_e}_G(\mathbb{F}_q)=\bigsqcup_{\substack{\bar{C} \\ \bar{C}^{q\tau^{-e}}=\bar{C}}}\mathcal{H}^{\zeta_e}_{\bar{C}}(\mathbb{F}_q).$$ where the union runs over $\mathbb{F}_q$-rational Nielsen tuples of type $e$. Let $\bar{C}$ be a Nielsen tuple. $\bar{C}=[\bar{C}]_G$ and $\mathcal{H}_{\bar{C}}\cong\mathcal{H}_{[\bar{C}]_G}$ under the isomorphism $\mathcal{H}_G\cong\mathcal{H}_{G,G}$. By Theorem \[thm\_decomp\] , $\bar{C}^{q\tau^{-e}}=\bar{C}$ if and only if the corresponding component $\mathcal{H}^{\zeta_e}_{\bar{C}}$ is defined over $\mathbb{F}_q$. Applying Theorem \[thm\_decomp\] for $\mathcal{H}_G$, we get the desired decomposition. Connected Components of Hurwitz Schemes ======================================= Given an $N_G$-cover $(f,\mathcal{S})$ defined over $\mathbb{F}_q$, by definition we have a surjective morphism $\tilde{\phi}:\Pi_1(\mathbb{P}^1_{\mathbb{F}_q}\backslash B,b)\rightarrow \rho(N)$ which is unique up to composition with an inner automorphism of $N$ [@we Remark 3.4.1]. If $L/\mathbb{F}_q(t)$ is the $N$-extension corresponding to $ker(\tilde{\phi})$, then $L=\mathbb{F}_q(Y)$ for some connected (not necessarily geometrically connected) curve $Y/\mathbb{F}_q$ and the extension $\mathbb{F}_q(Y)/\mathbb{F}_q(t)$ corresponds to a Galois $N-$cover $g:Y\rightarrow\mathbb{P}^1_{\mathbb{F}_q}$. Now, let $g':Y'\rightarrow \mathbb{P}^1_{\mathbb{F}_q}$ be the degree-$n$ cover associated to one-point stabilizer $H$. The [*discriminant*]{} of the $N_G$-cover $(f,\mathcal{S})$ is the number $q^{r(f,\mathcal{S})}$ where $r(f,\mathcal{S})$ is the degree of the ramification divisor of the degree-$n$ cover $g':Y'\rightarrow \mathbb{P}^1_{\mathbb{F}_q}$. Define our counting function $\mathcal{Z}_{G,N}(\mathbb{F}_q,X)$ to be the number of isomorphism classes of $N_G$-covers $(f,\mathcal{S})$ defined over ${\mathbb{F}_q}$ with $q^{r(f,\mathcal{S})}<X$. For a given $k$-tuple of conjugacy classes $\bar{C}=(C_1,...,C_k)$, set $|\bar{C}|:=k$ and $r(\bar{C}):=\sum_{i=1}^k ind(C_i)$ where $ind(C)$ is the index of a representative of the conjugacy class $C$. Let $\Sigma_{r,e}$ be the set of Nielsen tuples $\bar{C}$ of type $e$ with $r(\bar{C})=r$; and let $\Sigma_r$ be the set of Nielsen tuples $\bar{C}$ with $r(\bar{C})=r$. Let $\Sigma^N_r$ be the set of $G-N$-Nielsen tuples $[\bar{C}]_N$ with $r(\bar{C})=r$. Finally, let $\mathcal{H}_{{[\bar{C}]_N}}(\mathbb{F}_q)^N$ denote the subset of $\mathcal{H}_{{[\bar{C}]_N}}(\mathbb{F}_q)$ consisting of all the points in $\mathcal{H}_{{[\bar{C}]_N}}(\mathbb{F}_q)$ parameterizing $N_G$-covers. We will need the following counting functions: 1. $h(q,r):=\sum_{[\bar{C}]_N\in \Sigma^N_r}|\mathcal{H}_{{[\bar{C}]_N}}(\mathbb{F}_q)^N|$. 2. $h_1(q,r,e):=\sum_{\bar{C}\in \Sigma_{r,e}} |\mathcal{H}^{\zeta_e}_{\bar{C}}(\mathbb{F}_q)|$ and $h_1(q,r):=\sum_{\substack{1\leq e\leq d'\\ (e,d')=1}} h_1(q,r,e)$. 3. $h_2(q,r,e):=\sum_{\bar{C}} q^{|\bar{C}|}$ where the sum runs over all geometrically connected components of $\mathcal{H}^{\zeta_e}_{\bar{C}}$’s defined over $\mathbb{F}_q$ where $\bar{C}\in \Sigma_{r,e}$ and $h_2(q,r):=\sum_{\substack{1\leq e\leq d'\\ (e,d')=1}} h_2(q,r,e)$. 4. $h_3(q,r,e):=\sum_{\bar{C}\in \Sigma_{r,e}} q^{|\bar{C}|}$ and $h_3(q,r):=\sum_{\substack{1\leq e\leq d'\\ (e,d')=1}} h_3(q,r,e)$. Recall that we want to count the $N$-covers defined over $\mathbb{F}_q$. More precisely, we want to compute $\mathcal{Z}_{G,N}(\mathbb{F}_q,X)$ which is asymptotic to $\sum_{q^r<X}h(q,r)$ by Theorem \[thm\_decomp\]. On the other hand, by Proposition \[prop\_dec\], we have $\sum_{q^r<X}h(q,r)\asymp \sum_{q^r<X}h_1(q,r)$. Thus, we get: \[lemma\_h1\] We have $$\mathcal{Z}_{G,N}(\mathbb{F}_q,X)\asymp \sum_{q^r<X}h_1(q,r).$$ Observe that $\sum_{q^r<X}h_2(q,r)$ is a good approximation to the desired sum $\sum_{q^r<X}h_1(q,r)$ on heuristic grounds. Therefore, in this section, our aim is to compute the sum $\sum_{q^r<X}h_2(q,r)$ (and then we will use the heuristic); this is not easy. If $\mathcal{H}^{\zeta_e}_{\bar{C}}$ were geometrically connected for all $\mathbb{F}_q$-rational of type $e$ Nielsen tuples $\bar{C}$ and for all relevant $e$, then the sum $\sum_{q^r<X}h_2(q,r)$ would be equal to $\sum_{q^r<X}h_3(q,r)$ and so we would reduce the problem to computing $\sum_{q^r<X}h_3(q,r)$. Unfortunately, in general, this is not the case. Note that computing the sum $\sum_{q^r<X}h_3(q,r)$ boils down to computing the connected components of Hurwitz spaces and this is a very old combinatorial problem with a rich history, going all the way back to Hurwitz [@hr] and Clebsch [@cl]. In this section, we will see that there are “many” $\mathbb{F}_q$-rational Nielsen tuples $\bar{C}$ such that $\mathcal{H}^{\zeta_e}_{\bar{C}}$ possesses a geometrically connected component defined over $\mathbb{F}_q$ for all $e$ and the number of these components is bounded by positive constant depending only on the group $N$. Thus, we will reduce the problem to computing $\sum_{q^r<X}h_3(q,r)$. More precisely, the purpose of this section is to prove the following proposition. \[prop\_main\] Let $e$ be such that $1\leq e\leq d'$ and $(e,d')=1$. There exist positive constants $m, c_1$ depending on $N$ such that $$\sum_{r<R-m}h_3(q,r,e)<\sum_{r<R}h_2(q,r,e) < c_1 \sum_{r<R}h_3(q,r,e).$$ By the lemma below, we have the right-hand-side inequality. \[lemma\_upb\][@el2 Lemma 3.3] There exists a constant $c_1$ such that $n(\bar{C})< c_1$ for all $\bar{C}$ where $n(\bar{C})$ is the number of $\Pi_1(U_k)$-orbits in $Ni({\bar{C}})$. As for the inequality on the left-hand-side, we need a result controlling the geometrically connected components of Hurwitz spaces. The first such theorem along the lines presented here is attributed to Conway and Parker– the first version of such a theorem to appear in print is due to Fried and Völklein [@fr]. \[lemma\_cp\](Conway-Parker) [@fr Appendix] Let $\tilde{G}'$ be a finite group such that the Schur multiplier $M(\tilde{G}')$ is generated by commutators. Then, there exists a constant $K$ such that for any Nielsen tuple $\bar{E}$ of $\tilde{G}'$ which contains at least $K$ copies of each nontrivial conjugacy classes of $\tilde{G}'$ the corresponding Hurwitz space $\mathcal{H}_{\bar{E}}$ is geometrically connected. We will need the next two technical lemmas to prove Proposition \[prop\_main\]. First of them is needed to apply Lemma \[lemma\_cp\] to our setting. \[lemma\_diag\] There exists a commutative diagram of finite groups $$\xymatrix{\tilde{N}\ar@{>>}[r]^\pi & N \\ \tilde{G}\ar@/_/[rr]_\pi\ar@{>>}[r]^{\pi'}\ar@{^{(}->}[u] & \tilde{G}'\ar@{>>}[r] & G\ar@{^{(}->}[ul] }$$ such that the Schur multiplier $M(\tilde{G}')$ is generated by commutators, where the vertical maps are inclusions and horizontal maps are surjections. By Lemma 1 of [@fr], there exists an extension of groups $$\xymatrix{1\ar[r] & M'\ar[r] & \tilde{G}'\ar[r] & G\ar[r] & 1}$$ such that Schur multiplier $M(\tilde{G}')\cong M'$ is generated by commutators. Let $M:=Ind_N^G(M')$ be the induced $N$-module. Then, by Shapiro’s lemma, $H^2(G,M')$ is isomorphic to $H^2(N,M)$. Let $$\xymatrix{1\ar[r] & M\ar[r] & \tilde{N}\ar[r]^\pi & N\ar[r] & 1}$$ be an extension which corresponds to the extension above via the isomorphism $H^2(G,M')\cong H^2(N,M)$, and let $\tilde{G}=\pi^{-1}(G)$. Now, the evaluation morphism $M\rightarrow M'$ defined by $f\mapsto f(1)$ induces the following surjective morphism of groups $\pi':\tilde{G}\twoheadrightarrow \tilde{G}'$. One can easily check that these extensions fits into the desired commutative diagram above and complete the proof. Let $\pi:\tilde{N}\rightarrow N$ be a surjective morphism of finite groups with $\tilde{G}:=\pi^{-1}(G)$ as in Lemma \[lemma\_diag\] . Clearly, $\tilde{G}$ is normal subgroup of index $d$ with cyclic quotient. Let’s fix and denote $\tilde{\tau}\in \tilde{N}$ generating $\tilde{T}:=\tilde{N}/\tilde{G}$ with $\pi(\tilde{\tau})=\tau$. Let $\tilde{T}':=\tilde{N}/\tilde{G}Cen_{\tilde{N}}\tilde{G}$. Simplifying the notation, $\tilde{\tau}$ will also denote its image in $\tilde{T}$ and $\tilde{\tau}_1$ will denote its image in $\tilde{T}'$. Note that the map $\bar{\pi}:\tilde{T}'\rightarrow T'$ induced by $\pi$ is an isomorphism. Note also that $\tilde{N}$ acts on geometrically connected Galois $\tilde{G}$-covers and this induces an action of $\tilde{T}'$ on $\mathcal{H}_{\tilde{G}}$. Namely, if a geometrically connected Galois $G$-cover is induced by a morphism $\phi:\pi_1(\mathbb{P}^1_{\bar{\mathbb{F}}_q}\backslash B,b)\rightarrow \tilde{G}$ then the action of $x\in N$ is defined by $\phi^x(\gamma):=x^{-1}\phi(\gamma)x$ for all $\gamma\in \pi_1(\mathbb{P}^1_{\bar{\mathbb{F}}_q}\backslash B,b)$. One defines the cocycles $\tilde{\zeta_e}\in H^1(\mathbb{F}_q, \tilde{T}')$ by $Frob_q\mapsto \tilde{\tau}_1^e$ for all $e$ with $1\leq e \leq d'$. Here is the second lemma we need to prove Proposition \[prop\_main\]: \[lemma\_main\] For every $e$ with $1\leq e\leq d'$ and $(e,d')=1$ there exists a finite set of $\mathbb{F}_q$-rational of type $e$ Nielsen tuples $\bar{D}_1,...,\bar{D}_r$ in $\tilde{G}$ such that for any given $\mathbb{F}_q$-rational of type $e$ Nielsen tuple $\bar{D}$, there exists $D_i$ which makes $\mathcal{H}_{\bar{D}+\bar{D}_i}(\bar{\mathbb{F}}_q)$ nonempty. We can replace $\tilde{\tau}$ with $\tilde{\tau}^e$ in the proof below so we may assume that $e=1$. Let $H\leq \tilde{G}$ be the subgroup generated by $g_1...g_k$ such that $\bar{D}^q=\bar{D}^{\tilde{\tau}}$ where $\bar{D}$ is the Nielsen tuple of $(g_1,...,g_k)$. We will show that for all $h\in H$, there exists a $k$-tuple $\bar{g}=(g_1,..,g_k)\in \tilde{G}^k$ such that -$g_1...g_k=h$ -$\tilde{G}=<g_1,..,g_k>$ -$\bar{g}$ represents a $\mathbb{F}_q$-rational Nielsen tuple $\bar{D}$ in $\tilde{G}$ with $\bar{D}^q=\bar{D}^{\tilde{\tau}}$. It suffices to show for $h=1$: If $\bar{g}=(g_1,...,g_k)$ is such a tuple with $g_1...g_k=1$ and if $\bar{h}=(h_1,...,h_k)$ is a $k$-tuple having product $h$ and representing a Nielsen tuple $\bar{D'}$ with $\bar{D'}^q=\bar{D'}^{\tilde{\tau}}$, then we can just concatenate $\bar{g}$ with $\bar{h}$ and thus, get such a tuple having multiple $h$. So, let $\bar{g}=(h_1,...,h_s)$ be a generating set for $\tilde{G}$ which represents a Nielsen tuple $\bar{D}_0$. Let $\bar{D}$ be the concatenation of the tuples in $\tilde{T}_1$-orbit of $\bar{D}_0$. Obviously, $\bar{D}^q=\bar{D}^{\tilde{\tau}}$. Now, if $(g_1,...,g_k)\in \bar{D}$ with $x=g_1...g_k$ then take $|x|$-multiple of $(g_1,...,g_k)$ where $|x|$ denotes the order of $x$. This completes the proof of the claim. Let $1\leq e\leq d'$ with $(e,d')=1$ and $t_e=q\tau^{-e}$. $t_e$ acts on the set of conjugacy classes of $G$ and $t_e^r$ acts trivially for some positive integer $r$ since $(q,|G|)=1$. Therefore, one can put a group structure on the set $T_e:=\{t_e^i| i=1,...,r\}$. Given a conjugacy class $\mathcal{O}$ in $G$, we will denote the $T_e$-orbit of $\mathcal{O}$ by $\mathcal{C}_e(\mathcal{O})$ and define $$\bar{\mathcal{O}}_e:=\sum_{\bar{C}\in \mathcal{C}_e(\mathcal{O})}\bar{C}.$$ Notice that, given a Nielsen tuple $\bar{C}$, the condition $\bar{C}^q=\bar{C}^{\tau^e}$ is equivalent to the condition $\bar{C}^{q\tau^{-e}}=\bar{C}$ so $\bar{\mathcal{O}}_{e}$ is the “smallest” $\mathbb{F}_q$-rational Nielsen tuple of type $e$ containing $\mathcal{O}$. Thus, given a Nielsen tuple $\bar{C}$ in $G$ with $\bar{C}^q=\bar{C}^{\tau^e}$, we can write $$\bar{C}=\sum_{\bar{\mathcal{O}}_e}a_{\mathcal{O}}\bar{\mathcal{O}}_e.$$ Likewise, we can consider the set $\tilde{T}_e=\{\tilde{t}^i_e|i=1,...,s\}$ where $\tilde{t}_e=q\tilde{\tau}^{-e}$ and the Nielsen tuple $\bar{\tilde{\mathcal{O}}}_e$ where $\tilde{\mathcal{O}}$ is a conjugacy class of $\tilde{G}$ which projects down to $\mathcal{O}$. We follow [@el2] closely to prove Proposition \[prop\_main\]:  \ [*Proof*]{} of Proposition \[prop\_main\]. We want to prove the inequality on the left-hand-side. Let $e$ be such that $1\leq e\leq d'$ and $(e,d')=1$. Let $\tilde{N}$, $\tilde{G}$ and $\tilde{G}'$ be as in Lemma \[lemma\_diag\] and let $\bar{D}_1,...,\bar{D}_r$ be the Nielsen tuples in $\tilde{G}$ as in Lemma \[lemma\_main\]. For each conjugacy class $\mathcal{O}$ in $G$ fix a conjugacy class $\tilde{\mathcal{O}}$ in $\tilde{G}$ with $\pi(\tilde{\mathcal{O}})=\mathcal{O}$. Note that $\pi(\bar{\tilde{\mathcal{O}}}_e)=b_{\mathcal{O}}\bar{\mathcal{O}}_e$ for some $b_{\mathcal{O}}$. Given a Nielsen tuple $\bar{C}$ in $G$ with $\bar{C}^q=\bar{C}^{\tau^e}$, we can write $\bar{C}=\sum_{\bar{\mathcal{O}}_e}a_{\mathcal{O}}\bar{\mathcal{O}}_e$ and the following Nielsen tuple in $\tilde{G}$ $$\bar{D}:=\sum_{\bar{\mathcal{O}}_e} \lceil \frac{a_{\mathcal{O}}}{b_{\mathcal{O}}} \rceil \bar{\tilde{\mathcal{O}}}_e$$ is $\mathbb{F}_q$-rational of type $e$ and $\pi(\bar{D})=\bar{C}+\bar{C}'$ where $\bar{C}'$ can be drawn from a finite set of Nielsen tuples $\bar{C}_1',...,\bar{C}_s'$ with $\bar{C}_i'^q=\bar{C}_i'^{\tau^e}$ for all $i$. Now, fix a Nielsen tuple $\bar{B}$ in $\tilde{G}$ with $\bar{B}^q=\bar{B}^{\tilde{\tau}^e}$ such that its projection $\pi'(\bar{B})$ in $\tilde{G}'$ contains at least $K$ nontrivial conjugacy classes of $\tilde{G}$ where $K$ is the constant in Lemma \[lemma\_cp\]. By Lemma \[lemma\_main\], there exists an $i$ such that $\bar{D}+ \bar{B}+\bar{D}_i$ is $\mathbb{F}_{q}$-rational of type $e$ and $\mathcal{H}^{\tilde{\zeta}_e}_{\bar{D}+ \bar{B}+\bar{D}_i}(\bar{\mathbb{F}}_q)$ is nonempty and, by Theorem \[thm\_decomp\], $\mathcal{H}^{\tilde{\zeta}_e}_{\bar{D}+ \bar{B}+\bar{D}_i}$ defined over $\mathbb{F}_q$. The projection $\pi(\bar{D}+ \bar{B}+\bar{D}_i)$ is $\mathbb{F}_q$-rational of type $e$ and it can be expressed as $\bar{C}+\bar{C}_j+n.\bar{1}$ where $\bar{C}_j$ can be drawn from a finite set of Nielsen tuples $\bar{C}_1,...,\bar{C}_k$ and $\bar{1}$ denotes the trivial conjugacy class. We now claim that $\mathcal{H}^{\zeta_e}_{\bar{C}+\bar{C}_j}$ has an $\mathbb{F}_q$-rational geometrically connected component. Once again note that $\mathcal{H}_{G,G}\cong \mathcal{H}_G/\mathbb{F}_q$ where $\mathcal{H}_G$ is the coarse moduli scheme of geometrically connected Galois $G$-covers. So, for any Galois $\tilde{G}$-cover $Y\rightarrow \mathbb{P}^1$ we have the canonically associated Galois $G$-cover $Y/U\rightarrow \mathbb{P}^1$ where $U:=\text{ker}(\pi:\tilde{G}\twoheadrightarrow G)$. This defines a morphism of schemes$/\mathbb{F}_q$ $$\pi_*:\mathcal{H}^{\tilde{\zeta}_e}_{\bar{D}+ \bar{B}+\bar{D}_i} \rightarrow \mathcal{H}^{\zeta_e}_{\bar{C}+\bar{C}_j}.$$ Notice that this map factors through (possibly over $\bar{\mathbb{F}}_q$) the natural map, which is induced likewise, $$\pi'_*:\mathcal{H}^{\tilde{\zeta}_e}_{\bar{D}+ \bar{B}+\bar{D}_i} \rightarrow \mathcal{H}_{\pi'(\bar{D}+ \bar{B}+\bar{D}_i)}.$$ Since $\pi'(\bar{B})$ contains at least $K$ nontrivial conjugacy classes of $\tilde{G}'$, by Lemma \[lemma\_cp\], the image of $\pi'_*$ is geometrically connected and so is the image of $\pi_*$. The image of $\pi_*$ is the $\mathbb{F}_{q}$-rational geometrically connected component we want. Define $h_2(q,\bar{C})$ to be the number of $\mathbb{F}_q$-rational geometrically connected components of $\mathcal{H}^{\zeta_e}_{\bar{C}}$ multiplied by $q^{|\bar{C}|}$. By the discussion above, $h_2(q,\bar{C}+\bar{C}_j)\geq q^{|\bar{C}+\bar{C}_j|}$ for some $j$. For each $\mathbb{F}_q$-rational of type $e$ Nielsen tuple $\bar{C}$, fix such a $\bar{C}_j$ and set $\bar{C}^{pr}:=\bar{C}+\bar{C}_j$. Thus, we have $$\sum_{\substack{\bar{C}\\ r(\bar{C})<R}} h_2(q,\bar{C}^{pr})\geq \sum_{\substack{\bar{C}\\ r(\bar{C})<R}} q^{|\bar{C}^{pr}|} > \sum_{r<R} h_3(q,r,e)$$ and $$\sum_{\substack{\bar{C}\\r(\bar{C})<R}} h_2(q,\bar{C}^{pr}) \leq \sum_{\substack{\bar{C}\\r(\bar{C})<R+m}} h_2(q,\bar{C})= \sum_{r<R+m} h_2(q,r,e)$$ where $m$ is the supremum of $r(\bar{C}_j)$’s and the sums run over $\mathbb{F}_q$-rational of type $e$ Nielsen tuples $\bar{C}$. This completes the proof. Malle’s Conjecture ================== Let $\mathcal{Z}'_{G,N}(\mathbb{F}_q,X)=\sum_{q^r<X}h_2(q,r)$. Observe that heuristic $(A)$ implies $h_1(q,r)=h_2(q,r)$ (this indeed is the only point where we use the heuristic). So, by lemma \[lemma\_h1\], $\mathcal{Z}_{G,N}(\mathbb{F}_q,X)$ has the same asymptotic order with the counting function $\mathcal{Z}'_{G,N}(\mathbb{F}_q,X)$ on heuristic grounds. More precisely, we have \[lemma\_change\] Assume heuristic $(A)$. If $G$ splits in $N$, then we have $$\mathcal{Z}_{G,N}(\mathbb{F}_q,X)\asymp \mathcal{Z}'_{G,N}(\mathbb{F}_q,X).$$ In this section, our main purpose is to compute asymptotic order of the counting function $\mathcal{Z}'_{G,N}(\mathbb{F}_q,X)$. In detail: we will define the constant $b(G,N,\mathbb{F}_q)$ and prove Theorem \[main\_thm\]. We will write a conjecture for the counting function $\mathcal{Z}_N(k,X)$ for any global field $k$ and, using Theorem \[main\_thm\], we will show that our conjecture gives the right asymptotic in some important cases. We will need the following lemma from Tauberian theory, for a proof see [@el2 Lemma 2.3]: \[lemma\_an\] Suppose $\{a_n\}$ is a sequence of real numbers with $a_n=0$ whenever $n$ is not a power of $q$, and suppose $$\sum_{r=1}^\infty a_{q^r} q^{-rs}$$ considered as a formal power series, is a rational function $f(t)$ of $t=q^s$. Let a be a positive real number. If $f(t)$ has no poles with $|t|\geq q^a$, then $$\sum_{n=1}^{X}a_n\ll X^a.$$ If $f(t)$ has a pole of order $b$ at $t=q^a$ and no other poles with $|t|\geq q^a$, then $$\sum_{n=1}^X a_n\asymp X^a(log X)^{b-1}.$$ Let $1\leq e\leq d'$ with $(e,d')=1$ and $t_e=q\tau^{-e}$. Note that $t_e$ acts on the set of conjugacy classes of $G$ and that $t_e^r$ acts trivially for some positive integer $r$. Set $T_e:=\{t_e^i| i=1,...,r\}$. Given a conjugacy class $\mathcal{O}$ in $G$, we denote the $T_e$-orbit of $\mathcal{O}$ by $\mathcal{C}_e(\mathcal{O})$ and define $\bar{\mathcal{O}}_e:=\sum_{\bar{C}\in \mathcal{C}_e(\mathcal{O})}\bar{C}$. Let $d'=[N:GCen_N(G)]$ and $d''=[GCen_N(G):G]$ *i.e.* $d=d'd''$. Denote the set of $G-$conjugacy classes of minimal-index elements of $G$ by $\mathcal{C}(G)$. Given $e$ with $1\leq e\leq d'$ and $(e,d')=1$, define $$\mathcal{C}_e(G,N,\mathbb{F}_q):=\{ \bar{\mathcal{O}}_e: \mathcal{O}\in \mathcal{C}(G)\} \text{\indent and \indent} b_e(G,N,\mathbb{F}_q):=|\mathcal{C}_e(G,N,\mathbb{F}_q)|.$$ Finally, set $$b(G,N,\mathbb{F}_q)=max\{b_e(G,N,\mathbb{F}_q)|1\leq e\leq d'\text{ and } (e,d')=1\}.$$ By Lemma \[lemma\_change\], the following theorem (respectively, the related corollaries) is just another way to state Theorem \[main\_thm\] (respectively, the corollaries) in the introduction: \[thm\_main\] If $G$ splits in $N$, then we have $$\mathcal{Z}'_{G,N}(\mathbb{F}_q,X)\asymp X^{a(G)} (logX)^{b(G,N,\mathbb{F}_q)-1}.$$ By Proposition \[prop\_main\], we have $$\sum_{q^r<X} h_2(q,r)\asymp \sum_{q^r<X} h_3(q,r)=\sum_{\substack{1\leq e\leq d' \\ (e,d')=1}}\sum_{q^r<X} h_3(q,r,e).$$ On the other hand, for every $e$ with $1\leq e\leq d'$ and $(e,d')=1$, we have the factorization: $$\sum_{r=1}^{\infty} h_3(q,r,e)q^{-rs}=\sum_{\substack{\bar{C} \\ \bar{C}^q=\bar{C}^{\tau^e}}} q^{|\bar{C}|q^{-r(\bar{C})s}} = \prod_{\bar{\mathcal{O}}_e} \frac{1}{1-q^{|\bar{\mathcal{O}}_e|(1-ind(\bar{\mathcal{O}}_e)s)}}$$ where the product is indexed by $T_e$-orbits $\bar{\mathcal{O}}_e$ of conjugacy classes $\mathcal{O}$ of $G$, $ind(\bar{\mathcal{O}}_e)$ is the index of an element of $\mathcal{O}$ and $|\bar{\mathcal{O}}_e|$ denotes the number of conjugacy classes in the orbit $\bar{\mathcal{O}}_e$. By Lemma \[lemma\_an\], $$\sum_{q^r<X} h_3(q,r,e) \asymp X^{a(G)} (logX)^{b_e(G,N,\mathbb{F}_q)-1}$$ and so $$\sum_{\substack{1\leq e\leq d' \\ (e,d')=1}}\sum_{q^r<X} h_3(q,r,e) \asymp X^{a(G)} (logX)^{b(G,N,\mathbb{F}_q)-1}$$ for sufficiently large $X$. Putting them all together, we get $$\sum_{q^r<X} h_2(q,r)\asymp X^{a(G)} (logX)^{b(G,N,\mathbb{F}_q)-1}$$ and this completes the proof. As a special case, we get the result of Ellenberg-Venkatesh [@el2]: \[cor\_ell\] $\mathcal{Z}'_{N,N}(\mathbb{F}_q,X)\asymp X^{a(N)}(logX)^{b(N,\mathbb{F}_q)-1}.$ We just need to show that $b(N,N,\mathbb{F}_q)=b(N,\mathbb{F}_q)$. Using the notation above, we have $d=e=1$ and $\tau=1$ in $N$. So, $\mathcal{C}_e(N,N,\mathbb{F}_q)=\mathcal{C}(N)/G_{\mathbb{F}_q}$ and $b(N,N,\mathbb{F}_q)=b_e(G,N,\mathbb{F}_q)=|\mathcal{C}(N)/G_{\mathbb{F}_q}|=b(N,\mathbb{F}_q)$. Let $N=(\langle(123)\rangle\oplus \langle(456)\rangle)\rtimes \langle(14)(25)(36)\rangle\leq S_6$ and let $\mathcal{Z}_N(\mathbb{Q},\mathbb{Q}(\zeta_3), X)$ be the number of isomorphism classes of $N$-extensions $K/\mathbb{Q}$ containing $\mathbb{Q}(\zeta_3)$ such that $d_{K/\mathbb{Q}}<X$. In [@kl2], Klüners shows that $\mathcal{Z}_N(\mathbb{Q}, X)\asymp X^{1/2}logX$ contradicting with Malle’s conjecture which predicts $\mathcal{Z}_N(\mathbb{Q},X)\asymp X^{1/2}$. Indeed, he proves that $\mathcal{Z}_N(\mathbb{Q},\mathbb{Q}(\zeta_3), X)\asymp X^{1/2}logX$. With evident modifications, one gets the same result in function field case for $q=2$ $(mod3)$ *i.e.* $\mathcal{Z}_N(\mathbb{F}_q,\mathbb{F}_{q^2},X)\asymp X^{1/2}logX$. The following corollary shows that our theorem gives the right asymptotic for the number of $N$-extensions in this case. \[cor\_coun\] $\mathcal{Z}'_N(\mathbb{F}_q,X)\asymp X^{1/2}logX$ where $N=(\langle(123)\rangle\oplus \langle(456)\rangle)\rtimes \langle(14)(25)(36)\rangle$ and $q=2$ $(mod3)$. Given an $N$-extension $K/\mathbb{F}_q(t)$, maximal constant intermediate subfield in its Galois closure $\widehat{K/\mathbb{F}_q(t)}$ might be $\mathbb{F}_q(t)$, $\mathbb{F}_{q^2}(t)$ or $\mathbb{F}_{q^6}(t)$ corresponding to the normal subgroups of $N$, respectively, $N$, $G_1:=\langle(123)\rangle\oplus \langle(456)\rangle$ or $G_2:=\langle(123)(456)\rangle$. Therefore, we have $$\mathcal{Z}'_N(\mathbb{F}_q(t),X)\asymp \mathcal{Z}'_{N,N}(\mathbb{F}_q,X)+\mathcal{Z}'_{G_1,N}(\mathbb{F}_q,X)+\mathcal{Z}'_{G_2,N}(\mathbb{F}_q,X).$$ Using Theorem \[thm\_main\], one can easily see that $$\mathcal{Z}'_{N,N}(\mathbb{F}_q,X)\asymp X^{1/2} \text{\indent and \indent} \mathcal{Z}'_{G_2,N}(\mathbb{F}_q,X)\asymp X^{1/4}.$$We will show that $\mathcal{Z}'_{G_1,N}(\mathbb{F}_q,X)\asymp X^{1/2}logX$. Using the notation above, we have: $$\mathcal{C}(G)=\{(123),(132),(456),(465)\},$$ $a(G)=1/2$, $d=d'=2$, $e=1$ and $\tau=(14)(25)(36)$. We have two $T_e$-orbits: $\{(123),(465)\}$ and $\{(132),(456)\}$. Therefore, $b(G_1,N,\mathbb{F}_q)=b_e(G_1,N,\mathbb{F}_q)=2$. Hence, we are done. Another interesting example in $\cite{kl2}$ is $N\cong (C_3\wr C_3)\times C_2\subseteq S_{18}$ where $C_3$ denotes the cyclic group of order $3$. Klüners shows that $\mathcal{Z}_N(\mathbb{Q},X)<<X^{1/4}$. In function field case, our main result gives us the exact asymptotic: \[cor\_exam\] $\mathcal{Z}'_N(\mathbb{F}_q(t),X)\asymp X^{1/4}$ where $q=2$ $(mod3)$ and $N\cong (C_3\wr C_3)\times C_2\subseteq S_{18}$. Let’s write $N=C_3\wr C_3 \times C_2= (\langle g_1\rangle \times \langle g_2\rangle \times \langle g_3\rangle)\rtimes \langle x\rangle \times \langle y\rangle$. We will just consider the normal subgroups $G$ with $a(G)=a(N)$. Given a $N$-extension $K/\mathbb{F}_q(t)$, $N$ has four such normal subgroups $G$ with cyclic quotient corresponding to possible maximal constant subextensions in the Galois closure of $K/\mathbb{F}_q(t)$. So, we have four cases. [*Case $1$*]{}: $G=N$. In this case, $\mathcal{C}(G)=\{ \tilde{g}_1,\tilde{g}_1^2 \}$ where $\tilde{g}_i$ is the conjugacy class of $g_i$. So, $a(N)=1/4$. Since $q=2$ $(mod3)$ and $d=d'=e=1$, there is only one $T_e$-orbit, namely $\{ \tilde{g}_1,\tilde{g}_1^2 \}$. Thus, $b(N,N,\mathbb{F}_q)=1$ and $\mathcal{Z}'_{G,N}(\mathbb{F}_q,X)\asymp X^{1/4}.$ [*Case $2$*]{}: $G=C_3\wr C_3$. We have $[N:G]=2$, $a(G)=1/4$ and $d'=e=1$ since $y\in Cen_N(G)$. With the notation above, we have $\tau=y$ and $\mathcal{C}(G)=\{ \tilde{g}_1,\tilde{g}_1^2 \}$. Since $\tau\in Cen_N{G}$, the only $T_e$-orbit is $\{ \tilde{g}_1,\tilde{g}_1^2 \}$ and so $b(G,N,\mathbb{F}_q)=1$. Hence, $\mathcal{Z}'_{G,N}(\mathbb{F}_q,X)\asymp X^{1/4}$ [*Case $3$*]{}: $G=C_3\times C_3\times C_3$. We have $[N:G]=6$, $a(G)=1/4$, $d'=3$ and $\tau=xy$. We have $\mathcal{C}(G)=\{ g_1,g_1^2,g_2,g_2^2,g_3,g_3^2 \}$ and $e=1$ or $e=2$. One can easily see $$\mathcal{C}_1(g_i^{\epsilon})=\mathcal{C}_2(g^{\epsilon}_i)=\{g_1, g_1^2, g_2, g_2^2, g_3, g_3^2\}$$ for $i=1,2,3$ and $\epsilon=1,2.$ Therefore, there is only one $T_e$-orbit and $b(G,N,\mathbb{F}_q)=1$. Hence, $\mathcal{Z}'_{G,N}(\mathbb{F}_q,X)\asymp X^{1/4}.$ [*Case $4$*]{}: $G=C_3\times C_3\times C_3 \times C_2$. Using the same argument in the previous case, we see that $\mathcal{Z}'_{G,N}(\mathbb{F}_q,X)\asymp X^{1/4}$. QED The following corollary shows that our result coincides with Malle’s conjecture for abelian groups, which is a result of Wright [@wr] in the case the base field is $\mathbb{Q}$. By Corollary \[cor\_ell\] and Lemma \[lemma\_change\], we just need to show that extensions without a constant subextension is “more” than the extensions with a constant subextension (whose corresponding subgroup $G$ splits in $N$). \[cor\_abel\] Let $N$ be an abelian group. Then, $\mathcal{Z}'_{G,N}(\mathbb{F}_q,X)\ll \mathcal{Z}'_{N,N}(\mathbb{F}_q,X)$ for any normal subgroup $G$ with a cyclic complement. Let $G$ be such a subgroup. Then, $a(G)\leq a(N)$. If $a(G) < a(N)$, then we are done. Assume $a(G)=a(N)$. Since $N$ is abelian, $d'=e=1$ and $\mathcal{C}_1(G,N,\mathbb{F}_q)=\mathcal{C}(G)/G_{\mathbb{F}_q}$. On the other hand, $\mathcal{C}(N,N,\mathbb{F}_q)=\mathcal{C}(N)/G_{\mathbb{F}_q}$ and $\mathcal{C}(G)\subseteq \mathcal{C}(N)$. So, $b(G,N,\mathbb{F}_q)\leq b(N,N,\mathbb{F}_q)$ and we are done. Note that one can revise Malle’s conjecture (in other words, the constant $b(N,\mathbb{F}_q)$ in the conjecture) for function fields by just taking the maximum of all the constants $b(G,N,\mathbb{F}_q)$ over all normal subgroups $G\leq N$ with cyclic quotient and $a(G)=a(N)$. Notice that the constant $b(G,N,\mathbb{F}_q)$ is defined for any such subgroup $G$ (which is not necessarily split in $N$). Now, we want to conclude our paper with the statement of the revised conjecture in number field case. First, we will introduce some notation. Let $k$ be a number field. We need to consider normal subgroups $G\leq N$ with abelian quotient $N/G$. Let $k^c$ be the maximal cyclotomic extension of $k$ (in a fixed algebraic closure $\bar{k}$). Set $C:=Cen_N(G)$. Given a cocycle $\varphi \in {\rm Hom}(G(k^c/k),N/G)$ and a conjugacy class $\bar{g}$ in $G$, define $\varphi$[*-twisted action*]{} of $\sigma\in G_k$ on $\bar{g}$ by $$\sigma(\bar{g}):= \bar{g}^{\chi(\sigma)\varphi({\rm Res}(\sigma))^{-1}}$$ where ${\rm Res}:G_k\rightarrow G(k^c/k)$ is the restriction, $\chi$ is the cyclotomic character and $\varphi({\rm Res}(\sigma))^{-1}$ acts on the conjugacy class by conjugation. By analogy, we define $$b_{\varphi}(G,N,k):=|\mathcal{C}(G)/G_k|$$ where the $G_k$-action in question is the $\varphi$-twisted action. Finally, set $$b(G,N,k)=max\{b_{\varphi}(G,N,k): \varphi\in {\rm Hom}(G(k^c/k),N/G) \text{ and } \varphi \text{ is surjective} \}.$$ Based on our result, we propose the following correction to Malle’s conjecture for any finite transitive subgroup $N\subseteq S_n$:  \ Fixing $N$ and $k$, set $$b(N,k)=max\{ b(G,N,k): N/G \text{ is abelian, and } a(G)=a(N)\}.$$ Then, we have $$\mathcal{Z}_N(k,X)\asymp X^{a(N)} (logX)^{b(N,k)-1}.$$  \ We remark that there are counterexamples in [@kl2] which are different than the one in Corollary \[cor\_coun\], but in the same spirit. Using Theorem \[thm\_main\], one can show that our result gives the right asymptotic in these cases. Summing it up, in the case of function fields, corollaries of our result shows that the conjecture above coincides with the existing results on Malle’s conjecture and eliminates all the counterexamples known (to our knowledge) so far. Hence, it provides strong evidence in the favor of the conjecture above.  \ [100]{} M. Bhargava, The density of discriminants of quartic rings and fields, *Ann. of Math.*, 162 (2005), 1031-1063. M. Bhargava, The density of discriminants of quintic rings and fields, *Ann. of Math.*, to appear. A. Clebsch. Zur Theorie der Riemann’schen Flächen. Math. Ann. 6:216-230, 1872. Harold Davenport and Hans Arnold Heilbronn. On the density of discriminants of cubic fields. II. *Proc. Roy. Soc. London Ser. A*, 322(1551):405-420, 1971. P. Debes and J.C. Douai. Algebraic Covers: field of moduli versus field of definition. *Ann. Sci. Ecole Norm. Sup. (4)*, 30 (1997), no. 3, p 303-338. Jordan Ellenberg and Akshay Venkatesh. The number of extensions of a number field with fixed degree and bounded discriminant, *Ann. of Math.* 163 (2), 723–741 (2006) . Jordan Ellenberg and Akshay Venkatesh. Counting extensions of function fields with bounded discriminant and specified Galois group. *In Geometric Methods in Algebra and Number Theory*, volume 235 in *Progress in Mathematics*, pages 151-168. Birkhauser, 2005. M. D. Fried and H. Volklein. The inverse Galois problem and rational points on moduli spaces. *Math. Ann.*, Volume: 290(1991), no. 4, p. 2021-2027. A. Hurwitz. Uber Riemann’sche Flächen mit gegebenen Verzweigungspunkten. Math. Ann. 39:1-61, 1891. Jürgen Klüners and Gunter Malle. Counting nilpotent Galois extensions. *J. Reine. Angew. Math*., 572:1-26, 2004. Jürgen Klüners. A Counter Example to Malle’s Conjecture on the Asymptotics of Discriminants, *C. R. Acad. Sci. Paris, Ser. I* 340 (2005). Gunter Malle. On the distribution of Galois groups. *J. Number Theory*, Volume: 92 (2002), no. 2, p. 315-329. Gunter Malle. On the distribution of Galois groups II. *Exp. Math*., 13:129-135, 2004. S. Wewers. Construction of Hurwitz spaces. PhD Thesis, University Essen, 1998. David Wright. Distribution of discriminants of abelian extensions. *Proc. London Math. Soc.*, 58:17-50, 1989.  \ <span style="font-variant:small-caps;">department of mathematics, university of wisconsin, 480 Lincoln Dr Madison wi 53706</span> *E-mail address:* turkelli@math.wisc.edu

ArXiv

--- abstract: 'We report the degree of order of twenty-two Jackson Pollock’s paintings using *Hausdorff-Besicovitch fractal dimension*. Through the maximum value of each multi-fractal spectrum, the artworks are classify by the year in which they were painted. It has been reported that Pollock’s paintings are fractal and it increased on his latest works. However our results show that fractal dimension of the paintings are on a range of fractal dimension with values close to two. We identify this behavior as a fractal-order transition. Based on the study of disorder-order transition in physical systems, we interpreted the fractal-order transition through its dark paint strokes in Pollocks’ paintings, as structured lines following a power law measured by fractal dimension. We obtain self-similarity in some specific Pollock’s paintings, that reveal an important dependence on the scale of observation. We also characterize by its fractal spectrum, the called *Teri’s Find*. We obtained similar spectrums between *Teri’s Find* and *Number 5* from Pollock, suggesting that fractal dimension cannot be completely rejected as a quantitative parameter to authenticate this kind of artworks.' address: - 'Instituto de Física, Universidade Federal do Rio Grande do Sul, Caixa Postal 15051, 91501-970, Porto Alegre, RS, Brazil.' - 'Department of Applied Physics, CINVESTAV-IPN, Carr. Antigua a Progreso km.6, Cordemex, C.P.97310, Mérida, Yucatán, México.' - 'Department of Informatics, Universidad Politécnica de Puebla, $72640$, México.' author: - 'E.M. de la Calleja\*' - 'F. Cervantes' - 'J. de la Calleja' title: 'Order-Fractal transition in abstract paintings' --- Multi-fractal spectrum ,Order-fractal transition ,Abstract art Introduction ============ Fractality is present in many objects in nature, in structures generated by mathematical algorithms, in spacial interaction among populations, on distributions of particles in amorphous solids or in particles configurations created by computer simulations [@Barnsley; @Benoit; @Vicsek; @Meakin; @Spehar1]. It means that, we can measure fractal characteristics in a wide variety of two-dimension digital images. From the point of view of the analysis of physical systems, it is possible to identify fractal characteristics in Pollock’s paintings [@TayNat99; @Shamir] Since R. P. Taylor *et al.* presents its famous result, indicating that all the Jackson Pollock’s paintings are fractals [@TayNat99], many reports have been published confirming or questioning the fractal characteristic of abstract artworks [@Spehar; @TayMic03; @Coddington; @MureikaChaos; @Jones; @Jones1]. Based on fractal dimension, Taylor presented five criteria to describe the construction processes of the colored layers on the Pollock’s canvases [@TayGuz07]. He additionally concluded that it is possible to authenticate the paintings using fractal dimension [@TayNat99; @TayGuz07; @Tay10]. Taylor’s criteria have been tested by different methods, and there are many reports where conclude that it is not enough use fractal dimension to authenticate any drip painting [@Coddington; @MureikaChaos; @Jones; @Jones1; @Rockmoreauthen; @Mureikamulti]. However, the objections on these works have been addressed by others [@Irfan; @Hany], leaving the fractal dimension as a rigorous measure to characterize in a good way the Jackson Pollock’s paintings; used as an important parameter on the authentication process, demonstrating its useful to improve the analysis of complex abstract art [@Coddington; @Rockmoreauthen]. Pollock’s paintings were created by dripping, pouring, splashing or peeling layers of paint of different colors on canvas placed on the floor. Paint strokes above each others are painted, until the canvas is totally covered, or at least in large part of it. Pollock argued that he had control of the splattering or dripping over the paints, denying the accidental paintings [@Book-Pollock; @film]. This argument suggest that he was aware about the action in the tact, the movements to control the paint strokes, the flow of each paint, the velocity and the rhythm in his creative processes. For many researches it is surprising considering the apparent complexity of the strokes. The visual complexity of the linked paint strokes results by the method described above apparently poses a high degree of complexity, taking a definition that it is directly proportional to a high number of linked lines superimposed. For the case of Pollock’s paintings, we found that it is not completely true. If Pollock spills paint lines, adding one by one, in a particular way which only the artist knows, we can identify that the process perhaps is not completely in a random way. Certainly the creative process is unique, complex, and unrepeatable, specially in the case of Pollock. In this report we compare the physical complexity found in two-dimensional images from experimental phenomena and we used the same methodology to quantify the degree of order in physical structures. The paint strokes distribution on Pollock’s paintings has been compared to natural objects, however we related them to some fibrilar aggregation [@Jordens] and particle aggregation processes [@Calleja1; @Calleja3; @Cervantes; @Gonzales]. The distribution of paint strokes reflects some degree of disorder, but we assumed a high degree of order if the fractal dimension is close to two, According to previous results of treatment of physical images that describe order evolution in the framework of supercooled liquids [@Calleja1], granular materials [@Gonzales] and magnetic properties [@Carrillo2], we believe that the structure formed by paint strokes posses a high degree or order. Our results suggest an increasing order of the paint strokes distribution from a particular array of his paintings (See Table \[table-1\]), also our results indicate that the fractal property of Pollock’s paintings is in a specific range of values. It has been discussed that fractal dimension can not be used as a quantitative parameter to authenticate abstract art. However, we point out the case of *Teri’s Find*, which has been the famous case of discussion about the authentication process and has been evidenced the methodology by art historians [@Biro] to carry out the complex process of authenticating artworks. We tested the paint and interesting results between this painting and *Number 5* from Pollock, are discussed here. Multi-fractal spectrum ====================== Fractality is a geometrical, topological, structural and beauty property present in many natural, physics or simulated systems. This property can be recognized and calculated in many two-dimensional structures through digital images [@Barnsley; @Benoit; @Vicsek; @Gouyet]. In addition, there has been reported that fractal structures are result of kinetic aggregation or/and reaction processes [@Meakin; @Schmitt; @Ben-Jacob; @Naito]. The multi-fractal spectrum has been used as a measure of all the local fractal dimensions coexisting in spatial structures of a wide variety of physical, chemical and biological systems. This measure of complexity has been employed to characterize the structural transition of rheological fluids [@Calleja1; @Cervantes; @Calleja2], granular materials [@Gonzales; @Pusey], magnetic wall domains in boracite [@Carrillo2] and other complex systems [@Suzuki; @Gonzalez2]. The width of multi-fractal spectrum, calculated by the well-known box-counting method, is affected by the treatment applied on digital images [@Abry]. However there has been proposed different methods to analyze and treat complex digital images [@Amirshahi; @Kovalevsky; @Stork] to obtain a dependable measure. All these methods confirm the validity of many free computer programs to analyze, by fractal dimension, a wide variety of digital images from a extensive variety of systems [@Gouyet; @Redies; @Falconer; @Lopez], including the abstract artworks [@MureikaChaos; @Tay10; @Mureikamulti; @Abry; @Sarkar]. The multi-fractal spectrum generated by an infinite set of dimension measures the scaling structure as a function of the local pattern density. This give us information about the structural properties at different scales [@Halsey] and also describes the generalized dimensions [@Hentschel]. The standard scaling relation to relate the number of boxes to cover the set $N(\varepsilon)$ of size $\varepsilon$ is $$N(\varepsilon)\sim\varepsilon^{-D_{Q}}$$ where $\varepsilon$ acquired successively smaller values of length until the minimum value of $\varepsilon_{0}$. This defines the fractal dimension as $$D_{Q}=\lim_{\varepsilon \rightarrow 0}\frac{\ln N(\varepsilon)}{\ln (\varepsilon_{0}/\varepsilon)}$$ To calculate all the local fractal dimension [@Halsey; @Ott], we used the generalized box counting dimension [@Hentschel; @Feigenbaum; @Procaccia] defined as $$D_{Q}=\frac{1}{1-Q}\lim_{\varepsilon\rightarrow 0}\frac{ln I(Q,\varepsilon)}{ln (\varepsilon_{0}/\varepsilon)}$$ where $$I(Q,\varepsilon)=\sum_{i=1}^{N(\varepsilon)}[P_{(i,Q)}]^{Q}$$ We are taking into account the scaling exponent defined by Halsey et al. [@Halsey] as $P_{i,Q}^{Q}\sim\varepsilon_{i}^{\alpha Q}$ where $\alpha$ can take a width range of values measuring different regions of the set. When *Q=0* the generalized fractal dimension represents the classic fractal dimension [@TayMic03]. As the digital image in a gray scale is divided into pieces of size $\varepsilon$, it suggested that the number of times that $\alpha$ in $P_{i,q}$ takes a value between $\alpha'$ and $d\alpha'$ defined as $d\alpha'\rho(\alpha')\varepsilon^{-f(\alpha')}$ where $f(\alpha')$ is a continuous function. As *Q* represents different scaling indices, we can define $$I(Q,\varepsilon)=\sum_{i=1}^{N(\varepsilon)}[P_{(i,Q)}]^{Q}=\int d\alpha'\rho(\alpha')\varepsilon^{-f(\alpha')+Q\alpha'}$$ $\alpha_{i}$ is the Lipschitz-Hölder exponent, which characterizes the singularity strength in the *ith* box. The factor $\alpha_{i}$ quantifies the distribution of complexity in an spatial location. The multi-fractal spectrum is a set of overlapping self-similar configurations. In that way, we used the scaling relationship taking into account $f(\alpha)$ as a function to cover a length scales of observations. Defining the number of boxes as a function of the Lipschitz-Hölder exponent $N(\alpha)$, can be related to the box size $\varepsilon$ as $$N(\alpha)\sim\varepsilon^{-f(\alpha)}\label{eq1}$$ The multi-fractal spectrum shows a line of consecutive points for $Q\geq0$ that starts on the left side of the spectrum climbing up to the maximum value. The values for $Q\leq0$, are dotted on the right side of the spectrum descending until $Q=-10$, which is the minimum value for *Q*. The maximum value for the generalized dimension corresponds to $Q=0$, which correspond to the box counting dimension. To obtain the multi-fractal spectrum we use the plugin *FracLac* for ImageJ [@FracLac]. We select the case of $D_{f}=D_{Q=0}$ as the parameter to quantify the order in the digital images of Pollock’s paintings. In the plugin we select four grid positions that cover the total image. We apply a gray scale differential option to measure de fractal dimension, and the mode default sampling sizes was selected to plot the values of the spectrum. We define the smallest sampling element on $30$ px and $100\%$ as the maximum area on analysis of each image. In this report we present the left side of the spectrum, taking into account the correspondence with thermodynamic formalism [@Chhabra; @Chhabra-1] Pollock’s Multi-fractal Spectrum ================================ We selected twenty-two amazing Pollock’s paintings, painted during the called “dripping period”. This selection was done according to the distribution of darkness paint strokes on front, a reduce number of colored layers and its light background. All the paintings are classified into the movement called [*a*bstract expressionism]{}. In table (\[table-1\]) is presented the list of the selected Pollock’s paintings. -------- ----------- --------------------------------------------------- -------------- Number Year Painting Size (cm) 1 1946 Free Form [@Pollock1] 49x36 2 1947 Lucifer [@Pollock] 267.9x104.1 3 1947 Cathedral [@Pollock3] 89x181.6 4 1947 Enchanted Forest [@Pollock7] 221.3x114.6 5 1947 Reflection of the Big dipper [@Pollock] 111x91.5 6 1947 Undulating Paths [@Pollock] 114x86 7 1947-1950 Number 19 [@Pollock5] 78.4x57.4 8 1948 Summertime: Number 9A [@Pollocka] 84.8x555 9 1948 Number 26A; Black and White 1948 [@Pollock4] 208x121.7 10 1948 Number 23 [@Pollock2] 575x784 11 1948 Number 4(Gray and Red) [@Pollock3] 58x79 12 1948 Silver over Black,White,Yellow and Red [@Pollock] 61x80 13 1948 Number 14 Gray [@Pollock] 57x78.5 14 1949 Number 10 [@Pollock3] 46.04x272.41 15 1950 Autumn Rhythm; Number 30 [@Pollock] 525.8x266.7 16 1950 Number 32 [@Pollock] 457.5x269 17 1950 Number 29 [@Pollock3] 182.9x121.9 18 1950 Number 18 [@Pollock7] 56x56.7 19 1951 In Echo: Number 25 [@Undulating] 233.4x218.4 20 1912-1956 No. 15, 1950 [@Pollock6] 55.88x55.88 21 1951 Untitled,ink on Japanese paper [@Pollock9] 62.9x100.3 22 1951 Untitled [@Pollock8] 63.5x98.4 -------- ----------- --------------------------------------------------- -------------- : Pollock’s paintings\[table-1\] The multi-fractal spectrum is formed by consecutive points on the left side that grows up until the maximum value, represent all the localized fractal dimensions by different box sizes that means different scales of measure ($\varepsilon$). All the curves of the multi-fractal spectrum present similar width and length that grows from $\alpha=1.6$. Fig.(\[47-48\])*(a)* shows the left side of the multi-fractal spectrum for the selected Pollock’s artworks painted in $1947$. *(7)Number 19* reached the maximum value on the generalized fractal dimension, while *(2)Lucifer* reached the lowest one of this group. In Fig.(\[47-48\])*(b)* is presented the spectrums for the corresponding selected paintings from $1948$. In the group of paintings from $1948$, *(13)Number 14* obtains the highest value of $f(\alpha)$ among them, while *(9)Number 26A* gets the lowest. It is outstanding the case of *(6)Undulating Paths* where the number of local fractal dimensions is less than all others. Evidently the fractality of Pollock’s paintings is unquestionably, however, it is interesting the similarity between all the spectrums, taking into consideration that all the paintings are different, created in different moments and circumstances. ![In (a) is shown the left side of the multi-fractal spectrum of Pollock’s paintings selected from $1947$. In (b) are show the corresponding multi-fractal spectrum of Pollock’s canvas painted in $1948$.\[47-48\]](F-1a.eps "fig:"){width="45.00000%"} ![In (a) is shown the left side of the multi-fractal spectrum of Pollock’s paintings selected from $1947$. In (b) are show the corresponding multi-fractal spectrum of Pollock’s canvas painted in $1948$.\[47-48\]](F-1b.eps "fig:"){width="45.00000%"} ![The figure show the left side of the multi-fractal spectrum of Pollock’s paintings from $1950$ in (a), and from $1951$ in (b).\[50-51\]](F-2a.eps "fig:"){width="45.00000%"} ![The figure show the left side of the multi-fractal spectrum of Pollock’s paintings from $1950$ in (a), and from $1951$ in (b).\[50-51\]](F-2b.eps "fig:"){width="45.00000%"} Fig.(\[50-51\]) is shown the left side of the multi-fractal spectrum of the Pollock’s artworks painted in $1950$ and $1951$ respectively. In Fig.(\[50-51\])*(a)* a similar behavior of the spectrums between the paintings *(15)Autumn Rhythm*, *(16)Number 32* and *(18)Number 18* is observed. It is possible to recognize similar paint strokes distributed on these canvases, and clear differences with the painting *(17)Number 29*. The Pollock’s paintings created in these years correspond to a period where the technique developed by the artist, was completely dominated. However they exhibit a wide variety of type paint strokes, very different movements of the paint are visually recognizable and the strokes look so simple on some of them, for example in the cases of *(17)Number 29*, or very complex in other as *(15)Autumn Rhythm*. In Fig.(\[50-51\])*(b)* is presented the behavior of the left side of the spectrums of the selected Pollock’s artworks painted during $1951$. We can distinguish the maximum values of fractal dimension for each one. The painting called *(22)Untitled* reaches the higher value of $f(\alpha)$, while *(19)In Echho: number 25* obtains the minimum one. The shape of the spectrums with exception of *(22)Untitled*, are similar, obtaining approximately the same number of dimensions. It notice that the scale for the paintings from $1950$ begins in $\alpha>1.62$ and the spectrum from the paintings from $1951$ arise from $\alpha>1.5$. It represent differences on scale of observation between the groups of paintings where the complex behavior it shows on the spectrum. ![The figure shows the left side of the multi-fractal spectrum and a photograph of the Pollock’s artworks: *Free Form* painted in $1946$, *Summertime: Number 9A* painted in $1948$ and *Number 10* painted in $1949$.\[46-49\]](F-a.eps){width="98.00000%"} In Fig.(\[46-49\]) we show the left side of the multi-fractal spectrum of the paintings *(1)Free Form* painted in $1946$, *(8)Summertime: Number 9A* painted in $1948$ and *(14)Number 10* painted in $1949$. We can notice a notable difference between the spectrums obtained for these three paintings. The *(14)Number 10* spectrum grows for $\alpha>1.6$ while in the case of *Free Form* and *(8)Summertime: Number 9A* grows arise from $\alpha>1.5$. According to equation (\[eq1\]) the local fractal dimension is manifested for small values of $\alpha$ in comparison with the values for the other paintings presented in Fig.(\[47-48\]) and Fig.(\[50-51\]). The two large paintings: *(14)Number 10* and *(8)Summertime: Number 9A*, present similar number of local fractal dimensions and its spectrums grow from very close $\alpha$ exponent, reach very close maximum values. There is a notable difference of paint strokes between these paintings and *(1)Free Form* which has major density of paint strokes reflected in the magnitude of scale of $\alpha$. Order in Pollock’s paintings ============================ The paint strokes on Pollock’s canvases are apparently distributed randomly. However the maximum value of the multi-fractal spectrums presented above, can be associated with different degree of order [@Calleja1; @Cervantes; @Gonzales; @Carrillo2]. We obtained for an specific array of paintings listed in table (\[table-1\]), the evolution of the maximum value on the multi-fractal spectrum for each paint, corresponding to the Hausdorf-Besicovich dimension when $Q=0$. Based on the characterization by fractal dimension of liquid-solid transitions [@Calleja1; @Calleja3] we identify that a value of fractal dimension close to two, represents an ordered structure, on the contrary represent a disordered one. On the case of Pollock’s paintings, we measure the linked paint strokes and we obtain, according to the sequence of paintings listed on table (\[table-1\]) a fractal-order transition. ![The figure shows the evolution over the time of the fractal dimension of Pollock’s paintings (following the list on table (\[table-1\])). The paintings have different complexity within a range of values. The figure show the fractal-order transition for Pollock’s paintings grouped by year in which they were painted.\[FPA\]](F-6-2.eps){width="95.00000%"} Fig.(\[FPA\]) shows the evolution of the fractal dimension as a function of the selected Pollock’s paintings grouped by year. We can observe that fractal dimension increases, approaching to a dimension close to two, and it happens in all the groups of paintings. We can speculate that, if Jackson Pollock paints his artwork in the sequence on table (\[table-1\]), he added more drip lines that improve the homogenous distribution of them and it is reflecting on the value of the fractal dimension. For example, on paintings of $1947$, we observe that *(2)Lucifer* obtained the smallest fractal dimension than all and *(7)Number 19* obtain the highest. This is understood as a result of the distribution of darkness paint strokes that construct a disorder structure on the canvas and also a difference on the density of paint strokes on each painting. Visually can be see that *(2)Lucifer* look very similar to *(7)Number 19*, however there is high difference on $D_{f}$ of both paintings. For the paintings selected from $1948$, we observe an special case for *(8)Summertime: Number 9A* with respect to the others. This painting obtain the lowest fractal dimension of this group. We analyzed in detail this behavior and we found that the local fractal dimension depends on the size or section taken in the image. In Fig.(\[SP\]) is shown the fractal evolution of *(8)Summertime: Number 9A* and *(14)Number 10* for different fragments taken from the whole original image. We found that the paintings *(8)Summertime: Number 9A* and *(14)Number 10* reached a fractal dimension close to $1.6$ when is taken in to consideration the whole painting. This indicates according to our approach that they are the paintings with paint strokes structured with major disorder than all. We also found that, if we select a third part of the whole image and calculate the local fractal dimension using the same method, we obtained interest results reported in Fig.(\[SP\]). To do this, we divided the image into four fragments and we calculate its corresponding singularity spectrum for each part. Fig.(\[SP\]) shows that the Hausdorff-Besikovich fractal dimension increases in an inversely way as the size of the image. This could be interpreted as an evidence of self-similarity in Pollock’s paintings and the effect of the scale of observation. The same treatment was made on *(14)Number 10* and we obtain the same behavior that [(8)Summertime: Number 9A]{}. It is interesting that we did not find this behavior for all the others selected Pollock’s paintings, just on this two cases. ![The figure shows the evolution of fractal dimension ($D_{f}$) for four different fragments of *Summertime: Number 9A* and four different fragments of *Number 10*. There are presented the fragments of *Summertime: Number 9A*. As the section of the image decreases from (d) to (a), the local fractal dimension increases. The same behavior occurs on this Pollock’s paintings.\[SP\]](F-s.eps){width="95.00000%"} According to our analysis, the group of paintings from 1948: *(13)Number 14, Gray*, on Fig.(\[FPA\]) present the highest fractal dimension. This can be correlated to an homogeneous distribution of the paint strokes. Meanwhile, *(15)Autumn Rhythm* reaches the minimum $D_{f}$ of the $1950$ paintings group, i.e it was the painting with a disorder distribution of paint strokes. On the other hand *(18)Number 18* presents the highest degree of order with a fractal dimension close to two. For the $1951$ paintings group, *(22)Untitled* presents an homogeneous paint strokes distribution. This can be correlated to its higher local fractal dimension. The opposite case is represented by *(19)In Echo: Number 25*. This group of paintings present a dramatic evolution of fractality. This can be interpreted as the dripping technique was controlled and directed by Pollock in a more perfect way. These last four canvases were painted in the last years of Pollock’s life, when he improved and know perfectly his famous technique. The hypothesis that fractal dimension increases as a function of the year in which the paintings were painted, was tested in this report. We found effectively this kind of behavior. *Untitled* painted in $1951$ was the one with paint strokes distributed in a more ordered structure, according to our interpretation, while *(2)Lucifer* was the most fractal one, without considering the peculiar cases of *(8)Summertime: Number 9A* and *(14)Number 10*. Authentication ============== The fractal dimension has been questioned as an order parameter to authenticate abstract artworks [@MureikaChaos; @Jones; @Jones1]. However, it is unquestionable its use to characterize digital images from physical systems [@Barnsley; @Calleja1; @Cervantes; @Carrillo2]. Why it can not be used to characterized abstract art? We believe that Jackson Pollock developed a unique painting technique, and his creations are fantastic and unrepeatable. Evidently they present certain degree of complexity, however, it has been measured. Obviously the authentication processes require more than one parameter, however the fractal dimension seems to be adequate for his kind of artworks [@Shamir; @Irfan; @Hany]. ![The figure shows the multi-fractal spectrum of the called *Teri’s Find* and the Pollock *Number 5*. The spectrums are too similar between them.\[teri\]](F-t.eps){width="98.00000%"} We obtained the multi-fractal spectrum of the painting called *Teri’s Find* [@Teri; @Teri2]. Fig.(\[teri\])*(a)* is shown the complete multi-fractal spectrum of *Teri’s Find* not attributed to Pollock and *Number 5* recognized as painted by Jackson Pollock. We obtain that both spectrums look very similar. The unique difference is that the Pollock’s painting presents a longer multi-fractal spectrum than the*Ter’s Find*. However they agree on many of their values. Fig.(\[teri\])*(b)* shows details on the maximum values of both multi-fractal spectrums. *Teri’s Find* reaches $D_{f}=1.8477$ and *Number 5* obtains $D_{f}=1.8496$. Quantitatively also both spectrums are very close. Could be possible that *Teri’s Find* was painted by Pollock based on the curved of its multi-fractal dimension? or It is the fractal dimension insufficient to quantify abstract art? According to our results *Teri’s Find* presents a high degree of complexity in a similar way than Pollock’s paintings are, and taking into account the magnitude of its value, it is within the range of characteristic values found on Pollock’s artworks. Due to that, we support the idea that *Teri’s Find* was painted by Jackson Pollock. Conclusions =========== We report the multi-fractal behavior of twenty-two Pollock’s paintings which were selected by considering the apparent complexity of the paint strokes, the year in which they were painted, the density of darkness strokes on top and the light background. The left side of the spectrums is presented and was found that all the paintings have many local fractal dimensions. This suggest that the magnitude of fractal dimension depends on the scale of observation ($\varepsilon$). The length of all the left side of the spectrums grows from $\alpha=1.5$ to $\alpha=2$ with the notable exceptions of *(8) Summertime: Number 9A* and *(14) Number 10* which begins from $\alpha=1.5$. This result corroborates the self-similarity of the Pollock’s paintings. The visual complexity of the paintings was quantified by the maximum value of the multi-fractal spectrum as a function of the year in which the paintings were painted. We obtain a range of values between $D_{f}=1.78$ up to $D_{f}=1.88$. From the definition of order using in this report, the degree of order in Pollock’s paintings grows as a function of the year in which they were painted, taking into account the sequence of paintings listed in Table (\[table-1\]). This result could be an indication as a perfect knowledge of the dripping technique by the artist, a manifestation of control and perfection of his dripping technique and we can suggest that paint strokes were made consciously. The self-similarity was tested on *(8)Summertime:Number 9A* and *(14)Number 10*. The results reported in Fig.(\[SP\]) indicate that fractal dimension increases inversely as a function of the size of the digital image. Finally, *Teri’s Find* was tested by our method, and we found a similar multi-fractal spectrum between this not recognized painting, that could have been painted by Pollock and *Number 5*, painted by Pollock. We suggest that fractal dimension can not be definitely rejected as a parameter to authenticate abstract artworks. Despite, this statement has been tested by different methods. Many aspects have to be consider to analyze Pollock’s paintings, such as the right kind of paint and brush, the correct speed and movements not only of the hand even also the arm; all in exactly concordance with the knowledges and the individual projection of the artist. For all of that, the description of Pollock’s paintings needs to be made taking into account a combination of different scientific techniques, and fractal dimension can be a good quantitative parameter without invasive techniques. [99]{} M. F. Barnsley, , (Academic Press Profesional, United Sates of America, 1993). B. Mandelbrot, (W.H. Freeman and Company, New York, 1977). T. Vicsek (World Scientific, Singapore, 1989). P. Meakin, *The growth of fractal aggregates and their fractal measures, in Phase Transitions and Critical Phenomena* (edited by C. Domb and J.L. Lebowitz, Academic, London, 1988), pp. 336–489. 20. B. Spehar, C.W.G. Clifford, B.R. Newell, R.P. Taylor (2003) Universal aesthetic of fractals Computers $\&$ Graphics [**27**]{} 5, 813–820 doi:10.1016/S0097-8493(03)00154 R.P. Taylor, A.P. Micholich and D. Jonas, Nature [**399**]{}, 422 (1999). J. Alvarez-Ramirez, C. Ibarra-Valdez, and E. Rodriguez. Chaos, Solitons $\&$ Fractals, [**83**]{}, 97-104 (2016). R.P. Taylor, B. Spehar, P. Van Donkelaar and C.M. Hagerhall, Front. Hum. Neurosci. [**5**]{}60 (2011) doi: 10.3389/fnhum.2011.00060 R.P. Taylor, A.P. Micolich, and D. Jonas, Leonardo [**35**]{}, 203 (2003). J. Coddington J. Elton D. Rockmorec, and Y. Wan, Proc. SPIE 6810, Computer Image Analysis in the Study of Art, 68100F (29 February 2008). J.R. Mureika, Chaos [**15**]{}, 043702 (2005). K. Jones-Smith, H. Mathur, and L.M. Krauss, Phys. Rev. E [**79**]{}, 046111 (2009) 046111. K. Jones-Smith and H. Mathur, Revisiting Pollock’s drip paintings Arising from: R. P. Taylor, A. P. Micolich and D. Jonas, Nature [**399**]{}, 422 (1999) R.P. Taylor, R. Guzman, T.P. Martin, G.D.R. Hall, A.P. Micolich, D. Jonas, and C.A. Marlow, Pattern Recognition Letters [**28**]{}, 695-702 (2007) 695–702. R.P. Taylor, [*Chaos, Fractals, Nature: A New Look at Jackson Pollock*]{} (Fractals Research ISBN: 0-9791874-1-9, 2010). D. Rockmore, S. Lyu, and H. Farid, [*Proceedings of the National Academy of Sciences*]{}, 2004, 17006–17010 J.R. Mureika, C.C. Dyer, and G.C. Cupchik, Phys. Rev. E [**72**]{}, 046101 (2005). M. Irfan,and D.G. Stork, D. G.. In IS$\&$T/SPIE Electronic Imaging (pp. 72510Q-72510Q). International Society for Optics and Photonics.(2009, February) The arguments in K. Jones-Smith et al. “Fractal Analysis: Revisiting Pollock’s Paintings” Publication status: Nature, Brief Communication Arising, vol. 444, E9-10, (2006)\] were addressed in Taylor’s replies: Nature, vol. 444, E10-11 (2006) and arXiv:0712.1652v1 \[cond-mat.stat-mech\] 2007 (2009) P. Karmel, *Jackson Pollock: Interviews, Articles, and Reviews* (The Museum of Modern Art, New York, 1999). E. Harris, *Pollock y Pollock: la vida de un creador (España)*, Brant-Allen, Fred Berner Films, Pollock Films, Zeke Productions, United States of America (2000). S. Jordens, L. Isa, I. Usov and R. Mezzenga, Nature Communications [**4**]{} (2013). E. M. de la Calleja Mora, J. L. Carrillo, M. E. Mendoza and F. Donado, Eur. Phys. J. B. [**86**]{}, 1126 (2013). E.M. de la Calleja-Mora, Leandro B. Krott, M.C. Barbosa. Physica A: Statistical Mechanics and its Applications. [**449**]{} 18–26 (2016). F. Cervantes-Alvarez, J.J. Reyes-Salgado, V. Dossetti, and J.L. Carrillo, J. Phys. D: Appl. Phys. [**47**]{} 235303 (2014). J González-Gutiérrez, J L Carrillo-Estrada and J C Ruiz-Suárez, J. Stat. Mech. P12015 (2013). R. E. Moctezumaa, J. L. Carrillob, and M. E. Mendoza Rev. Mex. de Fıs. S [**58**]{} 48–53 (2012). P. P. Biró, [*Home page of the author*]{}, , 2013. J.F. Gouyet Physics and Fractal Structures Springer, New York (1996). C. G. Schmitt and A. J. Heymsfield, J. Atmos. Sci. [**67**]{}, 1605-1616 (2010). E. Ben-Jacob, O. Shochet, A. Tenenbaum, I. Cohen, A. Cziŕok, and T. Vicsek Nature [**368**]{}, 46 (1994). T. Naito, H. Yamamoto, K. Okuda, K. Konishi, H. Mayama, D. Yamaguchi, S. Koizumi, K. Kubo, and T. Nakamura, Eur. Phys. J. B [**86**]{}, 410 (2013). E.M. de la Calleja, J. L. Carrillo, F. Donado, Rev. Mex. de Fís. [**58**]{} 54-57 (2012). P. Pusey, (J.P. Hansen, D. Levesque, J. Zinn-Justin (Eds.), Elsevier Science, Amsterdam, 1991). M. Suzuki, Progr. of Theor. Phys. [**69**]{} 1 (1983). Jorge González-Gutiérrez, J.L. Carrillo-Estrada, and J. C. Ruiz-Suárez, Phys. Rev. E, 89, 052205 (2014). P. Abry, H. Wendt, S. Jaffard, Signal Processing, Elsevier, [**93**]{}, 554-572 (2013). S. A. Amirshahi, C. Redies, and J. Denzler, Computational Aesthetics (2013). V. A. Kovalevsky, Computer Vision Graphics and Image Processing [**46**]{}, 141-161 (1989). David G. Stork and Marco Duarte, IEEE Multimedia [**14**]{}, 14–18 (2007). C. Redies, S.A. Amirshahi, M. Koch, and J. Denzler, *Computer Vision–ECCV 2012. Workshops and Demonstrations* (Springer, 522–531). K. Falconer, (Techniques in Fractal Geometry, Wiley, New York, 1997). T. López, F..Rojas, R. Alexander-Katza, F. Galindo, A. Balankin, A. Buljanc, Journal of Solid State Chemistry [**177**]{}, 1873-1885 (2004). N. Sarkar and B. B. Chaudhur, Pattern Recognition, [**25**]{}, 1035-1041 (1992). T. C. Halsey, M. H. Jensen, L. P. Kadanoff, I. Procaccia, and N. I. Shraiman Phys. Rev. A [**33**]{} 1141 (1986). G. E. Hentschel, and I. Procaccia Physica [**8D**]{} 435 (1983). E. Ott, , (Cambridge University Press, United States of America, 1993). M.J. Feigenbaum, M.H. Jensen, I. Procaccia, Phys. Rev. Lett. [**57**]{}, 1503 (1986). H.G.E. Hentschel and I. Procaccia, Physica D [**8**]{}, 435 (1983) T. Ferreira and W. Rasband, [*ImageJ user guide*]{}, , 2013. A. Chhabra, R. V. Jensen Phys. Rev. lett. [**62**]{} 1327 (1989). A. B. Chhabra, C. Meneveau, R. V. Jensen, and K. R. Sreenivasan, Phys. Rev. A [**40**]{}, 5284 (1989) http://funmozar.com/famous-paintings-of-jackson-pollock/ http://www.jackson-pollock.org/index.jsp http://www.wikiart.org/en/jackson-pollock/by-period/drip-period http://www.guggenheim.org/new-york/collections/collection-online/artists/963 https://expresandoabstract.wordpress.com/2013/12/31/el-mercado-del-arte-en-2013-las-mayores-subastas/jackson-pollock-number-19/ http://www.tate.org.uk/art/artworks/pollock-summertime-number-9a-t03977 http://www.artsunlight.com/artist-NP/N-P0005-Jackson-Pollock/N-P0005-031-number-a-black-and-white.html https://gerryco23.wordpress.com/2015/09/01/jackson-pollock-at-tate-liverpool-wrestling-with-a-blind-spot/ http://www.moma.org/collection/works/79251?locale=en http://collections.lacma.org/node/229998 http://elogedelart.canalblog.com/archives/2009/10/02/15285905.html http://www.moma.org/collection/works/37765?locale=en Who the Is Jackson Pollock?, Directed by Harry Moses Documental Produced by Don Hewitt, Steven Hewitt, Michael Lynne. Starring: Teri Horton, Distributed by PicturehouseNovember 9, 2006 http://www.peterpaulbiro.com/Teri

ArXiv

--- author: - 'Nathan Berkovits$^{*}$, Max Guillen$^{*}$' title: 'Equations of motion from Cederwall’s pure spinor superspace actions' --- Introduction ============ Pure spinors $\lambda^\alpha$ in ten and eleven dimensions have been useful for constructing vertex operators and computing on-shell scattering amplitudes with manifest spacetime supersymmetry in super-Yang-Mills, supergravity and superstring theory [@HOWE1991141; @Berkovits:2001rb; @Berkovits:2002uc; @Gomez:2013sla]. After including non-minimal variables $(\bar\lambda_\alpha, r_\alpha)$, pure spinors have also been useful for constructing BRST-invariant off-shell actions for these maximally supersymmetric theories [@Cederwall:2010tn; @Cederwall:2013vba; @Cederwall:2011vy]. These BRST-invariant actions have a very simple form and were constructed by Cederwall using superfields $\Psi(x^m, \theta^\alpha, \lambda^\alpha, \bar\lambda_\alpha, r_\alpha)$ which transform covariantly under spacetime supersymmetry and depend on both the usual superspace variables $(x^m, \theta^\alpha)$ and the non-minimal pure spinor variables $(\lambda^\alpha, \bar\lambda_\alpha, r_\alpha)$. Although the actions require a non-supersymmetric regulator to define integration over the non-minimal pure spinor variables, it is easy to show that the supersymmetry transformation of the regulator is BRST-trivial so the action is spacetime supersymmetric. However, since the superfields $\Psi$ can depend in a non-trivial manner on the non-minimal variables, it is not obvious how to show that the solutions to the equations of motion correctly describe the usual on-shell $D=10$ and $D=11$ superfields which depend only on the $(x^m, \theta^\alpha)$ superspace variables. In this paper, an explicit procedure will be given for extracting the usual on-shell $D=10$ and $D=11$ superfields from the equations of motion of the pure spinor actions for the cases of $D=10$ supersymmetric Born-Infeld and for $D=11$ supergravity. This procedure will be given explicitly to first order in the coupling constant in these two actions, but it is expected that the procedure generalizes to all orders in the coupling constant as well as to other types of actions constructed from pure spinor superfields [^1]\[foot1\]. The procedure consists in using BRST cohomology arguments to define a unique decomposition of the on-shell pure spinor superfield $\Psi(x^m, \theta^\alpha, \lambda^\alpha, \bar\lambda_\alpha, r_\alpha)$ into the sum of two terms as $$\Psi(x^m, \theta^\alpha, \lambda^\alpha, \bar\lambda_\alpha, r_\alpha) = \tilde\Psi(x^m, \theta^\alpha, \lambda^\alpha) + \Lambda( x^m, \theta^\alpha, \lambda^\alpha, \bar\lambda_\alpha, r_\alpha)$$ where $\tilde\Psi(x^m, \theta^\alpha, \lambda^\alpha)$ is independent of the non-minimal variables and $\Lambda$ is constructed from the superfields in $\tilde\Psi$ and the non-minimal variables. Since $\tilde\Psi$ will have a fixed ghost number $g$ ($g$=1 for $D=10$ super-Born-Infeld and $g=3$ for $D=11$ supergravity), it can be expanded as $\tilde\Psi = \lambda^\alpha \tilde A_\alpha (x,\theta)$ or $\tilde\Psi = \lambda^\alpha\lambda^\beta\lambda^\gamma \tilde C_{\alpha\beta\gamma} (x,\theta)$, and it will be shown to first order in the coupling constant that $\tilde A_\alpha (x,\theta)$ and $\tilde C_{\alpha\beta\gamma}(x,\theta)$ correctly describe the on-shell spinor gauge superfield of $D=10$ super-Born-Infeld and the on-shell spinor 3-form superfield of $D=11$ supergravity. We expect it should be possible to generalize this procedure to all orders in the coupling constant and to other types of pure spinor actions, but there is an important issue concerning these pure spinor actions which needs to be further investigated. If the superfields $\Psi$ in these actions are allowed to have poles of arbitrary order in the non-minimal pure spinor variables, the cohomology arguments used to define the on-shell superfields become invalid. This follows from the well-known property of non-minimal pure spinor variables that one can construct a state $\xi (\lambda, \bar\lambda, \theta, r)$ satisfying $Q\xi =1$ if $\xi$ is allowed to have poles of order $\lambda^{-11}$ in $D=10$ or poles of order $\lambda^{-23}$ in $D=11$. And if $\xi$ is allowed in the Hilbert space of states, all BRST cohomology becomes trivial since any state $V$ satisfying $QV=0$ can be expressed as $V = Q (\xi V)$. So in order for these actions to correctly describe the on-shell superfields, one needs to impose restrictions on the possible pole dependence of the superfields $\Psi$. But since the pole dependence of the product of superfields can be more singular than the pole dependence of individual superfields, it is not obvious how to restrict the pole dependence of the superfields in a manner which is consistent with the non-linear BRST transformations of the action. In section 2 of this paper, the $D=10$ pure spinor superparticle and the pure spinor actions for $D=10$ super-Maxwell and super-Yang-Mills will be reviewed. And in section 3, these actions will be generalized to abelian $D=10$ supersymmetric Born-Infeld constructed in terms of a non-minimal pure spinor superfield $\Psi$. The super-Born-Infeld equations of motion take the simple form $$\begin{aligned} Q\Psi + k(\lambda\gamma^{m}\hat{\chi}\Psi)(\lambda\gamma^{n}\hat{\chi}\Psi)\hat{F}_{mn}\Psi &=& 0\end{aligned}$$ where $k$ is the dimensionful coupling constant and $\hat\chi^\alpha$ and $\hat F_{mn}$ are operators depending in a complicated manner on the non-minimal variables. After expanding $\Psi$ in powers of $k$ as $\Psi= \sum_{i=0}^{\infty}k^{i}\Psi_{i}$, one finds that $\Psi_0$ satisfies the equation $Q\Psi_0=0$ with the super-Maxwell solution $\Psi_0 = \lambda^\alpha A_\alpha (x,\theta)$, and $\Psi_1$ can be uniquely decomposed as $$\Psi_1(x, \theta, \lambda, \bar\lambda, r) = \tilde\Psi_1(x, \theta, \lambda) + \Lambda(A_{\alpha}, \lambda, \bar\lambda, r)$$ where $\tilde\Psi_1$ satisfies [@Cederwall:2001td; @Chang:2014nwa] $$\begin{aligned} \label{eq01} Q\tilde\Psi_1 + (\lambda\gamma^{m}\chi)(\lambda\gamma^{n}\chi){F}_{mn} = 0\end{aligned}$$ and $\chi^\alpha$ and $F_{mn}$ are the linearized spinor and vector field-strengths constructed from the super-Maxwell superfield in $\Psi_0$. It is straightforward to show that correctly describes the first-order correction of Born-Infeld to the super-Maxwell equations. In section 4 of this paper, the $D=11$ pure spinor superparticle and the pure spinor action for linearized $D=11$ supergravity will be reviewed. And in section 5, this action will be generalized to the complete $D=11$ supergravity action constructed in terms of a non-minimal pure spinor superfield $\Psi$. The supergravity equations of motion take the form $$\begin{aligned} Q\Psi + \frac{\kappa}{2}(\lambda\Gamma_{ab}\lambda)R^{a}\Psi R^{b}\Psi + \frac{\kappa}{2}\Psi \{Q , T\}\Psi - \kappa^{2}(\lambda\Gamma_{ab}\lambda)T\Psi R^{a}\Psi R^{b}\Psi &=& 0\end{aligned}$$ where $\kappa$ is the dimensionful coupling constant and $R^a$ and $T$ are operators depending in a complicated manner on the non-minimal variables. After expanding $\Psi$ in powers of $\kappa$ as $\Psi= \sum_{i=0}^{\infty}\kappa^{i}\Psi_{i}$, one finds that $\Psi_0$ satisfies the equation $Q\Psi_0=0$ with the linearized supergravity solution $\Psi_0 = \lambda^\alpha\lambda^\beta\lambda^\gamma C_{\alpha\beta\gamma}(x, \theta)$, and $\Psi_1$ can be uniquely decomposed as $$\Psi_1(x, \theta, \lambda, \bar\lambda, r) = \tilde\Psi_1(x, \theta, \lambda) + \Lambda(C_{\alpha\beta\delta}, \lambda, \bar\lambda, r)$$ where $\tilde\Psi_1$ satisfies $$\begin{aligned} \label{eq02} Q\tilde\Psi_1 + \frac{1}{2}(\lambda\Gamma_{ab}\lambda)\Phi^{a}\Phi^{b} &=& 0\end{aligned}$$ and $\Phi^a \equiv \lambda^\alpha E_\alpha^{(0)P}\hat{E}_P{}^a (x,\theta)$ is constructed from the linear deformation of the supergravity supervielbein and the background value of its respective inverse. It is straightforward to show that correctly describes the first-order correction to the linearized supergravity equations. Finally, Appendices \[apA\] and \[apB\] will contain some useful gamma matrix identities in $D=10$ and $D=11$, and Appendix \[apC\] will explain the relation of the $D=11$ supergravity superfields $\Psi$ and $\Phi^a$. Ten-dimensional Pure Spinor Superparticle and Super Yang-Mills ============================================================== In this section we will review the pure spinor description for the ten-dimensional superparticle and its connection with ten-dimensional super-Maxwell. We will then discuss the generalization to the non-abelian case. $D=10$ Pure spinor superparticle -------------------------------- The ten-dimensional pure spinor superparticle action is given by [@Berkovits:2001rb; @Berkovits:2002zk] $$\begin{aligned} S &=& \int d\tau \left[P_{m}\partial_{\tau}X^{m} + p_{\mu}\partial_{\tau}\theta^{\mu} + w_{\mu}\partial_{\tau}\lambda^{\mu}\right]\end{aligned}$$ where $X^{m}$ is a ten-dimensional coordinate, $\theta^{\mu}$ is a ten-dimensional Majorana-Weyl spinor, $\lambda^{\mu}$ is a bosonic ten-dimensional Weyl spinor satisfying $\lambda\gamma^{m}\lambda = 0$; and $P_{m}$, $p_{\mu}$, $w_{\mu}$ are the conjugate momenta relative to $X^{m}$, $\theta^{\mu}$, $\lambda^{\mu}$ respectively. We are using Greek/Latin letters from the middle of the alphabet to denote ten-dimensional Majorana-Weyl spinor/vector indices. Furthermore, $(\gamma^{m})^{\mu\nu}$ and $(\gamma^{m})_{\mu\nu}$ are $16\times 16$ symmetric real matrices satisfying $(\gamma^{m})^{\mu\nu}(\gamma^{n})_{\nu\sigma} + (\gamma^{n})^{\mu\nu}(\gamma^{m})_{\nu\sigma} = 2\eta^{mn}\delta^{\mu}_{\sigma}$. The BRST operator is given by $$\begin{aligned} Q_{0} &=& \lambda^{\mu}d_{\mu}\end{aligned}$$ where $d_{\mu} = p_\mu - (\gamma^m \theta)_\mu P_m$ are the fermionic constraints of the $D=10$ Brink-Schwarz superparticle [@BRINK1981310]. The physical spectrum is defined as the cohomology of the BRST operator $Q_{0}$. One can show that the ten-dimensional super-Maxwell physical fields are described by ghost number one states: $\Psi = \lambda^{\mu}A_{\mu}$. This can be easily seen since states in the cohomology satisfy the equation of motion and gauge invariance $$\begin{aligned} (\gamma^{mnpqr})^{\mu\nu}D_{\mu}A_{\nu} &=& 0\nonumber\\ \delta A_{\mu} &=& D_{\mu}\Lambda\label{eeeq104}\end{aligned}$$ where $D_\mu = {\partial\over{\partial \theta^\mu}}- (\gamma^m\theta)_\mu \partial_m$. These are indeed the superspace constraints describing ten-dimensional super-Maxwell. It can be shown that the remaining non-trivial cohomology is found at ghost number 0, 2 and 3 states; describing the super-Maxwell ghost, antifields and antighost, respectively, as dictated by BV quantization. $D=10$ Super-Maxwell -------------------- In order to describe $D=10$ super-Maxwell from a well-defined pure spinor action principle, one should introduce non-minimal pure spinor variables [@Berkovits:2005bt]. These non-minimal variables were studied in detail in [@Berkovits:2006vi] and consist of a pure spinor $\bar{\lambda}_{\mu}$ satisfying $\bar{\lambda}\gamma^{m}\bar{\lambda} = 0$, a fermionic spinor $r_{\mu}$ satisfying $\bar{\lambda}\gamma^{m}r = 0$ and their respective conjugate momenta $\bar{w}^{\mu}$, $s^{\mu}$. The non-minimal BRST operator is defined as $Q = Q_{0} + r_{\mu}\bar{w}^{\mu}$, so that these non-minimal variables will not affect the BRST cohomology. This means that one can always find a representative in the cohomology which is independent of non-minimal variables. Note that it will be assumed that the dependence on the non-minimal variables of the states is restricted to diverge slower than $(\lambda\bar\lambda)^{-11}$ when $\lambda^\mu \to 0$. Without this restriction, any BRST-closed operator is BRST-trivial since $Q(\xi V) = V$ where $\xi \equiv (\lambda\bar\lambda +r\theta)^{-1} (\bar\lambda\theta)$. Since the gauge transformation $\delta \Psi = Q\Lambda$ of super-Maxwell is linear, this restriction is easy to enforce by imposing a similar restriction on the gauge parameter $\Lambda$. However, for the non-linear gauge transformations discussed in the following sections for the super-Yang-Mills, supersymmetric Born-Infeld, and supergravity actions, it is unclear how to enforce this restriction. We shall ignore this subtlety here, but it is an important open problem to define the allowed set of states and gauge transformations for $\Psi$ and $\Lambda$ in these nonlinear actions. Let $\mathcal{S}_{SM}$ be the following pure spinor action $$\begin{aligned} \label{eeeq102} \mathcal{S}_{SM} &=& \int [dZ] \,\Psi Q \Psi\end{aligned}$$ where $[dZ] = [d^{10}x][d^{16}\theta][d\lambda][d\bar{\lambda}][d r] N$ is the integration measure, $\Psi$ is a pure spinor superfield (which can also depend on non-minimal variables) and $Q$ is the non-minimal BRST-operator. Let us explain what $[dZ]$ means. Firstly, $[d^{10}x][d^{16}\theta]$ is the usual measure on ordinary ten-dimensional superspace. The factors $[d\lambda][d\bar{\lambda}][d r]$ are given by $$\begin{aligned} \left[d\lambda\right]\lambda^{\mu}\lambda^{\nu}\lambda^{\rho} &=& (\epsilon T^{-1})^{\mu\nu\rho}_{\hspace{5mm}\sigma_{1}\ldots\sigma_{11}}d\lambda^{\sigma_{1}}\ldots d\lambda^{\sigma_{11}}\nonumber\\ \left[d\bar{\lambda}\right]\bar{\lambda}_{\mu}\bar{\lambda}_{\nu}\bar{\lambda}_{\rho} &=& (\epsilon T)_{\mu\nu\rho}^{\hspace{5mm}\sigma_{1}\ldots\sigma_{11}}d\bar{\lambda}_{\sigma_{1}}\ldots d\bar{\lambda}_{\sigma_{11}}\nonumber\\ \left[d r\right] &=& (\epsilon T^{-1})^{\mu\nu\rho}_{\hspace{5mm}\sigma_{1}\ldots\sigma_{11}}\bar{\lambda}_{\mu}\bar{\lambda}_{\nu}\bar{\lambda}_{\rho}(\frac{\partial}{\partial r_{\sigma_{1}}})\ldots (\frac{\partial}{\partial r_{\sigma_{11}}})\end{aligned}$$ where the Lorentz-invariant tensors $(\epsilon T)_{\mu\nu\rho}{}^{\sigma_{1}\ldots\sigma_{11}}$ and $(\epsilon T^{-1})^{\mu\nu\rho}{}_{\sigma_{1}\ldots \sigma_{11}}$ were defined in [@Berkovits:2006vi]. They are symmetric and gamma-traceless in $(\mu,\nu,\rho)$ and are antisymmetric in $[\sigma_{1},\ldots,\sigma_{11}]$. $N = e^{-Q(\bar{\lambda} \theta)} = e^{(-\bar{\lambda}\lambda - r \theta)}$ is a regularization factor. Since the measure converges as $\lambda^{8}\bar{\lambda}^{11}$ when $\lambda \to 0$, the action is well-defined as long as the integrand diverges slower than $\lambda^{-8}\bar{\lambda}^{-11}$. One can easily see that the equation of motion following from is given by $$\begin{aligned} Q\Psi = 0\end{aligned}$$ and since the measure factor $[dZ]$ picks out the top cohomology of the ten-dimensional pure spinor BRST operator, the transformation $\delta \Psi = Q\Lambda$ is a symmetry of the action . Therefore, describes $D=10$ super-Maxwell. $D=10$ Super Yang-Mills ----------------------- Let us define $\mathcal{S}_{SYM}$ to be $$\begin{aligned} \label{eq222} \mathcal{S}_{SYM} &=& \int [dZ] \,Tr(\frac{1}{2}\Psi Q\Psi + \frac{g}{3}\Psi \Psi \Psi)\end{aligned}$$ where $[dZ]$ is the measure discussed above, $\Psi$ is a Lie-algebra valued generic pure spinor superfield, $Q$ is the non-minimal BRST operator and $g$ is the coupling constant. For $SU(n)$ gauge group, expand $\Psi$ in the form: $\Psi = \Psi^{a}T^{a}$, where $T^{a}$ are the Lie algebra generators and $a = 1,\ldots , n^{2}-1$. Using the conventions: $[T^{a}, T^{b}] = f^{abc}T^{c}$ with $f^{abc}$ totally antisymmetric, and $Tr(T^{a}T^{b})=\delta^{ab}$, one can rewrite as follows $$\begin{aligned} \mathcal{S}_{SYM} &=& \int [dZ](\frac{1}{2}\Psi^{a}Q\Psi^{a} + \frac{g}{6}f^{abc}\Psi^{a}\Psi^{b}\Psi^{c})\end{aligned}$$ The e.o.m following from this action is given by: $$\begin{aligned} \label{eq225} Q\Psi^{a} + \frac{g}{2}f^{abc}\Psi^{b}\Psi^{c} &=& 0\end{aligned}$$ or in compact form $$\begin{aligned} \label{eq223} Q\Psi + g\Psi\Psi &=& 0\end{aligned}$$ It turns out that is invariant under the BRST symmetry $$\begin{aligned} \delta\Psi^{a} &=& Q\Lambda^{a} + f^{abc}\Psi^{b}\Lambda^{c}\end{aligned}$$ or in compact form $$\begin{aligned} \label{eq224} \delta\Psi = Q\Lambda + [\Psi , \Lambda]\end{aligned}$$ Since the equations , describe on-shell $D=10$ Super Yang-Mills on ordinary superspace [@Berkovits:2002zk], one concludes that the action describes $D=10$ Super Yang-Mills on a pure spinor superspace. Pure Spinor Description of Abelian Supersymmetric Born-Infeld ============================================================= In this section, we review the construction of the pure spinor action for supersymmetric abelian Born-Infeld and deduce the equations of motion on minimal pure spinor superspace to first order in the coupling. Physical operators ------------------ In order to deform the quadratic super-Maxwell action to the supersymmetric Born-Infeld action, Cederwall introduced the ghost number -1 pure spinor operators [@Cederwall:2011vy] $$\begin{aligned} \hat{A}_{\mu} &=& -\frac{1}{(\lambda\bar{\lambda})}[\frac{1}{8}(\gamma^{mn}\bar{\lambda})_{\mu}N_{mn} + \frac{1}{4}\bar{\lambda}_{\mu}N]\nonumber\\ \hat{A}_{m} &=& -\frac{1}{4(\lambda\bar{\lambda})}(\bar{\lambda}\gamma_{m}D) + \frac{1}{32(\lambda\bar{\lambda})^{2}}(\bar{\lambda}\gamma_{m}^{\hspace{2mm}np}r)N_{np}\nonumber\\ \hat{\chi}^{\mu} &=& \frac{1}{2(\lambda\bar{\lambda})}(\gamma^{m}\bar{\lambda})^{\mu}\Delta_{m}\nonumber\\ \hat{F}_{mn} &=& -\frac{1}{4(\lambda\bar{\lambda})}(r\gamma_{mn}\hat{\chi}) = \frac{1}{8(\lambda\bar{\lambda})}(\bar{\lambda}\gamma_{mn}^{\hspace{4mm}p}r)\Delta_{p}\label{eeq6}\end{aligned}$$ where $\Delta_{m}$ is defined by $$\begin{aligned} \Delta_{m} &=& \partial_{m} + \frac{1}{4(\lambda\bar{\lambda})}(r\gamma_{m}D) - \frac{1}{32(\lambda\bar{\lambda})^{2}}(r\gamma_{mnp}r)N^{np}\label{eeq5}\end{aligned}$$ These operators are constructed to satisfy $$\begin{aligned} \left[ Q, \hat{A}_{\mu}\right] &=& -D_{\mu} - 2(\gamma^{m}\lambda)_{\mu}\hat{A}_{m}\nonumber\\ \{Q, \hat{A}_{m}\} &=& \partial_{m} - (\lambda\gamma_{m}\hat{\chi})\nonumber\\ \left[ Q, \hat{\chi}^{\mu}\right] &=& -\frac{1}{2}(\gamma^{mn}\lambda)^{\mu}\hat{F}_{mn}\nonumber\\ \{Q, \hat{F}_{mn}\} &=& 2(\lambda\gamma_{[m}\partial_{n]}\hat{\chi})\end{aligned}$$ which mimic the superspace equations of motion of $D=10$ Super-Maxwell $$\begin{aligned} D_{\alpha}\Psi_0 + Q_{0} A_{\mu} + 2(\gamma^{m}\lambda)_{\mu}A_{m} &=& 0\nonumber\\ \partial_{m}\Psi_0 - Q_{0} A_{m} - (\lambda\gamma_{m}\chi) &=& 0\nonumber\\ Q_{0}\chi^{\mu} + \frac{1}{2}(\lambda\gamma_{mn})^{\mu}F^{mn} &=& 0\nonumber\\ Q_{0}F_{mn} - 2(\lambda\gamma_{[m}\partial_{n]}\chi) &=& 0 \label{eeq1}\end{aligned}$$ with $\Psi_0 = \lambda^{\mu}A_{\mu}$. If one acts with these operators on $\Psi_0$, they satisfy $$\begin{aligned} \hat{A}_{\mu}\Psi_0 = A_{\mu}\hspace{2mm},\hspace{2mm} \hat{A}_{m}\Psi_0 = A_{m}\hspace{2mm},\hspace{2mm} \hat{\chi}^{\mu}\Psi_0 = \chi^{\mu}\hspace{2mm},\hspace{2mm} \hat{F}_{mn}\Psi_0 = F_{mn}\label{eeq2} \end{aligned}$$ up to BRST-exact terms and certain “shift-symmetry terms” defined in [@Cederwall:2011vy; @Cederwall:2013vba]. For example, the operator $\hat{A}_{\mu}$ acts as $$\begin{aligned} \label{eeq3} \hat{A}_{\mu}\Psi_0 &=& A_{\mu} - \frac{1}{2(\lambda\bar{\lambda})}(\lambda\gamma^{m})_{\mu}(\bar{\lambda}\gamma_{m}A)\end{aligned}$$ where the shift symmetry is $\delta A_\mu = (\lambda\gamma_m)_\mu \phi^m$ for any $\phi^m$. For $\hat{A}_{m}$ one finds that $$\begin{aligned} \hat{A}_{m}\Psi_0 &=& A_{m} - (\lambda\gamma_{m}\rho) + Q[\frac{1}{4(\lambda\bar{\lambda})}(\bar{\lambda}\gamma_{m}A)]\end{aligned}$$ where the on-shell relation $D_{(\mu}A_{\nu)} = -(\gamma^{m})_{\mu\nu}A_{m}$ has been used, $\delta A_m = (\lambda \gamma_m)_\mu \rho^\mu$ is the shift symmetry, and $$\label{eeq4} \rho^{\mu} = \frac{1}{2(\lambda\bar{\lambda})}(\bar{\lambda}\gamma^{m})^{\mu}A_{m} + \frac{1}{8(\lambda\bar{\lambda})^{2}}(\bar{\lambda}\gamma^{m})^{\mu}(r\gamma_{m}A)$$ Analogously, one can show a similar behavior for the other operators $\hat{\chi}^{\mu}$, $\hat{F}_{mn}$. $D=10$ Abelian supersymmetric Born-Infeld ----------------------------------------- The deformation to the linearized action consistent with BRST symmetry is given by [@Cederwall:2011vy] $$\begin{aligned} \label{eeq12} \mathcal{S}_{SBI} &=& \int [dZ] \left[\frac{1}{2}\Psi Q\Psi + \frac{k}{4}\Psi(\lambda\gamma^{m}\hat{\chi}\Psi)(\lambda\gamma^{n}\hat{\chi}\Psi)\hat{F}_{mn}\Psi\right]\end{aligned}$$ which is invariant under the BRST transformation $$\begin{aligned} \delta\Psi &=& Q\Lambda + k(\lambda\gamma^{m}\hat{\chi}\Psi)(\lambda\gamma^{n}\hat{\chi}\Psi)\hat{F}_{mn}\Lambda + 2k(\lambda\gamma^{m}\hat{\chi}\Psi)(\lambda\gamma^{n}\hat{\chi}\Lambda)\hat{F}_{mn}\Psi\end{aligned}$$ for any ghost number 0 pure spinor superfield $\Lambda$. Note that $k$ is a dimensionful parameter related to the string tension by $k = \alpha\ensuremath{'}^{2}$. The equation of motion coming from is $$\begin{aligned} \label{eeq8} Q\Psi + k(\lambda\gamma^{m}\hat{\chi}\Psi)(\lambda\gamma^{n}\hat{\chi}\Psi)\hat{F}_{mn}\Psi &=& 0\end{aligned}$$ which can be written in terms of $\Delta_{m}$ as follows $$\begin{aligned} \label{eeq9} Q\Psi + \frac{k}{8(\lambda\bar{\lambda})^{2}}(\bar{\lambda}\gamma^{mnp}r)(\Delta_{m}\Psi)(\Delta_{n}\Psi)(\Delta_{p}\Psi) &=& 0\label{eeq9}\end{aligned}$$ Since the equation of motion of for $\Psi$ depends explicitly on the non-minimal variables, it is not obvious how to extract from $\Psi$ the Born-Infeld superfield $\tilde A_\mu(x,\theta)$ which should be independent of the non-minimal variables. However, it will now be argued that there is a unique decomposition of the solution to as $$\begin{aligned} \label{eq1000} \Psi(x,\theta,\lambda,\bar{\lambda},r) &=& \lambda^\mu \tilde A_\mu (x,\theta) + \Lambda (\tilde A_\mu, \lambda, \bar\lambda, r)\end{aligned}$$ where $\tilde A_\mu(x,\theta)$ is the on-shell Born-Infeld superfield and $\Lambda$ depends on $\tilde A_\mu$ and on the non-minimal variables. This will be explicitly shown here to the leading Born-Infeld correction to super-Maxwell, and work is in progress on extending this to the complete Born-Infeld solution. As mentioned in footnote \[foot1\], a similar procedure was used in [@Chang:2014nwa] for the abelian and non-abelian Born-Infeld solutions. To extract this leading-order correction to super-Maxwell from , we will first expand the pure spinor superfield $\Psi$ in positive powers of $k$: $$\begin{aligned} \label{eq100} \Psi(x,\theta,\lambda,\bar{\lambda},r) &=& \sum_{i=0}^{\infty}k^{i}\Psi_{i}\end{aligned}$$ The replacement of in gives us the following recursive relations $$\begin{aligned} Q\Psi_{0} &=& 0 \label{eq101}\\ Q\Psi_{1} &=& -\frac{1}{8(\lambda\bar{\lambda})^{2}}(\bar{\lambda}\gamma^{mnp}r)\Delta_{m}\Psi_{0}\Delta_{n}\Psi_{0}\Delta_{p}\Psi_{0}\label{eq102}\\ Q\Psi_{2} &=& -\frac{3}{8(\lambda\bar{\lambda})^{2}}(\bar{\lambda}\gamma^{mnp}r)\Delta_{m}\Psi_{1}\Delta_{n}\Psi_{0}\Delta_{p}\Psi_{0}\\ \vdots\nonumber\end{aligned}$$ To determine $\Lambda$ in , first note that has the solution $\Psi_0 = \lambda^\mu A_{0\,\mu}$ where $ A_{0\,\mu}$ is the super-Maxwell superfield which is independent of the non-minimal variables. However, the solution $\Psi_1$ to must depend on the non-minimal variables because the right-hand side of depends on these variables. To decompose the solution $\Psi_1$ to the form $$\begin{aligned} \label{eq1001} \Psi_1(x,\theta,\lambda,\bar{\lambda},r) &=& \lambda^\mu A_{1\,\mu} (x,\theta) + \Lambda,\end{aligned}$$ note that implies $$\begin{aligned} \label{eeeq99} Q(\frac{1}{(\lambda\bar{\lambda})^{2}}(\bar{\lambda}\gamma^{mnp}r)\Delta_{m}\Psi_{0}\Delta_{n}\Psi_{0}\Delta_{p}\Psi_{0}) = 0\end{aligned}$$ Since any BRST-closed expression can be expressed in terms of minimal variables up to a BRST-trivial term, there must exist a term $\Lambda$ such that $$\begin{aligned} \label{eqeqe34} -\frac{1}{8(\lambda\bar{\lambda})^{2}}(\bar{\lambda}\gamma^{mnp}r)\Delta_{m}\Psi_{0}\Delta_{n}\Psi_{0}\Delta_{p}\Psi_{0} &=& Q\Lambda + F(\Psi_{0})\end{aligned}$$ where $F({\Psi_{0}})$ is independent of non-minimal variables. This equation determines $\Lambda$ and $F(\Psi_0)$ up to the shift $$\begin{aligned} \label{eqeqe33} \delta \Lambda = H(\Psi_0) + Q\Omega, \quad \delta F(\Psi_0) = - Q H(\Psi_0)\end{aligned}$$ where $H(\Psi_0)$ only depends on the minimal variables. But the BRST-trivial shift $F(\Psi_0) \to F(\Psi_0) - QH(\Psi_0)$ can be cancelled by a redefinition of the field $\Psi_0 \to \Psi_0 - k H(\Psi_0)$. So the ambiguity in defining $\Lambda$ in does not affect the physical spectrum. In order to find $\Lambda$ and $F(\Psi_{0})$ in , first write $\Delta_{m}$ in the more convenient form $$\begin{aligned} \Delta_{m} &=& \partial_{m} - \{Q , \hat{A}_{m}\} + \bar{\lambda}\gamma_{m}\hat{\xi}\end{aligned}$$ where $\hat{\xi}_{\mu}$ is an operator depending on $N_{mn}, D_{\mu}$, etc. Although it is not complicated to determine $\hat{\xi}_{\mu}$, this will not be relevant for our purposes as we will see later. Using the on-shell relation $\partial_{m}A_{\mu} - D_{\mu}A_{m} = (\gamma_{m}\chi)_{\mu}$, one finds that $$\begin{aligned} \Delta_{m}\Psi &=& (\lambda\gamma_{m}\chi) + \lambda\gamma_{m}Q\rho + \bar{\lambda}\gamma_{m}\hat{\xi}\Psi\end{aligned}$$ So $$\begin{aligned} -\frac{1}{8}(\bar{\lambda}\gamma^{mnp}r)\Delta_{m}\Psi_{0}\Delta_{n}\Psi_{0}\Delta_{p}\Psi_{0}&=& -\frac{1}{8(\lambda\bar{\lambda})^{2}}(\bar{\lambda}\gamma^{mnp}r)[(\lambda\gamma_{m}\chi)(\lambda\gamma_{n}\chi)(\lambda\gamma_{p}\chi)\nonumber\\ && + 3(\lambda\gamma_{m}Q\rho)(\lambda\gamma_{n}\chi)(\lambda\gamma_{p}\chi)\nonumber\\ && + 3(\lambda\gamma_{m}Q\rho)(\lambda\gamma_{n}Q\rho)(\lambda\gamma_{p}\chi)\nonumber\\ &&+ (\lambda\gamma_{m}Q\rho)(\lambda\gamma_{n}Q\rho)(\lambda\gamma_{p}Q\rho)]\end{aligned}$$ The first term $H_{1} = -\frac{1}{8(\lambda\bar{\lambda})^{2}}(\bar{\lambda}\gamma_{mnp}r)(\lambda\gamma^{m}\chi)(\lambda\gamma^{n}\chi)(\lambda\gamma^{p}\chi)$ will provide us the term independent of non-minimal variables: $$\begin{aligned} H_{1} &=& -\frac{1}{8(\lambda\bar{\lambda})^{2}}(\bar{\lambda}\gamma_{mnp}r)(\lambda\gamma^{m}\chi)(\lambda\gamma^{n}\chi)(\lambda\gamma^{p}\chi)\nonumber\\ &=& \frac{1}{4(\lambda\bar{\lambda})}(r\gamma^{mn}\chi)(\lambda\gamma_{m}\chi)(\lambda\gamma_{n}\chi) - \frac{1}{8(\lambda\bar{\lambda})^{2}}(\lambda r)(\bar{\lambda}\gamma_{mn}\chi)(\lambda\gamma^{m}\chi)(\lambda\gamma^{n}\chi)\nonumber\\ &=& Q[\frac{1}{4(\lambda\bar{\lambda})}(\bar{\lambda}\gamma_{mn}\chi)(\lambda\gamma^{m}\chi)(\lambda\gamma^{n}\chi)] - F_{mn}(\lambda\gamma^{m}\chi)(\lambda\gamma^{n}\chi)\end{aligned}$$ where the identity was used. Analogous computations show us that the other terms are Q-exact: $$\begin{aligned} H_{2} &=& -\frac{3}{8(\lambda\bar{\lambda})^{2}}(\bar{\lambda}\gamma^{mnp}r)(\lambda\gamma_{m}Q\rho)(\lambda\gamma_{n}\chi)(\lambda\gamma_{p}\chi)\nonumber\\ &=& Q[\frac{6}{8(\lambda\bar{\lambda})}(\bar{\lambda}\gamma_{mn}Q\rho)(\lambda\gamma^{m}\chi)(\lambda\gamma^{n}\chi)]\end{aligned}$$ $$\begin{aligned} H_{3} &=& -\frac{3}{8(\lambda\bar{\lambda})^{2}}(\bar{\lambda}\gamma^{mnp}r)(\lambda\gamma_{m}Q\rho)(\lambda\gamma_{n}Q\rho)(\lambda\gamma_{p}\chi)\nonumber\\ &=& Q[\frac{6}{8(\lambda\bar{\lambda})}(\bar{\lambda}\gamma_{mn}Q\rho)(\lambda\gamma^{m}Q\rho)(\lambda\gamma^{n}\chi)]\end{aligned}$$ $$\begin{aligned} H_{4} &=& -\frac{1}{8(\lambda\bar{\lambda})^{2}}(\bar{\lambda}\gamma^{mnp}r)(\lambda\gamma_{m}Q\rho)(\lambda\gamma_{n}Q\rho)(\lambda\gamma_{p}Q\rho)\nonumber\\ &=& Q[\frac{2}{8(\lambda\bar{\lambda})}(\bar{\lambda}\gamma_{mn}Q\rho)(\lambda\gamma^{m}Q\rho)(\lambda\gamma^{n}Q\rho)]\end{aligned}$$ Hence, one obtains $$\begin{aligned} -\frac{1}{8}(\bar{\lambda}\gamma^{mnp}r)\Delta_{m}\Psi_{0}\Delta_{n}\Psi_{0}\Delta_{p}\Psi_{0} &=& Q[\Lambda] - F_{mn}(\lambda\gamma^{m}\chi)(\lambda\gamma^{n}\chi)\end{aligned}$$ where $\Lambda$ is defined by the expression $$\begin{aligned} \Lambda &=& \frac{1}{4(\lambda\bar{\lambda})}(\bar{\lambda}\gamma_{mn}\chi)(\lambda\gamma^{m}\chi)(\lambda\gamma^{n}\chi) + \frac{3}{4(\lambda\bar{\lambda})}(\bar{\lambda}\gamma_{mn}Q\rho)(\lambda\gamma^{m}\chi)(\lambda\gamma^{n}\chi)\nonumber\\ && + \frac{3}{4(\lambda\bar{\lambda})}(\bar{\lambda}\gamma_{mn}Q\rho)(\lambda\gamma^{m}Q\rho)(\lambda\gamma^{n}\chi) + \frac{1}{4(\lambda\bar{\lambda})}(\bar{\lambda}\gamma_{mn}Q\rho)(\lambda\gamma^{m}Q\rho)(\lambda\gamma^{n}Q\rho)\end{aligned}$$ Now, let us define the field $\tilde\Psi = \Psi_{0} + k(\Psi_{1} - \Lambda)$ which satisfies to first order in $k$ the equation of motion $$\begin{aligned} Q\tilde{\Psi} &=&Q (\Psi_0 + k(\Psi_1 - \Lambda)) = - k F_{mn}(\lambda\gamma^{m}\chi)(\lambda\gamma^{n}\chi)\label{eeq11}\end{aligned}$$ where $F_{mn}$, $\chi^{\mu}$ are the usual super-Maxwell superfields constructed from $A_{0\,\mu}$. Since the equation does not involve non-minimal variables, the solution is $$\begin{aligned} \tilde \Psi = \lambda^\mu \tilde {A}_{\mu} \end{aligned}$$ where $\tilde A_\mu\equiv A_{0\,\mu} + k A_{1\,\mu}$ satisfies $$\begin{aligned} \lambda^{\mu}\lambda^{\nu}\left [D_{\mu}\tilde{A}_{\nu} + k(\gamma^{m}\chi)_{\mu}(\gamma^{n}\chi)_{\nu}F_{mn}\right] &=& 0\end{aligned}$$ This equation of motion coincides, at first order in $k$, with the abelian supersymmetric Born-Infeld equations of motion [@BERGSHOEFF1987371; @JAMESGATES1987172; @Berkovits:2002ag]. So it has been shown to first order in $k$ that $$\begin{aligned} \Psi &=& \lambda^\mu \tilde A_\mu + k\Lambda\end{aligned}$$ where $\tilde A_\mu(x, \theta)$ is the on-shell Born-Infeld superfield and $\Lambda$ depends on $A_{0\,\mu}$ and on the non-minimal variables. Eleven-Dimensional Pure Spinor Superparticle and Supergravity ============================================================= In this section we review the eleven-dimensional pure spinor superparticle and its connection with linearized eleven-dimensional supergravity. $D=11$ Pure spinor superparticle -------------------------------- The eleven-dimensional pure spinor superparticle action is given by [@Berkovits:2002uc; @PhysRevD.97.066002] $$\begin{aligned} S &=& \int d\tau \left[P_{m}\partial_{\tau}X^{m} + P_{\mu}\partial_{\tau}\theta^{\mu} + w_{\alpha}(\partial_{\tau}\lambda^{\alpha}+ \partial_\tau Z^M \Omega_{M\beta}{}^\alpha \lambda^\beta)\right]\end{aligned}$$ where $X^{m}$ is an eleven-dimensional coordinate, $\theta^{\mu}$ is an eleven-dimensional Majorana spinor, $Z^M=(X^m, \theta^\mu)$, $\lambda^{\alpha}$ is a bosonic eleven-dimensional Majorana spinor satisfying $\lambda\Gamma^{a}\lambda = 0$; $P_{m}$, $P_{\mu}$, $w_{\alpha}$ are the conjugate momenta relative to $X^{m}$, $\theta^{\mu}$, $\lambda^{\alpha}$ respectively, and $\Omega_{M\beta}{}^\alpha$ is the spin connection of the background. We are using Greek/Latin letters from the beginning of the alphabet to denote tangent-space eleven-dimensional spinor/vector indices, and Greek/Latin letters from the middle of the alphabet to denote coordinate-space eleven-dimensional spinor/vector indices. Furthermore, capital letters from the beginning of the alphabet will denote tangent-space indices (both spinor and vector) and capital letters from the middle of the alphabet will denote coordinate-space indices (both spinor and vector). Finally, $(\Gamma^{a})^{\alpha\beta}$ and $(\Gamma^{a})_{\beta\delta}$ are $32\times 32$ symmetric matrices satisfying $(\Gamma^{a})^{\alpha\beta}(\Gamma^{b})_{\beta\delta}$ + $(\Gamma^{b})^{\alpha\beta}(\Gamma^{a})_{\beta\delta} = 2\eta^{ab}\delta^{\alpha}_{\delta}$. The BRST operator is given by $$\begin{aligned} Q_{0} &=& \lambda^{\alpha}d_{\alpha}\end{aligned}$$ where $$\begin{aligned} d_\alpha &=& E_{\alpha}^{\hspace{2mm}M} (P_M +\Omega_{M\beta}{}^\gamma w_\gamma \lambda^\beta)\end{aligned}$$ In a flat Minkowski background, $d_{\alpha} = P_{\alpha} - (\Gamma^{m}\theta)_{\alpha}P_{m}$ are the fermionic constraints of the $D=11$ Brink-Schwarz-like superparticle. The physical spectrum is defined as the cohomology of the BRST operator $Q_{0}$. One can show that the eleven-dimensional linearized supergravity physical fields are described by ghost number three states: $\Psi = \lambda^{\alpha}\lambda^{\beta}\lambda^{\delta}C_{\alpha\beta\delta}$ [@Berkovits:2002uc] where the physical state condition imposes the following equations of motion and gauge transformations for $C_{\alpha\beta\delta}$ $$\begin{aligned} \label{eqeqe1} D_{(\alpha}C_{\beta\delta\epsilon)} &=& (\Gamma^{a})_{(\alpha\beta}C_{\vert a\vert \delta\epsilon)}\nonumber\\ \delta C_{\alpha\beta\delta} &=& D_{(\alpha}\Lambda_{\beta\delta)}\end{aligned}$$ for some superfield $\Lambda_{\beta\delta}$. These are the superspace constraints describing eleven-dimensional linearized supergravity [@BRINK1980384]. It can be shown that the remaining non-trivial cohomology is found at ghost number 0, 1, 2, 4, 5, 6 and 7 states; describing the ghosts, antifields and antighosts as dictated by BV quantization of $D=11$ linearized supergravity. $D=11$ Linearized Supergravity ------------------------------ In order to describe $D=11$ linearized supergravity from a pure spinor action principle, one should introduce eleven-dimensional non-minimal pure spinor variables [@Cederwall:2013vba]. These non-minimal variables were studied in detail in [@Cederwall:2012es; @Cederwall:2009ez] and consist of a pure spinor $\bar{\lambda}_{\alpha}$ satisfying $\bar{\lambda}\Gamma^{a}\bar{\lambda} = 0$, a fermionic spinor $r_{\alpha}$ satisfying $\bar{\lambda}\Gamma^{a}r = 0$ and their respective conjugate momenta $\bar{w}^{\alpha}$, $s^{\alpha}$. The non-minimal BRST operator is defined as $Q = Q_{0} + r_{\alpha}\bar{w}^{\alpha}$, so that these non-minimal variables will not affect the BRST cohomology. Let $\mathcal{S}_{LSG}$ be the following pure spinor action $$\begin{aligned} \label{eeeeq102} \mathcal{S}_{LSG} &=& \int [dZ] \,\Psi Q \Psi\end{aligned}$$ where $[dZ] = [d^{11}x][d^{32}\theta][d\lambda][d\bar{\lambda}][d r] N$ is the integration measure, $\Psi$ is a pure spinor superfield (which, in general, can also depend on non-minimal variables) and $Q$ is the non-minimal BRST-operator. Let us explain what $[dZ]$ means. Firstly, $[d^{11}x][d^{32}\theta]$ is the usual measure on ordinary eleven-dimensional superspace. The factors $[d\lambda][d\bar{\lambda}][d r]$ are given by $$\begin{aligned} \left[d\lambda\right]\lambda^{\alpha_{1}}\ldots\lambda^{\alpha_{7}} &=& (\epsilon T^{-1})^{\alpha_{1}\ldots\alpha_{7}}_{\hspace{9mm}\beta_{1}\ldots\beta_{23}}d\lambda^{\beta_{1}}\ldots d\lambda^{\beta_{23}}\nonumber\\ \left[d\bar{\lambda}\right]\bar{\lambda}_{\alpha_{1}}\ldots\bar{\lambda}_{\alpha_{7}} &=& (\epsilon T)_{\alpha_{1}\ldots\alpha_{7}}^{\hspace{9mm}\beta_{1}\ldots\beta_{23}}d\bar{\lambda}_{\beta_{1}}\ldots d\bar{\lambda}_{\beta_{23}}\nonumber\\ \left[d r\right] &=& (\epsilon T^{-1})^{\alpha_{1}\ldots\alpha_{7}}_{\hspace{9mm}\beta_{1}\ldots\beta_{23}}\bar{\lambda}_{\alpha_{1}}\ldots\bar{\lambda}_{\alpha_{7}}(\frac{\partial}{\partial r_{\beta_{1}}})\ldots (\frac{\partial}{\partial r_{\beta_{23}}})\end{aligned}$$ The Lorentz-invariant tensors $(\epsilon T)_{\alpha_{1}\ldots\alpha_{7}}^{\hspace{9mm}\beta_{1}\ldots\beta_{23}}$ and $(\epsilon T^{-1})^{\alpha_{1}\ldots \alpha_{7}}_{\hspace{9mm}\beta_{1}\ldots\beta_{23}}$ were defined in [@Cederwall:2009ez]. They are symmetric and gamma-traceless in $(\alpha_{1},\ldots,\alpha_{7})$ and are antisymmetric in $[\beta_{1},\ldots ,\beta_{23}]$. $N$ is a regularization factor which is given by $N = e^{-\lambda\bar{\lambda} - r\theta}$. Since the measure converges as $\lambda^{16}\bar{\lambda}^{23}$ when $\lambda\to 0$, the action is well-defined if the integrand diverges slower than $\lambda^{-16}\bar{\lambda}^{-23}$. One can easily see that the equation of motion following from is given by $$\begin{aligned} Q\Psi = 0\end{aligned}$$ and since the measure factor $[dZ]$ picks out the top cohomology of the eleven-dimensional pure spinor BRST operator, the transformation $\delta \Psi = Q\Lambda$ is a symmetry of the action , that is a gauge symmetry of the theory. Therefore, describes $D=11$ linearized supergravity. Pure Spinor Description of Complete $D=11$ Supergravity ======================================================= As discussed in [@Cederwall:2013vba; @Cederwall:2010tn] , the pure spinor BRST-invariant action for complete $D=11$ supergravity is given by $$\begin{aligned} \label{eq105} \mathcal{S}_{SG} &=& {1\over{\kappa^2}}\int [dZ][\frac{1}{2}\Psi Q\Psi + \frac{1}{6}(\lambda\Gamma_{ab}\lambda)(1 - \frac{3}{2}T\Psi)\Psi R^{a}\Psi R^{b}\Psi]\end{aligned}$$ which is invariant under the BRST symmetry $$\begin{aligned} \delta\Psi &=& Q\Lambda + (\lambda\Gamma_{ab}\lambda)R^{a}\Psi R^{b}\Lambda + \frac{1}{2}\Psi\{Q,T\}\Lambda - \frac{1}{2}\Lambda\{Q,T\}\Psi - 2(\lambda\Gamma_{ab}\lambda)T\Psi R^{a}\Psi R^{b}\Lambda\nonumber\\ && - (\lambda\Gamma_{ab}\lambda)(T\Lambda)R^{a}\Psi R^{b}\Psi\end{aligned}$$ for any ghost number 2 pure spinor superfield $\Lambda$. Here $\kappa$ is the gravitational coupling constant, and $R^{a}$ and $T$ are ghost number -2 and -3 operators respectively, defined by the relations [@Cederwall:2009ez; @Cederwall:2010tn] $$\begin{aligned} R^{a} &=& -8[\frac{1}{\eta}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})\partial_{b} + \frac{1}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)(\lambda\Gamma_{bcd}D)\nonumber\\ && - \frac{4}{\eta^{3}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)(\bar{\lambda}\Gamma^{ef}r)(\lambda\Gamma_{fb}\lambda)(\lambda\Gamma_{cde}w)\nonumber\\ && + \frac{4}{\eta^{3}}(\bar{\lambda}\Gamma^{ac}\bar{\lambda})(\bar{\lambda}\Gamma^{de}r)(\bar{\lambda}\Gamma^{bf}r)(\lambda\Gamma_{fb}\lambda)(\lambda\Gamma_{cde}w)]\label{eeq13}\\ T &=& \frac{512}{\eta^{3}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}r)(rr)N_{ab},\end{aligned}$$ and $\eta \equiv (\lambda \Gamma^{ab}\lambda)(\bar\lambda\Gamma_{ab}\bar\lambda)$. Note that the action is invariant under the shift symmetry $\delta R^{a} = (\lambda\Gamma^{a}{\cal O})$ for any operator ${\cal O}$. The equation of motion coming from the action is $$\begin{aligned} Q\Psi + \frac{1}{2}\Psi\{Q,T\}\Psi + \frac{1}{2}(\lambda\Gamma_{ab}\lambda)(1 - 2T\Psi)R^{a}\Psi R^{b}\Psi &=& 0\end{aligned}$$ To compare with the linearized equations, it is convenient to rescale $\Psi \to \kappa \Psi$ so that $\kappa$ drops out of the quadratic term in the action, and the e.o.m. takes the form $$\begin{aligned} Q\Psi + \frac{\kappa}{2}(\lambda\Gamma_{ab}\lambda)R^{a}\Psi R^{b}\Psi + \frac{\kappa}{2}\Psi \{Q , T\}\Psi - \kappa^{2}(\lambda\Gamma_{ab}\lambda)T\Psi R^{a}\Psi R^{b}\Psi &=& 0\label{eq106}\end{aligned}$$ In order to find the superspace equations of motion, we expand the pure spinor superfield $\Psi$ in positive powers of $\kappa$ $$\begin{aligned} \Psi &=& \sum_{n=0}^{\infty} \kappa^{n}\Psi_{n}\end{aligned}$$ where $\Psi_{0}$ is the linearized solution satisfying $Q\Psi_{0} = 0$, which describes linearized 11D supergravity. The recursive relations that one finds from equation are: $$\begin{aligned} Q\Psi_{0} &=& 0\\ Q\Psi_{1} + \frac{1}{2}(\lambda\Gamma_{ab}\lambda)R^{a}\Psi_{0} R^{b}\Psi_{0} + \frac{1}{2}\Psi_{0} \{Q , T\}\Psi_{0} &=& 0\label{eq107}\\ \vdots\nonumber\end{aligned}$$ The procedure will now be the same as that applied to the Born-Infeld case: We will first write the non-minimal contribution to as a BRST-exact term $Q\Lambda$. We will then define a new superfield $\tilde\Psi = \Psi-\Lambda$, which will satisfy the equation $Q\tilde\Psi = G(\Psi_{0})$ where $G(\Psi_{0})$ is independent of non-minimal variables. We will finally identify $\tilde C_{\alpha\beta\gamma} = C_{0\,\alpha\beta\gamma} + \kappa C_{1\,\alpha\beta\gamma}$ in $\tilde \Psi = \lambda^\alpha\lambda^\beta \lambda^\gamma \tilde C_{\alpha\beta\gamma}$ as the first-order correction to the linearized D=11 superfield. To find $\Lambda$ and $G(\Psi_0)$, the first step will be to write $R^{a}\Psi_{0}$ in terms of a superfield $\Phi^a(x, \theta, \lambda)$ depending only on minimal variables as $$\begin{aligned} \label{eeq15} R^{a}\Psi_{0} &=& \Phi^{a}(x, \theta, \lambda) + Q(f^{a}) + \lambda\Gamma^a{\cal O}\end{aligned}$$ where $\lambda \Gamma^a {\cal O}$ is the shift symmetry of $R^a$. To linearized order in the supergravity deformation of the background, the superfield $\Phi^a$ can be expressed in terms of the super-vielbein $E_A{}^P$ and its inverse $E_P{}^A$ as $$\begin{aligned} \label{eeqq19} \Phi^a &=& \lambda^\alpha E_\alpha^{(0)P} \hat E_P{}^a\end{aligned}$$ where $E_A{}^P$ and $E_P{}^A$ have been expanded around their background values $\hat E_A{}^P$ and $\hat E_P{}^A$ as $$\begin{aligned} E_A{}^P &=& \hat E_A{}^P + \kappa E_A^{(0)P} + \kappa^2 E_A^{(1)P} + ...,\nonumber\\ E_P{}^{A} &=& \hat E_P{}^A + \kappa E_P^{(0)A} + \kappa^2 E_P^{(1)A} + ... .\end{aligned}$$ For example, if one is expanding around the Minkowski space background, $\hat E_a{}^p = \delta_a^p$, $\hat E_\alpha{}^\mu = \delta_\alpha^\mu$ and $\hat E_\alpha{}^m = - (\Gamma^m\theta)_\alpha$. Note that $E^{(0)P}_\alpha \hat E_P{}^a + \hat E_\alpha{}^P E^{(0) a}_P =0$, so one can also express $\Phi^a$ to linearized order in the deformation as $$\begin{aligned} \label{eeqq20} \Phi^a &=& - \lambda^\alpha \hat E_\alpha{}^P E_P^{(0)a}\end{aligned}$$ Since all of the supergravity fields are contained in $\Psi_0$, one should be able to describe $\Phi$ in terms of $\Psi_0$. As discussed in [@Cederwall:2009ez], this relation is given by and it will be explicitly shown in Appendix \[apC\] that $$\begin{aligned} f^{a} &=& -\frac{24}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)(\lambda\Gamma_{bcd})^{\delta}C_{\delta\alpha\beta}\lambda^{\alpha}\lambda^{\beta} - \frac{24}{\eta}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})C_{b\alpha\beta}\lambda^{\alpha}\lambda^{\beta}\end{aligned}$$ Plugging eq. in implies that $$\begin{aligned} Q\Psi_{1} + \frac{1}{2}(\lambda\Gamma_{ab}\lambda)[\Phi^{a} + Qf^{a}][\Phi^{b} + Qf^{b}] - Q\left[\frac{1}{2}\Psi_{0}T\Psi_{0}\right] &=& 0,\end{aligned}$$ which implies that $$\begin{aligned} \label{eeq19} Q(\Psi_1 - \Lambda) = - \frac{1}{2}(\lambda\Gamma_{ab}\lambda)\Phi^{a}\Phi^{b}\end{aligned}$$ where $$\Lambda = \frac{1}{2}\Psi_{0}T\Psi_{0} + (\lambda\Gamma_{ab}\lambda)\Phi^{a}f^{b} - \frac{1}{2}(\lambda\Gamma_{ab}\lambda)f^{a}Qf^{b}.$$ Hence one can define the superfield $\tilde{\Psi}$: $$\begin{aligned} \tilde{\Psi} = \Psi_{0} + \kappa(\Psi_1 - \Lambda)\end{aligned}$$ which will satisfy the following e.o.m at linear order in $\kappa$ $$\begin{aligned} Q\tilde{\Psi} &=& - \frac{\kappa}{2}(\lambda\Gamma_{ab}\lambda)\Phi^{a}\Phi^{b}\end{aligned}$$ which implies $$\begin{aligned} \label{eq99} \lambda^{\alpha}\lambda^{\beta}\lambda^{\delta}\lambda^{\epsilon} [D_{\alpha}\tilde{C}_{\beta\delta\epsilon} +\frac{\kappa}{2}(\Gamma_{ab})_{\alpha\beta}E_{\delta}^{(0)P} \hat E_P{}^a E_{\epsilon}^{(0)Q}\hat E_Q{}^b] =0 \end{aligned}$$ where $\tilde\Psi = \lambda^\alpha\lambda^\beta\lambda^\delta \tilde C_{\alpha\beta\delta}$. This equation of motion will now be shown to coincide with the $D=11$ supergravity equations of motion at first order in $\kappa$. The non-linear $D=11$ supergravity equations of motion can be expressed using pure spinors as $$\begin{aligned} \label{eqeqe181} \lambda^\alpha \lambda^\beta \lambda^\gamma \lambda^\delta H_{\alpha\beta\gamma\delta} =0\end{aligned}$$ where we use the standard transformation rule from curved to tangent-space indices for the 4-form superfield strength: $$\begin{aligned} \label{eqeqe301} H_{\alpha\beta\delta\epsilon} &=& E_{\alpha}^{\hspace{2mm}M}E_{\beta}^{\hspace{2mm}N}E_{\delta}^{\hspace{2mm}P}E_{\epsilon}^{\hspace{2mm}Q}H_{MNPQ}\end{aligned}$$ and $H_{MNPQ} = \nabla_{[M} C_{NPQ]}$. Furthermore, implies that one can choose conventional constraints (by appropriately defining $C_{\alpha\beta a}$ and $C_{\alpha a b}$) so that $$H_{\alpha\beta\gamma\delta}= H_{\alpha\beta\gamma a}= 0, \quad H_{\alpha\beta a b}= -{1\over{12}} (\Gamma_{ab})_{\alpha\beta}.$$ This is expected since there are no physical supergravity fields with the dimensions of $H_{\alpha\beta\gamma\delta}$, $H_{\alpha\beta\gamma a}$ and $H_{\alpha\beta a b}$. To perform an expansion in $\kappa$ and compare with , define $$\begin{aligned} \hat{H}_{\alpha\beta\gamma\delta} &=& \hat{E}_{\alpha}^{\hspace{2mm}M}\hat{E}_{\beta}^{\hspace{2mm}N}\hat{E}_{\gamma}^{\hspace{2mm}P}\hat{E}_{\delta}^{\hspace{2mm}Q}H_{MNPQ}.\end{aligned}$$\[eqeq000\] Equation implies that $$0 = \lambda^\alpha \lambda^\beta \lambda^\gamma \lambda^\delta (\hat H_{\alpha\beta\gamma\delta} +4\kappa \hat E_\alpha{}^M \hat E_\beta{}^N \hat E_\gamma{}^P E_\delta^{(0)Q} H_{MNPQ}$$ $$\begin{aligned} +6 \kappa^2 \hat E_\alpha{}^M \hat E_\beta{}^N E_\gamma^{(0) P} E_\delta{}^{(0)Q} H_{MNPQ} + ...) \end{aligned}$$\[eqq180\] $$= \lambda^\alpha \lambda^\beta \lambda^\gamma \lambda^\delta (\hat H_{\alpha\beta\gamma\delta} +4\kappa \hat E_\alpha{}^M \hat E_\beta{}^N \hat E_\gamma{}^P E_\delta^{(0)Q} E_M{}^A E_N{}^B E_P{}^C E_Q{}^D H_{ABCD}$$ $$\begin{aligned} +6 \kappa^2 \hat E_\alpha{}^M \hat E_\beta{}^N E_\gamma^{(0) P} E_\delta^{(0)Q} E_M{}^A E_N{}^B E_P{}^C E_Q{}^D H_{ABCD} + ...) \end{aligned}$$ $$= \lambda^\alpha \lambda^\beta \lambda^\gamma \lambda^\delta (\hat H_{\alpha\beta\gamma\delta} +12\kappa^2 \hat E_\gamma{}^P E_P^{(0) a} E_\delta^{(0)Q}\hat E_Q{}^b H_{\alpha\beta a b}$$ $$\begin{aligned} +6 \kappa^2 E_\gamma^{(0) P}\hat E_P{}^a E_\delta^{(0)Q} \hat E_Q{}^b H_{\alpha \beta a b} + ...) \end{aligned}$$ $$\begin{aligned} \label{eqq22} &=& \lambda^\alpha \lambda^\beta \lambda^\gamma \lambda^\delta (\hat H_{\alpha\beta\gamma\delta} + {1\over 2} \kappa^2 E_\gamma^{(0) P} \hat E_P{}^a E_\delta^{(0)Q} \hat E_Q{}^b (\Gamma_{ab})_{\alpha \beta} + ...)\end{aligned}$$ where $...$ denotes terms higher-order in $\kappa$. Since $$\lambda^\alpha \lambda^\beta \lambda^\gamma \lambda^\delta\hat H_{\alpha\beta\gamma\delta} = \kappa \lambda^\alpha \lambda^\beta \lambda^\gamma \lambda^\delta D_\alpha \tilde C_{\beta\gamma\delta},$$ equation for the back-reaction to $\hat H_{\alpha\beta\gamma\delta}$ coincides with . Acknowledgments =============== MG acknowledges FAPESP grant 2015/23732-2 for financial support and NB acknowledges FAPESP grants 2016/01343-7 and 2014/18634-9 and CNPq grant 300256/94-9 for partial financial support. $D=10$ gamma matrix identities {#apA} ============================== In $D=10$ dimensions, one has chiral and antichiral spinors which have been denoted here by $\chi^{\mu}$ and $\chi_{\mu}$ respectively. The product of two spinors can be decomposed into two forms depending on the chiralities of the spinors used: $$\begin{aligned} \xi_{\mu}\chi^{\nu} &=& \frac{1}{16}\delta^{\nu}_{\mu}(\xi\chi) - \frac{1}{2!16}(\gamma^{mn})^{\nu}_{\hspace{2mm}\mu}(\xi\gamma_{mn}\chi) + \frac{1}{4!16}(\gamma^{mnpq})^{\nu}_{\hspace{2mm}\mu}(\xi\gamma_{mnpq}\chi)\\ \xi^{\mu}\chi^{\nu} &=& \frac{1}{16}\gamma_{m}^{\mu\nu}(\xi\gamma^{m}\chi) + \frac{1}{3!16}(\gamma_{mnp})^{\mu\nu}(\xi\gamma^{mnp}\chi) + \frac{1}{5!32}\gamma^{\mu\nu}_{mnpqr}(\xi\gamma^{mnpqr}\chi) \label{ap4}\end{aligned}$$ The 1-form and 5-form are symmetric, and the 3-form is antisymmetric. Furthermore, it is true that $(\gamma^{mn})^{\mu}_{\hspace{2mm}\nu} = -(\gamma^{mn})_{\nu}^{\hspace{2mm}\mu}$, $(\gamma^{mnpq})^{\mu}_{\hspace{2mm}\nu} = (\gamma^{mnpq})_{\nu}^{\hspace{2mm}\mu}$. Two particularly useful identities are: $$\begin{aligned} (\gamma^{m})_{(\mu\nu}(\gamma_{m})_{\rho)\sigma} &=& 0\\ (\gamma^{m})^{\mu}_{\hspace{2mm}\nu}(\gamma_{m})^{\rho}_{\hspace{2mm}\sigma} &=& 4(\gamma^{m})^{\mu\rho}(\gamma_{m})_{\nu\sigma} - 2\delta^{\mu}_{\nu}\delta^{\rho}_{\sigma} - 8\delta^{\mu}_{\sigma}\delta^{\rho}_{\nu} \label{ap1}\end{aligned}$$ From \[ap1\] we can deduce the following: $$\begin{aligned} (\gamma^{mn})^{\mu}_{\hspace{2mm}\nu}\gamma_{mnp}^{\rho\sigma} &=& 2(\gamma^{m})^{\mu\rho}(\gamma_{pm})^{\sigma}_{\hspace{2mm}\nu} + 6\gamma_{p}^{\mu\rho}\delta^{\sigma}_{\nu} - (\rho \leftrightarrow \sigma)\label{ap2} \\ (\gamma^{mn})^{\mu}_{\hspace{2mm}\nu}(\gamma_{mnp})_{\rho\sigma} &=& -2\gamma^{m}_{\nu\sigma}(\gamma_{pm})^{\mu}_{\hspace{2mm}\rho} + 6(\gamma_{p})_{\nu\sigma}\delta^{\mu}_{\rho} - (\rho \leftrightarrow \sigma)\\ \gamma_{mnp}^{\mu\nu}(\gamma^{mnp})^{\rho\sigma} &=& 12[\gamma_{m}^{\mu\sigma}(\gamma^{m})^{\nu\rho} - \gamma_{m}^{\mu\rho}(\gamma^{m})^{\nu\sigma}] \label{ap5}\\ \gamma^{\mu\nu}_{mnp}\gamma^{mnp}_{\rho\sigma} &=& 48(\delta^{\mu}_{\rho}\delta^{\nu}_{\sigma} - \delta^{\mu}_{\sigma}\delta^{\nu}_{\rho})\label{ap6}\end{aligned}$$ $D=11$ gamma matrix identities {#apB} ============================== In $D=11$ dimensions, one has Majorana spinors and an antisymmetric tensor $C_{\alpha\beta}$ (and its inverse) which can be used to raise and lower spinor indices. The product of two spinors can be decomposed into the form $$\begin{aligned} \label{app7} \chi^{\alpha}\psi^{\beta} &=& -\frac{1}{32}C^{\alpha\beta}(\chi\psi) + \frac{1}{32}(\Gamma^{a})^{\alpha\beta}(\chi\Gamma_{a}\psi) - \frac{1}{2!.32}(\Gamma^{ab})^{\alpha\beta}(\chi\Gamma_{ab}\psi) + \frac{1}{3!.32}(\Gamma^{abc})^{\alpha\beta}(\chi\Gamma_{abc}\psi)\nonumber\\ && - \frac{1}{4!.32}(\Gamma^{abcd})^{\alpha\beta}(\chi\Gamma_{abcd}\psi) + \frac{1}{5!.32}(\Gamma^{abcde})^{\alpha\beta}(\chi\Gamma_{abcde}\psi)\end{aligned}$$ The 1-form, 2-form and 5-form are symmetric; and the 0-form, 3-form and 4-form are antisymmetric. The crucial identity in eleven dimensions is $$\begin{aligned} \label{ap7} (\Gamma^{ab})_{(\alpha\beta}(\Gamma_{b})_{\delta\epsilon)} &=& 0\end{aligned}$$ One can find analogous formulae to - for $D=11$ dimensions. However, they do not enter into any computations of this paper, therefore we will not list them. From and the pure spinor constraint, one can find the following useful pure spinor identities $$\begin{aligned} (\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\Gamma_{b}\bar{\lambda})_{\alpha} &=& 0 \label{app1}\\ (\bar{\lambda}\Gamma^{[ab}\bar{\lambda})(\bar{\lambda}\Gamma^{c]d}\bar{\lambda}) &=& 0 \label{app2}\\ (\bar{\lambda}\Gamma^{[ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd]}\bar{\lambda}) &=& 0 \label{app3}\\ (\bar{\lambda}\Gamma^{[ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd]}r) &=& 0 \label{app4} $$ If $a$ is a shift-symmetry index, there exists a very useful identity which states the following $$\begin{aligned} \label{ap8} (\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\lambda\Gamma_{cb}\lambda) &=& \frac{1}{2}\delta^{a}_{c}\eta\end{aligned}$$ This can be easily seen from the following argument. Eqn. implies the relation $$\begin{aligned} -(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\lambda\Gamma_{b}\Gamma_{c}\lambda) &=& 2(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma_{b}\Gamma_{c}\lambda) + 2(\bar{\lambda}\Gamma^{ab}\Gamma_{c}\lambda)(\bar{\lambda}\Gamma_{c}\lambda)\end{aligned}$$ which can be rewritten in the more convenient form $$\begin{aligned} -(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\lambda\Gamma_{b}\Gamma_{c}\lambda) &=& \lambda\Gamma^a\xi_c + 4\delta^a_c(\lambda\bar{\lambda})^2 - 4\delta^a_c (\bar{\lambda}\Gamma^b \lambda)(\bar{\lambda}\Gamma_b \lambda)\end{aligned}$$ where $\xi_c^\alpha$ is defined as follows $$\begin{aligned} \xi_c^\alpha &=& -2(\bar{\lambda}\Gamma_c)^\alpha (\bar\lambda \lambda) - 2(\bar{\lambda}\Gamma^b)^\alpha (\bar{\lambda}\Gamma_b \Gamma_c \lambda) + 2(\bar{\lambda}\Gamma_{bc})^\alpha (\lambda\Gamma^b \bar{\lambda}) + 4\bar{\lambda}^\alpha (\lambda\Gamma_c \bar\lambda)\end{aligned}$$ The use of allows us to write $$\begin{aligned} -(\lambda\bar{\lambda})^2 &=& -\frac{1}{64}\eta + \frac{1}{3840}(\lambda\Gamma^{abcde}\lambda)(\bar{\lambda}\Gamma_{abcde}\bar{\lambda})\\ (\bar{\lambda}\Gamma^a \lambda)(\bar{\lambda}\Gamma_a \lambda) &=& -\frac{7}{64}\eta - \frac{1}{3840}(\lambda\Gamma^{abcde}\lambda)(\bar{\lambda}\Gamma_{abcde}\bar{\lambda})\end{aligned}$$ Therefore, $$\begin{aligned} (\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\lambda\Gamma_{cb}\lambda) &=& \frac{1}{2}\delta^a_c \eta + \lambda\Gamma^a\xi_c\end{aligned}$$ Relation between $\Psi$ and $\Phi^{a}$ {#apC} ====================================== At linearized level, there exists a simple relation between $\Psi$ and $\Phi^{a}$. To find this relation, define $$\begin{aligned} \label{eqeqe18} \hat {H}_{ABCD} &=& \hat{E}_{A}^{\hspace{2mm}M}\hat{E}_{B}^{\hspace{2mm}N}\hat{E}_{C}^{\hspace{2mm}P}\hat{E}_{D}^{\hspace{2mm}Q} {H}_{MNPQ}\end{aligned}$$ as in . Using the conventions $$\begin{aligned} \label{eeq17} H_{\alpha\beta\delta\gamma} = 0\hspace{2mm},\hspace{2mm}H_{a\alpha\beta\delta} = 0\hspace{2mm},\hspace{2mm}H_{ab\alpha\beta} = -\frac{1}{12}(\Gamma_{ab})_{\alpha\beta}\hspace{2mm},\hspace{2mm}H_{abc\alpha} = 0, $$ one finds that $$\begin{aligned} \lambda^\alpha \lambda^\beta\lambda^\gamma\hat{H}_{a\alpha\beta\gamma} &=& \lambda^\alpha \lambda^\beta\lambda^\gamma\hat E_a{}^M \hat E_\alpha{}^N \hat E_\beta{}^P \hat E_\gamma{}^Q E_M{}^A E_N{}^B E_P{}^C E_Q{}^D H_{ABCD}\end{aligned}$$ $$\begin{aligned} &=& 3\kappa \lambda^\alpha \lambda^\beta\lambda^\gamma \hat E_\alpha{}^N E_N^{(0)b} H_{ab \beta\gamma} + ...\end{aligned}$$ $$\begin{aligned} &=& \frac{1}{4}\kappa \Phi^b \lambda^\beta\lambda^\gamma (\Gamma_{ab})_{\beta\gamma} + ...\end{aligned}$$ where $...$ denotes terms of order $\kappa^2$ and $\Phi^b = -\lambda^\alpha \hat E_\alpha{}^N E_N^{(0)b}$. Since $$\lambda^\alpha\lambda^\beta\lambda^\gamma\hat H_{a\alpha\beta\gamma} = \kappa(\partial_a \Psi_0 - 3Q(\lambda^{\alpha}\lambda^\beta C_{a\alpha\beta})) ,$$ one obtains the relation $$\begin{aligned} \label{eqeqe305} \partial_{a}\Psi_{0} &=& \frac{1}{4}(\lambda\Gamma_{ab}\lambda)\Phi^{b} + 3Q(\lambda^{\alpha}\lambda^\beta C_{a\alpha\beta})\end{aligned}$$ The use of equation and the linearized e.o.m $$\begin{aligned} \label{eeq20} D_{\alpha}\Psi_{0} + 3Q_{0}(C_{\alpha\beta\delta})\lambda^{\beta}\lambda^{\delta} = -6(\Gamma^{a}\lambda)_{\alpha}C_{a\beta\delta}\lambda^{\beta}\lambda^{\delta}\end{aligned}$$ allows us to compute the action of $R^{a}$ on $\Psi_{0}$ in the form displayed in . To see this, it will be useful to express $R^{a}$ in the more convenient way [@Cederwall:2010tn] $$\begin{aligned} R^{a} &=& -8[\frac{1}{\eta}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})\partial_{b} + \frac{1}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)(\lambda\Gamma_{bcd}D)\nonumber\\ && - \{Q , \frac{1}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)\}(\lambda\Gamma_{bcd}w)]\end{aligned}$$ Therefore, $$\begin{aligned} R^{a}\Psi_{0} &=& -8[\frac{1}{\eta}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})\partial_{b}\Psi_{0} + \frac{1}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)(\lambda\Gamma_{bcd}D\Psi_{0})\nonumber\\ && + 3\{Q , \frac{1}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)\}(\lambda\Gamma_{bcd})^{\alpha}C_{\alpha\beta\delta}\lambda^{\beta}\lambda^{\delta}]\nonumber\\ &=& -8[\frac{1}{\eta}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})\partial_{b}\Psi_{0} - \frac{3}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)(\lambda\Gamma_{bcd})^{\alpha} (Q C_{\alpha\beta\delta})\lambda^{\beta}\lambda^{\delta}\nonumber\\ && - \frac{6}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)(\lambda\Gamma_{bcd}\Gamma^{e}\lambda)C_{e\alpha\beta}\lambda^{\alpha}\lambda^{\beta}\nonumber\\ && + 3\{Q , \frac{1}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)\}(\lambda\Gamma_{bcd})^{\alpha}C_{\alpha\beta\delta}\lambda^{\beta}\lambda^{\delta}]\nonumber\\ &=& -8\{\frac{1}{\eta}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})\partial_{b}\Psi_{0} + Q\left[ \frac{3}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)(\lambda\Gamma_{bcd})^{\alpha}C_{\alpha\beta\delta}\lambda^{\beta}\lambda^{\delta}\right]\nonumber\\ && - \frac{6}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)[-2(\lambda\Gamma_{bd}\lambda)\eta^{e}_{c} + (\lambda\Gamma_{cd}\lambda)\eta^{e}_{b}]C_{e\alpha\beta}\lambda^{\alpha}\lambda^{\beta}\}\nonumber\\ &=& -8\{\frac{1}{\eta}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})\partial_{b}\Psi_{0} + Q\left[ \frac{3}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)(\lambda\Gamma_{bcd})^{\alpha}C_{\alpha\beta\delta}\lambda^{\beta}\lambda^{\delta}\right]\nonumber\\ && + \frac{6}{\eta}(\bar{\lambda}\Gamma^{ab}r)C_{b\alpha\beta}\lambda^{\alpha}\lambda^{\beta} - \frac{6}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda}) (\bar{\lambda}\Gamma^{cd}r)(\lambda\Gamma_{cd}\lambda) C_{b\alpha\beta}\lambda^{\alpha}\lambda^{\beta}\}\nonumber\\ &=& -8\{\frac{1}{\eta}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})\partial_{b}\Psi_{0} + Q\left[ \frac{3}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)(\lambda\Gamma_{bcd})^{\alpha}C_{\alpha\beta\delta}\lambda^{\beta}\lambda^{\delta}\right]\nonumber\\ && + Q[\frac{3}{\eta}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})C_{b\alpha\beta}\lambda^{\alpha}\lambda^{\beta}] - \frac{3}{\eta}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})Q[C_{b\alpha\beta}\lambda^{\alpha}\lambda^{\beta}]\}\nonumber\\ &=& -\frac{2}{\eta}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\lambda\Gamma_{bc}\lambda)\Phi^{c} + Q\left[ -\frac{24}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)(\lambda\Gamma_{bcd})^{\alpha}C_{\alpha\beta\delta}\lambda^{\beta}\lambda^{\delta} - \frac{24}{\eta}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})C_{b\alpha\beta}\lambda^{\alpha}\lambda^{\beta}\right]\nonumber\\ &=& \Phi^{a} + Q\left[ -\frac{24}{\eta^{2}}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})(\bar{\lambda}\Gamma^{cd}r)(\lambda\Gamma_{bcd})^{\alpha}C_{\alpha\beta\delta}\lambda^{\beta}\lambda^{\delta} - \frac{24}{\eta}(\bar{\lambda}\Gamma^{ab}\bar{\lambda})C_{b\alpha\beta}\lambda^{\alpha}\lambda^{\beta}\right]\end{aligned}$$ Notice that in order for the normalization factor of $\Phi^{a}$ to be one after applying $R^{a}$ on $\Psi$, one should choose the conventions used for $R^{a}$ in and those displayed in . [10]{} P. Howe, “Pure spinor lines in superspace and ten-dimensional supersymmetric theories,” [[ *Physics Letters B*]{} [**258**]{} (1991) no. 1, 141 – 144](http://dx.doi.org/https://doi.org/10.1016/0370-2693(91)91221-G). <http://www.sciencedirect.com/science/article/pii/037026939191221G>. N. Berkovits, “[Covariant quantization of the superparticle using pure spinors]{},” [[*JHEP*]{} [**09**]{} (2001) 016](http://dx.doi.org/10.1088/1126-6708/2001/09/016), [[arXiv:hep-th/0105050 \[hep-th\]]{}](http://arxiv.org/abs/hep-th/0105050). N. Berkovits, “[Towards covariant quantization of the supermembrane]{},” [[*JHEP*]{} [**09**]{} (2002) 051](http://dx.doi.org/10.1088/1126-6708/2002/09/051), [[arXiv:hep-th/0201151 \[hep-th\]]{}](http://arxiv.org/abs/hep-th/0201151). H. Gomez and C. R. Mafra, “[The closed-string 3-loop amplitude and S-duality]{},” [[*JHEP*]{} [**10**]{} (2013) 217](http://dx.doi.org/10.1007/JHEP10(2013)217), [[arXiv:1308.6567 \[hep-th\]]{}](http://arxiv.org/abs/1308.6567). M. Cederwall, “[D=11 supergravity with manifest supersymmetry]{},” [[*Mod. Phys. Lett.*]{} [**A25**]{} (2010) 3201–3212](http://dx.doi.org/10.1142/S0217732310034407), [[arXiv:1001.0112 \[hep-th\]]{}](http://arxiv.org/abs/1001.0112). M. Cederwall, “[Pure spinor superfields – an overview]{},” [[*Springer Proc. Phys.*]{} [**153**]{} (2014) 61–93](http://dx.doi.org/10.1007/978-3-319-03774-5_4), [[arXiv:1307.1762 \[hep-th\]]{}](http://arxiv.org/abs/1307.1762). M. Cederwall and A. Karlsson, “[Pure spinor superfields and Born-Infeld theory]{},” [[*JHEP*]{} [ **11**]{} (2011) 134](http://dx.doi.org/10.1007/JHEP11(2011)134), [[arXiv:1109.0809 \[hep-th\]]{}](http://arxiv.org/abs/1109.0809). C.-M. Chang, Y.-H. Lin, Y. Wang, and X. Yin, “[Deformations with Maximal Supersymmetries Part 2: Off-shell Formulation]{},” [[*JHEP*]{} [**04**]{} (2016) 171](http://dx.doi.org/10.1007/JHEP04(2016)171), [[arXiv:1403.0709 \[hep-th\]]{}](http://arxiv.org/abs/1403.0709). M. Cederwall, B. E. W. Nilsson, and D. Tsimpis, “[D = 10 superYang-Mills at O(alpha-prime\*\*2)]{},” [[*JHEP*]{} [**07**]{} (2001) 042](http://dx.doi.org/10.1088/1126-6708/2001/07/042), [[arXiv:hep-th/0104236 \[hep-th\]]{}](http://arxiv.org/abs/hep-th/0104236). N. Berkovits, “[ICTP lectures on covariant quantization of the superstring]{},” [*ICTP Lect. Notes Ser.*]{} [**13**]{} (2003) 57–107, [[arXiv:hep-th/0209059 \[hep-th\]]{}](http://arxiv.org/abs/hep-th/0209059). L. Brink and J. H. Schwarz, “Quantum superspace,” [[ *Physics Letters B*]{} [**100**]{} (1981) no. 4, 310 – 312](http://dx.doi.org/https://doi.org/10.1016/0370-2693(81)90093-9). <http://www.sciencedirect.com/science/article/pii/0370269381900939>. N. Berkovits, “[Pure spinor formalism as an N=2 topological string]{},” [[*JHEP*]{} [**10**]{} (2005) 089](http://dx.doi.org/10.1088/1126-6708/2005/10/089), [[arXiv:hep-th/0509120 \[hep-th\]]{}](http://arxiv.org/abs/hep-th/0509120). N. Berkovits and N. Nekrasov, “[Multiloop superstring amplitudes from non-minimal pure spinor formalism]{},” [[*JHEP*]{} [**12**]{} (2006) 029](http://dx.doi.org/10.1088/1126-6708/2006/12/029), [[arXiv:hep-th/0609012 \[hep-th\]]{}](http://arxiv.org/abs/hep-th/0609012). E. Bergshoeff, M. Rakowski, and E. Sezgin, “Higher derivative super [Yang]{}-[Mills]{} theories,” [[ *Physics Letters B*]{} [**185**]{} (1987) no. 3, 371 – 376](http://dx.doi.org/https://doi.org/10.1016/0370-2693(87)91017-3). <http://www.sciencedirect.com/science/article/pii/0370269387910173>. S. J. Gates and S. Vashakidze, “On d = 10, n = 1 supersymmetry, superspace geometry and superstring effects (i),” [[ *Nuclear Physics B*]{} [**291**]{} (1987) 172 – 204](http://dx.doi.org/https://doi.org/10.1016/0550-3213(87)90470-6). <http://www.sciencedirect.com/science/article/pii/0550321387904706>. N. Berkovits and V. Pershin, “[Supersymmetric Born-Infeld from the pure spinor formalism of the open superstring]{},” [[*JHEP*]{} [**01**]{} (2003) 023](http://dx.doi.org/10.1088/1126-6708/2003/01/023), [[arXiv:hep-th/0205154 \[hep-th\]]{}](http://arxiv.org/abs/hep-th/0205154). M. Guillen, [“Equivalence of the 11d pure spinor and [Brink]{}-[Schwarz]{}-like superparticle cohomologies,”[*Phys. Rev. D*]{} [**97**]{} (Mar, 2018) 066002](http://dx.doi.org/10.1103/PhysRevD.97.066002). <https://link.aps.org/doi/10.1103/PhysRevD.97.066002>. L. Brink and P. Howe, “Eleven-dimensional supergravity on the mass shell in superspace,” [[ *Physics Letters B*]{} [**91**]{} (1980) no. 3, 384 – 386](http://dx.doi.org/https://doi.org/10.1016/0370-2693(80)91002-3). <http://www.sciencedirect.com/science/article/pii/0370269380910023>. M. Cederwall and A. Karlsson, “[Loop amplitudes in maximal supergravity with manifest supersymmetry]{},” [[*JHEP*]{} [**03**]{} (2013) 114](http://dx.doi.org/10.1007/JHEP03(2013)114), [[arXiv:1212.5175 \[hep-th\]]{}](http://arxiv.org/abs/1212.5175). M. Cederwall, “[Towards a manifestly supersymmetric action for 11-dimensional supergravity]{},” [[*JHEP*]{} [**01**]{} (2010) 117](http://dx.doi.org/10.1007/JHEP01(2010)117), [[arXiv:0912.1814 \[hep-th\]]{}](http://arxiv.org/abs/0912.1814). [^1]: A similar procedure was used by Chang, Lin, Wang and Yin in [@Chang:2014nwa] to find the on-shell solution to abelian and non-abelian $D=10$ supersymmetric Born-Infeld. We thank Martin Cederwall for informing us of their work after the first version of our preprint was submitted.

ArXiv

--- abstract: 'We show that any homomorphism from the homeomorphism group of a compact $2$-manifold, with the compact-open topology, or equivalently, with the topology of uniform convergence, into a separable topological group is automatically continuous.' author: - Christian Rosendal date: November 2005 title: 'Automatic continuity in homeomorphism groups of compact $2$-manifolds' --- Introduction ============ It is well-known and easy to see that for any compact metric space $(X,d)$, its group of homeomorphisms is a separable complete metric group when equipped with the topology of uniform convergence or equivalently with the compact open topology. In fact, a compatible right-invariant metric on ${\rm Homeo}(X,d)$ is given by $d_\infty(g,f)=\sup_{x\in X}d(g(x),f(x))$, and a complete metric by $d_\infty'(g,f)=d_\infty(g,f)+d_\infty(g{^{-1}}, f{^{-1}})$. We denote by $B(x,{\epsilon})$ the open ball of radius ${\epsilon}$ around $x$ and by ${\overline}B(x,{\epsilon})$ the corresponding closed ball. If $g\in {\rm Homeo}(X,d)$, we denote by ${\rm supp}^\circ(g)$ the open set $\{x\in X{ \; \big| \;}g(x)\neq x\}$ and by ${\rm supp}(g)$ its closure, which we call the [*support*]{} of $g$. We intend to show here that in the case of compact $2$-manifolds, this group topology is intrinsically given by the underlying discrete or abstract group, in the sense that any homomorphism $\pi$ from this group into a separable group is continuous. Let $M$ be a compact $2$-manifold and $\pi:{\rm Homeo}(M){\rightarrow}H$ a homomorphism into a separable group. Then $\pi$ is automatically continuous when ${\rm Homeo}(M)$ is equipped with the compact-open topology. Let us first note the following simple fact, which helps to clear up the situation. Suppose $G$ is a topological group. Then the following conditions are equivalent. 1. Any homomorphism $\pi:G{\rightarrow}{\rm Homeo}([0,1]^{\mathbb N})$ is continuous, 2. any homomorphism $\pi:G{\rightarrow}H$ into a separable group is continuous. As $[0,1]^{\mathbb N}$ is a compact metric space, its homeomorphism group is a (completely metrisable) separable group in the compact-open topology, so (1) is a special case of (2). For the other implication, suppose that (1) holds and let $H$ be separable. Let $N$ be the closed normal subgroup of $H$ consisting of all elements that cannot be separated from the identity by an open set and let $H/N$ be the quotient topological group, which is Hausdorff and separable, and, in particular, any non-empty open set covers the group by countably many translates. However, it is an old result (see I.I. Guran [@guran]) that for Hausdorff groups this condition is equivalent to being topologically isomorphic to a subgroup of a direct product of separable metric groups, or equivalently, second countable Hausdorff groups (by the Birkhoff-Kakutani metrisation Theorem). Also, a result of Uspenskiĭ [@uspenskij] states that any separable metric group is topologically isomorphic to a subgroup of ${\rm Homeo}([0,1]^{\mathbb N})$, and we can therefore, see $H/N$ as a subgroup of some power of ${\rm Homeo}([0,1]^{\mathbb N})$. Thus, as a mapping into the Tikhonov product is continuous if and only if the composition with each coordinate projection is continuous, $\pi$ composed with the quotient mapping is continuous, and hence by the choice of $N$, also $\pi$ is continuous. However, we shall not use this result in any way, but instead simplify matters by not be working with arbitrary homomorphisms, but rather with arbitrary subsets of the group satisfying a certain algebraic largeness condition. Let $G$ be a group and $W\subseteq G$ be a symmetric set. We say that $W$ is [*countably syndetic*]{} if there are countably many left-translates of $W$ whose union cover $G$. Moreover if $G$ is a topological group, we say that $G$ is [*Steinhaus*]{} if for some $k\geq 1$ and all symmetric, countably syndetic $W\subseteq G$, $\rm Int(W^k)\neq {\emptyset}$. It is not hard to prove (see, e.g., [@rossol]) that Steinhaus groups satisfy the equivalent conditions of the above fact, and this is the condition that we will verify. Note however the order of quantification; the $k$ is universal for all symmetric, countably syndetic $W$. Indeed, the group ${\rm Homeo}_+(S^1)$ equipped with the trivial topology $\tau=\{{\emptyset},{\rm Homeo}_+(S^1)\}$ satisfies the condition when we have inversed the quantifiers, but the identity homomorphism into itself equipped with the compact-open topology is obviously discontinuous. It is instructive also to consider from which groups one can construct discontinuous homeomorphisms. Of course the first case that comes to mind is $({\mathbb R},+)$, on which one can with the help of a Hamel basis, i.e., a basis for ${\mathbb R}$ as a ${\mathbb Q}$-vector space, construct discontinuous automorphisms, and, in fact, construct group isomorphisms between ${\mathbb R}$ and ${\mathbb R}^2$. The proof ========= Commutators ----------- We shall first prove a general lemma about homeomorphisms of ${\mathbb R}^n$. \[commutator\]Suppose that $g\in {\rm Homeo}({\mathbb R}^n)$ has compact support. Then there are $f,h\in {{\rm Homeo}({\mathbb R}^n)}$ with compact support such that $g=[f,h]=fhf{^{-1}}h{^{-1}}$. Fix some open ball $U_0\subseteq {\mathbb R}^n$ containing the support of $g$ and let $(U_m)$ be a sequence of disjoint open balls such that for some distinct $x_0$ and $x_1$ in ${\mathbb R}^n$, the sequences $({\overline}U_m)_{m\geq 0}$ and $({\overline}U_{-m})_{m\geq 0}$ converge in the Vietoris topology to $x_0$ and $x_1$ respectively. We can now find a shift $h\in {{\rm Homeo}({\mathbb R}^n)}$ with compact support, i.e., such that $h[U_m]=U_{m+1}$ and define our $f$ by letting $f|U_m=h^mgh^{-m}|U_m$ for $m\geq 0$ and setting $f={\rm id}$ everywhere else. We now see that for $m>0$, $$hf{^{-1}}h{^{-1}}|U_m=h(h^{m-1}g{^{-1}}h^{-m+1})h{^{-1}}|U_m= h^mg{^{-1}}h^{-m}|U_m,$$ and for $m\leq 0$, $$hf{^{-1}}h{^{-1}}|U_m =h\,{\rm id}\,h{^{-1}}|U_m={\rm id}|U_m,$$ while $hf{^{-1}}h{^{-1}}={\rm id}$ everywhere else. Therefore, $f\cdot hf{^{-1}}h{^{-1}}|U_m={\rm id}|U_m$ for $m>0$, $f\cdot hf{^{-1}}h{^{-1}}|U_0=f|U_0=g|U_0$, $f\cdot hf{^{-1}}h{^{-1}}|U_m={\rm id}|U_m$ for $m<0$, and $fhf{^{-1}}h{^{-1}}={\rm id}$ everyhere else. This shows that $g=[f,h]=fhf{^{-1}}h{^{-1}}$. We notice that in the proof above we used $f$ and $h$ with slightly bigger support than $g$. I believe it is an open problem whether this can be avoided and indeed it seems to be a much harder problem. We can restate the problem as follows. Can every homeomorphism of $[0,1]^n$ that fixes the boundary pointwise be written as a commutator of $f$ and $h$ that also fixes the boundary pointwise? What happes if we replace pointwise by setwise? Let us mention that the first question has a positive answer in dimension $1$ as, for example, the group of orientation preserving homeomorphisms of $[0,1]$ has a comeagre conjugacy class [@kecros]. The above result slightly strengthens a result of Mather [@mather] saying that the homology groups of the group of homeomorphisms ${\mathbb R}^n$ with compact support vanish. One can of course also extend the lemma to $[0,\infty[\times {\mathbb R}^{n-1}$ and thus also improve the result of Rybicki [@rybicki]. Countably syndetic sets ----------------------- We will now prove some properties of coutably syndetic sets in the homeomorphism groups of arbitrary manifolds. These results will allow us to completely solve our problem for compact two-dimensional manifolds and provide techniques for higher dimensions. So let $M$ be a manifold of dimension $n$ and fix a compatible complete metric $d$ on $M$. In the following we fix a countably syndetic symmetric subset $W\subseteq {{\rm Homeo}(M)}$ and a sequence $k_m\in {{\rm Homeo}(M)}$ such that $\bigcup_mk_mW={{\rm Homeo}(M)}$. \[patched\] For all distinct $y_1,\ldots,y_p\in M$ and ${\epsilon}>0$, there are ${\epsilon}>\delta>0$ and $z_i\in B(y_i,{\epsilon})$ such that if $g\in {{\rm Homeo}(M)}$ has support contained in $D=\bigcup_{i=1}^p {\overline}B(z_i,\delta)$, then $g\in W^{16}$. We notice that it is enough to find $z_i\in B(y_i,{\epsilon})$ and open neighbourhoods $U_i$ of $z_i$ such that if $g\in {{\rm Homeo}(M)}$ has support contained in $\bigcup_iU_i$, then $g\in W^{16}$. We choose some open neighbourhood of $y_i$, $E_i\subseteq B(y_i,{\epsilon})$, that is homeomorphic to $]0,{\epsilon}[^2$. We also suppose that the sets $E_i$ are $4{\epsilon}$-separated. We will also temporarily transport the standard euclidian metric from $]0,{\epsilon}[^2$ to each of the sets $E_i$. As we will be working separately on each of $E_i$, this will not cause a problem. Thus in the following, the notation $B(x,\beta)$ will refer to the balls in the transported euclidian metric, which we denote by $d$. \[semi-patched\] For all $u_i\in E_i$ and $\gamma>0$ such that $d(u_i,\partial E_i)>2\gamma$, there are $\gamma>\alpha>0$ and $x_i\in \partial B(u_i,\gamma)$ such that if $g\in {{\rm Homeo}(M)}$ has support contained in $A=\bigcup_{i=1}^p {\overline}B(x_i,\alpha)\cap{\overline}B(u_i,\gamma)$, then there is an $h\in W^2$ with support contained in $\bigcup_{i=1}^p{\overline}B(u_i,\gamma)$ such that $g|A=h|A$. Let $u_1,\ldots,u_p$ be given. We fix for each $i\leq p$ a sequence of distinct points $x^i_m\in \partial B(u_i,\gamma)$ converging to some point $x^i_\infty\in\partial B(u_i,\gamma)$ and choose a sequence $\frac{\gamma}{2}>\alpha_m>0$ such that $B(x^i_m,\alpha_m)\cap B(x^i_l,\alpha_l)={\emptyset}$ for any $m\neq l$. Thus, as $\alpha_m{\rightarrow}0$, we have that if $g_m\in {{\rm Homeo}(M)}$ has support only in $$A_m=\big({\overline}B(x^1_m,\alpha_m)\cap {\overline}B(u_1,\gamma)\big)\cup\ldots \cup \big({\overline}B(x^p_m,\alpha_m)\cap {\overline}B(u_p,\gamma)\big)$$ for each $m\geq 0$, then there is a homeomorphism $g\in {{\rm Homeo}(M)}$, whose support is contained in $C={\overline}B(u_1,\gamma)\cup\ldots\cup {\overline}B(u_p,\gamma)$, such that $g| A_m=g_m|A_m$. We claim that for some $m_0\geq 0$, if $g\in {{\rm Homeo}(M)}$ has support contained in $A_{m_0}$, then there is an element $h\in k_{m_0}W$, with support contained in $C$, such that $g|A_{m_0}=h|A_{m_0}$. Assume toward a contradiction that this is not the case. Then for every $m$ we can find some $g_m\in {{\rm Homeo}(M)}$ with support contained in $A_m$ such that for all $h\in k_mW$, if ${\rm supp}(h)\subseteq C$, then $g_m|A_m\neq h|A_m$. But then letting $g\in {{\rm Homeo}(M)}$ have support in $C$ and agree with each $g_m$ on $A_m$ for each $m$, we see that if $h\in k_mW$ has support in $C$, then $g$ disagrees with $h$ on $A_m$. Therefore, $g$ cannot belong to any $k_mW$, contradicting that these cover ${{\rm Homeo}(M)}$. Suppose that $m_0$ has been chosen as above and denote $x^i_{m_0}$ by $x_i$, $A_{m_0}$ by $A$, and $\alpha_{m_0}$ by $\alpha$. Then for any $g\in {{\rm Homeo}(M)}$ with support contained in $A$, there is an element $h\in W^2$ with support contained in $C$ such that $g|A=h|A$ for all $i\leq p$. To see this, it is enough to notice that we can find $h_0,h_1\in k_{m_0}W$, with ${\rm supp}(h_0),{\rm supp}(h_1)\subseteq C$, such that $g|A=h_1|A$ and ${\rm id}|A=h_0|A$. But then $h_0{^{-1}}h_1\in (k_{m_0}W){^{-1}}k_{m_0}W=W{^{-1}}W=W^2$ and $g|A={\rm id}\,g|A=h_0{^{-1}}h_1|A$. We will first apply Sublemma \[semi-patched\] to the situation where $u_i=y_i$ and $\gamma>0$ is sufficiently small. We thus obtain $\gamma>\alpha>0$ and $x_i\in \partial B(y_i,\gamma)$ such that if $g\in {{\rm Homeo}(M)}$ has support contained in $A=\bigcup_{i=1}^p {\overline}B(x_i,\alpha)\cap{\overline}B(y_i,\gamma)$, then there is an $h\in W^2$ with support contained in $\bigcup_{i=1}^p{\overline}B(y_i,\gamma)$ such that $g|A=h|A$. Now pick $y'_i\in B(x_i,\alpha)\cap B(y_i,\gamma)$ and $\gamma'>0$ such that $B(y'_i,2\gamma')\subseteq B(x_i,\alpha)\cap B(y_i,\gamma)$. We now apply Lemma \[semi-patched\] once again to this new situation, in order to obtain $\gamma'>\alpha'>0$ and $x'_i\in \partial B(y'_i,\gamma')$ such that if $g\in {{\rm Homeo}(M)}$ has support contained in $A'=\bigcup_{i=1}^p {\overline}B(x'_i,\alpha')\cap{\overline}B(y'_i,\gamma')$, then there is an $h\in W^2$ with support contained in $\bigcup_{i=1}^p{\overline}B(y'_i,\gamma')$ such that $g|A'=h|A'$. Now clearly there is a homeomorphism $a\in {{\rm Homeo}(M)}$ whose support is contained in $A=\bigcup_{i=1}^p {\overline}B(x_i,\alpha)\cap{\overline}B(y_i,\gamma)$ such that $a[B(y'_i,\gamma')]=B(x'_i,\alpha')$ and $$a[{\overline}B(y'_i,\gamma')\cap {\overline}B(x'_i,\alpha')]={\overline}B(y'_i,\gamma')\cap {\overline}B(x'_i,\alpha'),$$ and hence we can also find such an $a$ in $W^2$, except that its support may now be all of $\bigcup_{i=1}^p{\overline}B(y_i,\gamma)$. We therefore have that if $g\in {{\rm Homeo}(M)}$ has support contained in $A'$, then $a{^{-1}}ga$ also has support contained in $A'$, and so there is an $h\in W^2$ with support contained in $\bigcup_{i=1}^p{\overline}B(y'_i,\gamma')$ such that $a{^{-1}}ga|A'=h|A'$. But then $g|A'=aha{^{-1}}|A'$, while $${\rm supp}(aha{^{-1}})=a[{\rm supp}(h)]\subseteq a[\bigcup_{i=1}^p{\overline}B(y'_i,\gamma')]=\bigcup_{i=1}^p{\overline}B(x'_i,\alpha').$$ We now notice that $aha{^{-1}}\in W^6$, and thus that if $g\in {{\rm Homeo}(M)}$ has support contained in $A'=\bigcup_{i=1}^p {\overline}B(x'_i,\alpha')\cap{\overline}B(y'_i,\gamma')$, then there is some $f\in W^6$ with support contained in $\bigcup_{i=1}^p{\overline}B(x'_i,\alpha')$ such that $g|A'=f|A'$. Now suppose finally that $g\in {{\rm Homeo}(M)}$ is any homeomorphism having support contained in $\bigcup_{i=1}^p B(x'_i,\alpha')\cap B(y'_i,\gamma')$. Since the sets $B(x'_i,\alpha')\cap B(y'_i,\gamma')$ are homeomorphic to ${\mathbb R}^n$, working separately on each of these sets and noticing that $g$ has compact support, we can invoke Lemma \[commutator\] to write $g$ as a commutator $[b,c]$ for some $b,c\in {{\rm Homeo}(M)}$ whose supports are contained in $\bigcup_{i=1}^p B(x'_i,\alpha')\cap B(y'_i,\gamma')\subseteq A'$. Find now $h\in W^2$ agreeing with $b$ on $A'$ and with support contained in $\bigcup_{i=1}^p{\overline}B(y'_i,\gamma')$, and, similarly, find $f\in W^6$ agreeing with $c$ on $A'$ and with support contained in $\bigcup_{i=1}^p{\overline}B(x'_i,\alpha')$. Then the set of common support of $h$ and $f$ is included in $A'$ on which they agree with $b$ and $c$ respectively, and we have therefore that $[h,f]=hfh{^{-1}}f{^{-1}}=bcb{^{-1}}c{^{-1}}=g$. In other words, $g\in W^{16}$. We can therefore finish the proof by choosing some $z_i\in B(x'_i,\alpha')\cap B(y'_i,\gamma')$ and letting $U_i=B(x'_i,\alpha')\cap B(y'_i,\gamma')$. Circular orders --------------- In order to simplify notation, we will consider [*circular orders*]{} on finite sets. Since we are really just interested in simplifying notation, let me just say what a circular order is in terms of an example, namely, $S^1$. For $x,y,z$ distinct points on $S^1$, $y$ is said to be between $x$ and $z$, in symbols $B(x,y,z)$, if going counterclockwise around $S^1$ from $x$ to $y$, one does not pass through $z$. Thus a circular order is just a circular betweeness relation. When $B$ is a circular order on a finite set ${\mathbb F}$, we denote for each $x\in {\mathbb F}$ its immediate successor and immediate predecessor, i.e., the first elements encountered by going respectively counterclockwise and clockwise around ${\mathbb F}$, by $x^+$ and $x^-$. So, e.g., $(x^+)^-=x$. A quantitative annulus theorem ------------------------------ Fix three points $v_0,v_1,v_2\in {\mathbb R}^2$ such that for $i\neq j$, $d(v_i,v_j)=1$, and denote by $\triangle$ the $2$-cell consisting of the points lying within the triangle $\triangle v_0v_1v_2$. Suppose also that the barycenter of $\triangle$ lies at the origin, so that for all $\lambda>0$, $\lambda \triangle$ and $\triangle$ are concentric triangles, the former with sidelengths $\lambda$. Let $\phi:(1-2\eta)\triangle{\rightarrow}\triangle$ be a homeomorphic embedding satisfying $$\sup_{x\in(1-2\eta)\triangle}d(x,\phi(x))<\frac\eta{100},$$ where $\eta<\frac1{1000}$. Then there is a homeomorphism $\psi:\triangle{\rightarrow}\triangle$ that is the identity outside of $(1-\eta)\triangle$, with $\sup_{x\in\triangle}d(x,\psi(x))<100\eta$, and such that $\psi\circ\phi|_{(1-2\eta)\triangle}={\rm id}$. Let $\partial(1-\eta)\triangle$ be the boundary of $(1-\eta)\triangle$ and pick a finite set of points ${\mathbb F}$ containing $(1-\eta)v_0,(1-\eta)v_1,(1-\eta)v_2$ and lying in $\partial(1-\eta)\triangle$, such that when ${\mathbb F}$ is equipped with the circular order obtained from going counterclockwise around $\partial(1-\eta)\triangle$, we have $d(x,x^+)\in ]20\eta,21\eta[$ for all $x\in {\mathbb F}$. As $\triangle$ is equilateral, $d(x,y)>20\eta$ for all $x\neq y$ in ${\mathbb F}$. Let now $C=\phi[\partial(1-2\eta)\triangle]$ be the image of the boundary of $(1-2\eta)\triangle$, so $C$ is a simple closed curve. Choose also for each $x\in {\mathbb F}$ a point $\hat x\in C$ such that the distance $d(x,\hat x)$ is minimal. Since $\sup_{x\in (1-2\eta)\triangle} d(x,\phi(x))<\frac\eta{100}$ and $$\frac \eta3<d(x,\partial(1-2\eta)\triangle)<\frac{2\eta}3$$ for all $x\in \partial(1-\eta)\triangle$, also $d(x,\hat x)<\eta$ and $d(C,\partial(1-\eta)\triangle)>\frac\eta4$. For all $x\in {\mathbb F}$, denote by $\alpha_x$ the straight (oriented) line segment from $x$ to $\hat x$ and by $\beta_x$ the straight line segment from $x$ to $x^+$. We also let $\gamma'_x$ be the shortest path in $\partial(1-2\eta)\triangle$ from $\phi{^{-1}}(\hat x)$ to $\phi{^{-1}}(\widehat{x^+})$ and put $\gamma_x=\phi[\gamma'_x]$. By definition of $\hat x$, $\alpha_x$ intersects $C$ exactly in $\hat x$, intersects $\partial(1-\eta)\triangle$ in exactly $x$, and therefore $\alpha_x$ and $\gamma_y$ intersect only if $y=x^-$ or $y=x$. Similarly, none of the paths $\beta_x$ and $\gamma_y$ intersect as they lie in $\partial(1-\eta)\triangle$ and $C$ respectively. Therefore, for any $x\in {\mathbb F}$, ${\mathcal}C_x=\alpha_x\centerdot\gamma_x\centerdot\bar\alpha_{x^+}\centerdot\bar\beta_x$ is a simple closed curve beginning and ending at $x$. Here $\bar\alpha$ denotes the reverse path of $\alpha$ and $\centerdot$ the concatenation of paths. By the Schönflies Theorem, ${\mathbb R}^2\setminus {\mathcal}C_x$ has exactly two components, one unbounded and the other $U_x$ bounded, homeomorphic with ${\mathbb R}^2$ and with boundary ${\mathcal}C_x$. Moreover, as the diameter of ${\mathcal}C_x$ is bounded by $30\eta$, ${\mathcal}C_x$ intersects $\partial(1-\eta)\triangle$ in exactly $\beta_x$, and the diameter of $\partial(1-\eta)\triangle\setminus\beta_x$ is $1-\eta>30\eta$, this means that $\partial(1-\eta)\triangle\setminus\beta_x$ lies in the unbounded component. Therefore, if $R_x=\overline U_x=U_x\cup {\mathcal}C_x$, we have for $x\neq y$ $$\begin{split} R_x\cap R_y =\left \{ \begin{array}{ll} {\emptyset}\;\;\;\;\;\;\;\;\;\;\;\;&\textrm{ if } y\neq x^+ \textrm { and } y\neq x^-\\ \alpha_y\;\;\;\;\;\;\;\;\;\;\;\;&\textrm{ if } y= x^+\\ \alpha_x\;\;\;\;\;\;\;\;\;\;\;\;&\textrm{ if } y= x^- \\\end{array}\right. \end{split}$$ We can now define $\psi:\triangle{\rightarrow}\triangle$ by letting $\psi=\phi{^{-1}}$ on $\phi[(1-2\eta)\triangle]$, $\psi={\rm id}$ on $\triangle\setminus (1-\eta)\triangle$, and, moreover, along the boundaries of $R_x$ construct $\psi$ as follows: $\psi[\alpha_x]$ is the straight line segment from $x$ to $\phi{^{-1}}(\hat x)$, $\psi[\gamma_x]=\gamma_x'$, and $\psi[\beta_x]=\beta_x$. Then $$\psi[{\mathcal}C_x]=\psi[\alpha_x\centerdot\gamma_x\centerdot\bar\alpha_{x^+}\centerdot\bar\beta_x]= \psi[\alpha_x]\centerdot\psi[\gamma_x]\centerdot\overline{\psi[\alpha_{x^+}]}\centerdot\overline{\psi[\beta_x]}= \psi[\alpha_x]\centerdot\gamma'_x\centerdot\overline{\psi[\alpha_{x^+}]}\centerdot\overline\beta_x$$ is the boundary of a region $K_x$ homeomorphic to the unit disk $D^2$ and hence, by Alexander’s Lemma, the homeomorphism $\psi$ from ${\mathcal}C_x=\alpha_x\centerdot\gamma_x\centerdot\bar\alpha_{\hat x}\centerdot\bar\beta_x$ to $\psi[\alpha_x]\centerdot\gamma'_x\centerdot\overline{\psi[\alpha_{x^+}]}\centerdot\overline\beta_x$ extends to the regions that they bound, i.e., to a homeomorphism of $R_x$ with $K_x$. This finishes the description of $\psi$ and it therefore only remains to see that $\sup_{x\in\triangle}d(x,\psi(x))<100\eta$. Since $\psi=\phi{^{-1}}$ on $\phi[(1-2\eta)\triangle]$ and $\psi={\rm id}$ on $\triangle\setminus(1-\eta)\triangle$ it is enough to consider what $\psi$ does to $x\in (1-\eta)\triangle\setminus\phi[(1-2\eta)\triangle]\subseteq \bigcup_{x\in {\mathbb F}}R_x$. Now, $\psi[R_x]=K_x$ for all $x\in {\mathbb F}$, and hence it is enough to show that no points in $R_x$ and in $K_x$ are more than $100\eta$ apart. But ${\rm diam}(R_x)<30\eta$ and ${\rm diam}(K_x)<40\eta$, while $R_x\cap K_x\neq {\emptyset}$, which gives the desired result. This finishes the proof. Patching along a triangulation of a compact $2$-manifold -------------------------------------------------------- As ${{\rm Homeo}(M)}$ is a separable complete metric group it is not covered by countably many nowhere dense sets (this is the Baire category theorem) and hence $W$ must be dense in some non-empty open set, whereby $W{^{-1}}W=W^2$ is dense in some neighbourhood of the identity in ${{\rm Homeo}(M)}$. So fix some $\eta_1>0$ such that $W^2$ is dense in $$V_{\eta_1}=\{g\in {{\rm Homeo}(M)}{ \; \big| \;}d_\infty(g,{\rm id})<\eta_1\}.$$ It is a well-known fact, first proved rigorously by Tibor Rado, that any compact $2$-manifold can be triangulated. So from now on, we assume that $M$ is a fixed compact $2$-manifold and we pick a triangulation $\{T_1,\ldots, T_m\}$ of $M$ with corresponding homeomorphisms $\chi_i:\triangle{\rightarrow}T_i$. By further triangulating each $T_i$, we can suppose that the diameter of $T_i$ is less than $\frac{\eta_1}{10}$ for all $i$. Moreover, by first modifying the $\chi_i$ along each edge of $\triangle$ and then extending to the interior of $\triangle$ by Alexander’s Lemma, we can suppose that the following holds. If $T_i=\chi_i[\triangle]$ and $T_j=\chi_j[\triangle]$ have an edge in common, then $\chi_i$ and $\chi_j$ agree along this edge, i.e., if $\chi_i(v_a)=\chi_j(v_\alpha)$ and $\chi_i(v_b)=\chi_j(v_\beta)$, then for all $t\in [0,1]$, $\chi_i(tv_a+(1-t)v_b)=\chi_j(tv_\alpha+(1-t)v_\beta)$. \[triangle-support\]For all $0<\eta<1$, if $h\in {\rm Homeo}(M)$ has support contained in $$\bigcup_{i=1}^m\chi_i[(1-\eta)\triangle],$$ then $h\in W^{20}$. Let $y_i=\chi_i(\vec 0)$ and choose ${\epsilon}>0$ such that $\overline B(y_i,{\epsilon})\subseteq \chi_i[(1-\eta)\triangle]$ for all $i\leq m$. By Lemma \[patched\], we can find some $0<\delta<{\epsilon}$ and $z_i\in B(y_i,{\epsilon})$ such that if $g\in {\rm Homeo}(M)$ has support contained in $\bigcup_{i=1}^m\overline B(z_i,\delta)$ then $g\in W^{16}$. As $W^2$ is dense in $V_{\eta_1}$, we can find an $f\in W^2$ such that for every $i\leq m$, $f[\chi_i[(1-\eta)\triangle]]\subseteq \overline B(z_i,\delta)$ and thus if $h$ is given as in the statement of the lemma, ${\rm supp}(fhf{^{-1}})=f[{\rm supp}(h)]\subseteq \bigcup_{i=1}^m\overline B(z_i,{\epsilon})$ and thus $g=fhf{^{-1}}\in W^{16}$, whence $h\in W^{20}$. \[fixing triangles\] Let $\delta,\eta>0$, $\eta<\frac1{1000}$ be such that for $i\leq m$ and $x,y\in \triangle$, $$d(x,y)<100\eta{\rightarrow}d(\chi_i(x),\chi_i(y))<\delta.$$ Then there is an $\alpha>0$ such that for all $g\in V_\alpha$ there is $\psi\in V_{\delta}\cap W^{20}$ whose support is contained in $\bigcup_{i=1}^m\chi_i[ {(1-\eta)\triangle}]$ and such that for all $i\leq m$, $$\psi\circ g|_{\chi_i[(1-2\eta)\triangle]}={\rm id}.$$ Fix $\delta$ and $\eta$ as in the Lemma. Then for any continuous $\phi:\triangle {\rightarrow}\triangle$ such that $\sup_{x\in \triangle} d(x,\phi(x))<100\eta$, we have for every $i\leq m$, $$\sup_{y\in T_i}d(y,\chi_i\circ\phi\circ\chi_i{^{-1}}(y))=\sup_{x\in \triangle}d(\chi_i(x),\chi_i\circ\phi(x))<\delta.$$ Now pick some $\alpha>0$ such that for $g\in V_\alpha$ and $i\leq m$, we have $$g\circ\chi_i[(1-2\eta)\triangle]\subseteq \chi_i[\triangle]=T_i,$$ whereby $\chi_i{^{-1}}\circ g\circ\chi_i:(1-2\eta)\triangle{\rightarrow}\triangle$, and such that $$\sup_{x\in (1-2\eta)\triangle}d(x,\chi_i{^{-1}}\circ g\circ\chi_i(x))<\frac\eta{100}.$$ By the quantitative annulus theorem we can therefore find some homeomorphism $\psi_i:\triangle{\rightarrow}\triangle$ that is the identity outside of $(1-\eta)\triangle$, satisfies $\sup_{x\in \triangle}d(x,\psi_i(x))<100\eta$, and $$\psi_i\circ\chi_i{^{-1}}\circ g\circ \chi_i|_{(1-2\eta)\triangle}={\rm id}.$$ This implies that for each $i\leq m$, $\chi_i\circ \psi_i\circ\chi_i{^{-1}}:T_i{\rightarrow}T_i$ is a homeomorphism that is the identity outside of $\chi_i[(1-\eta)\triangle]$, $\sup_{x\in T_i}d(x,\chi_i\circ \psi_i\circ\chi_i{^{-1}}(x))<\delta$, and $$\chi_i\circ \psi_i\circ\chi_i{^{-1}}\circ g|_{\chi_i[{(1-2\eta)\triangle}]}={\rm id}.$$ We can therefore define $\psi=\bigcup_{i=1}^m \chi_i\circ \psi_i\circ\chi_i{^{-1}}\in {\rm Homeo}(M)$ and notice that $\psi\in V_\delta$ and $\psi\circ g|_{\chi_i[ {(1-2\eta)\triangle}]}={\rm id}$ for every $i\leq m$. We see that $\psi$ has its support contained within the set $\bigcup_{i=1}^m\chi_i[ {(1-\eta)\triangle}]$ and thus, by Lemma \[triangle-support\], $\psi$ belongs to $W^{20}$. Fix some $0<\tau<\frac{1}{100}$. We now define the following set of points in $\triangle$: For distinct $i,j=0,1,2$, we put $w_{ij}=(1-10\tau)v_i+10\tau v_j$, $w_{ij}^+=(1-9\tau)v_i+9\tau v_j$, $u_{ij}=(1-\tau)w_{ij}$ and $u_{ij}^+=(1-\tau)w^+_{ij}$. So $w_{ij},w_{ij}^+\in \partial\triangle$, while $u_{ij},u_{ij}^+\in\partial(1-\tau)\triangle$. We also define a number of paths as follows: - $\alpha_{ij}$ is the straight line segment from $u_{ij}$ to $w_{ij}$. - $\beta_{ij}$ is the straight line segment from $w_{ij}$ to $w^+_{ij}$. - $\gamma_{ij}$ is the straight line segment from $u^+_{ij}$ to $w^+_{ij}$. - $\zeta_{ij}$ is the straight line segment from $u_{ij}$ to $u^+_{ij}$. - $\kappa_{ij}$ is the straight path from $w_{ij}$ to $w_{ji}$. - $\omega_{ij}$ is the straight path from $u_{ij}$ to $u_{ji}$. - $\xi_{0}$ is the shortest path in $\partial(1-\tau)\triangle$ from $u^+_{02}$ to $u^+_{01}$. - $\xi_{1}$ is the shortest path in $\partial(1-\tau)\triangle$ from $u^+_{10}$ to $u^+_{12}$. - $\xi_{2}$ is the shortest path in $\partial(1-\tau)\triangle$ from $u^+_{21}$ to $u^+_{20}$. - $\theta_{0}$ is the shortest path in $\partial\triangle$ from $w^+_{02}$ to $w^+_{01}$. - $\theta_{1}$ is the shortest path in $\partial\triangle$ from $w^+_{10}$ to $w^+_{12}$. - $\theta_{2}$ is the shortest path in $\partial\triangle$ from $w^+_{21}$ to $w^+_{20}$. We thus see that $${\mathcal}C_{ij}=\kappa_{ij}\centerdot\overline \alpha_{ji}\centerdot \omega_{ji}\centerdot\alpha_{ij}$$ is a simple closed curve bounding a closed region $R_{ij}=R_{ji}\subseteq \triangle$, $${\mathcal}C^+_{ij}=\overline\beta_{ij}\centerdot\kappa_{ij}\centerdot \beta_{ji}\centerdot\overline\gamma_{ji}\centerdot\overline\zeta_{ji}\centerdot\omega_{ji}\centerdot\zeta_{ij}\centerdot\gamma_{ij}$$ is a simple closed curve bounding a closed region $R^+_{ij}=R^+_{ji}\subseteq \triangle$ that contains $R_{ij}$. Notice however that the preceding definitions depend on the choice of $\tau$, which is therefore also the case for the following lemma. \[band-support\] If $\phi\in {\rm Homeo}(M)$ has support contained in $\bigcup_{l=1}^m\bigcup_{0\leq i<j\leq 2}\chi_l[R_{ij}^+]$, then $\phi\in W^{20}$. We notice that for distinct $l,l'$, $\chi_l[R^+_{ab}]\cap \chi_{l'}[R^+_{a'b'}]\neq {\emptyset}$ if and only if the triangles $T_l$ and $T_{l'}$ have the edge $\chi_l[\overline{v_av_b}]=\chi_{l'}[\overline{v_{a'}v_{b'}}]$ in common. Moreover, in this case, the set $\chi_l[R^+_{ab}]\cup \chi_{l'}[R^+_{a'b'}]$ is homeomorphic to the unit disk $D^2$ and is contained in an open set homeomorphic to ${\mathbb R}^2$. So let $A_1,\ldots,A_{\frac{3m}{2}}$ be an enumeration of all the closed sets $\chi_l[R^+_{ab}]\cup \chi_{l'}[R^+_{a'b'}]$ with $\chi_l[R^+_{ab}]$ and $\chi_{l'}[R^+_{a'b'}]$ overlapping and let $U_i\subseteq M$ be an open set containing $A_i$, homeomorphic to ${\mathbb R}^2$. We can suppose that the $U_i$ are all pairwise disjoint. Moreover, as the diameter of each $T_j$ is at most $\frac{\eta_1}{10}$, the diameter of each $A_i$ is at most $\frac{\eta_1}5$. The proof is now very much the same as the proof of Lemma \[triangle-support\]. Let $y_i\in A_i$ and choose $0<{\epsilon}<\frac{\eta_1}{5}$ such that $\overline B(y_i,{\epsilon})\subseteq U_i$ for all $i\leq m$. By Lemma \[patched\], we can find some $0<\delta<{\epsilon}$ and $z_i\in B(y_i,{\epsilon})$ such that if $g\in {\rm Homeo}(M)$ has support contained in $\bigcup_{i=1}^m\overline B(z_i,\delta)$ then $g\in W^{16}$. As $W^2$ is dense in $V_{\eta_1}$, we can find an $f\in W^2$ such that for every $i\leq \frac{3m}2$, $f[A_i]\subseteq \overline B(z_i,\delta)$ and thus if $\phi$ is given as in the statement of the lemma, $${\rm supp}(f\phi f{^{-1}})=f[{\rm supp}(\phi)]\subseteq \bigcup_{i=1}^m\overline B(z_i,{\epsilon}),$$ and thus $g=f\phi f{^{-1}}\in W^{16}$, whence $\phi\in W^{20}$. \[vertices left\] There is a $\nu>0$ such that if $g\in V_\nu$ and $g$ is the identity on $\bigcup_{i=1}^m\chi_i[ {(1-\tau)\triangle}]$, then there is a $\phi\in W^{20}$ such that $\phi\circ g$ is the identity on $$\bigcup_{i=1}^m\chi_i[ {(1-\tau)\triangle}]\cup \bigcup_{l=1}^m\;\bigcup_{0\leq i<j\leq 2}\chi_l[R_{ij}].$$ Consider the closed set $M_0=M\setminus{\rm Int}(\bigcup_{i=1}^m\chi_i[ {(1-\tau)\triangle}])$ and the closed subgroup $H=\{g\in {\rm Homeo}(M){ \; \big| \;}g|_{ \bigcup_{i=1}^m\chi_i[ {(1-\tau)\triangle}]}={\rm id}\}$. Assume that $T_l$ and $T_{l'}$ have an edge in common, i.e., $\chi_l(v_a)=\chi_{l'}(v_{a'})$ and $\chi_l(v_b)=\chi_{l'}(v_{b'})$ for some $a,a',b,b'$. Then $\chi_l[R_{ab}]\cup \chi_{l'}[R_{a'b'}]\subseteq {\rm Int}_{M_0}(\chi_l[R^+_{ab}]\cup \chi_{l'}[R^+_{a'b'}])$. Therefore, we can find some $\nu>0$, not depending on the particular choice of $l,l',a,a',b,b'$, such that for all such choices of $l,l',a,a',b,b'$ and $g\in V_\nu\cap H$ we have $$\begin{aligned} \label{g} g[\chi_l[R_{ab}]\cup \chi_{l'}[R_{a'b'}]]\subseteq {\rm Int}_{M_0}(\chi_l[R^+_{ab}]\cup \chi_{l'}[R^+_{a'b'}]).\end{aligned}$$ Fix some $g\in V_\nu\cap H$. Assume now that $\chi_l[\triangle]$ and $\chi_k[\triangle]$ have an edge in common. For concreteness we can suppose that, e.g., $\chi_l(v_0)=\chi_k(v_1)$ and $\chi_l(v_1)=\chi_k(v_2)$. As the covering mappings $\chi_i$ were supposed to agree along their edges, this implies that $\chi_l[\beta_{01}]=\chi_k[\beta_{12}]$, $\chi_l[\kappa_{01}]=\chi_k[\kappa_{12}]$, and $\chi_l[\beta_{10}]=\chi_k[\beta_{21}]$. Also, as $g\in H$, $g$ is the identity on the paths $\chi_l[\zeta_{01}],\chi_l[\omega_{01}],\chi_l[\zeta_{10}], \chi_k[\zeta_{12}],\chi_k[\omega_{12}]$ and $\chi_k[\zeta_{21}]$. By consequence, $\chi_l[\zeta_{01}]\centerdot \chi_l[\gamma_{01}]\centerdot\chi_k[\overline\gamma_{12}]\centerdot\chi_k[\overline\zeta_{12}]$ and $\chi_l[\alpha_{01}]\centerdot \chi_k[\overline\alpha_{12}]$ are paths from $\chi_l(u_{01})$ to $\chi_k(u_{12})$ only intersecting in their endpoints. Similarly, $\chi_l[\zeta_{10}]\centerdot \chi_l[\gamma_{10}]\centerdot\chi_k[\overline\gamma_{21}]\centerdot\chi_k[\overline\zeta_{21}]$ and $\chi_l[\alpha_{10}]\centerdot\chi_k[\overline\alpha_{21}]$ are paths from $\chi_l(u_{10})$ to $\chi_k(u_{21})$ only intersecting in their endpoints. This shows that $${\mathcal}K=\chi_l[\zeta_{01}]\centerdot\chi_l[\gamma_{01}]\centerdot\chi_k[\overline\gamma_{12}]\centerdot\chi_k[\overline\zeta_{12}]\centerdot \chi_k[\alpha_{12}]\centerdot\chi_l[\overline\alpha_{01}]$$ is a simple closed curve and thus, by the Schönflies Theorem, bounds a region $A$ homeomorphic to the unit disk $D^2$. Similarly, $${\mathcal}K'=\chi_l[\zeta_{10}]\centerdot\chi_l[\gamma_{10}]\centerdot\chi_k[\overline\gamma_{21}]\centerdot\chi_k[\overline\zeta_{21}]\centerdot \chi_k[\alpha_{21}]\centerdot\chi_l[\overline\alpha_{10}]$$ is a simple closed curve and thus bounds a region $A'$ homeomorphic to the unit disk $D^2$. Now, as $\chi_l[\alpha_{01}]\centerdot \chi_k[\overline\alpha_{12}]\subseteq \chi_l[R_{01}]\cup \chi_{k}[R_{12}]$, by condition \[g\] on $g$, $$g[\chi_l[\alpha_{01}]\centerdot \chi_k[\overline\alpha_{12}]]\subseteq{\rm Int_{M_0}}(\chi_l[R^+_{01}]\cup \chi_{k}[R^+_{12}])$$ and hence intersects $\chi_l[\zeta_{01}]\centerdot \chi_l[\gamma_{01}]\centerdot\chi_k[\overline\gamma_{12}]\centerdot\chi_k[\overline\zeta_{12}]$ only in their common endpoints. Thus, $${\mathcal}L=\chi_l[\zeta_{01}]\centerdot \chi_l[\gamma_{01}]\centerdot\chi_k[\overline\gamma_{12}]\centerdot\chi_k[\overline\zeta_{12}] \centerdot g[\chi_k[\alpha_{12}]]\centerdot g[\chi_l[\overline\alpha_{01}]]$$ is a simple closed curve bounding a region $B$ homeomorphic to $D^2$. Similarly, $${\mathcal}L'=\chi_l[\zeta_{10}]\centerdot \chi_l[\gamma_{10}]\centerdot\chi_k[\overline\gamma_{21}]\centerdot\chi_k[\overline\zeta_{21}] \centerdot g[\chi_k[\alpha_{21}]]\centerdot g[\chi_l[\overline\alpha_{10}]]$$ bounds a region $B'$ homeomorphic to $D^2$. We now have two decompositions of $\chi_l[R^+_{01}]\cup \chi_{k}[R^+_{12}]$. 1. $A\cup \big[\chi_l[R_{01}]\cup \chi_{k}[R_{12}]\big]\cup A'$. 2. $B\cup g[\chi_l[R_{01}]\cup \chi_{k}[R_{12}]]\cup B'$. Here $A$ and $\chi_l[R_{01}]\cup \chi_{k}[R_{12}]$ overlap along the edge $\chi_l[\alpha_{01}]\centerdot \chi_k[\overline\alpha_{12}]$, $\chi_l[R_{01}]\cup \chi_{k}[R_{12}]$ and $A'$ overlap along $\chi_l[\alpha_{10}]\centerdot \chi_k[\overline\alpha_{21}]$, while $A\cap A'={\emptyset}$. Similarly, $B$ and $g[\chi_l[R_{01}]\cup \chi_{k}[R_{12}]]$ overlap along the edge $g[\chi_l[\alpha_{01}]]\centerdot g[\chi_k[\overline\alpha_{12}]]$, $g[\chi_l[R_{01}]\cup \chi_{k}[R_{12}]]$ and $B'$ overlap along $g[\chi_l[\alpha_{10}]]\centerdot g[\chi_k[\overline\alpha_{21}]]$, while $B\cap B'={\emptyset}$. We can now define a homeomorphism $\varphi_{lk}: \chi_l[R^+_{01}]\cup \chi_k[R^+_{12}]{\rightarrow}\chi_l[R^+_{01}]\cup \chi_k[R^+_{12}]$, by first setting $\varphi_{lk}=g{^{-1}}$ on $g[\chi_l[R_{01}]\cup \chi_k[R_{12}]]$, and then let $\varphi_{lk}$ send $B$ to $A$, while fixing each point of $\chi_l[\zeta_{01}]\centerdot \chi_l[\gamma_{01}]\centerdot\chi_k[\overline\gamma_{12}]\centerdot\chi_k[\overline\zeta_{12}]$ and be $g{^{-1}}$ on $g[\chi_l[\alpha_{01}]\centerdot \chi_k[\overline\alpha_{12}]]$. Similarly for $B'$ and $A'$. This can be done for all pairs of $\chi_l$ and $\chi_k$ with a common edge, and we thus produce homeomorphisms $\varphi_{lk}$ on all of the regions, similar to $\chi_l[R^+_{01}]\cup \chi_k[R^+_{12}]$, that fix each point of the boundary curve $$\chi_l[\omega_{10}]\centerdot \chi_l[\zeta_{10}]\centerdot \chi_l[\gamma_{01}]\centerdot\chi_k[\overline\gamma_{12}]\centerdot\chi_k[\overline\zeta_{12}]\centerdot \chi_k[\omega_{12}]\centerdot \chi_k[\zeta_{21}]\centerdot\chi_k[\gamma_{21}]\centerdot \chi_l[\overline\gamma_{10}]\centerdot\chi_l[\overline\zeta_{10}].$$ Pasting all of these $\varphi_{lk}$ together and extending to all of $M$ by setting $\phi={\rm id}$ elsewhere, we obtain a homeomorphism $\phi\in {\rm Homeo}(M)$ whose support is contained in $\bigcup_{l=1}^m\;\bigcup_{0\leq i<j\leq 2}\chi_l[R^+_{ij}]$, while being the inverse of $g$ on $\bigcup_{l=1}^m\;\bigcup_{0\leq i<j\leq 2}\chi_l[R_{ij}]$. By Lemma \[band-support\], $\phi\in W^{20}$, which finishes the proof. We are now ready to finish the proof of the Theorem using the preceding sequence of lemmas. Let $y_1,\ldots,y_p\in M$ be the vertices of the triangulation and choose for each $i\leq p$ a neighbourhood $U_i$ of $y_i$ homeomorphic to ${\mathbb R}^2$. Find also $0<{\epsilon}<\eta_1$ such that $\overline B(y_i,{\epsilon})\subseteq U_i$ for all $i$. By Lemma \[patched\], there are $0<\delta_0<{\epsilon}$, $z_i\in B(y_i,{\epsilon})$, such that if $g\in{\rm Homeo}(M)$ has support contained in $\bigcup_{i=1}^p\overline B(z_i,\delta_0)$, then $g\in W^{16}$. As $y_i,z_i\in U_i{\simeq}{\mathbb R}^2$, we can, as $W^2$ is dense in $V_{\eta_1}$, find some $h_0\in W^2$ such that $h_0(y_i)\in U'_i\subseteq \overline B(z_i,\delta_0)$, where $U_i'$ is a neighbourhood of $z_i$ homeomorphic to ${\mathbb R}^2$. Therefore, there is some $g_0\in W^{16}$ such that $g_0h_0(y_i)=z_i$. This shows that if $f\in {\rm Homeo}(M)$ has support contained in $U=(g_0h_0){^{-1}}[\bigcup_{i=1}^p]$, then $(g_0h_0){^{-1}}f(g_0h_0)$ has support contained in $\bigcup_{i=1}^p B(z_i,\delta_0)$ and hence belongs to $W^{16}$. So $f$ belongs to $W^{52}$. We notice that $U$ is an open set containing $y_1,\ldots, y_p$. Recall now the definition of the paths $\alpha_{ij}, \beta_{ij}$, etc. and also the fact that these paths all depend on the choice of $0<\tau<1$. For a fixed choice of $\tau$, we define the following simple closed curves in $\triangle$ $$\begin{split} {\mathcal}F_0^\tau= \beta_{02}\centerdot\theta_{0}\centerdot\overline\beta_{01}\centerdot\overline\alpha_{01}\centerdot\zeta_{01}\centerdot\overline\xi_{0}\centerdot\overline\zeta_{02}\centerdot\alpha_{02},\\ {\mathcal}F^\tau_1=\beta_{10}\centerdot\theta_{1}\centerdot\overline\beta_{12}\centerdot\overline\alpha_{12}\centerdot\zeta_{12}\centerdot\overline\xi_{1}\centerdot\overline\zeta_{10}\centerdot\alpha_{10},\\ {\mathcal}F_2^\tau=\beta_{21}\centerdot\theta_{2}\centerdot\overline\beta_{20}\centerdot\overline\alpha_{20}\centerdot\zeta_{20}\centerdot\overline\xi_{2}\centerdot\overline\zeta_{21}\centerdot\alpha_{21}.\\ \end{split}$$ Moreover, we let $F_0^\tau, F_1^\tau, F^\tau_2$ be the closed regions that they enclose. We notice that $F_i^\tau$ converges in the Vietoris topology to $\{v_i\}$ when $\tau{\rightarrow}0$, and thus for some $\tau>0$, we have for all $i=0,1,2$ and $l=1,\ldots, m$, $\chi_l[F^\tau_i]\subseteq U$. So fix this $\tau$ and denote $F^\tau_i$ by $F_i$. We notice that $$\triangle=(1-\tau)\triangle\cup \bigcup_{0\leq i<j\leq 2}R_{ij} \cup \bigcup_{i=0,1,2}F_i.$$ By consequence, if $f\in {\rm Homeo}(M)$ is the identity on $$\bigcup_{i=1}^m\chi_i[ {(1-\tau)\triangle}]\cup \bigcup_{l=1}^m\;\bigcup_{0\leq i<j\leq 2}\chi_l[R_{ij}],$$ then $f$ has support contained in $\bigcup_{l=1}^m\; \bigcup_{i=0,1,2}\chi_l[F_i]\subseteq U$, and hence $f\in W^{52}$. Find now a $\nu>0$ as in the statement of Lemma \[vertices left\]. Then if $g\in V_\nu$ and $g$ is the identity on $\bigcup_{i=1}^m\chi_i[ {(1-\tau)\triangle}]$, then there is a $\phi\in W^{20}$ such that $\phi\circ g$ is the identity on $$\bigcup_{i=1}^m\chi_i[ {(1-\tau)\triangle}]\cup \bigcup_{l=1}^m\;\bigcup_{0\leq i<j\leq 2}\chi_l[R_{ij}],$$ and hence belongs to $W^{52}$. But then also $g\in W^{72}$. Fix $\delta<\frac\nu2$ and find an $\eta>0$ satisfying $\eta<\frac1{1000}$, $\eta<\frac\nu2$, and such that for $i\leq m$ and $x,y\in \triangle$, $$d(x,y)<100\eta{\rightarrow}d(\chi_i(x),\chi_i(y))<\delta.$$ By Lemma \[fixing triangles\], we can find an $0<\alpha<\frac\nu2$ such that for all $h\in V_\alpha$ there is $\psi\in V_\delta\cap W^{20}$ such that for all $i\leq m$, $$\psi\circ h|_{\chi_i[(1-2\eta)\triangle]}={\rm id}.$$ In particular, $\psi\circ h\in V_{\delta}V_{\alpha}\subseteq V_{\delta+\alpha}\subseteq V_\nu$ and is the identity on $\bigcup_{i=1}^m\chi_i[ {(1-\tau)\triangle}]$, whereby $\psi\circ h\in W^{72}$ and $h\in W^{92}$. This shows that $V_\alpha\subseteq W^{92}$ and thus $W^{92}$ contains an open neighbourhood of the identity in ${\rm Homeo}(M)$ and hence we have proved that ${\rm Homeo}(M)$ is Steinhaus, which finishes the proof of the Theorem. Department of mathematics,\ University of Illinois at Urbana-Champaign,\ 273 Altgeld Hall, MC 382,\ 1409 W. Green Street,\ Urbana, IL 61801,\ USA.\ `rosendal@math.uiuc.edu` [99]{} Mather, John N.: The vanishing of the homology of certain groups of homeomorphisms. Topology 10 1971 297–298. Rybicki, Tomasz: Commutators of homeomorphisms of a manifold. Univ. Iagel. Acta Math. No. 33 (1996), 153–160. Kechris, A.S.; Rosendal, C.: Turbulence, amalgamation, and generic automorphisms of homogeneous structures. To appear in Proc. London Math. Soc. Guran, I.I.: Topological groups similar to Lindelöf groups. (Russian) Dokl. Akad. Nauk SSSR 256 (1981), no. 6, 1305–1307. Uspenskiĭ, V. V.: A universal topological group with a countable basis. (Russian) Funktsional. Anal. i Prilozhen. 20 (1986), no. 2, 86–87. Rosendal, C.; Solecki, S.: Automatic continuity of group homomorphisms and discrete groups with the fixed point on metric compacta property. To appear in Israel Journal of Math.

ArXiv

\#1 ”) \] Introduction ============ The search for the mechanism which causes superconductivity in the copper oxide materials is an ongoing effort which has yet to reach a consensus. One factor which any model should account for is that the critical temperature tends to be higher in systems with multiple adjacent CuO$_2$ layers; and even in systems, such as YBCO, in which a CuO layer is adjacent to the CuO$_2$ layers $T_{\rm c}$ seems to be enhanced. It is generally believed that the superconducting condensate resides in the CuO$_2$ planes, although one interpretation of the observed large $x$-$y$ anisotropy of the zero temperature magnetic penetration depth (a factor of $\sim1.6$) in YBCO indicates that there is as much condensate in the CuO chains as in the CuO$_2$ planes (ie, the condensate in the chains only contributes to the penetration depth for current in the direction parallel to the chains, ie the $b$-direction).[@basov] Since it is believed that whatever mechanism is responsible for superconductivity in the copper oxides is intrinsic to the CuO$_2$ planes some other mechanism for creating superconducting condensate on the CuO chains is required. In this paper we derive a Hamiltonian for a layered system and, making a simplifying assumption that there is no pairing between electrons which reside in different layers, derive a pair of coupled [bcs]{} equations for a system of two layers, each with possibly different dispersion and pairing interactions. In this model Cooper pairs can scatter between the layers so that, as in a two band model,[@mmp] even if there is no pairing interaction in one of the layers there will still be a condensate in that layer due to the interlayer interaction. Although we make a particular choice for the pairing interaction (which is motivated by the nearly antiferromagnetic Fermi liquid model which leads naturally to a $d$-wave gap for single CuO$_2$ planes.[@chi]) a $d$-wave solution is also found for other types of non-isotropic pairing interactions.[@odonovan1] One result of having gap nodes cross the Fermi surface is a low temperature linear behaviour of the magnetic penetration depth, $\lambda_{ii}^{-2}$, as is observed in YBCO.[@basov; @bonn; @hardy2; @bonn2] Here we only try to model the low temperature behaviour and relative magnitude[@basov] in the $x$ and $y$ directions of the magnetic penetration depth. The $T\sim T_{\rm c}$ behaviour is only reproduced for values of $2\mit\Delta_{\rm max}/T_{\rm c}$ about $1.5$ times higher[@scalapino] than the value of 4.4 found in the [bcs]{} weak coupling approximation. In the two layer model that we study here a higher value of $2\mit\Delta_{\rm max}/T_{\rm c}$ is obtained which is closer to that observed in experiments such as [arpes]{}[@ding] or current-imaging tunneling spectroscopy ([cits]{})[@edwards] that measure the absolute magnitude of the order parameter. We find that the presence of the chains destroys the tetragonal symmetry of the CuO$_2$ planes and shifts the $d$-wave gap nodes in the CuO$_2$ plane off the diagonals in agreement with an earlier model.[@odonovan3] In this case the gap contains an admixture of $s$- and $d$-wave symmetry. Calculations of the $c$-axis Josephson tunneling current show that the positive and negative parts of the order parameter do not cancel, as for $d$-wave pairing in a tetragonal system, and that the Josephson junction resistance-tunneling current products, $RJ(T=0)$, are in the range of 0.1-3.0meV, in agreement with the experiments of Sun [*et al*]{},[@sun]. We also calculate both the normal and superconducting density of states ([dos]{}) for the CuO$_2$ planes and CuO chains separately since some surface probes, such as [cits]{},[@edwards] can measure them separately. Finally the Knight shift is calculated separately for both the planes and the chains. In section II we introduce our Hamiltonian and derive a set of coupled [bcs]{} equations for planes and chains and other necessary formulas, particularly the expression for the magnetic penetration depth in this model. In section III we present the the solutions of these [bcs]{} equations as well as the results of calculations of the magnetic penetration depth, electronic density of states, $c$-axis Josephson tunneling, and Knight shift. Section IV contains a short discussion and conclusion. Formalism ========= In this section we will present a Hamiltonian in which multiple layers are coupled through the pairing interactions between adjacent layers. We will then make the assumption that there is no interlayer pairs (ie, that each Cooper pair resides in only one of the layers) and that there is no single particle interlayer hopping. The Hamiltonian of this special case is then diagonalized and coupled [bcs]{} equations derived. We then give expressions for the magnetic penetration depth in this model, the Knight shift, the quasi-particle density of states and the DC Josephson junction resistance-tunneling current product for a $c$-axis tunnel junction. The general Hamiltonian is: $$\begin{aligned} H =&&\sum_{{\bf k},\alpha\beta}{ \varepsilon_{{\bf k},\alpha\beta} \left(a_{{\bf k}\uparrow,\alpha}^\dagger a_{{\bf k}\uparrow,\beta} +a_{{\bf k}\downarrow,\beta}^\dagger a_{{\bf k}\downarrow,\alpha}\right) } \nonumber \\ &-&\frac{1}{\Omega} \sum_{{\bf k},{\bf q},\alpha\beta\gamma\delta}{ V_{{\bf k},{\bf q},\alpha\beta\gamma\delta}a_{{\bf k}\uparrow,\alpha}^\dagger a_{-{\bf k}\downarrow,\beta}^\dagger a_{{\bf q}\uparrow,\gamma} a_{-{\bf q}\downarrow,\delta} },%\nonumber \label{hamilton.eq}\end{aligned}$$ where the greek indices enumerate the layers, the $a_{{\bf k},\alpha}^\dagger$ ($a_{{\bf k},\alpha}$) create (destroy) electrons of momentum ${\bf k}$ in layer $\alpha$ ([**k**]{} is in units of $a^{-1}$ where $a$ is the lattice parameter), $\varepsilon_{{\bf k},\alpha\beta}$ is the electron dispersion, and $V_{{\bf k},{\bf q},\alpha\beta\gamma\delta}$ is the pairing interaction. Performing a mean field analysis, we get: $$\begin{aligned} H =&&\sum_{{\bf k},\alpha\beta}{\varepsilon_{{\bf k},\alpha\beta} \left(a_{{\bf k}\uparrow,\alpha}^\dagger a_{{\bf k}\uparrow,\beta} +a_{{\bf k}\downarrow,\beta}^\dagger a_{{\bf k}\downarrow,\alpha}\right) } \nonumber \\ &-&\sum_{{\bf k},\alpha\beta}{\left(\mit\Delta_{{\bf k},\alpha\beta}a_{{\bf k}\uparrow,\alpha}^\dagger a_{-{\bf k}\downarrow,\beta}^\dagger+ {\rm H.c.}\right)}+C,\end{aligned}$$ where $C$ is a constant, H.c. indicates the Hermitian conjugate, $\mit\Delta_{{\bf k},\alpha\beta}\equiv \Omega^{-1}\sum_{{\bf q},\gamma\delta}{V_{{\bf k},{\bf q},\alpha\beta\gamma\delta}\chi_{{\bf q},\gamma\delta}}$ are the order parameters and $\chi_{{\bf q},\gamma\delta}\equiv \left<a_{{\bf q}\uparrow,\gamma} a_{-{\bf q}\downarrow,\delta}\right>$ are the pair susceptibilities. Writing this in Nambu spinor notation, we get: $$\begin{aligned} H=\sum_{{\bf k},\alpha\beta}A_{{\bf k},\alpha}^\dagger \hat H_{{\bf k},\alpha\beta}A_{{\bf k},\beta}\end{aligned}$$ where $A_{{\bf k},\alpha}^\dagger\equiv \left[\begin{array}{cc} a_{{\bf k}\uparrow,\alpha}^\dagger & a_{-{\bf k}\downarrow,\alpha} \end{array}\right]$ and: $$\hat H_{{\bf k},\alpha\beta}\equiv\left[ \begin{array}{cc} \varepsilon_{{\bf k},\alpha\beta} & \mit\Delta_{{\bf k},\alpha\beta} \\ \mit\Delta_{{\bf k},\alpha\beta}^\dagger & -\varepsilon_{{\bf k},\alpha\beta} \end{array}\right]. \label{hamiltonian2.eq}$$ In this model the magnetic penetration depth is given by the expression:[@atkinson] $$\begin{aligned} \label{pd1.eq} \lambda_{ij}^{-2}&&=\frac{4\pi e^2}{\hbar^2c^2}\frac{1}{\Omega}\sum_{{\bf k},\alpha\beta} \hat\gamma^{(i)}_{{\bf k},\alpha\beta}\hat\gamma^{(j)}_{{\bf k},\beta\alpha} \left(\hat G_{{\bf k},\alpha\beta}\big|_{\mit\Delta=0}-\hat G_{{\bf k},\alpha\beta}\right)\end{aligned}$$ where: $$\begin{aligned} \hat G_{{\bf k},\alpha\beta}\equiv&&\frac{\partial f(E_{{\bf k},\alpha})}{\partial E_{{\bf k},\alpha}}\delta_{\alpha\beta} \nonumber\\ &+&\frac{f(E_{{\bf k},\alpha})-f(E_{\bf k,{\beta}})}{E_{{\bf k},\alpha}-E_{\bf k,{\beta}}}(1-\delta_{\alpha\beta}),\nonumber\\ \hat\gamma^{(i)}_{{\bf k},\alpha\beta}\equiv&&\sum_{\gamma\delta}\hat U_{{\bf k},\alpha\gamma}^\dagger\frac{\partial \varepsilon_{{\bf k},\gamma\delta}} {\partial k_i}\hat U_{{\bf k},\delta\beta},\nonumber\end{aligned}$$ $\pm E_{{\bf k},\alpha}$ are the eigenvalues of Eq. (\[hamiltonian2.eq\]), $f(x)$ is the Fermi function, $\delta_{\alpha\beta}$ is a Kronecker delta and $\hat U_{{\bf k},\alpha\beta}$ is the unitary matrix which diagonalizes Eq. (\[hamiltonian2.eq\]). The quantities $e$, $\hbar$, and $c$ are the electron charge, Planck’s constant and the speed of light respectively. We now make the following simplification: $V_{{\bf k},{\bf q},\alpha\gamma} \equiv V_{{\bf k},{\bf q},\alpha\alpha\gamma\gamma} = V_{{\bf k},{\bf q},\alpha\beta\gamma\delta}\delta_{\alpha\beta}\delta_{\gamma\delta}$ and $\varepsilon_{{\bf k},\alpha}\equiv\varepsilon_{{\bf k},\alpha\alpha}=\varepsilon_{{\bf k},\alpha\beta}\delta_{\alpha\beta}$. This means that there is only intralayer pairing and no interlayer pairing (ie, $\mit\Delta_{{\bf k},\alpha}\equiv\mit\Delta_{{\bf k},\alpha\beta}\delta_{\alpha\beta}$ and $\chi_{{\bf q},\gamma}\equiv\chi_{{\bf q},\gamma\delta}\delta_{\alpha\beta}$ are both diagonal in the greek indices) and there is no single particle interlayer hopping. This Hamiltonian has the same form as that for a two band model studied by one of us [@chi] in an earlier publication and is similar to that studied by others.[@klemm; @klemm2; @combescot] Interlayer pairing[@combescot; @kettemann; @wheatley] has also been studied. The Hamiltonian has eigenvalues given by $E_{{\bf k},\alpha}=\pm\sqrt{ \varepsilon_{{\bf k},\alpha}^2+\mit\Delta_{{\bf k},\alpha}^2}$ and is diagonalized by the unitary matrix: $$\label{unitary.eq} \hat U_{{\bf k},\alpha}=\left[ \begin{array}{cc} u_{{\bf k},\alpha} & v_{{\bf k},\alpha} \\ -v_{{\bf k},\alpha} & u_{{\bf k},\alpha} \end{array}\right],$$ where: $$\begin{aligned} u_{{\bf k},\alpha}&\equiv&\sqrt{\frac{1}{2} \left(1+\frac{\varepsilon_{{\bf k},\alpha}}{E_{{\bf k},\alpha}}\right)} \nonumber\\ v_{{\bf k},\alpha}&\equiv&\sqrt{\frac{1}{2} \left(1-\frac{\varepsilon_{{\bf k},\alpha}}{E_{{\bf k},\alpha}}\right)}\nonumber\end{aligned}$$ are the usual [bcs]{} coherence factors. Using this unitary transformation (\[unitary.eq\]) we can evaluate the pair susceptibilities to get: $$\begin{aligned} \label{sus.eq} \chi_{{\bf q},\alpha}&\equiv& \left<a_{{\bf q}\uparrow,\alpha} a_{-{\bf q}\downarrow,\alpha}\right> \nonumber \\ &=&\frac{\mit\Delta_{{\bf q},\alpha}}{2E_{{\bf q},\alpha}} \tanh \left( \frac{E_{{\bf q},\alpha}}{2 k_{\rm B}T} \right),\end{aligned}$$ where $T$ is the temperature and $k_{\rm B}$ is Boltzmann’s constant. Note that if we had included the interlayer pairing from Eq. \[hamilton.eq\] we would have susceptibilities of the form $\left<a_{{\bf q}\uparrow,\alpha} a_{-{\bf q}\downarrow,\beta}\right>$ with $\alpha\not =\beta$ and both the eigenvalues and the unitary matrix (\[unitary.eq\]) would be much more complicated. For a bilayer system (ie, $\alpha=1,2$) the [bcs]{} equations are: $$\begin{aligned} \label{bcs.eq} \mit\Delta_{{\bf k},1}&=& \frac{1}{\Omega}\sum_{{\bf q}}{\left( V_{{\bf k},{\bf q},11}\chi_{{\bf q},1} +V_{{\bf k},{\bf q},12}\chi_{{\bf q},2} \right)} \nonumber \\ \mit\Delta_{{\bf k},2}&=& \frac{1}{\Omega}\sum_{{\bf q}}{\left( V_{{\bf k},{\bf q},12}\chi_{{\bf q},1} + V_{{\bf k},{\bf q},22}\chi_{{\bf q},2} \right)},\end{aligned}$$ where we have taken $V_{{\bf k},{\bf q},12}=V_{{\bf k},{\bf q},21}$, although in general only $V_{{\bf k},{\bf q},12}=V_{{\bf k},{\bf q},21}^\dagger$ is required. Noting that $\chi_{{\bf q},2}$ changes sign (see Eq. \[sus.eq\]) with $\mit\Delta_{{\bf k},2}$ we see that this set of equations (\[bcs.eq\]) is unchanged by the substitution $\{\mit\Delta_{{\bf k},2},V_{{\bf k},{\bf q},12}\}\rightarrow \{-\mit\Delta_{{\bf k},2},-V_{{\bf k},{\bf q},12}\}$ which means that the overall sign of $V_{{\bf k},{\bf q},12}$ only affects the relative sign of the order parameters in the two layers and not their magnitudes. This is interesting because it means that the effect on $T_{\rm c}$ of having an interlayer interaction is independent of whether this interaction is attractive or repulsive, although some calculated properties (eg, $c$-axis Josephson tunneling current) still depend upon the relative sign of the interlayer interaction. [*It is important to emphasize that any interlayer interaction, either attractive or repulsive, tends to enhance $T_{\rm c}$ and that this is consistent with the observation that $T_{\rm c}$ is higher in materials with multiple adjacent CuO layers.*]{} This well known result can be easily shown by examining the coupled [bcs]{} equations (\[bcs.eq\]) near $T\sim T_c$. In this limit we can write: $$\begin{aligned} {\mit\Delta}_{{\bf k},\alpha} &=& {\mit\Delta}_\alpha \eta_{\bf k}\\ V_{{\bf k},{\bf q},\alpha\beta} &=& V_{\alpha\beta} \eta_{\bf k}\eta_{\bf q}\end{aligned}$$ where $\mit\Delta_\alpha$ and $V_{\alpha\beta}$ are numbers and $\eta_{\bf k}$ is a normalized function which could be taken to be $d$-wave and corresponds to the highest $T_c$. The coupled [bcs]{} equations (6) can then be written as: $$\begin{aligned} {\mit\Delta}_1 &=& {\mit\Delta}_1 V_{11}\frac{1}{\Omega}\sum_{{\bf q}}{ \frac{\left(\eta_{\bf q}\right)^2}{\varepsilon_{{\bf q},1}} \tanh \left( \frac{\varepsilon_{{\bf q},1}}{2 k_{\rm B}T_c} \right)}\nonumber\\ &&+{\mit\Delta}_2 V_{12}\frac{1}{\Omega}\sum_{{\bf q}}{ \frac{\left(\eta_{\bf q}\right)^2}{\varepsilon_{{\bf q},2}} \tanh \left( \frac{\varepsilon_{{\bf q},2}}{2 k_{\rm B}T_c} \right)}\\ {\mit\Delta}_2 &=& {\mit\Delta}_1 V_{12}\frac{1}{\Omega}\sum_{{\bf q}}{ \frac{\left(\eta_{\bf q}\right)^2}{\varepsilon_{{\bf q},1}} \tanh \left( \frac{\varepsilon_{{\bf q},1}}{2 k_{\rm B}T_c} \right)}\nonumber\\ &&+{\mit\Delta}_2 V_{22}\frac{1}{\Omega}\sum_{{\bf q}}{ \frac{\left(\eta_{\bf q}\right)^2}{\varepsilon_{{\bf q},2}} \tanh \left( \frac{\varepsilon_{{\bf q},2}}{2 k_{\rm B}T_c} \right)}.\end{aligned}$$ We now assume that $\mit\Delta_1$ is the dominant superconducting channel when $V_{12}=V_{21}=0$ and obtain assuming an infinite band with cutoff $\omega_C$: $$\begin{aligned} \mit\Delta_1 &=& \left(\mit\Delta_1\lambda_{11}+\mit\Delta_2\lambda_{12}\right)\ln\left(\frac{1.13\omega_C}{T_c}\right) \label{silliness1}\\ \mit\Delta_2 &=& \left(\mit\Delta_1\lambda_{21}+\mit\Delta_2\lambda_{22}\right)\ln\left(\frac{1.13\omega_C}{T_c}\right), \label{silliness2}\end{aligned}$$ where $\lambda_{ij}\equiv V_{ij}\times$ the density of electronic states at the Fermi surface. Substitution of equation (\[silliness2\]) into (\[silliness1\]) leads to a quadratic in $\ln\left(1.13\omega_C/T_c\right)$ with solution: $$T_c=1.13 e^{1/\tilde{\lambda}}$$ with: $$\tilde{\lambda}=\frac{1}{2}\left[\lambda_{11}+\lambda_{22}+\sqrt{(\lambda_{11}-\lambda_{22})^2+4\lambda_{12}\lambda_{21}}\right].$$ This result is well known and is given by equation (6.3) on page 105 of Mechanisms of Conventional and High $T_c$ Superconductivity.[@kresin] It is also found as equation (40) of H. Chi and Carbotte.[@chi] We note that $\lambda_{12}$, whatever its sign, increases $\tilde{\lambda}$ and so increases $T_c$. If we had taken $V_{{\bf k},{\bf q},12}$ as complex the symmetry would be $\{\mit\Delta_{{\bf k},2},V_{{\bf k},{\bf q},12}\}\rightarrow \{\mit\Delta_{{\bf k},2}e^{-\imath \phi},V_{{\bf k},{\bf q},12}e^{\imath \phi}\}$ where $V_{{\bf k},{\bf q},12}=|V_{{\bf k},{\bf q},12}|e^{\imath \phi}$, and the relative phase between the layers would no longer be $\pm 1$. We note that by performing the unitary transformation $\hat S^\dagger \hat H \hat S$ where: $$\begin{aligned} \hat S\equiv\frac{1}{\sqrt{2}}\left[\begin{array}{rrrr} 1&1&\ \ 0&0\\ 1&-1&0&0\\ 0&0&1&1\\ 0&0&1&-1 \end{array} \right],\end{aligned}$$ and making the substitutions $V_{11}=V_{22}= V_{\parallel}+V_{\perp}$, $V_{12}= V_{\parallel}-V_{\perp}$, $\varepsilon_{1}= \varepsilon+t$ and $\varepsilon_{2}= \varepsilon-t$ we obtain both the Hamiltonian and [bcs]{} equations used by Liu [*et al*]{}.[@liu] Our work differs from theirs in that we allow both the dispersion and the interaction to be different in the two layers. This is important not only because we are able to model systems such as YBCO in which there are CuO$_2$ planes and CuO chains, but also because the order parameter in each of the layers may differ in sign even in two identical layers.[@combescot] Other workers have studied models in which the electrons in the pairs reside in different layers[@kettemann] (ie, in which only $\chi_{{\bf k},12}$ is non-zero) as well as models in which there is no intralayer interaction[@combescot; @wheatley] (ie, in which only $V_{{\bf k,q},12}$ is non-zero). After solving the set of coupled [bcs]{} equations (\[bcs.eq\]) at $T=0$ using a [fft]{} technique[@odonovan1; @odonovan3; @odonovan2] we approximate the order parameters, $\mit\Delta_{{\bf k},1}$ and $\mit\Delta_{{\bf k},2}$, with: $$\begin{aligned} \mit\Delta_{{\bf k},\alpha} =&&\left(\mit\Delta_\alpha^{(s_\circ)} \eta_{\bf k}^{(s_\circ)} +\mit\Delta_\alpha^{(s_x)}\eta_{\bf k}^{(s_x)}+\mit\Delta_\alpha^{(d)}\eta_{\bf k}^{(d)}\right) \nonumber\\ &&\times \tanh \left(1.74\sqrt{T_{\rm c}/T-1}\right), \label{harmonics.eq}\end{aligned}$$ where the $\eta_{\bf k}^{(\cdot)}$ are the three lowest harmonics given by: $$\begin{aligned} \eta_{\bf k}^{(s_\circ)}&=& 1\nonumber \\ \eta_{\bf k}^{(s_x)}&=& \cos(k_x)+\cos(k_y)\nonumber \\ \eta_{\bf k}^{(d)}&=& \cos(k_x)-\cos(k_y),\end{aligned}$$ and the $\mit\Delta_\alpha^{(\cdot)}$ are their amplitudes. The amplitudes of the higher harmonics are all very much smaller in magnitude and the gap nodes and the maximum magnitude of the gap, which are the most important features of the order parameter, are essentially unchanged by this approximation. We also calculate the magnetic penetration depth which in this system, since the $\hat\gamma^{(i)}_{{\bf k},\alpha\beta}$ are diagonal in the greek indices, is given by the simplified expression: $$\begin{aligned} \label{pd2.eq} \lambda_{ij}^{-2}=&&\frac{4\pi e^2}{\hbar^2c^2}\frac{1}{\Omega}\sum_{{\bf k},\alpha} \frac{\partial \varepsilon_{{\bf k},\alpha}}{\partial k_i} \frac{\partial \varepsilon_{{\bf k},\alpha}}{\partial k_j} \nonumber\\&&\times \left( \frac{\partial f(\varepsilon_{{\bf k},\alpha})}{\partial \varepsilon_{{\bf k},\alpha}} -\frac{\partial f(E_{{\bf k},\alpha})}{\partial E_{{\bf k},\alpha}} \right)\end{aligned}$$ which is the usual expression[@odonovan2] summed over the layers. The curvature of the penetration depth curve, $\lambda_{ii}^{-2}(T)$, (and also its low temperature slope) is governed by the ratio $2\mit\Delta_{\rm max}/T_{\rm c}$, where $\mit\Delta_{\rm max}$ is the maximum value of the order parameter in the first Brillouin zone, and is close to a straight line for the $d$-wave [bcs]{} value of $2\mit\Delta_{\rm max}/T_{\rm c}=$4.4. The presence of the interlayer interaction increases this ratio and makes the $\lambda_{ii}^{-2}(T)$ curve have a downward curvature. Experimental measurements of both the ratio $2\mit\Delta_{\rm max}/T_{\rm c}$[@edwards; @maggio] as well as the penetration depth in high quality crystals of both YBCO[@basov] and BSCO[@jacobs] indicate that this ratio is quite high in the [htc]{} materials – on the order of 6 or 7. Other quantities calculated are the Knight shift which is given by: $$K(T)\propto \frac{1}{\Omega}\sum_{\bf k} \frac{\partial f(E_{\bf k})}{\partial E_{\bf k}}, \label{ks.eq}$$ the normal state electronic [dos]{} which is given by: $$\begin{aligned} N(\omega)&=&\frac{1}{\Omega}\sum_{\bf k}\delta (\varepsilon_{\bf k}-\omega) \nonumber\\ &=&\lim_{\mit\Gamma\to 0}\frac{1}{\pi\Omega}\sum_{\bf k}\frac{\mit\Gamma}{(\varepsilon_{\bf k}-\omega)^2+\mit\Gamma^2}, \label{dos.eq}\end{aligned}$$ the superconducting electronic [dos]{} which is given by: $$\begin{aligned} N(\omega)&=&\frac{1}{\Omega}\sum_{\bf k}\delta (E_{\bf k}-\omega) \nonumber\\ &=&\lim_{\mit\Gamma\to 0}\frac{1}{\pi\Omega}\sum_{\bf k}\frac{\mit\Gamma}{(E_{\bf k}-\omega)^2+\mit\Gamma^2}, \label{scdos.eq}\end{aligned}$$ and the $c$-axis Josephson junction resistance-tunneling current product, $RJ(T)$, through a superconductor-insulator-superconductor junction for incoherent $c$-axis tunneling is given by the relation [@ambegaokar]: $$\begin{aligned} \label{joe.eq} RJ(T)=&&\frac{2\pi T}{N^L(0) N^R(0) \pi^2}\sum_n { A^L(\omega_n)A^R(\omega_n)},\end{aligned}$$ where: $$\begin{aligned} A^{L(R)}(\omega_n)\equiv\frac{1}{\Omega}\sum_{\bf k} \frac{\mit\Delta^{L(R)}_{\bf k}}{(\varepsilon^{L(R)}_{\bf k})^2+(\mit\Delta^{L(R)}_{\bf k})^2+(\omega_n )^2}.\end{aligned}$$ in which the superscript L(R) indicates on which side of the junction the dispersion and order parameter are on, the sum over $\omega_n\equiv\pi T(2n-1)$ is for all Matsubara frequencies, $R$ is the resistance of the junction and $N^{L(R)}(0)$ is the normal state electronic [dos]{} given by equation (\[dos.eq\]). If the tunneling were coherent the matrix element (which is incorporated into $R$) would have a $({\bf k-k^\prime})$ dependence, and the sums over ${\bf k}$-space wouldn’t be separable. Results ======= In this section we make an explicit choice for the dispersions and interactions and then present the results of our numerical solutions to the coupled [bcs]{} equations (\[bcs.eq\]) as well as the results of our calculations of the magnetic penetration depth, densities of states, Knight shift and Josephson current. As we wish to model YBCO we will want to account for both the CuO$_2$ planes as well as the CuO chains. Further, we will assume that we do not have a pairing interaction in the chains, but only in the planes as well as an interlayer interaction. This means that all of the order parameter in the chains is due to the interlayer interaction. We note that although our solution technique[@odonovan1] allows the order parameters to be complex and to have a relative phase between layers we find that in the models studied here, to within an overall phase, the order parameters are all real with a relative phase of $\pm 1$. For the dispersions, $\varepsilon_{{\bf k},\alpha}$, we use: $$\begin{aligned} \label{disp.eq} \varepsilon_{{\bf k},\alpha}= &-&2t_\alpha\left[(1+\epsilon_\alpha)\cos(k_x)+\cos(k_y)\right.\nonumber\\ &-&\left.2B_\alpha\cos(k_x)\cos(k_y)-(2-2B_\alpha-\mu_\alpha)\right],\end{aligned}$$ where the parameters $\{t_\alpha, \epsilon_\alpha, B_\alpha, \mu_\alpha\}$ are chosen so that the Fermi surface and bandwidth are close approximations to those observed experimentally [@gofron] (see Fig. \[fermi.fig\]). In order to model YBCO we chose $\{100,0,0.45,0.51\}$ for the planes and $\{-50,-0.9,0,1.2\}$ for the chains. In both dispersions $t_\alpha$, which sets the overall energy scale, is in units of meV. For the interactions, $V_{{\bf k},{\bf q},\alpha\beta}$, we chose an [mmp]{} form:[@mmp] $$\begin{aligned} V_{{\bf k},{\bf q},\alpha\beta}=g_{\alpha\beta} \frac{-\chi_\circ}{1+\xi_\circ^2|{\bf k-q-Q}|^2},\end{aligned}$$ where $\chi_\circ$ is a constant that sets the scale of the susceptibility, $\xi_\circ$ is the magnetic coherence length, ${\bf Q}\equiv(\pi,\pi)$ is the commensurate nesting vector, and $g_{\alpha\beta}$ is the coupling to the conduction electrons, the size of which can be fixed to get a desired value of the critical temperature and can be considered to contain $\chi_\circ$. The remaining parameter, $\xi_\circ$, is given in reference [@mmp] and will not be varied in this work. In this paper we set $g_{22}=0$ ie, there is no intrinsic pairing in the chains. This means that any superconductivity in the chains is induced by the interlayer interaction, $g_{12}$, since we have set the hopping between layers to zero. The effect of an interlayer hopping has been extensively studied in works by Atkinson and Carbotte[@atkinson] as well as others.[@klemm; @klemm2; @kettemann; @wheatley; @liu] We solve the coupled [bcs]{} equations (\[bcs.eq\]) using a [fft]{} technique.[@odonovan1] In Fig. \[ybco.fig\] we plot the lowest three Fourier components (\[harmonics.eq\]) of the zero temperature order parameter as a function of the interlayer interaction (higher Fourier components are all much smaller in magnitude) for two different intralayer interactions (upper and lower frames) in the planes (left frames) and chains (right frames). The values plotted in Fig. \[ybco.fig\] are the amplitudes $\mit\Delta_\alpha^{(s_\circ)}$, $\mit\Delta_\alpha^{(s_x)}$ and $\mit\Delta_\alpha^{(d)}$ given by (\[harmonics.eq\]) with $\alpha=1$ for the plane layers (right frames) and $\alpha=2$ for the chain layers (left frames). For the orthorhombic system studied here all three of these harmonics belong to the same irreducible representation of the crystal point group except for $g_{12}=0$ when the tetragonal CuO$_2$ layer is decoupled from the orthorhombic CuO layer. Recent [cits]{} measurements[@edwards2] show that the gap has a magnitude of $\sim 20$meV in the chains and $\sim 30$meV in the planes which would indicate that $g_{12}$ is large. For the first choice of intralayer interaction (upper frames), $g_{11}=26.2$, and there is no order parameter in the chains when there is no interlayer interaction (ie, $g_{12}=0$) and the order parameter in the planes is pure $d$-wave. As the interlayer interaction is increased from zero, $s$-wave components appear in the planes and all three components appear in the chains. This “$s+d$ mixing” is caused by the breaking of the tetragonal symmetry upon the introduction of the chains; there is no relative phase between the $s$- and $d$-wave components within either the planes or chains but there can be a relative phase between the order parameter in the planes and chains. In the range of $g_{12}$ explored here the $d$-wave component in the plane remains dominant but for sufficiently strong interlayer interaction the isotropic $s$-wave component eventually dominates[@liechtenstein] (ie, the gap nodes disappear). For interaction parameters $\{g_{11},g_{12}, g_{22}\} = \{26.2,10,0\}$ the critical temperature is 100K and the maximum value of the gap in the Brillouin zone is 27.5meV in the planes and 8.0meV in the chains, while the maximum values on the Fermi surfaces are approximately 22meV and 7meV respectively. The ratio $2\mit\Delta_{\rm max}/T_{\rm c}$ is 6.4 in the planes and 1.9 in the chains. For the second choice of intralayer interaction (lower frames), $g_{11}=9.18$, there is no order parameter in either the chains or the planes when there is no interlayer interaction (ie, $g_{12}=0$). As the interlayer interaction is increased $d$-wave and then $s$-wave components of the order parameter appear in both the planes and chains. Again, there is no relative phase between the $s$- and $d$-wave components within either the planes or chains but there can be a relative phase between the order parameter in the planes and chains. In the range of $g_{12}$ explored here, the $d$-wave component is dominant. At approximately $g_{12}=15$ the gap nodes no longer cross the Fermi surface in the chains; the feature at $g_{12}\sim 25$ coincides with the gap nodes leaving the Brillouin zone and isotropic $s$-wave becoming dominant. For interaction parameters $\{g_{11},g_{12}, g_{22}\} = \{9.18,20,0\}$ the critical temperature is again 100K and the maximum value of the gap in the Brillouin zone is now 32.8meV in the planes and 20.1meV in the chains, while the maximum values on the Fermi surfaces are approximately 27meV and 17meV respectively. The ratio $2\mit\Delta_{\rm max}/T_{\rm c}$ is 7.6 in the planes and 4.7 in the chains. Note that for $g_{12}>0$ all of the $s$-wave components of the order parameters in both the planes and chains have the same relative sign and the $d$-wave components have opposite signs, while for $g_{12}<0$ all of the relative signs are reversed but that the magnitudes of the components are insensitive to the sign of $g_{12}$ as noted after Eq. \[bcs.eq\]. In Fig. \[dos.fig\] we plot the density of states ([dos]{}) for the planes (left frames) and chains (right frames) calculated using the lowest three harmonics (\[harmonics.eq\]) of the solution to the [bcs]{} equations (\[bcs.eq\]) with two sets of interaction parameters. The dotted curves are the normal state [dos]{} ([nsdos]{}) and the solid curves are the superconducting [dos]{} ([scdos]{}). The insets show the Fermi surface (dashed curves) and gap nodes (solid curves) in the first Brillouin zone (with $(\pi,\pi)$ at the center). The peak in the [nsdos]{} (dotted curves) is the van Hove singularity located at $2t_\alpha(2-\mu_\alpha-4B_\alpha\pm\epsilon_\alpha)$ which is at -62meV in the plane layers (left frames) and at 10 and -170meV in the chain layers (right frames). They are caused by the saddle points in the electron dispersions, $\varepsilon_{{\bf k},\alpha}$, at $(0,\pm\pi)$ and $(\pm\pi,0)$. These van Hove singularities are shifted by the presence of the superconducting order parameter (solid curves) by an amount that depends upon the value of the order parameter at the saddle points;[@zhou] in frame (c) these values are very different and the van Hove singularity is split, in frame (a) these values are almost the same and no splitting is evident. An interesting feature is that the low energy behaviour ($\omega\sim 0$) of the [scdos]{} is governed by the [*smallest*]{} local [*maxima*]{} of the gap on the Fermi surface when the gap nodes cross the Fermi surface and by the [*minima*]{} of the gap on the Fermi surface when there is no gap nodes which cross the Fermi surface. In Fig. \[angles.fig\] the magnitude of the gap along the Fermi surface is plotted. In Fig. \[angles.fig\]a-c one can see that there are two different local maxima of the gap on the Fermi surface and these maxima are (to first order[@zhou]) manifested as twin peaks in the [scdos]{} (Fig. \[dos.fig\]a-c); these peaks are distinct from the van Hove singularities which are also present in the normal [dos]{} (dotted curves in Fig. \[dos.fig\]) and which are slightly shifted in the superconducting state.[@zhou] In frame (a) of Fig. \[angles.fig\] the local maxima of the gap on the Fermi surface are 16 and 18meV; in (b) they are 1 and 7meV, and in (c) they are 25 and 3meV. In (d) one can see that there are no gap nodes which cross the Fermi surface; the maximum and minimum value of the gap on the Fermi surface are 17 and 4meV respectively. In Fig. \[dos.fig\]d the finite gap in the [scdos]{} corresponds to the minimum of the gap on the Fermi surface and the peak to the maximum. The Josephson junction resistance-tunneling current product, $RJ(T=0)$, for a $c$-axis YBCO-Pb junction, given by (\[joe.eq\]) with $\mit\Delta^L_{\bf k}$ and $\varepsilon^L_{\bf k}$ appropriate for Pb,[@odonovan3] are $\pm0.25$meV and $\pm2.2$meV for the planes and 2.2meV and 3.4meV for the chains for the two choices of $g_{\alpha\beta}$ made, in agreement with earlier calculations.[@odonovan3] The relative sign is due to the relative sign of the $s$-wave components (ie, the only part which contributes) of the order parameters. The actual $c$-axis Josephson junction resistance-tunneling current product, $RJ(T)$, for a junction made with untwinned YBCO would be some weighted average of the plane and chain $c$-axis tunneling currents depending upon the relative abundance of chains and planes in the top layer of the YBCO. For a twinned sample with both twins equally abundant there would be zero net tunneling current, although there is evidence that for single crystals of YBCO there can be up to a 5:1 ratio in the relative abundance of the two twin orientations.[@gaulin] We note that due to the different magnitudes of the order parameters in the two layers the model presented here is consistent with the observed $\pi$ shifts in corner junctions[@wollman; @brawner; @tsuei; @mathai] for both attractive and repulsive interlayer interactions, $g_{12}$. In Fig. \[pd.fig\] we have plotted the magnetic penetration depth (left frames) and the Knight shift (right frames) calculated with the lowest three harmonics (\[harmonics.eq\]) of the solutions of the [bcs]{} equations (\[bcs.eq\]) for the two choices of interaction parameters. In the penetration depth frames (left) the solid curve is for the $x$-direction (along the chains) and the dashed curve is for the $y$-direction (perpendicular to the chains). The dotted curve is $1-(T/T_{\rm c})^2$ and is plotted for comparison. The ratio $\lambda_{yy}/\lambda_{xx}$ at zero temperature is 1.37 for both interaction parameter choices since the zero temperature penetration depth is a normal state property (ie, the second term in Eq. \[pd2.eq\] does not contribute at zero temperature). The zero temperature penetration depth is largely governed by the bandwidth (ie, $4t_\alpha(2-\epsilon_\alpha)$) – the larger the bandwidth the larger the zero temperature penetration depth. As pointed out above, the curvature of the penetration depth curve, $\lambda_{ii}^{-2}(T)$, is largely governed by the ratio $2\mit\Delta_{\rm max}/T_{\rm c}$ and is a straight line for the $d$-wave [bcs]{} value of $4.4$. The presence of the chain layer and the interlayer interaction increases this ratio in the plane layer but it remains low in the chain layer due to the absence of an interaction in this layer. It is this lower value that makes $\lambda^{-2}_{yy}(T)$ (along the chains) have upward curvature (solid curves). One can see (Fig. \[pd.fig\]a,c) that the in-plane penetration depth perpendicular to the chains (dashed curve) closely resembles that observed experimentally in high quality crystals,[@basov; @jacobs] and is largely determined by the presense of gap nodes crossing the Fermi surface, which cause the low temperature linear behaviour, and the ratio $2\mit\Delta_{\rm max}/T_{\rm c}$ which, for values above $\sim4.4$, make the curve of $\lambda_{ii}^{-2}(T)$ have downward curvature. The penetration depth along the chains (solid curves), however, has an overall upward curvature due to the low values of $2\mit\Delta_{\rm max}/T_{\rm c}$ in the chains. It is the component of the penetration depth due to the chains that makes the overall $\lambda^{-2}_{yy}(T)$ have downward curvature. The component of the penetration depth due to the chains perpendicular to the chains does not contribute significantly to the overall penetration depth in this direction, $\lambda^{-2}_{yy}$, is due almost entirely to the CuO$_2$ layers. For a single plane model[@odonovan2] we would have $\{g_{11},g_{12}, g_{22}\} =\{30,0,0\}$, $2\mit\Delta_{\rm max}/T_{\rm c}$ would be 4.5, and the penetration depth would closely resemble the straight line $1-(T/T_{\rm c})$.[@odonovan2] The Knight shift (Fig. \[pd.fig\]b,d), which is calculated independently for the planes (solid curves) and chains (dashed curves), has a low temperature power law behaviour when the gap nodes cross the Fermi surface (planes, both figures and chains in upper figure) and an exponential behaviour when the gap is finite over all the Fermi surface (chains, lower figure). When these quantities are measured[@scalapino] the distinction between a power law and exponential behaviour rests upon the choice zero and so is not a reliable indicator of the presence of gap nodes on the Fermi surface. Conclusion and Discussion ========================= We have derived a general expression for the Hamiltonian in a multi-layer system and then made a simplification and have explicitly diagonalized the Hamiltonian. A set of two coupled [bcs]{} equations are then derived for this simplified system which is subsequently solved numerically by a [fft]{} technique. This technique, unlike others,[@wheatley; @combescot] makes no assumptions about the functional form (and hence the symmetry) of the order parameter in either layer nor any relationship between the order parameters in the different layers. Using the three lowest harmonics (\[harmonics.eq\]) of the solutions found for the coupled [bcs]{} equations (\[bcs.eq\]) the magnetic penetration depth (\[pd2.eq\]), normal and superconducting density of states (\[dos.eq\],\[scdos.eq\]), Knight shift (\[ks.eq\]) and $c$-axis Josephson resistance-tunneling current products (\[joe.eq\]) were calculated. Our choice of electron dispersion relations were made so as to approximate the YBCO system in which there are layers consisting of CuO$_2$ planes as well as layers which contain CuO chains. Our choice of interactions was made so that there is a pairing interaction in the CuO$_2$ layer as well as an interlayer interaction, but no pairing interaction in the CuO layer. The solution of the [bcs]{} equations is predominately of a $d$-wave character, but because the tetragonal symmetry is broken by the presence of the chains there is some mixture of $s$-wave order parameter with no relative phase between the components, although the relative sign of the order parameter in the planes and chains may be $\pm 1$. Further, we find that due to a symmetry in the set of coupled [bcs]{} equations derived, the sign of the interlayer interaction affects only the relative sign of the order parameter in the two layers and not their absolute magnitudes (although some properties could be affected by this relative sign). We also find that any interlayer interaction strongly enhances the zero temperature order parameter, and hence the critical temperature. This is consistent with the observation that $T_{\rm c}$ is higher in materials with multiple adjacent CuO$_2$ layers. Furthermore, we find that the presence of gap nodes in only one of the layers is enough to produce a low temperature linear behaviour for the penetration depth, although if the minimum gap in the chains is too large the $\lambda^{-2}_{xx}(T)$ and $\lambda^{-2}_{yy}(T)$ curves can cross. Our calculation of the magnetic penetration depth gives a form that is similar to that measured experimentally perpendicular to the chains, but not along the chains due to the small value of $2\mit\Delta_{\rm max}/T_{\rm c}$ in the chains. This leads us to speculate that there may be intrinsic pairing in the CuO chains of the same order as in the CuO$_2$ planes since the $2\mit\Delta_{\rm max}/T_{\rm c}$ ratios must both be large (6 to 7) for the low temperature slope of the $\lambda^{-2}_{ii}(T)$ curves to be approximately equal as is observed in experiments.[@basov; @jacobs] This would tend to support the simple single band orthorhombic model previously proposed by us.[@odonovan3; @odonovan2] Our calculation of the superconducting density of states indicate that a surface probe may measure very different results depending upon whether the top layer is CuO chains or CuO$_2$ planes. Depending upon the interlayer pairing strength the CuO chain layer may have a very narrow “$d$-wave” type gap or a finite “isotropic $s$-wave” type gap. Our calculation of the $c$-axis Josephson resistance-tunneling current products, $RJ(T=0)$, for a YBCO-Pb junction for several choices of $g_{\alpha\beta}$ range from 0.18meV to 0.50meV for the planes and 2.35meV to 3.06meV for the chains, with possibly a relative sign between the chain and plane layers due to the relative sign of the $s$-wave components of the order parameters. The actual $c$-axis Josephson resistance-tunneling current product for a junction made with untwinned YBCO would be some weighted average of the plane and chain results depending upon the relative abundance of chains and planes in the top layer of the YBCO. For a twinned sample with both twins equally abundant there would be zero net tunneling current although there is evidence that for single crystals of YBCO there can have up to a 5:1 ratio in the relative abundance of the two twin orientations.[@gaulin] Acknowledgements {#acknowledgements .unnumbered} ================ Research supported in part by the Natural Sciences and Engineering Research Council of Canada ([nserc]{}) and by the Canadian Institute for Advanced Research ([ciar]{}). We would like to thank W. A. Atkinson and A. M. Westgate for discussions and insights. Electonic address: [odonovan@mcmaster.ca]{} D.N. Basov, R. Liang, D.A. Bonn, W.N. Hardy, B. Dabrowski, M. Quijada, D.B. Tanner, J.P. Rice, D.M. Ginsberg and T. Timusk Phys. Rev. Lett. [**74**]{}, 598 (1995). A.J. Millis, H. Monien and D. Pines, Phys. Rev. [**B 42**]{}, 167 (1990). H. Chi and J.P. Carbotte, Physics [**C 169**]{}, 55 (1990). C. O’Donovan and J.P. Carbotte, Physica [**C 252**]{} 87 (1995). D.A. Bonn, P. Dosanjh, R. Liang and W.N. Hardy, Phys. Rev. Lett., 2390 (1992). W.N. Hardy [*et al*]{}, Phys. Rev. Lett. [**70**]{}, 3999 (1993). D.A. Bonn, S. Kamal, K. Zhang, R. Liang, D.J. Baar, E. Kleit and W.N. Hardy, Phys. Rev. [**B 50**]{}, 4051 (1994). N. Bulut and D.J. Scalapino, Phys. Rev. [**B 45**]{}, 2371 (1992). H. Ding, J.C. Campuzano et al., Phys. Rev. Lett. [**50**]{}, 1333 (1994). H.L. Edwards, J.T. Markert, and A.L. de Lozanne, Phys. Rev.Lett. [**69**]{}, 2967 (1992). C. O’Donovan, D. Branch, J.P. Carbotte and J. Preston, Phys. Rev. [**B 51**]{}, 6588 (1995). A.G. Sun et al., Phys. Rev. Lett. [**72**]{}, 2267 (1994). W.A. Atkinson and J.P. Carbotte, Phys. Rev. [**B 52**]{}, 10601 (1995). R.A. Klemm and S.H. Liu, Physics [**C 191**]{}, 383 (1992). S.H. Liu and R.A. Klemm, Phys. Rev. [**B 45**]{}, 415 (1992). R. Combescot and X. Leyranas, Phys. Rev. Lett. [**75**]{}, 3732 (1995). S. Kettemann and K.B. Efetov, Phys. Rev. [**B 46**]{}, 8515 (1992). T. Xiang and J.M. Wheatley, Phys. Rev. Lett. [**76**]{} 134 (1996). D.Z. Liu, K. Levin and J. Maly, Phys. Rev. [**B 51**]{}, 8680 (1995). C. O’Donovan and J.P. Carbotte, Phys. Rev. [**B 52**]{}, 4548 (1995). I. Maggio-Aprile, Ch. Renner, A. Erb, E. Walker and O. Fisher, Phys. Rev. Lett. [**75**]{} 2754 (1995). T. Jacobs, S. Sridhar, Qiang Li, G.D. Gu and N. Koshizuka, Phys. Rev. Lett. [**75**]{}, 4516 (1995). Mechanisms of Conventional and High $T_c$ Superconductivity, V.Z. Kresin, H. Morawitz and S.A. Wolf, Oxford University Press (1993). V. Ambegaokar and A. Baratoff, Phys. Rev. Lett. [**10**]{}, 486 (1963); [**11**]{}, 104 (1964). K. Gofron, J.C. Campuzano, A.A. Abrikosov, M. Lindroos, A. Bansil, H. Ding, D. Koelling, and B. Dabrowski, Phys. Rev. Lett. [**73**]{}, 3302 (1994). H.L. Edwards, D.J. Derro, A.L. Barr, J.T. Markert, and A.L. de Lozanne, Phys. Rev. Lett. [**75**]{}, 1387 (1992). A.I. Liechtenstein, I.I. Mazin and O.K. Anderson, Phys. Rev. Lett. [**74**]{}, 2303 (1995). C. Zhou and H.J. Schulz, Phys. Rev. [**B 45**]{}, 7397 (1992). B. Gaulin, private communication. D.A. Wollman, D.J. Van Harlingen, W.C. Lee, D.M. Ginsberg and A.J. Leggett, Phys. Rev. Lett. [**71**]{}, 2134 (1993). D.A. Brawner and H.R. Ott, Phys. Rev. [**B 50**]{}, 6530 (1994). C.C. Tsuei [*et al.,*]{} Phys. Rev. Lett. [**73**]{}, 593 (1994). A. Mathai, Y. Gim, R.C. Black, A. Amar and F.C. Wellstood, Phys. Rev. Lett. [**74**]{}, 4523 (1995). [c c]{} &\ =0.5 & =0.5\ \ =0.5 & =0.5\ ------ ------ =0.5 =0.5 =0.5 =0.5 ------ ------ ------ ------ =0.5 =0.5 =0.5 =0.5 ------ ------

ArXiv

--- author: - 'J. D. Anderson' - 'G. Schubert' - 'V. Trimble' - 'M. R. Feldman' title: 'Measurements of Newton’s gravitational constant and the length of day' --- Introduction ============ Newton’s gravitational constant, $G$, is one of a handful of universal constants that comprise our understanding of fundamental physical processes [@Mohr2012] and plays an essential role in our understanding of gravitation, whether previously in Newton’s attractive gravitational force between two massive bodies $m_1,m_2$ of magnitude [@Newton1687] $$F = \frac{Gm_1 m_2}{r^2},$$ where $r$ is their separation distance, or currently as the proportionality constant in the interaction between energy-momentum content $T_{ab}$ (the stress-energy tensor) and space-time curvature $G_{ab}$ (Einstein tensor) in Einstein’s general relativity [@Einstein1916; @Wald1984] $$G_{ab} = R_{ab} - \frac{1}{2}g_{ab}R= 8\pi G T_{ab},$$ in units where the local speed of light in vacuum $c=1$. Yet, experimental determination of Newton’s gravitational constant remains a challenging endeavor. As reviewed in [@Speake2014], several measurements over the last thirty years appear to give inconsistent values for $G$, of course an issue for our understanding of this universal constant. Our purpose with this letter is to inform the reader of a one-to-one correlation between an apparent temporal periodicity in measurements of $G$, generally thought to result from inconsistency in measurements, with recently reported oscillatory variations in measurements of LOD [@Holme2013]. LOD refers to the excess of the duration of the day (observed period of rotation of the Earth) relative to a standard unit and is calculated by taking the difference between atomic time (TAI) and universal time (UT1) divided by the aforementioned standard unit of $86400$ SI s [@Ray1996]. Variations in LOD can be used to determine changes in the Earth’s rotation rate effectively providing a means to examine geophysical and atmospheric processes [@Peltier2007]. For the following discussion, we emphasize that our $G$ analysis and LOD analysis (a verification of the procedures employed in [@Holme2013]) are very much independent of one another with the determined fitting parameters for both the period and phase of the periodicities in these measurements coinciding in near perfect agreement. Although we recognize that the one-to-one correlation between the fit to the $G$ measurements and the LOD periodicity of 5.9 years could be fortuitous, we think this is unlikely, given the striking agreement shown in Fig. \[PlotL1\]. Furthermore, after taking into account this fitted oscillatory trend in the $G$ measurements, we obtain agreement amongst the different experiments mentioned in [@Speake2014] with a weighted mean value for $G$ of $( 6.673899 \pm 0.000069 ) \times 10^{-11}$ m$^3$ kg$^{-1}$ s$^{-2}$. ![Result of the comparison of the CODATA set of $G$ measurements with a fitted sine wave (solid curve) and the 5.9 year oscillation in LOD daily measurements (dashed curve), scaled in amplitude to match the fitted $G$ sine wave. The acronyms for the measurements follow the convention used by CODATA, with the inclusion of a relatively new BIPM result from Quinn [*et al.*]{} [@Quinn2013] and another measurement LENS-14 from the MAGIA collaboration [@Rosi2014] that uses a new technique of laser-cooled atoms and quantum interferometry, rather than the macroscopic masses of all the other experiments. The green filled circle represents the weighted mean of the included measurements, along with its one-sigma error bar, determined by minimizing the L1 norm for all 13 points and taking into account the periodic variation.[]{data-label="PlotL1"}](PlotL1.eps){width="8.0cm"} Methods ======= In the July 2014 issue of Physics Today, Speake and Quinn [@Speake2014] lay out the problem and review the history of seemingly inconsistent measurements of the gravitational constant $G$. They plot twelve $G$ determinations, along with one-sigma error bars, extending from an experiment by Luther and Towler at the National Bureau of Standards (NBS) in 1982 [@Luther1982] to their own at BIPM in 2001 and 2007 (the latter of which was published in 2013) [@Quinn2001; @Quinn2013], two measurements in good agreement with each other, but not with the other 10 measurements. Though the vertical scale of years when the measurements were made is not linear, there is a striking appearance of a periodicity running through these values, characterized by a linear drift which suddenly reverses direction and then repeats more than once. With this pattern in mind, we compute a periodogram for the measured $G$ values versus estimated dates of when the experiments were run. A single clear period of 5.9 years emerges. The data for our $G$ analysis were obtained directly from Table XVII in the 2010 CODATA report published in 2012 [@Mohr2012]. There are 11 classical measurements made at the macroscopic level. To those we added two more recent data points, another macroscopic measurement, which we label BIPM-13, and the first ever quantum measurement with cold atoms, labeled LENS-14. Next we used our best estimates of when the experiments were run, not the publication dates, for purposes of generating a measured $G$ value versus date data file, with one-sigma errors included too. These dates were obtained from the respective articles. This gives us the best data set possible, defined by the measured $G$ values used for the CODATA recommendation plus two more published after 2012. We fit with the raw standard errors, $\sigma_i$, provided with each of the $G$ measurements and used a numerical minimization of the L1 and L2 norms of the weighted residuals, $r_i/\sigma_i$, where the residuals are about a fitting model of a single sine wave, $a_0 + a_1\cos{\omega t}+b_1\sin{\omega t}$, four parameters in all with 13 measurements. Results for the fit to the 13 measured $G$ values are summarized in Fig. \[PlotL1\]. The L2 minimization is equivalent to a weighted least squares fit, yet the L1 minimization (solid line in Fig. \[PlotL1\]) is a more robust estimator that discriminates against outliers. Both yield excellent fits with a suggestion that two measurements at Moscow [@Karagioz1999] and from the MAGIA collaboration [@Rosi2014] are outliers. However, the Moscow value is known to suffer from an unexplained temporal drift [@Karagioz1999] and the cold-atom value could be fundamentally different ($G$ at the quantum level). Still, we refrain from speculating further on the cold-atom outlier until more microscopic measurements of $G$ are obtained by different experimental groups. The other 11 measurements are consistent with the L1 fitting curve at the one-sigma level or better. Figure \[PlotL1\] appears to provide convincing evidence that there exists a 5.9 year periodicity in the macroscopic determinations of $G$ in the laboratory with variations at the level of $\Delta G/G \sim 2.4 \times 10^{-4}$ about a mean value of $6.673899 \times 10^{-11}$ m$^3$ kg$^{-1}$ s$^{-2}$, close to the value recommended by CODATA in 2010 [@Mohr2012] but with a much smaller standard error of 10.3 ppm instead of the CODATA recommended error of 120 ppm. The most accurate determination by the Washington group [@Gundlach2000] with a standard error of 14 ppm now falls squarely on the fitting curve. Because the two BIPM measurements were made at the peak of the fitting curve, they now not only agree, but they are consistent with all other measurements. Notably, the measurement with a simple pendulum gravity gradiometer at JILA is no longer biased to an unacceptably small value, but like the BIPM measurements it falls right on the fitting curve, but at the minimum of the sine wave. The Huazhong measurement is also at the minimum of the curve. Results ======= With the 5.9 year periodicity in the $G$ measurements accepted, the question arises as to what could be the cause and what does it mean. The only thing we can think of is a correlation with a 5.9 year periodicity in the Earth’s LOD, published by Holme and de Viron last year [@Holme2013]. The International Earth Rotation and Reference Systems Service (IERS), established in 1987, maintains downloadable data files containing daily values of several parameters related to Earth orientation and rotation. The files extend from 1962 January 01, when the Consultative Committee on International Radio (CCIR) established Universal Time Coordinated (UTC) as the standard for time keeping, to the most current date available. We extract two rotation files, the first is the difference UT1-UTC in seconds and the second the LOD, also expressed in seconds, along with daily estimates of standard errors for both. There is also a piecewise constant file in integer seconds for the standard of atomic time TAI minus UTC. By differencing these two files the phase of the Earth rotation is obtained as measured against a uniform atomic time. This difference can be thought of as a continuous phase function $\phi (t)$ in radians sampled once per day at the beginning of the day. It can be expressed in SI seconds, the units on the IERS files, by multiplying by the conversion factor $86400/2 \pi $. It essentially provides the time gained or lost over the years by a poor mechanical clock, the Earth, which runs slow with a loss of about 33 s over the 52 years of the downloaded file. Because of its name and units of seconds only, the second file LOD is more difficult to interpret. It is also the gain or loss of time by the Earth, but only over the current day, and because of definitions there is a reversal in sign. When expressed as a continuous function of the Earth’s rotational frequency $\nu ( t )$, it is simply $\nu_0 - \dot{\phi} / 2 \pi$, where $ \nu_0$ is an adopted frequency of rotation with sidereal period of 86164.098903697 s. The quantity $\dot{\phi}/ ( 2 \pi \nu_0 )$ is small and can be taken to the first order in all calculations. Formally, the spectral density of frequency is related to the spectral density of phase by $S_{\mathrm{LOD}} ( f) = ( 2 \pi f )^2 S_{\mathrm{UT1}} ( f )$, where $f$ is the Fourier frequency. However, a separate computation of the spectrum for each file shows that before 1994 either file can be used for analysis, but after the introduction of Global Positioning (GPS) data in 1993, the LOD data become more accurate by a factor of seven or more. This conclusion is consistent with the standard errors included with the data files of LOD and UT1-UTC. We show our estimate of the spectral density for the LOD data in Fig. \[PlotLOD\], obtained by weighted least squares and SVD, but this time with 850 Fourier coefficients, 430 degrees of freedom, and 19169 observations. The spectral resolution is $0.019$ yr$^{-1}$, which we oversample by a factor of four, and the frequency cut off is 2 yr$^{-1}$, far short of the Nyquist frequency of 0.5 d$^{-1}$. A window function is not applied to the data. It introduces undesirable artifacts into the low-frequency noise spectrum of interest and does little to isolate spectral lines. The Gaussian window produces a hint of a line at 5.9 yr, but only a hint. We proceed to an analysis of the data in the time domain. ![One-sided power spectral density per unit frequency for LOD data over the years 1962 to 2014. The white-noise floor is indicated by the horizontal solid line and corresponds to a standard deviation of 0.54 ms d$^{-1}$, achieved by introduction of GPS data in 1993 and consistent with the daily estimates of standard error archived with the LOD data. The upper dashed curve corresponds to mean spectral density for the numerical time derivative of the UT1 data, dependent on VLBI data from radio sources on the sky. For the low end of the spectrum the LOD and UT1 data both indicate a $f^{-2}$ random walk, which with only 52 years of data can be confused with a drift in the Earth’s rotation. At the high end, the underlying spectrum indicates white LOD noise, but with a rich spectrum from tidal torques and atmospheric loading at higher frequencies not plotted. Although there is power in the region, there is no suggestion of a single spectral line from the 5.9 year oscillation, a term which must be extracted by analysis in the time domain [@Holme2013].[]{data-label="PlotLOD"}](PlotLOD.eps){width="8.0cm"} The 5.9 year periodicity in the LOD data is plotted by Holme and de Viron in Figure 2 of their paper [@Holme2013]. Their plot looks in phase with the fit to the 13 $G$ values, but in order to obtain an independent check on the reality of the signal and for purposes of having a numerical sine wave extending into 2014, we first smooth the LOD data with a Gaussian filter with a radius of 600 days and a standard deviation of 200 days. As a result, the high-frequency noise at a period of one year and shorter is practically eliminated, and with little effect on the low-frequency noise spectrum. Next we fit a cubic spline to the smoothed data with a selection of knots or segments for the cubic polynomials done by eye, such that the fitting curve is sufficiently smooth but with a negligible effect on the 5.9 year periodicity. The resulting LOD residuals are fit with a sine wave of fixed 5.9 year period which is then subtracted from the smoothed data. The same procedure is applied to the new smoothed data and the procedure repeated four times with the knots for the spline at closer spacing with each iteration. The final result is the pure sine wave plotted as a dashed curve in Fig. \[PlotL1\]. It agrees with the periodic signal found by Holme and de Viron. A removal of the fitted spline representation of the random walk, and also the sine wave, from the smoothed data is all that is needed in order to reduce the LOD residuals about the fit to a one-sigma noise level of $4.8$ $\mu$s d$^{-1}$. The amplitude of the fitted periodic signal is $92.64\pm 0.18$ $\mu$s d$^{-1}$, reduced from the amplitude of $150$ $\mu$s d$^{-1}$ [@Holme2013] by the Gaussian smoothing, but with a well-determined period of $5.90076 \pm 0.00074$ yr. With 99% confidence the period lies between 5.898 and 5.903 yr. The phasing of the sine wave is as shown in Fig. \[PlotL1\] with a standard error of 0.25 yr. ![Result of the comparison of our $G$ data set with the monthly mean of the total sunspot number, appropriately scaled. The black curves represent solar activity as reflected in the international sunspot number.[]{data-label="Plotssn"}](Plotssn.eps){width="8.0cm"} The correlation between LOD and $G$ measurements in Fig. \[PlotL1\] is most likely of terrestrial origin, but the period of 5.9 years is also close to one-half the principal period of solar activity. References [@Djurovic1996] and [@Rio2003] discuss in greater detail that a possible correlation between solar activity and LOD measurements is not unexpected. Solar activity has an effect on mass distribution in the atmosphere which ultimately affects the Earth’s axial moment of inertia. It is feasible that this effect occurs at longer periods in the 5.9-year range, as well as at much shorter periods, on the order of days, for which models exist [@Holme2013]. Consequently, we plot in Fig. \[Plotssn\] the monthly mean of the total sunspot number and also a 13-month smoothing curve, both shown in black. The two curves, again scaled to the magnitude of the $G$ data, are taken directly from freely available downloads of data archived at www.sidc.be by WDC-SILSO, Royal Observatory of Belgium, Brussels. The smoothing is done by a standard tapered-boxcar approach and is generally regarded as a good measure of solar activity. Although the $G$ measurements show a general agreement with solar cycle 23, which peaked around 2002, the long and unexpected minimum that followed, and lasted until about 2010, is at odds with the rise in $G$ values during that minimum. There is also a negative correlation between the measurement from 1982 at the National Bureau of Standards, labeled NIST-82, and the sunspot number. It seems that solar activity can be disregarded as a cause of the variations in $G$ measurements. Conclusions =========== Over the relatively short time span of 34 years considered here, variations in the rotation of the Earth can be considered either a random walk or possibly a drift. Over much longer time scales the rotation must be slowing because of the transfer of spin angular momentum to orbital angular momentum caused by tidal friction of the Moon. Similarly, a real increase in $G$ should pull the Earth into a tighter ball with an increase in angular velocity and a shorter day due to conservation of angular momentum, contrary to the correlation shown in Fig. \[PlotL1\]. Thus, we do not expect that this behavior necessarily points to a real variation in $G$ but instead to some yet-to-be determined mechanism affecting both measurements in a similar manner. Importantly, if the observed effect is connected with a centrifugal force acting on the experimental apparatus, changes in LOD are too small by a factor of about $10^5$ to explain the changes in $G$ for the following reason. The Earth’s angular velocity $\omega_E$ is by definition $$\omega_E = \omega_0 ( 1 - \mathrm{LOD} ),$$ where $\omega_0$ is an adopted sidereal frequency equal to $72921151.467064$ picoradians per second and the LOD is in ms d$^{-1}$ (www.iers.org). The total centrifugal acceleration is given by $$a_c = r_s \omega_0^2 \bigg[ 1 - 2 A \sin\bigg(\frac{2\pi}{P} (t-t_0)\bigg) \bigg],$$ where $A$ is the amplitude $0.000150/86400$ of the 5.9 year sinusoidal LOD variation and $r_s$ is the distance of the apparatus from the Earth’s spin axis. The maximum percentage variation of the LOD term is $ 3.47 \times 10^{-9}$ of the steady-state acceleration, while $\Delta G/G$ is $2.4 \times 10^{-4}$, hence even the full effect of the acceleration with no experimental compensation changes $G$ by only $10^{-5}$ of the amplitude in Fig. \[PlotL1\]. Perhaps instead, the effect is connected with changing torques on the Earth’s mantle due to changing motions in the core. Changes of circulation in the core must be accompanied by changes in density variations in the core causing variations in the gravitational acceleration $g$ in the laboratory. At least this mechanism links both LOD and gravitational changes to changes in the core although we do not immediately see how either of these mechanisms could affect measurements of $G$ in the laboratory given the torsion balance schemes employed. The least likely explanation is a new-physics effect that could make a difference in the macroscopic and microscopic determinations of $G$. Perhaps a repetition of the single 2014 quantum measurement over the next decade or so can show consistency with a constant value, although if the variations in $G$ measurements are caused by an unknown inertial or frame effect, not by systematic experimental error, it likely applies at both the macroscopic and the microscopic levels. The gravitational parameter for the Sun, $GM_{\odot}$, is known to ten significant figures from orbital motions in the Solar System (ssd.jpl.nasa.gov/?constants). The universal constant $G$ does not vary at that scale, although Krasinsky and Brumberg [@Krasinsky2004; @Anderson2010] report a detection of an unexplained secular increase in the astronomical unit (AU) over the years 1976 to 2008, which can be interpreted as an increase in $GM_{\odot}$ proportional to the cube of the AU. However the effect on $G$, if real, is at the level of an increase of 3 parts in $10^{12}$ per year and undetectable with laboratory measurements of $G$. Nevertheless, the increase in $GM_{\odot}$ is not explainable as an increase of the solar mass by accretion as opposed to the mass radiated away by solar luminosity [@Anderson2010]. Apparently, there does seem to be a secular or very long period (greater than 20000 years) $G$ variation in the Solar System, but of order $10^{-6}$ smaller than the variation shown in Fig. \[PlotL1\]. [0]{} . . . . . . . in . . . . . . . . . . . . . Appendix ======== Since the recent publication of this article suggesting that there is a strong correlation with coefficient 0.99764 between measured values of the gravitational constant G as adopted by CODATA [@Mohr2012] and the 5.9-year oscillation of the length of day (LOD) [@Holme2013], Schlamminger et al. [@Schlamminger2015] point out that the dates assigned to some of the measurements do not agree with the intervals over which the measurements were made. They provide in their TABLE II a corrected and augmented list of measurements carried out over the last 35 years. There are 19 independent measurements at average times accurate to about 20% of the total interval for each experiment’s operation. The BIPM-01 and BIPM-13 values used by CODATA are separated into measurements in the electrostatic servo and Cavendish modes. The HUST-09 data are separated into two segments, one in 2007 with an interval of 60 days and another in 2008 with an interval of 39 days. Three measurement intervals, not included in the CODATA set, from the University of California Irvine (UCI) are appended [@Newman2014]. A detailed discussion of the CODATA TR&D-96 measurement is included for a measurement interval of 3835 days from 1985 to 1995. As this measurement is relatively uncertain (see Fig. \[PlotL1\]), and in addition there is little or no evidence of a variation over 10.5 years, we exclude it from the revised data set. ![Two-period fit to 18 revised $G$ measurements recommended by Schlamminger, Gundlach and Newman [@Schlamminger2015]. The L1 norm of the weighted residuals is minimized at a value of 28.81, with the seven parameters of the fitting model and their one-sigma errors from the converged covariance matrix given in Table I. The revised weighted mean and its uncertainty is indicated by the green dot.[]{data-label="newfit"}](newfit.eps){width="8.0cm"} The fit to the remaining 18 points is shown in Fig. \[newfit\], where two periods are included, the original period of about 5.9 years and an annual term suggested by FIG. 3 in Schlamminger et al. [@Schlamminger2015]. The amplitude of the periodic fitting curve is reduced significantly, but it has advantages not apparent in the fit to the uncorrected data of Fig. \[PlotL1\], in particular an excellent fit to the most accurate measurements at the maximum between 2000 and 2003, and a reconciliation of the comparably accurate JILA-10 and UCI-14c measurements, which without reconciliation differ by 16 sigma. The LENS-14 measurement is no longer an obvious outlier, but is negatively biased by 2.7 sigma from the fitting curve, no worse than the positive bias in the four BIPM measurements, with only BIPM-13c greater than three sigma. The only obvious outlier is HUST-01a, with a negative six-sigma bias from the curve. There are seven parameters in the revised fitting model with two periods, as opposed to the simple sine wave of Fig. \[PlotL1\] with four parameters. Values from a minimization of the L1 norm and standard errors from the converged covariance matrix are given in Table 1. The fitting model is given by, $$\begin{aligned} \label{GMdl} G = a_0 + a_1 \cos\bigg( \frac{2 \pi t}{P_1}\bigg) + b_1 \sin \bigg( \frac{2 \pi t}{P_1} \bigg) \nonumber \\ + a_2 \cos \bigg( \frac{2 \pi t}{P_2} \bigg) + b_2 \sin \bigg( \frac{2 \pi t}{P_2} \bigg).\end{aligned}$$ The two-period model is no longer in phase with the LOD sine wave, and as a consequence the sample correlation coefficient is reduced to 0.860. However, if the LOD sine wave is shifted earlier in phase by 174 days, the correlation coefficient is 0.944. Nevertheless, this makes the interpretation of a possible correlation of the $G$ measurements with LOD more problematic, with the similar periods near 5.9 years possibly a coincidence. Parameter Value Standard Error ----------- ----------- ---------------- $\rm a_0$ 6.673488 0.000071 $\rm a_1$ 0.000084 0.000031 $\rm b_1$ 0.000150 0.000072 $\rm P_1$ 1.023087 0.000042 $\rm a_2$ -0.001116 0.000091 $\rm b_2$ -0.000126 0.000070 $\rm P_2$ 5.911615 0.000028 : Parameters of the fitting model of Eq. \[GMdl\]. The periods $\rm P_1$ and $\rm P_2$ are in years and the coefficients are in the units of $G$, or $\rm 10^{-11}~m^3~kg^{-1}~s^{-2}$.

ArXiv

--- abstract: 'The problem of unsupervised learning and segmentation of hyperspectral images is a significant challenge in remote sensing. The high dimensionality of hyperspectral data, presence of substantial noise, and overlap of classes all contribute to the difficulty of automatically clustering and segmenting hyperspectral images. We propose an unsupervised learning technique called spectral-spatial diffusion learning (DLSS) that combines a geometric estimation of class modes with a diffusion-inspired labeling that incorporates both spectral and spatial information. The mode estimation incorporates the geometry of the hyperspectral data by using diffusion distance to promote learning a unique mode from each class. These class modes are then used to label all points by a joint spectral-spatial nonlinear diffusion process. A related variation of DLSS is also discussed, which enables active learning by requesting labels for a very small number of well-chosen pixels, dramatically boosting overall clustering results. Extensive experimental analysis demonstrates the efficacy of the proposed methods against benchmark and state-of-the-art hyperspectral analysis techniques on a variety of real datasets, their robustness to choices of parameters, and their low computational complexity.' author: - 'James M. Murphy' -   Mauro Maggioni bibliography: - 'HSI\_LearningIEEE.bib' title: Unsupervised Clustering and Active Learning of Hyperspectral Images with Nonlinear Diffusion --- Introduction {#Intro} ============ Machine Learning for Hyperspectral Data --------------------------------------- Hyperspectral imagery (HSI) has emerged as a significant data source in a variety of scientific fields, including medical imaging [@Lu2014], chemical analysis [@Wang2016], and remote sensing [@Eismann2012]. Hyperspectral sensors capture reflectance at a sequence of localized electromagnetic ranges, allowing for precise differentiation of materials according to their spectral signatures. Indeed, the power of hyperspectral imagery for material discrimination has led to its proliferation, making manual analysis of hyperspectral data infeasible in many cases. The large data size of HSI, combined with their high dimensionality, demands innovative methods for storage and analysis. In particular, efficient machine learning algorithms are needed to automatically process and glean insight from the deluge of hyperspectral data now available. The problem of *HSI classification*, or supervised segmentation, is to label each pixel in a given HSI as belonging to a particular class, given a training set of labeled samples (pixels) from each class. A variety of statistical and machine learning techniques have been used for HSI classification, including nearest-neighbor and nearest subspace methods [@Ma2010; @Li2014], support vector machines [@Melgani2004; @Fauvel2008], neural networks [@Ratle2010; @Chen2016; @Liang2016] and regression methods [@Qian2013; @Li2013_1]. These methods are design to perform well especially when the number of labeled training pixels is large. The process of labeling pixels typically requires an expert and it is costly. This motivates the design of machine learning techniques that require little or no labeled training data. So on the other end of the spectrum from classification, we have the problem of [*[HSI clustering]{}*]{}, or [*[unsupervised]{}*]{} segmentation, which has the same goal as HSI classification, but no labeled training data is available. This is considerably more challenging, and is an ill-posed problem unless further assumptions are made, for example about the distribution of the data and how it relates to the unknown labels. Recent techniques for hyperspectral clustering include those based on particle swarm optimization [@Paoli2009], Gaussian mixture models (GMM) [@Acito2003], nearest neighbor clustering [@Cariou2015], total variation methods [@Zhu2017], density analysis [@Chen2017], sparse manifold models [@Elhamifar2011; @Elhamifar2013], hierarchical nonnegative matrix factorization (HNMF) [@Gillis2015], graph-based segmentation [@Meng2017], and fast search and find of density peaks clustering (FSFDPC) [@Chen2017; @Rodriguez2014; @Zhang2016]. Another interesting modality is *active learning* for HSI classification. This is a supervised technique where a small, automatically but carefully chosen set of pixels is labeled, as opposed to the standard supervised learning setting, in which the labels are usually randomly selected. Active learning can lead to high quality classification results with significantly fewer labeled samples than in the case of randomly selected training data. Since far fewer training points are available in the active learning setting, the structure of the data may be analyzed with unsupervised learning, in order to decide which data points to query for labels. Thus, active learning may be understood as a form of *semisupervised learning* that exploits both global structure of the data—learned without supervision—and a small number of supervised training data points. A variety of active learning methods have been successfully deployed in remote sensing [@Tuia2011], including those based on relevance feedback [@Demir2015], region-based heuristics [@Stumpf2014], exploration-based heuristics [@Tuia2011_Using], belief propagation [@Li2013_2], support vector machines [@Tuia2009], and regression [@Li2010]. Machine learning for HSI suffers from several major challenges. First, the dimensionality of the data to be analyzed is high: it is not uncommon for the number of spectral bands in an HSI to exceed $200$. The corresponding sampling complexity for such a high number of dimensions renders classical statistical methods inapplicable. Second, clusters in HSI are typically nonlinear in the spectral domain, rendering methods that rely on having linear clusters ineffective. Third, there is often significant noise and between-cluster overlap among HSI classes, due to the materials being imaged and poor sensing conditions. Finally, HSI images may be quite large, requiring machine learning methods with computational complexity essentially linear in the number of pixels. This article addresses the problems of HSI clustering and, relatedly, active learning, which overcome these significant challenges. The methods we propose combine density-based methods with geometric learning through diffusion geometry [@Coifman2005; @Coifman2006] in order to identify class modes. This information is then used to propagate labels on training data to all data points through a nonlinear process that incorporates both spectral and spatial information. The use of data-dependent diffusion maps for mode detection significantly improves over current state-of-the-art methods experimentally, and also enjoys robust theoretical performance guarantees [@Murphy2018]. The use of diffusion distances exploits low-dimensional structures in the data, which allows the proposed method to handle data that is high-dimensional but intrinsically low-dimensional, even when nonlinear and noisy. Moreover, the spectral-spatial labeling scheme takes advantage of the geometric properties of the data, and greatly improves the empirical performance of clustering when compared to labeling based on spectral information alone. In addition, the proposed unsupervised method assigns to each data point a measure of confidence for the unsupervised label assignment. This leads naturally to an active learning algorithm in which points with low confidence scores are queried for training labels, which then propagate through the remaining data. The proposed algorithms enjoy nearly linear computational complexity in the number of pixels in the HSI and in the number of spectral dimensions, thus allowing for its application to large scenes. Extensive empirical results, including comparisons with many state-of-the-art techniques, for our method applied to HSI clustering and active learning are in Sections \[subsec:HyperspectralClustering\] and \[subsec:ActiveLearning\], respectively. Overview of Proposed Method {#s:overview} --------------------------- The proposed unsupervised clustering method is provided with data $X=\{x_{n}\}_{n=1}^N\subset\mathbb{R}^{D}$ (for HSI, $N$ = number of pixels and $D$ = number of spectral bands) and the number $K$ of classes, and outputs labels {$y_{n}\}_{n=1}^{N},$ each $y_{n}\in\{1,\dots,K\}$, by proceeding in two steps: - [**[Mode Identification]{}**]{}: This step consists first in performing density estimation and analyzing the geometry of the data to find $K$ [*[modes]{}*]{} $\{x_{i}^{*}\}_{i=1}^{K}$, one for each class. - [**[Labeling Points]{}**]{}: Once the modes are learned, they are assigned a unique label. Remaining points are labeled in a manner that preserves spectral and spatial proximity. By a mode, we mean a point of high density within a class, that is representative of the entire class. We assume $K$ is known, but otherwise we have no access to labeled data; in Section \[sec:Future\] we discuss a method for estimating $K$. One of the key contributions of this article is to measure similarities in the spectral domain not with the widely used Euclidean distance or distances based on angles (correlations) between points, but with *diffusion distance* [@Coifman2005; @Coifman2006], which is a data-dependent notion of distance that accounts for the geometry—linear or nonlinear—of the distribution of the data. The motivation for this approach is to attain robustness with respect to the shape of the distributions corresponding to the different classes, as well as to high-dimensional noise. The modes, suitably defined via density estimation, are robust to noise, and the process we use to pick only one mode per class is based on diffusion distances. The labeling of the points from the modes respects the geometry of the data, by incorporating proximity in both spectral and spatial domains. We model $X$ as samples from a distribution $\mu=\sum_{i=1}^{K}w_{i}\mu_{i},$ where each $\mu_{i}$ corresponds to the probability distribution of the spectra in class $i$, and the nonnegative weights $\{w_{i}\}_{i=1}^{K}$ correspond to how often each class is sampled, and satisfy $\sum_{i=1}^{K} w_{i}=1$. More precisely, sampling $x\sim\mu$ means first sampling $Z\sim\text{Multinomial}(w_{1},\dots,w_{K})$, then sampling from $\mu_{i}$ conditioned on the event $Z=i\in\{1,\dots,K\}$. ### Mode Identification The computation of the modes is a significant aspect of the proposed method, which we now summarize for a general dataset $X$, consisting of $K$ classes. The mode identification algorithm outputs a point $x_{i}^{*}$ (“mode”) for each $\mu_{i}$. We make the assumption that modes of the constituent classes can be characterized as a set of points $\{x_{i}^{*}\}_{i=1}^{K}$ such that 1. the empirical density of each $x_{i}^{*}$ is relatively high; 2. the diffusion distance between pairs $x_{i}^{*},x_{i'}^{*}$, for $i\neq i'$, is relatively large. The first assumption is motivated by the fact that points of high density ought to have nearest neighbors corresponding to a single class; the modes should thus produce neighborhoods of points that with high confidence belong to a single class. However, there is no guarantee that the $K$ densest points will correspond to the $K$ unique classes: some classes may have a multimodal distribution, meaning that the class has several modes, each with potentially higher density than the densest point in another class. The second assumption addresses this issue, requiring that modes belonging to different distributions are far away in diffusion distance. [.15]{} ![\[fig:NonlinearComparison\] In this $2$-dimensional example, data is drawn from two distributions $\mu_1$ and $\mu_2$. $\mu_1$ is a mixture of two isotropic Gaussians with means at $(0,1)$ and $(1,0)$, respectively, connected by a set of points uniformly sampled from a nonlinear, parabolic shape. $\mu_2$ is an isotropic Gaussian with mean at $(0,0)$. Samples of uniform background noise are added and labeled according to their nearest neighbor among the two clusters. The data is plotted and colored by cluster in subfigure (a). We plot the distances from the point $(0,1)$ in the Euclidean and diffusion distances in subfigures (b), (c), respectively. The “parabolic rectangle” acts as a “bridge” between the two Gaussians and causes the high density regions near $(0,1)$ and $(1,0)$ to be closer in diffusion distance than they would be in the usual Euclidean distance. The bridge is overcome efficiently with diffusion distance, because there are many paths with short edges connecting the high density regions across this bridge.](Images/ToyExamples/ParabolicExample-crop.pdf "fig:"){width="\textwidth"} [.15]{} ![\[fig:NonlinearComparison\] In this $2$-dimensional example, data is drawn from two distributions $\mu_1$ and $\mu_2$. $\mu_1$ is a mixture of two isotropic Gaussians with means at $(0,1)$ and $(1,0)$, respectively, connected by a set of points uniformly sampled from a nonlinear, parabolic shape. $\mu_2$ is an isotropic Gaussian with mean at $(0,0)$. Samples of uniform background noise are added and labeled according to their nearest neighbor among the two clusters. The data is plotted and colored by cluster in subfigure (a). We plot the distances from the point $(0,1)$ in the Euclidean and diffusion distances in subfigures (b), (c), respectively. The “parabolic rectangle” acts as a “bridge” between the two Gaussians and causes the high density regions near $(0,1)$ and $(1,0)$ to be closer in diffusion distance than they would be in the usual Euclidean distance. The bridge is overcome efficiently with diffusion distance, because there are many paths with short edges connecting the high density regions across this bridge.](Images/ToyExamples/ParabolicEuclideanDistances-crop.pdf "fig:"){width="\textwidth"} [.15]{} ![\[fig:NonlinearComparison\] In this $2$-dimensional example, data is drawn from two distributions $\mu_1$ and $\mu_2$. $\mu_1$ is a mixture of two isotropic Gaussians with means at $(0,1)$ and $(1,0)$, respectively, connected by a set of points uniformly sampled from a nonlinear, parabolic shape. $\mu_2$ is an isotropic Gaussian with mean at $(0,0)$. Samples of uniform background noise are added and labeled according to their nearest neighbor among the two clusters. The data is plotted and colored by cluster in subfigure (a). We plot the distances from the point $(0,1)$ in the Euclidean and diffusion distances in subfigures (b), (c), respectively. The “parabolic rectangle” acts as a “bridge” between the two Gaussians and causes the high density regions near $(0,1)$ and $(1,0)$ to be closer in diffusion distance than they would be in the usual Euclidean distance. The bridge is overcome efficiently with diffusion distance, because there are many paths with short edges connecting the high density regions across this bridge.](Images/ToyExamples/ParabolicDiffusionDistances-crop.pdf "fig:"){width="\textwidth"} Enforcing that these modes are far apart in diffusion distance has several advantages over enforcing they are far apart in Euclidean distance. Importantly, it leads, empirically, to a unique mode from each class. This is true even when certain classes are multimodal. Moreover, diffusion distances are robust with respect to the shape of the support of the distribution, and are thus suitable for identifying nonlinear clusters. An instance of these advantages of diffusion distance is illustrated in the toy example Figure \[fig:NonlinearComparison\], with the results of the proposed mode detection algorithm in Figure \[fig:ToyExample\]. We postpone the mathematical and algorithmic details to Section \[subsec:DD\]. ### Labeling Points At this stage we assume that we found exactly one mode $x_{i}^{*}$ for each class, to which a unique and arbitrary class label is assigned. The remaining points are now labeled in a two-stage scheme, which takes into account both spectral and spatial information. It is known that the incorporation of spatial information with spectral information has the potential to improve machine learning of hyperspectral images, compared to using spectral information alone [@Fauvel2008; @Li2013_2; @Zhang2016; @Tarabalka2009; @Benedetto2012; @Fauvel2013; @Cahill2014; @Cloninger2014; @Wang2014; @Benedetto2016]. Spatial information is computed for each pixel by constructing a neighborhood of some fixed radius in the spatial domain, and considering the labels within this neighborhood. For a given point, let *spectral neighbor* refer to a near neighbor with distances measured in the spectral domain, and let *spatial neighbor* refer to a near neighbor with distances measured in the spatial domain. In the first stage, a point is given the same label as its nearest spectral neighbor of higher density, unless that label is sufficiently different from the labels of the point’s nearest spatial neighbors, in which case the point is left unlabeled. This produces an incomplete labeling in which we expect the labeled points to be far from the spectral and spatial boundaries of the classes, since these are points that are unlikely to have conflicting spectral and spatial labels. The first stage thus labels points using only spectral information, though spatial information may prevent a label from being assigned. In the second stage we label each of the points left unlabeled in the first stage, by assigning the *consensus label* of its nearest spatial neighbors (see Section \[subsec:AlgorithmDescription\]), if it exists, or otherwise the label of its nearest spectral neighbor of higher density. In this way the yet unlabeled points, typically near the spatial and spectral boundaries of the classes, benefit from the spatial information in the already labeled points, which are closer to the centers of the classes. The second stage thus labels points using both spectral and spatial information. Figure \[fig:SpectralSpatialLabeling\] shows an instance of this two-stage labeling process. This method of clustering combines the diffusion-based learning of modes with the joint spectral-spatial labeling of pixels and is called *spectral-spatial diffusion learning* (DLSS), detailed in Section \[subsec:AlgorithmDescription\]. We contrast it with another novel method we propose, called *diffusion learning (DL)*, in which modes are learned as in DLSS, but the labeling proceeds simply by enforcing that each point has the same spectral label as its nearest spectral neighbor of higher density. DL therefore disregards spatial information, while DLSS makes significant use of it, particularly in the second stage of the labeling. Our experiments show that while both DL and DLSS perform very well, DLSS is generally superior. [.24]{} ![\[fig:SpectralSpatialLabeling\] An example of the two-stage spectral-spatial labeling process, performed on the Indian Pines dataset used for experiments in Section \[subsubsec:IP\]. In subfigure (a), the partial labeling from the first stage is shown. After mode identification, points are labeled with the same label as their nearest spectral neighbor of higher density, unless that label is different from the consensus label in the spatial domain, in which case a point is left unlabeled. This leads to points far from the centers of the classes staying unlabeled after the first stage. In the second stage, unlabeled points are assigned labels by the same rule, unless there is a clear consensus in the spatial domain, in which case the unlabeled point is given the consensus spatial label; the results of this second stage appear in subfigure (b). For visual clarity, here and throughout the paper, pixels without ground truth (GT) labels are masked out.](Images/ToyExamples/IP_Pass1-crop.pdf "fig:"){width="\textwidth"} [.24]{} ![\[fig:SpectralSpatialLabeling\] An example of the two-stage spectral-spatial labeling process, performed on the Indian Pines dataset used for experiments in Section \[subsubsec:IP\]. In subfigure (a), the partial labeling from the first stage is shown. After mode identification, points are labeled with the same label as their nearest spectral neighbor of higher density, unless that label is different from the consensus label in the spatial domain, in which case a point is left unlabeled. This leads to points far from the centers of the classes staying unlabeled after the first stage. In the second stage, unlabeled points are assigned labels by the same rule, unless there is a clear consensus in the spatial domain, in which case the unlabeled point is given the consensus spatial label; the results of this second stage appear in subfigure (b). For visual clarity, here and throughout the paper, pixels without ground truth (GT) labels are masked out.](Images/ToyExamples/IP_FinalLabels-crop.pdf "fig:"){width="\textwidth"} Major Contributions ------------------- We propose a clustering algorithm for HSI with several significant innovations. First, diffusion distance is proposed to measure distance between high-density regions in hyperspectral data, in order to determine class modes. Our experiments show that this distance efficiently differentiates between points belonging to the same cluster and points in different clusters. This correct identification of modes from each cluster is essential to any clustering algorithm incorporating an analysis of modes. Compared to state-of-the-art fast mode detection algorithms, the proposed method enjoys excellent empirical performance; theoretical performance guarantees are beyond the scope of the present article and will be discussed in a forthcoming article [@Murphy2018]. A second major contribution of the proposed HSI clustering algorithm is the incorporation of spatial information through the labeling process. Labels for points are determined by diffusing in the spectral domain from the labeled modes, unless spatial proximity is violated. By not labeling points that would violate spatial regularity, the proposed algorithm first labels points that, with high confidence, are close to the spectral modes of the distributions. Only after labeling all of these points are the remaining points, further from the modes, labeled. This enforces a spatial regularity which is natural for HSI, because under mild assumptions, a pixel in an HSI is likely to have the same label as the most common label among its nearest spatial neighbors [@Fauvel2008; @Li2013_2; @Zhang2016; @Tarabalka2009; @Benedetto2012; @Fauvel2013; @Cahill2014; @Cloninger2014; @Wang2014; @Benedetto2016]. In both stages, DLSS takes advantage of the geometry of the dataset by using data-adaptive diffusion processes, greatly improving empirical performance. The proposed methods are $O(ND\log(N))$ in the number of points ($N$) and ambient dimension of the data ($D$) when the intrinsic dimension of the data is small, and thus have near optimal complexity, suitable for the big data setting. A third major contribution is the introduction of an *active learning* scheme based on distances of points to the computed modes. In the context of active learning, the user is allowed to label only a very small number of points, to be chosen parsimoniously. We propose an unsupervised method for determining which points to label in the active learning setting. We note that pixels that are equally far in diffusion distance from their nearest two modes are likely to be near class boundaries, and hence to be the most challenging pixels to label by the proposed unsupervised method. Our active learning method requires the labels of only the pixels whose distances to their nearest two modes are closest. The proposed active learning method builds naturally on the fully unsupervised method, since the computation of distances to nearest mode are already computed by the DL and DLSS algorithms, and hence the computational complexity of the proposed active learning method does not differ significantly from the fully unsupervised method. Our experiments show that this method can dramatically improve labeling accuracy with a number of labels $\ll 1\%$ of the total pixels. This work is detailed in Section \[subsec:ActiveLearning\]. Unsupervised Learning Algorithm and Active Learning Variation {#sec:Algorithm} ============================================================= Motivating Example and Approach ------------------------------- A key aspect of our algorithm is the method for identifying the modes of the classes in the HSI data. This is challenging because of the high ambient dimension of the data, potential overlaps between distributions at their tails, along with differing densities, sampling rates, and distribution shapes. [.24]{} ![\[fig:ToyExample\]Learned modes with Euclidean distances and diffusion distances. The Euclidean and diffusion distances from $(0,1)$ are shown in subfigures (b), (c) of Figure \[fig:NonlinearComparison\], while the corresponding learned modes are labeled, with nearby points colored red in subfigures (a), (b), of the present figure. Notice that the proposed diffusion learning method, using diffusion distances, correctly learns $M_{1}, M_{2}$ from different clusters (b), while using Euclidean distances leads assigning both $M_{1}, M_{2}$ to the same cluster (a), which would lead to poor clustering results.](Images/ToyExamples/ParabolicEuclideanModes-crop.pdf "fig:"){width="\textwidth"} [.24]{} ![\[fig:ToyExample\]Learned modes with Euclidean distances and diffusion distances. The Euclidean and diffusion distances from $(0,1)$ are shown in subfigures (b), (c) of Figure \[fig:NonlinearComparison\], while the corresponding learned modes are labeled, with nearby points colored red in subfigures (a), (b), of the present figure. Notice that the proposed diffusion learning method, using diffusion distances, correctly learns $M_{1}, M_{2}$ from different clusters (b), while using Euclidean distances leads assigning both $M_{1}, M_{2}$ to the same cluster (a), which would lead to poor clustering results.](Images/ToyExamples/ParabolicDiffusionModes-crop.pdf "fig:"){width="\textwidth"} Consider the simplified example in Figure \[fig:ToyExample\], showing the same data set as that in Figure \[fig:NonlinearComparison\]. The points of high density lie close to the center of $\mu_2$, and close to the two ends of the support of $\mu_1$. After computing an empirical density estimate, the distance between high density points is computed. If Euclidean distance is used to remove spurious modes, i.e. modes corresponding to the same distribution, then the learned modes $M_{1}, M_{2}$ both correspond to $\mu_2$; see subfigure (a) of Figure \[fig:ToyExample\]. When diffusion distance is used rather than Euclidean distance, the learned modes $M_{1}, M_{2}$ correspond to two different classes; see subfigure (b) of Figure \[fig:ToyExample\]. This is because the modes on the opposite ends of the support of $\mu_2$ are far in Euclidean distance but relatively close in diffusion distance. Furthermore, the substantial region of low density between the two distributions forces the diffusion distance between them to be relatively large. This suggests that diffusion distance is more useful than Euclidean distance for comparing high density points for the determination of modes, under the assumption that multimodal regions have modes that are connected by regions of not-too-low density. The results of the proposed clustering algorithm, as well a low-dimensional representation of diffusion distances, appears in Figure \[fig:ToyExample\_Learning\]. In the low-dimensional embedding corresponding to diffusion distance coordinates, the parabolic segment is linear and compressed, enabling the correct learning of modes. Labels are then assigned according to these modes in the diffusion coordinates, which can be projected back onto the original data to yield a clustering of the original data. [.16]{} ![\[fig:ToyExample\_Learning\]In Subfigure (a), the data from Figure \[fig:NonlinearComparison\] is represented in a new coordinate system, given by the second and third eigenfunctions of a Markov transition matrix. In this coordinate system, the natural Euclidean distance is equal to the diffusion distance on the original image. It is seen that the two ends of the parabolic segment are much closer in this embedding than in the original data, owing to the many short paths connecting them. The learned modes are labeled in this low-dimensional embedding as in Figure \[fig:ToyExample\], subfigure (b). In subfigure (b) of the present figure, points are labeled according to the proposed algorithm based on diffusion distance and the learned modes. Subfigure (c) shows the labels projected onto the original data, which conforms closely with the cluster structure in the data and the labels in Figure \[fig:NonlinearComparison\] (a).](Images/ToyExamples/ParabolicDiffusionEmbedding-crop.pdf "fig:"){width="\textwidth"} [.16]{} ![\[fig:ToyExample\_Learning\]In Subfigure (a), the data from Figure \[fig:NonlinearComparison\] is represented in a new coordinate system, given by the second and third eigenfunctions of a Markov transition matrix. In this coordinate system, the natural Euclidean distance is equal to the diffusion distance on the original image. It is seen that the two ends of the parabolic segment are much closer in this embedding than in the original data, owing to the many short paths connecting them. The learned modes are labeled in this low-dimensional embedding as in Figure \[fig:ToyExample\], subfigure (b). In subfigure (b) of the present figure, points are labeled according to the proposed algorithm based on diffusion distance and the learned modes. Subfigure (c) shows the labels projected onto the original data, which conforms closely with the cluster structure in the data and the labels in Figure \[fig:NonlinearComparison\] (a).](Images/ToyExamples/ParabolicLearnedLabels-crop.pdf "fig:"){width="\textwidth"} [.15]{} ![\[fig:ToyExample\_Learning\]In Subfigure (a), the data from Figure \[fig:NonlinearComparison\] is represented in a new coordinate system, given by the second and third eigenfunctions of a Markov transition matrix. In this coordinate system, the natural Euclidean distance is equal to the diffusion distance on the original image. It is seen that the two ends of the parabolic segment are much closer in this embedding than in the original data, owing to the many short paths connecting them. The learned modes are labeled in this low-dimensional embedding as in Figure \[fig:ToyExample\], subfigure (b). In subfigure (b) of the present figure, points are labeled according to the proposed algorithm based on diffusion distance and the learned modes. Subfigure (c) shows the labels projected onto the original data, which conforms closely with the cluster structure in the data and the labels in Figure \[fig:NonlinearComparison\] (a).](Images/ToyExamples/ParabolicLearnedLabels_OriginalSpace-crop.pdf "fig:"){width="\textwidth"} Diffusion Distance {#subsec:DD} ------------------ We now present an overview of diffusion distances. Additional analysis and comments on implementation appear in [@Coifman2005; @Coifman2006]. Diffusion processes on graphs lead to a data-dependent notion of distance, known as *diffusion distance*. This notion of distance has been applied to a variety of application problems, including analysis of stochastic and dynamical systems [@Coifman2005; @Nadler2006; @Coifman2008; @Singer2008], semisupervised learning [@Belkin2003; @SMC:GeneralFrameworkAdaptiveRegularization], data fusion [@Lafon2006; @Czaja2016], latent variable separation [@Lederman2015_1; @Lederman2015_2], and molecular dynamics [@Rohrdanz2011; @Zheng2011]. *Diffusion maps* provide a way of computing and visualizing diffusion distances, and may be understood as a type of nonlinear dimension reduction, in which data in a high number of dimensions may be embedded in a low-dimensional space by a nonlinear coordinate transformation. In this regard, diffusion maps are related to nonlinear dimension reduction techniques such as isomap [@Tenenbaum2000], Laplacian eigenmaps [@Belkin2003], and local linear embedding [@Roweis2000], among several others. Consider a discrete set $X=\{x_{n}\}_{n=1}^N\subset\mathbb{R}^{D}$. The diffusion distance [@Coifman2005; @Coifman2006] between $x,y\in X$, denoted $d_{t}(x,y)$, is a notion of distance that incorporates and is uniquely determined by the underlying geometry of $X$. The distance depends on a time parameter $t$, which enjoys an interpretation in terms of diffusion on the data. The computation of $d_{t}$ involves constructing a weighted, undirected graph $\mathcal{G}$ with vertices corresponding to the $N$ points in $X$, and weighted edges given by the $N\times N$ weight matrix $$\begin{aligned} \label{eqn:W}W(x,y):=\begin{cases}e^{-\frac{\|x-y\|_{2}^{2}}{\sigma^{2}}}, & x\in NN_{k}(y) \\ 0, & \text{else} \end{cases}\,,\end{aligned}$$for some suitable choice of $\sigma$ and with $NN_{k}(x)$ the set of $k$-nearest neighbors of $y$ in $X$ with respect to Euclidean distance. A fast nearest neighbors algorithm yields $W$ in quasilinear time in $N$ for $k$ small (see Section \[subsec:CC\] for details). The degree of $x$ is $\deg(x):=\sum_{y\in X}W(x,y).$ A Markov diffusion, representing a random walk on $\mathcal{G}$ (or $X$) has $N\times N$ transition matrix $P(x,y)={W(x,y)}\big/{\deg(x)}\,.$ For an initial distribution $\mu\in \mathbb{R}^{N}$ on $X$, the vector $\mu P^{t}$ is the probability over states at time $t\ge 0$. As $t$ increases, this diffusion process on $X$ evolves according to the connections between the points encoded by $P$. This Markov chain has a stationary distribution $\pi$ s.t. $\pi P=\pi$, given by $\pi(x)={\deg(x)}/{\sum_{y\in X}\deg(y)}$. The *diffusion distance at time $t$* is $$d^{2}_{t}(x,y):=\sum\nolimits_{u\in X} (P^{t}(x,u)- P^{t}(y,u))^{2}{d\mu(u)}/{\pi(u)}\,. \label{e:diffdist}$$ The computation of $d_{t}(x,y)$ involves summing over all paths of length $t$ connecting $x$ to $y$, so $d_{t}(x,y)$ is small if $x,y$ are strongly connected in the graph according to $P^{t}$, and large if $x,y$ are weakly connected in the graph. The eigendecomposition of $P$ allows to derive fast algorithms to compute $d_{t}$: the matrix $ P$ admits a spectral decomposition (under mild conditions, see [@Coifman2006]) with eigenvectors $\{\Phi_{n}\}_{n=1}^{N}$ and eigenvalues $\{\lambda_{n}\}_{n=1}^{N}$, where $1=\lambda_{1}\ge |\lambda_{2}|\ge \dots\ge|\lambda_{N}|$. The diffusion distance can then be written as $$\begin{aligned} \label{eqn:DD_eigen}d_{t}^2(x,y)={\sum\nolimits_{n=1}^{N}\lambda_{n}^{2t}(\Phi_{n}(x)-\Phi_{n}(y))^{2}}\,.\end{aligned}$$ The weighted eigenvectors $\{\lambda_{n}^{t}\Phi_{n}\}_{n=1}^{N}$ are new data-dependent coordinates of $X$, which are in fact close to being geometrically intrinsic [@Coifman2005]. Euclidean distance in these new coordinates is diffusion distance on $\mathcal{G}$. Diffusion distances are parametrized by $t$, which measures how long the diffusion process on $\mathcal{G}$ has run when the distances are computed. Small values of $t$ allow a small amount of diffusion, which may prevent the interesting geometry of $X$ from being discovered, but provide detailed, fine scale information. Large values of $t$ allow the diffusion process to run for so long that the fine geometry may be washed out. In this work an intermediate regime is typically when the diffusion geometry of the data is most useful; in all our experiments we set $t=30$. The choices of $\sigma,k,t$ in the construction of $W$ are in general important, see Section \[subsec:ParameterAnalysis\]. Note that under the mild condition that the underlying graph $\mathcal{G}$ is connected, $|\lambda_n|< 1$ for $n>1$. Hence, $|\lambda_{n}^{2t}|\ll1$ for large $t$ and $n> 1$, so that the sum (\[eqn:DD\_eigen\]) may approximated by its truncation at some suitable $2\le M\ll N$. In our experiments, $M$ was set to be the value at which the decay of the eigenvalues $\{\lambda_{n}\}_{n=1}^{N}$ begins to decrease; this is a standard heuristic for diffusion maps. The subset $\{\lambda_{n}^{t}\Phi_{n}\}_{n=1}^{M}$ used in the computation of $d_{t}$ is a dimension-reduced set of diffusion coordinates. The truncation also enables us to compute only a few eigenvectors, reducing computational complexity, see Section \[subsec:CC\]. In this sense, the mapping $$\begin{aligned} \label{eqn:DiffusionMapsDR}x\mapsto (\lambda_{1}^{t}\Phi_{1}(x),\lambda_{2}^{t}\Phi_{2}(x),\dots,\lambda_{M}^{t}\Phi_{M}(x))\end{aligned}$$ is a dimension reduction mapping of the ambient space $\mathbb{R}^{D}$ to $\mathbb{R}^{M}$. Unsupervised HSI Clustering Algorithm Description {#subsec:AlgorithmDescription} ------------------------------------------------- We now discuss the proposed HSI clustering algorithm in detail; see Figure \[fig:Flowchart\] for a flowchart representation. \ 0.2cm [max width=0.45]{} (Init) [Input: $X=\{x_{n}\}_{n=1}^{N}$, $K$; Parameters: $t, r_{s}$]{}; (ComputeEstimator) [Compute $p, \rho_{t},\mathcal{D}_{t}$ ( Equations (\[eqn:p\]), (\[eqn:D\_t\]))]{}; (ModeLabels)[Compute, label modes (Algorithm \[alg:modes\])]{}; (SortDensity)[Order unlabeled $\{x_{n}\}_{n=1}^{N}$ according to $p$]{}; (Iteration1)[For each unlabeled point:]{}; (SpectralLabel) [Compute $y^{\text{spectral}}_{n}$]{}; (SpatialLabel) [Compute $y^{\text{spatial}}_{n}$ (Equation (\[eqn:SpatialConsensus\]))]{}; (LabelAgreement) [$y_{n}^{\text{spatial}}$ exists\ and $y_{n}^{\text{spectral}}\neq y_{n}^{\text{spatial}}$?]{}; (NoConsensus)[$x_{n}$ not labeled]{}; (Consensus)[$y_{n}=y_{n}^{\text{spectral}}$]{}; (Iteration2)[For each still unlabeled point:]{}; (NewSpatialLabel) [Recompute $y^{\text{spatial}}_{n}$ (Equation (\[eqn:SpatialConsensus\]))]{}; (SpatialLabelExists?) [$y_{n}^{\text{spatial}}$ exists?]{}; (AssignSpatial)[$y_{n}=y_{n}^{\text{spatial}}$]{}; (AssignSpectral)[$y_{n}=y_{n}^{\text{spectral}}$]{}; (Init) – (ComputeEstimator); (ComputeEstimator)–(ModeLabels); (ModeLabels)–(SortDensity); (SortDensity)–node\[anchor=east\][Algorithm \[alg:labels\] (Stage 1)]{}(Iteration1); (Iteration1) – (SpectralLabel); (Iteration1) – (SpatialLabel); (SpectralLabel)–(LabelAgreement); (SpatialLabel)–(LabelAgreement); (LabelAgreement)– node\[anchor=south\][yes]{} (NoConsensus); (LabelAgreement)– node\[anchor=south\] [no]{} (Consensus); (LabelAgreement)–node\[anchor=east\][Algorithm \[alg:labels\] (Stage 2)]{}(Iteration2); (Iteration2) – (NewSpatialLabel); (NewSpatialLabel)–(SpatialLabelExists?); (SpatialLabelExists?)–node\[anchor=south\][yes]{}(AssignSpatial); (SpatialLabelExists?)–node\[anchor=south\][no]{}(AssignSpectral); Let $X=\{x_{n}\}_{n=1}^{N}\subset \mathbb{R}^{D}$ be the HSI, and let $K$ be the number of clusters. As described in Section \[s:overview\], our algorithm proceeds in two major steps: mode identification and labeling of points. The algorithm for learning the modes of the classes is summarized in Algorithm \[alg:modes\]. It first computes an empirical density for each point $x_n$ with a kernel density estimator $$\begin{aligned} \label{eqn:p} p(x_{n})={p_{0}(x_{n})}\big/{\sum\nolimits_{m=1}^{N}p_{0}(x_{m})},\end{aligned}$$ where $ p_{0}(x_n)=\sum\nolimits_{x_m\in NN_{k}(x_n)}e^{{-\|x_{n}-x_{m}\|_{2}^{2}}/{\sigma_{1}^{2}}} $. Here $\|x_{n}-x_{m}\|_{2}$ is the Euclidean distance in $\mathbb{R}^D$, and $NN_{k}(x_n)$ is the set of $k$-nearest neighbors to $x_n$, in Euclidean distance. The use of the Gaussian kernel density estimator is standard, enjoying strong theoretical guarantees [@Sheather1991; @Friedman2001] but certainly other estimators may be used. In our experiments we set $k=20$, though our method is robust to choosing larger $k$. The parameter $\sigma_{1}$ in the exponential kernel is set to be one twentieth the mean distance between all points (one could use the median instead in the presence of outliers). Once the empirical density $p$ is computed, the modes of the HSI classes are computed in a manner similar in spirit to [@Rodriguez2014], but employing diffusion distances. We compute the time-dependent quantity $\tilde\rho_{t}$ that assigns, to each pixel, the minimum diffusion distance between the pixel and a point of higher empirical density: $$\begin{aligned} \tilde\rho_{t}(x_n)=\begin{cases}\min\limits_{\{p(x_m)\ge p(x_n)\}}\!\!\!\!d_{t}(x_{n},x_{m}), & x_n\neq \operatorname*{arg\,max}_{i} p(x_i) \\ \max_{x_m}d_{t}(x_{n},x_{m}), & x_n = \operatorname*{arg\,max}_{i} p(x_i)\end{cases}\,,\end{aligned}$$ where $d_{t}(x_{m},x_{n})$ is the diffusion distance between $x_{m}, x_{n}$, at time $t$. In the following we will use the normalized quantity $\rho_{t}(x_n)=\tilde\rho_t(x_n)/\max_{x_m}\tilde\rho_{t}(x_m)$, which has maximum value $1$. The modes of the HSI are computed as the points $x_{1}^{*},\dots,x_{K}^{*}$ yielding the $K$ largest values of the quantity $$\begin{aligned} \label{eqn:D_t}\mathcal{D}_{t}(x_n)=p(x_n)\rho_{t}(x_n)\,.\end{aligned}$$ Such points should be both high density and far in diffusion distance from any other higher density points, and can therefore be expected to be modes of different cluster distributions. This method provably detects modes correctly under certain distributional assumptions on the data [@Murphy2018]. *Input*: $X, K$; $t$.\ Compute the empirical density $p(x_{n})$ for each $x_n\in X$.\ Compute $\{\rho_{t}(x_{n})\}_{n=1}^{N}$, the diffusion distance from each point to its nearest neighbor in diffusion distance of higher empirical density, normalized.\ Set the learned modes $\{x_{i}^{*}\}_{i=1}^{K}$ to be the $K$ maximizers of $\mathcal{D}_{t}(x_{n})=p(x_{n})\rho_{t}(x_{n})$.\ *Output:* $\{x_{i}^{*}\}_{i=1}^{K}, \{p(x_{n})\}_{n=1}^{N}, \{\rho_{t}(x_{n})\}_{n=1}^{N}$. Once the modes are detected, each is given a unique label. All other points are labeled using these mode labels in the following two-stage process, summarized in Algorithm \[alg:labels\]. In the first stage, running in order of decreasing empirical density, the *spatial consensus label* of each point is computed by finding all labeled points within distance $r_{s}\ge 0$ in the spatial domain of the pixel in question; call this set $NN^{s}_{r_{s}}(x_{n})$. If one label among $NN^{s}_{r_{s}}$ occurs with relative frequency $>1/2$, that label is the spatial consensus label. Otherwise, no spatial consensus label is given. In detail, let $L^{\text{spatial}}_{n}=\{y_{m} \ | \ x_{m}\in NN^{s}_{r_{s}}(x_{n}), x_{m}\neq x_{n}\}$ denote the labels of the spatial neighbors within radius $r_{s}$. Then the spatial consensus label of $x_i$ is $$\begin{aligned} \label{eqn:SpatialConsensus} y_{i}^{\text{spatial}}=\begin{cases}k, & \frac{|\{y_{n}| y_{n}=k, \ y_{n}\in L^{\text{spatial}}_{n}\}|}{|L_{n}^{\text{spatial}}|}>\frac12,\\ 0\ (\text{no label}), & \text{else}.\end{cases}\end{aligned}$$ After a point’s spatial consensus label is computed, its *spectral label* is computed as its nearest neighbor in the spectral domain, measured in diffusion distance, of higher density. The point is then given the overall label of the spectral label unless the spatial consensus label exists (i.e. is $\neq0$ in ) and differs from the spatial consensus label. In this case, the point in question remains unlabeled in the first stage. Note that points that are unlabeled are considered to have label 0 for the purposes of computing the spatial consensus label, so in the case that most pixels in the spatial neighborhood are unlabeled, the spatial consensus label will be 0. Hence, only pixels with many labeled pixels in their spatial neighborhood can have a consensus spatial label. In this first stage, a label is only assigned based on spectral information, though the spatial information may prevent a label from being assigned. Upon completion of the first stage, the dataset will be partially labeled; see Figure \[fig:SpectralSpatialLabeling\]. In the second stage, an unlabeled point is given the label of its spatial consensus label, if it exists, or otherwise the label of its nearest spectral neighbor of higher density. Thus, in the second stage, a label is assigned based on joint spectral-spatial information. *Input:* $\{x_{i}^{*}\}_{i=1}^{K}, \{p(x_{n})\}_{n=1}^{N}$, $\{\rho_{t}(x_{n})\}_{n=1}^{N}$; $r_{s}$.\ Assign each mode a unique label.\ *Stage 1*: Iterating through the remaining unlabeled points in order of decreasing density among unlabeled points, assign each point the same label as its nearest spectral neighbor (in diffusion distance) of higher density, unless the spatial consensus label exists and differs, in which case the point is left unlabeled.\ *Stage 2*: Iterating in order of decreasing density among unlabeled points, assign each point the consensus spatial label, if it exists, otherwise the same label as its nearest spectral neighbor of higher density.\ *Output:* Labels $\{y_{n}\}_{n=1}^{N}$. Points of high density are likely to be labeled according to their spectral properties. The reasons for this are twofold. First, these points are likely to be near the centers of distributions, and hence are likely to be in spatially homogeneous regions. Second, points of high density are labeled before points of low density, so it is unlikely for high density points to have many labeled points in their spatial neighborhoods. This means that the spatial consensus label is unlikely to exist for these points. Conversely, points of low density may be at the boundaries of the classes, and are hence more likely to be labeled by their spatial properties. The incorporation of spatial information into machine learning for HSI is justified by the fact that HSI images typically show some amount of spatial regularity, in that if a pixel’s nearest spatial neighbors all have the same class label, it is likely that the pixel has this same label, compared to the case in which the pixel’s nearest spatial neighbors have random labels [@Fauvel2008; @Li2013_2; @Zhang2016; @Tarabalka2009; @Benedetto2012; @Fauvel2013; @Cahill2014; @Cloninger2014; @Wang2014; @Benedetto2016]. The spatial information regularizes and improves performance, but it cannot take the place of the spectral information, as shall be seen in Section \[subsec:SpaceParameter\]: the spectral information is more discriminative than the spatial information, and is the more important of the two. The proposed method, combining Algorithms \[alg:modes\], \[alg:labels\] is called *spectral-spatial diffusion learning (DLSS)*. In our experimental analysis, the significance of the spectral-spatial labeling scheme is validated by comparing DLSS against a simpler method, called *diffusion learning (DL)*. This method learns class modes as in Algorithm \[alg:modes\], but labels all pixels simply by requiring each point have the same label as its nearest spectral neighbor of higher density. The expectation is that DLSS will generally outperform DL, due to the former’s incorporation of spatial data; this is confirmed by our experiments. Active Learning DLSS Variation {#subsec:ActiveLearning} ------------------------------ Both the DL and DLSS methods are unsupervised. We now present a variation of the DLSS method for active learning of hyperspectral images, where a few well-chosen pixels are automatically selected for labeling. The DLSS method labels points beginning with the learned class modes, and mistakes tend to be made on points that are near the class boundaries; in the active learning scheme the algorithm will ask for the labels of the points whose distances from their nearest two modes are closest. That is, points whose nearest mode is ambiguous will be labeled using training data, and all other points will be labeled as in the DLSS algorithm. More precisely, we fix a time $t$, and for each pixel $x_{n}$, let $x_{n_{1}}^{*}, x_{n_{2}}^{*}$ be the two modes closest to $x_{n}$ in diffusion distance $d_t$. We compute the quantity $$\begin{aligned} \label{eqn:ActiveQuantity}F_{t}(x_{n})=|d_{t}(x_{n},x_{n_{1}}^{*})-d_{t}(x_{n},x_{n_{2}}^{*})|.\end{aligned}$$If $F_{t}(x_{n})$ is close to 0, then there is substantial ambiguity as to the nearest mode to $x_{n}$. Suppose the user is afforded the labels of exactly $L$ points. Then the $L$ labels requested in our active learning regime are the $L$ minimizers of $F_{t}$. The proposed active learning scheme is summarized in Algorithm \[alg:ActiveLearning\]. To evaluate performance, we consider a range of $L$ values in our experiments. The active learning setting is most interesting when $\alpha=L/N$ is very small, where $N$ is the total number of pixels in the image. *Input:* $X,K$; $t,r_{s},L$.\ Compute the modes of the data using Algorithm \[alg:modes\].\ Give each mode a unique label.\ Compute, for each point $x_{n}$, $F_{t}(x_{n})$ as in .\ Label the $L$ minimizers of $F_{t}$ with ground truth labels.\ Label the remaining, unlabeled points as in steps 3, 4 in Algorithm \[alg:labels\].\ *Output:* Labels $\{y_{n}\}_{n=1}^{N}$.\ Note that the active learning algorithm can be iterated, by labeling points then recomputing the quantity to determine the most challenging points after some labels have been introduced [@Murphy2018_iterative]. Experiments {#sec:Experiments} =========== Algorithm Evaluation Methods and Experimental Data -------------------------------------------------- We consider several HSI datasets to evaluate the proposed unsupervised (Algorithms \[alg:modes\], \[alg:labels\]) and active learning (Algorithm \[alg:ActiveLearning\]) algorithms. For evaluation in the presence of ground truth (GT), we consider three quantitative measures, besides visual performance, namely: 1. *Overall Accuracy (OA)*: Total number of correctly labeled pixels divided by the total number of pixels. This method values large classes more than small classes. 2. *Average Accuracy (AA)*: The average, over classes, of the OA of each class. This method values small classes and large classes equally. 3. *Cohen’s* $\kappa$-*statistic* ($\kappa$): A measurement of agreement between two labelings, corrected for random agreement [@Banerjee1999]. Letting $a_{o}$ be the observed agreement between the labeling and the ground truth and $a_{e}$ the expected agreement between a uniformly random labeling and the ground truth, $\kappa={(a_{o}-a_{e})}/{(1-a_{e})}$. $\kappa=1$ corresponds to perfect overall accuracy, $\kappa\le0$ corresponds to labels no better than what is expected from random guessing. In order to perform quantitative analysis with these metrics and make consistent visual comparisons, the learned clusters are aligned with ground truth, when available. More precisely, let $S_{K}$ be the set of permutations of $\{1,2,\dots,K\}$. Let $\{C_{i}\}_{i=1}^{K}$ be the clusters learned from one of the clustering methods, and let $\{C_{i}^{GT}\}_{i=1}^{K}$ be the ground truth clusters. Cluster $C_{i}$ is assigned label $\hat{\eta}_{i}\in\{1,2,\dots,K\}$, with $\hat{\eta} = \operatorname*{arg\,max}_{\eta=(\eta_{1},\dots,\eta_{K})\in S_{K}} \sum_{i=1}^{K}|C_{\eta_{i}}\cap C_{i}^{GT}|$. We remark that while this alignment method maximizes the overall accuracy of the labeling and is most useful for visualization, better alignments for maximizing $AA$ and $\kappa$ may exist. We consider $4$ real HSI datasets to shed light on strengths and weaknesses of the proposed algorithm. These datasets are standard, have ground truth, and are publicly available[^1]. Experiments with active learning are performed for these same real HSI datasets with Algorithm \[alg:ActiveLearning\]. Additional experiments on synthetic and real HSI data are available, for conciseness, only in an appendix in the online preprint version. Note that some images are restricted to subsets in the spatial domain, which is noted in their respective subsections. This is because unsupervised methods for HSI struggle with data containing a large number of classes, due to variation within classes and similarity between certain end-members of different classes [@Zhu2017]. Hence, the Indian Pines, Pavia, and Kennedy Space Center datasets are restricted to reduce the number of classes and achieve meaningful clusters. The Salinas A dataset is considered in its entirety. The ground truth, when available, is often incomplete, i.e. not all pixels are labeled. For these datasets, labels are computed for all data, then the pixels with ground truth labels are used for quantitative and visual analysis. The number of class labels in the ground truth images were used as parameter $K$ for all clustering algorithms, though the proposed method automatically estimates the number of clusters; see Section \[sec:Future\]. Grayscale images of the projection of the data onto its first principal component and images of ground truth (GT), colored by class, for the Indian Pines, Pavia, Salinas A, and Kennedy Space Center datasets are in Figures \[fig:IP\], \[fig:Pavia\], \[fig:SalinasA\], and \[fig:KSC\], respectively. The projection onto the first principal component of the data is presented as a simple visual summary of the data, though it washes out the subtle information presented in individual bands. Since the proposed and comparison methods are unsupervised, experiments are performed on the entire dataset, including points without ground truth labels. The labels for pixels without ground truth are not accounted for in the quantitative evaluation of the algorithms tested. Note that additional experiments, not shown, were performed, using only the data with ground truth labels. These experiments consisted in restricting the HSI to the pixels with labels, which makes the clustering problem significantly easier. Quantitative results were uniformly better for all datasets and methods in these cases; the relative performances of the algorithms on a given dataset remained the same. Comparison Methods {#s:comparisonmethods} ------------------ We consider a variety of benchmark and state-of-the-art methods of HSI clustering for comparison. First, we consider the classic *$K$-means* algorithm [@Friedman2001] applied directly to $X$. This method is not expected to perform well on HSI data, due to the non-spherical shape of clusters, high dimensionality, and noise, all well-known problems for $K$-means. Several [*dimension reduction*]{} methods to reduce the dimensionality of the data, while preserving important discriminatory properties of the classes, as well as increasing the signal-to-noise ratio in the projected space, are also used as benchmarks for comparison with the proposed method. These methods first reduce the dimension of the data from $D$ to $K_{GT}\ll D$, where $K_{GT}$ is the number of classes, then run $K$-means on the reduced data. We consider linear dimension reduction via [*[principal component analysis]{} (PCA)*]{}; [*[independent component analysis]{} (ICA)*]{} [@Comon1994; @Hyvarinen2000], using the fast implementation [@Hyvarinen1999][^2]; and [*[random projections]{}*]{} via Gaussian random matrices, shown to be efficient in highly simplified data models [@Dasgupta2000; @Candes2006]. We also consider more computationally intensive methods for benchmarking. *DBSCAN* [@Ester1996] is a popular density-based clustering method, that although highly parameter-dependent, has proved useful for a variety of unsupervised tasks. *Spectral clustering* (SC) [@Ng2001; @VonLuxburg2007] has been applied with success in classification and clustering HSI [@Fauvel2013]. The spectral embedding consists of the top $K_{GT}$ row-normalized eigenvectors of the normalized graph Laplacian $L$; in this features space $K$-means is then run (see Section \[subsec:ParameterAnalysis\]). We also cluster with [*[Gaussian mixture models]{}*]{} (GMM) [@Acito2003; @Manolakis2001; @Kraut2001], with parameters determined by expectation maximization (EM). Finally we consider several recent, state-of-the-art clustering methods: *sparse manifold clustering and embedding (SMCE)* [@Elhamifar2011; @Elhamifar2013][^3], which fits the data to low-dimensional, sparse structures, and then applies spectral clustering; *hierarchical clustering with non-negative matrix factorization (HNMF)* [@Gillis2015][^4], which has shown excellent performance for HSI clustering when the clusters are generated from a single endmember; a graph-based method based on the Mumford-Shah segmentation [@Mumford1989][@Meng2017], related to spectral clustering, and called *fast Mumford-Shah (FMS)* in this article (we use a highly parallelized version[^5]); *fast search and find of density peaks clustering* (FSFDPC) algorithm [@Rodriguez2014], which has been shown effective in clustering a variety of data sets. Relationship Between Proposed Method and Comparison Methods ----------------------------------------------------------- The FSFDPC method has similarities with the mode estimation aspect of our work, in that both algorithms attempt to learn the modes of the classes via a density-based analysis, as described in, for example, [@Chen2017; @Rodriguez2014]. Our method is quite different, however: the proposed measure of distance between high density points is not Euclidean distance, but diffusion distance [@Coifman2005; @Coifman2006], which is more adept at removing spurious modes, due to its incorporation of the geometry of the data. This phenomenon is illustrated in Figures \[fig:NonlinearComparison\],\[fig:ToyExample\]. The assignment of labels from the modes is also quite different, as diffusion distances are used to determine spectral nearest neighbors, and spatial information is accounted for in our DLSS algorithm. FSFDPC, in contrast, assigns to each of the modes its own class label, and to the remaining points a label by requiring that each point has the same label as its Euclidean nearest neighbor of higher density. This means that FSFDPC only incorporates spectral information measured in Euclidean distance, disregarding spatial information. The benefits of both using diffusion distances to learn modes, and incorporating spatial proximities into the clustering process are very significant, as the experiments demonstrate. Both FSFDPC and the proposed algorithm have some similarities to DBSCAN which, however, performs poorly for data with clusters of differing densities, and is highly sensitive to its parameters. Note that FSFDPC was in fact proposed to improve on these drawbacks of DBSCAN [@Rodriguez2014]. The proposed DLSS and DL algorithms also share commonalities with spectral clustering, SMCE, and FMS in that these comparison methods compute eigenvectors of a graph Laplacian in order to develop a nonlinear notion of distance. This is related to computing the eigenvectors of the Markov transition matrix in the computation of diffusion maps. The proposed method, however, directly incorporates density into the detection of modes, which allows for more robust clustering compared to these methods, which work by simply applying $K$-means to the eigenvectors of the graph Laplacian. Moreover, our technique does not rely on any assumption about sparsity (unlike SMCE), and is completely invariant under distance-preserving transformations (it shares this property with SMCE), which could be useful if different imaging modalities (e.g. compressed modalities) were used. Additionally, our approach is connected to semisupervised learning techniques on graphs, where initial given labels are propagated by a diffusion process to other vertices (points); see [@SMC:GeneralFrameworkAdaptiveRegularization] and references therein. Here of course we have proceeded in an unsupervised fashion, replacing initial given labels by estimated modes of the clusters. Summary of Proposed and Comparison Methods ------------------------------------------ The experimental methods are summarized in Table \[tab:Methods\]. The two novel methods we proposed are the full spectral-spatial diffusion learning method (DLSS), as well as a simplified diffusion learning method (DL). We note that several algorithms were not implemented by the authors of this article: publicly available libraries were used when available. Links to these libraries are noted where appropriate. [max width=.49]{} Method D.R. Metric --------------------------------------------------------------------------- ------ ----------- $K$-means on full dataset No Euclidean $K$-means on PCA reduced dataset Yes Euclidean $K$-means on ICA reduced dataset Yes Euclidean $K$-means on data reduced by random projections Yes Euclidean DBSCAN [@Ester1996] No Euclidean Spectral clustering [@VonLuxburg2007] Yes Spectral Gaussian mixture models No Euclidean Sparse manifold clustering and embedding [@Elhamifar2011; @Elhamifar2013] Yes Spectral Hierarchical NMF [@Gillis2015] No Euclidean Fast Mumford Shah [@Meng2017] No Spectral FSFDPC [@Rodriguez2014] No Euclidean **Diffusion learning (DL)** Yes Diffusion **Spectral-spatial diffusion learning (DLSS)** Yes Diffusion : \[tab:Methods\]Methods used for experimental analysis, along with whether the method employs dimensionality reduction and which metric is used to compared points. The methods proposed in this article appear in bold. Note that the proposed methods employ dimension reduction, as illustrated in (\[eqn:DiffusionMapsDR\]). All experiments and subsequent analyses, except those involving FMS, were performed in MATLAB running on a 3.1 GHz Intel 4-Core i7 processor with 16 GB of RAM; code to reproduce all results is available on the authors’ website[^6]. Unsupervised HSI Clustering Experiments {#subsec:HyperspectralClustering} --------------------------------------- ### Indian Pine Dataset {#subsubsec:IP} The Indian Pines dataset used for experiments is a subset of the full Indian Pines datasets, consisting of three classes that are difficult to distinguish visually; see Figure \[fig:IP\]. This dataset is expected to be challenging due to the lack of clear separation between the classes. Results for Indian Pines appear in Figure \[fig:ResultsIP\] and Table \[tab:Summary\]. ![\[fig:IP\]The Indian Pines data is a $50\times 25$ subset of the full Indian Pines dataset. It contains 3 classes, one of which is not well-localized spatially. The dataset was captured in 1992 in Northwest IN, USA by the AVRIS sensor. The spatial resolution is 20m/pixel. There are $200$ spectral bands. Left: projection onto the first principal component of the data; right: ground truth (GT).](Images/IP_Images/IP_FirstPC-crop.pdf "fig:"){width=".24\textwidth"} ![\[fig:IP\]The Indian Pines data is a $50\times 25$ subset of the full Indian Pines dataset. It contains 3 classes, one of which is not well-localized spatially. The dataset was captured in 1992 in Northwest IN, USA by the AVRIS sensor. The spatial resolution is 20m/pixel. There are $200$ spectral bands. Left: projection onto the first principal component of the data; right: ground truth (GT).](Images/IP_Images/IP_GT-crop.pdf "fig:"){width=".24\textwidth"} [ .09]{} ![\[fig:ResultsIP\]Clustering results for Indian Pines dataset. The impact of the spectral-spatial labeling scheme is apparent, as the labels for the DLSS method are more spatially regular than those of the DL method. Note that the regions of difference between DL and DLSS are primarily near boundaries of classes and in very small interior regions. Near the boundaries of classes, pixels are likely to be far from the spectral class cores, and hence are more likely to be labeled based on spatial properties. The small interior regions are unlikely to be formed under the DLSS labeling regime, since these regions consist of points whose spectral label differs from their spatial consensus label. The simplified DL method performs second best, and in particular outperforms FSFDPC, which performs well among the comparison methods.](Images/IP_Images/IP_Kmeans-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsIP\]Clustering results for Indian Pines dataset. The impact of the spectral-spatial labeling scheme is apparent, as the labels for the DLSS method are more spatially regular than those of the DL method. Note that the regions of difference between DL and DLSS are primarily near boundaries of classes and in very small interior regions. Near the boundaries of classes, pixels are likely to be far from the spectral class cores, and hence are more likely to be labeled based on spatial properties. The small interior regions are unlikely to be formed under the DLSS labeling regime, since these regions consist of points whose spectral label differs from their spatial consensus label. The simplified DL method performs second best, and in particular outperforms FSFDPC, which performs well among the comparison methods.](Images/IP_Images/IP_PCA-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsIP\]Clustering results for Indian Pines dataset. The impact of the spectral-spatial labeling scheme is apparent, as the labels for the DLSS method are more spatially regular than those of the DL method. Note that the regions of difference between DL and DLSS are primarily near boundaries of classes and in very small interior regions. Near the boundaries of classes, pixels are likely to be far from the spectral class cores, and hence are more likely to be labeled based on spatial properties. The small interior regions are unlikely to be formed under the DLSS labeling regime, since these regions consist of points whose spectral label differs from their spatial consensus label. The simplified DL method performs second best, and in particular outperforms FSFDPC, which performs well among the comparison methods.](Images/IP_Images/IP_ICA-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsIP\]Clustering results for Indian Pines dataset. The impact of the spectral-spatial labeling scheme is apparent, as the labels for the DLSS method are more spatially regular than those of the DL method. Note that the regions of difference between DL and DLSS are primarily near boundaries of classes and in very small interior regions. Near the boundaries of classes, pixels are likely to be far from the spectral class cores, and hence are more likely to be labeled based on spatial properties. The small interior regions are unlikely to be formed under the DLSS labeling regime, since these regions consist of points whose spectral label differs from their spatial consensus label. The simplified DL method performs second best, and in particular outperforms FSFDPC, which performs well among the comparison methods.](Images/IP_Images/IP_RP-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsIP\]Clustering results for Indian Pines dataset. The impact of the spectral-spatial labeling scheme is apparent, as the labels for the DLSS method are more spatially regular than those of the DL method. Note that the regions of difference between DL and DLSS are primarily near boundaries of classes and in very small interior regions. Near the boundaries of classes, pixels are likely to be far from the spectral class cores, and hence are more likely to be labeled based on spatial properties. The small interior regions are unlikely to be formed under the DLSS labeling regime, since these regions consist of points whose spectral label differs from their spatial consensus label. The simplified DL method performs second best, and in particular outperforms FSFDPC, which performs well among the comparison methods.](Images/IP_Images/IP_DBSCAN-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsIP\]Clustering results for Indian Pines dataset. The impact of the spectral-spatial labeling scheme is apparent, as the labels for the DLSS method are more spatially regular than those of the DL method. Note that the regions of difference between DL and DLSS are primarily near boundaries of classes and in very small interior regions. Near the boundaries of classes, pixels are likely to be far from the spectral class cores, and hence are more likely to be labeled based on spatial properties. The small interior regions are unlikely to be formed under the DLSS labeling regime, since these regions consist of points whose spectral label differs from their spatial consensus label. The simplified DL method performs second best, and in particular outperforms FSFDPC, which performs well among the comparison methods.](Images/IP_Images/IP_SC-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![\[fig:ResultsIP\]Clustering results for Indian Pines dataset. The impact of the spectral-spatial labeling scheme is apparent, as the labels for the DLSS method are more spatially regular than those of the DL method. Note that the regions of difference between DL and DLSS are primarily near boundaries of classes and in very small interior regions. Near the boundaries of classes, pixels are likely to be far from the spectral class cores, and hence are more likely to be labeled based on spatial properties. The small interior regions are unlikely to be formed under the DLSS labeling regime, since these regions consist of points whose spectral label differs from their spatial consensus label. The simplified DL method performs second best, and in particular outperforms FSFDPC, which performs well among the comparison methods.](Images/IP_Images/IP_GMM-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![\[fig:ResultsIP\]Clustering results for Indian Pines dataset. The impact of the spectral-spatial labeling scheme is apparent, as the labels for the DLSS method are more spatially regular than those of the DL method. Note that the regions of difference between DL and DLSS are primarily near boundaries of classes and in very small interior regions. Near the boundaries of classes, pixels are likely to be far from the spectral class cores, and hence are more likely to be labeled based on spatial properties. The small interior regions are unlikely to be formed under the DLSS labeling regime, since these regions consist of points whose spectral label differs from their spatial consensus label. The simplified DL method performs second best, and in particular outperforms FSFDPC, which performs well among the comparison methods.](Images/IP_Images/IP_SMCE-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![\[fig:ResultsIP\]Clustering results for Indian Pines dataset. The impact of the spectral-spatial labeling scheme is apparent, as the labels for the DLSS method are more spatially regular than those of the DL method. Note that the regions of difference between DL and DLSS are primarily near boundaries of classes and in very small interior regions. Near the boundaries of classes, pixels are likely to be far from the spectral class cores, and hence are more likely to be labeled based on spatial properties. The small interior regions are unlikely to be formed under the DLSS labeling regime, since these regions consist of points whose spectral label differs from their spatial consensus label. The simplified DL method performs second best, and in particular outperforms FSFDPC, which performs well among the comparison methods.](Images/IP_Images/IP_NMF-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![\[fig:ResultsIP\]Clustering results for Indian Pines dataset. The impact of the spectral-spatial labeling scheme is apparent, as the labels for the DLSS method are more spatially regular than those of the DL method. Note that the regions of difference between DL and DLSS are primarily near boundaries of classes and in very small interior regions. Near the boundaries of classes, pixels are likely to be far from the spectral class cores, and hence are more likely to be labeled based on spatial properties. The small interior regions are unlikely to be formed under the DLSS labeling regime, since these regions consist of points whose spectral label differs from their spatial consensus label. The simplified DL method performs second best, and in particular outperforms FSFDPC, which performs well among the comparison methods.](Images/IP_Images/IP_Bertozzi-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsIP\]Clustering results for Indian Pines dataset. The impact of the spectral-spatial labeling scheme is apparent, as the labels for the DLSS method are more spatially regular than those of the DL method. Note that the regions of difference between DL and DLSS are primarily near boundaries of classes and in very small interior regions. Near the boundaries of classes, pixels are likely to be far from the spectral class cores, and hence are more likely to be labeled based on spatial properties. The small interior regions are unlikely to be formed under the DLSS labeling regime, since these regions consist of points whose spectral label differs from their spatial consensus label. The simplified DL method performs second best, and in particular outperforms FSFDPC, which performs well among the comparison methods.](Images/IP_Images/IP_FSFDPC-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsIP\]Clustering results for Indian Pines dataset. The impact of the spectral-spatial labeling scheme is apparent, as the labels for the DLSS method are more spatially regular than those of the DL method. Note that the regions of difference between DL and DLSS are primarily near boundaries of classes and in very small interior regions. Near the boundaries of classes, pixels are likely to be far from the spectral class cores, and hence are more likely to be labeled based on spatial properties. The small interior regions are unlikely to be formed under the DLSS labeling regime, since these regions consist of points whose spectral label differs from their spatial consensus label. The simplified DL method performs second best, and in particular outperforms FSFDPC, which performs well among the comparison methods.](Images/IP_Images/IP_DL-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsIP\]Clustering results for Indian Pines dataset. The impact of the spectral-spatial labeling scheme is apparent, as the labels for the DLSS method are more spatially regular than those of the DL method. Note that the regions of difference between DL and DLSS are primarily near boundaries of classes and in very small interior regions. Near the boundaries of classes, pixels are likely to be far from the spectral class cores, and hence are more likely to be labeled based on spatial properties. The small interior regions are unlikely to be formed under the DLSS labeling regime, since these regions consist of points whose spectral label differs from their spatial consensus label. The simplified DL method performs second best, and in particular outperforms FSFDPC, which performs well among the comparison methods.](Images/IP_Images/IP_DLSS-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsIP\]Clustering results for Indian Pines dataset. The impact of the spectral-spatial labeling scheme is apparent, as the labels for the DLSS method are more spatially regular than those of the DL method. Note that the regions of difference between DL and DLSS are primarily near boundaries of classes and in very small interior regions. Near the boundaries of classes, pixels are likely to be far from the spectral class cores, and hence are more likely to be labeled based on spatial properties. The small interior regions are unlikely to be formed under the DLSS labeling regime, since these regions consist of points whose spectral label differs from their spatial consensus label. The simplified DL method performs second best, and in particular outperforms FSFDPC, which performs well among the comparison methods.](Images/IP_Images/IP_GT-crop.pdf "fig:"){width="\textwidth"} The proposed methods, DL and DLSS, perform the best, with DLSS strongly outperforming the rest. For the average accuracy statistic, DBSCAN performs as well as DL, indicating that the clusters for this data are likely of comparable empirical density. The use of diffusion distances for mode detection and determination of spectral neighbors is evidently useful, as DL significantly outperforms FSFDPC, which has among the best quantitative performance of the comparison methods. Moreover, the use of the proposed spectral-spatial labeling scheme DLSS clearly improves over spectral-only labeling DL: as seen in Figure \[fig:ResultsIP\], DLSS correctly labels many small interior regions that DL labels incorrectly. ### Pavia Dataset The Pavia dataset used for experiments consists of a subset of the original dataset, and contains six classes, with one of them spread out across the image. As can be seen in Figure \[fig:Pavia\], the yellow class is small and diffuse, which is expected to add challenge to this example. Results appear in Table \[tab:Summary\]. Visual results appear in the online preprint version of this article. ![\[fig:Pavia\]The Pavia data is a $270\times 50$ subset of the full Pavia dataset. It contains 6 classes, some of which are not well-localized spatially. The dataset was captured by the ROSIS sensor during a flight over Pavia, Italy. The spatial resolution is 1.3 m/pixel. There are 102 spectral bands. Left: projection onto the first principal component of the data; right: ground truth (GT).](Images/Pavia_Images/Pavia_FirstPC-crop.pdf "fig:"){width=".24\textwidth"} ![\[fig:Pavia\]The Pavia data is a $270\times 50$ subset of the full Pavia dataset. It contains 6 classes, some of which are not well-localized spatially. The dataset was captured by the ROSIS sensor during a flight over Pavia, Italy. The spatial resolution is 1.3 m/pixel. There are 102 spectral bands. Left: projection onto the first principal component of the data; right: ground truth (GT).](Images/Pavia_Images/Pavia_GT-crop.pdf "fig:"){width=".24\textwidth"} The proposed methods give the best results, which also provide evidence of the value of both the diffusion learning stage and the spectral-spatial labeling scheme. The proposed DLSS algorithm makes essentially only two errors: the yellow-green class is slightly mislabeled, and the blue-green class in the bottom right is labeled completely incorrectly. However, both of these errors are made by all algorithms, often to a greater degree. Among the comparison methods, SMCE performs best; classical spectral clustering also performs well. ### Salinas A Dataset {#subsubsec:SalinasA} The Salinas A dataset (see Figure \[fig:SalinasA\]) consists of 6 classes arrayed diagonally. Some pixels in the original images have the same values, so some small Gaussian noise (variance $< 10^{-3})$ was added as a preprocessing step to distinguish these pixels. ![\[fig:SalinasA\]The Salinas A data consists of the full $86\times 83$ HSI. It contains 6 classes, all of which are well-localized spatially. The dataset was captured over Salinas Valley, CA, by the AVRIS sensor. The spatial resolution is 3.7 m/pixel. The image contains 224 spectral bands. Left: projection onto the first principal component of the data; right: ground truth (GT).](Images/SalinasA_Images/SalinasA_FirstPC-crop.pdf "fig:"){width=".24\textwidth"} ![\[fig:SalinasA\]The Salinas A data consists of the full $86\times 83$ HSI. It contains 6 classes, all of which are well-localized spatially. The dataset was captured over Salinas Valley, CA, by the AVRIS sensor. The spatial resolution is 3.7 m/pixel. The image contains 224 spectral bands. Left: projection onto the first principal component of the data; right: ground truth (GT).](Images/SalinasA_Images/SalinasA_GT-crop.pdf "fig:"){width=".24\textwidth"} [.09]{} ![\[fig:ResultsSalinasA\]Clustering results for Salinas A dataset. The proposed method, DLSS performs best, with the simplified DL method and benchmark spectral clustering also performing well. Notice that the spectral-spatial labeling scheme removes some of the mistakes in the yellow cluster, and also improves the labeling near some class boundaries. However, it is not able to fix the mislabeling of the light blue cluster in the lower right. Indeed, all methods split the cluster in the lower right of the image, indicating the challenging aspects of this dataset for unsupervised learning.](Images/SalinasA_Images/SalinasA_Kmeans-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![\[fig:ResultsSalinasA\]Clustering results for Salinas A dataset. The proposed method, DLSS performs best, with the simplified DL method and benchmark spectral clustering also performing well. Notice that the spectral-spatial labeling scheme removes some of the mistakes in the yellow cluster, and also improves the labeling near some class boundaries. However, it is not able to fix the mislabeling of the light blue cluster in the lower right. Indeed, all methods split the cluster in the lower right of the image, indicating the challenging aspects of this dataset for unsupervised learning.](Images/SalinasA_Images/SalinasA_PCA-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![\[fig:ResultsSalinasA\]Clustering results for Salinas A dataset. The proposed method, DLSS performs best, with the simplified DL method and benchmark spectral clustering also performing well. Notice that the spectral-spatial labeling scheme removes some of the mistakes in the yellow cluster, and also improves the labeling near some class boundaries. However, it is not able to fix the mislabeling of the light blue cluster in the lower right. Indeed, all methods split the cluster in the lower right of the image, indicating the challenging aspects of this dataset for unsupervised learning.](Images/SalinasA_Images/SalinasA_ICA-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![\[fig:ResultsSalinasA\]Clustering results for Salinas A dataset. The proposed method, DLSS performs best, with the simplified DL method and benchmark spectral clustering also performing well. Notice that the spectral-spatial labeling scheme removes some of the mistakes in the yellow cluster, and also improves the labeling near some class boundaries. However, it is not able to fix the mislabeling of the light blue cluster in the lower right. Indeed, all methods split the cluster in the lower right of the image, indicating the challenging aspects of this dataset for unsupervised learning.](Images/SalinasA_Images/SalinasA_RP-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsSalinasA\]Clustering results for Salinas A dataset. The proposed method, DLSS performs best, with the simplified DL method and benchmark spectral clustering also performing well. Notice that the spectral-spatial labeling scheme removes some of the mistakes in the yellow cluster, and also improves the labeling near some class boundaries. However, it is not able to fix the mislabeling of the light blue cluster in the lower right. Indeed, all methods split the cluster in the lower right of the image, indicating the challenging aspects of this dataset for unsupervised learning.](Images/SalinasA_Images/SalinasA_DBSCAN-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![\[fig:ResultsSalinasA\]Clustering results for Salinas A dataset. The proposed method, DLSS performs best, with the simplified DL method and benchmark spectral clustering also performing well. Notice that the spectral-spatial labeling scheme removes some of the mistakes in the yellow cluster, and also improves the labeling near some class boundaries. However, it is not able to fix the mislabeling of the light blue cluster in the lower right. Indeed, all methods split the cluster in the lower right of the image, indicating the challenging aspects of this dataset for unsupervised learning.](Images/SalinasA_Images/SalinasA_SC-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![\[fig:ResultsSalinasA\]Clustering results for Salinas A dataset. The proposed method, DLSS performs best, with the simplified DL method and benchmark spectral clustering also performing well. Notice that the spectral-spatial labeling scheme removes some of the mistakes in the yellow cluster, and also improves the labeling near some class boundaries. However, it is not able to fix the mislabeling of the light blue cluster in the lower right. Indeed, all methods split the cluster in the lower right of the image, indicating the challenging aspects of this dataset for unsupervised learning.](Images/SalinasA_Images/SalinasA_GMM-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![\[fig:ResultsSalinasA\]Clustering results for Salinas A dataset. The proposed method, DLSS performs best, with the simplified DL method and benchmark spectral clustering also performing well. Notice that the spectral-spatial labeling scheme removes some of the mistakes in the yellow cluster, and also improves the labeling near some class boundaries. However, it is not able to fix the mislabeling of the light blue cluster in the lower right. Indeed, all methods split the cluster in the lower right of the image, indicating the challenging aspects of this dataset for unsupervised learning.](Images/SalinasA_Images/SalinasA_SMCE-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![\[fig:ResultsSalinasA\]Clustering results for Salinas A dataset. The proposed method, DLSS performs best, with the simplified DL method and benchmark spectral clustering also performing well. Notice that the spectral-spatial labeling scheme removes some of the mistakes in the yellow cluster, and also improves the labeling near some class boundaries. However, it is not able to fix the mislabeling of the light blue cluster in the lower right. Indeed, all methods split the cluster in the lower right of the image, indicating the challenging aspects of this dataset for unsupervised learning.](Images/SalinasA_Images/SalinasA_NMF-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsSalinasA\]Clustering results for Salinas A dataset. The proposed method, DLSS performs best, with the simplified DL method and benchmark spectral clustering also performing well. Notice that the spectral-spatial labeling scheme removes some of the mistakes in the yellow cluster, and also improves the labeling near some class boundaries. However, it is not able to fix the mislabeling of the light blue cluster in the lower right. Indeed, all methods split the cluster in the lower right of the image, indicating the challenging aspects of this dataset for unsupervised learning.](Images/SalinasA_Images/SalinasA_Bertozzi-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsSalinasA\]Clustering results for Salinas A dataset. The proposed method, DLSS performs best, with the simplified DL method and benchmark spectral clustering also performing well. Notice that the spectral-spatial labeling scheme removes some of the mistakes in the yellow cluster, and also improves the labeling near some class boundaries. However, it is not able to fix the mislabeling of the light blue cluster in the lower right. Indeed, all methods split the cluster in the lower right of the image, indicating the challenging aspects of this dataset for unsupervised learning.](Images/SalinasA_Images/SalinasA_FSFDPC-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsSalinasA\]Clustering results for Salinas A dataset. The proposed method, DLSS performs best, with the simplified DL method and benchmark spectral clustering also performing well. Notice that the spectral-spatial labeling scheme removes some of the mistakes in the yellow cluster, and also improves the labeling near some class boundaries. However, it is not able to fix the mislabeling of the light blue cluster in the lower right. Indeed, all methods split the cluster in the lower right of the image, indicating the challenging aspects of this dataset for unsupervised learning.](Images/SalinasA_Images/SalinasA_DL-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsSalinasA\]Clustering results for Salinas A dataset. The proposed method, DLSS performs best, with the simplified DL method and benchmark spectral clustering also performing well. Notice that the spectral-spatial labeling scheme removes some of the mistakes in the yellow cluster, and also improves the labeling near some class boundaries. However, it is not able to fix the mislabeling of the light blue cluster in the lower right. Indeed, all methods split the cluster in the lower right of the image, indicating the challenging aspects of this dataset for unsupervised learning.](Images/SalinasA_Images/SalinasA_DLSS-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![\[fig:ResultsSalinasA\]Clustering results for Salinas A dataset. The proposed method, DLSS performs best, with the simplified DL method and benchmark spectral clustering also performing well. Notice that the spectral-spatial labeling scheme removes some of the mistakes in the yellow cluster, and also improves the labeling near some class boundaries. However, it is not able to fix the mislabeling of the light blue cluster in the lower right. Indeed, all methods split the cluster in the lower right of the image, indicating the challenging aspects of this dataset for unsupervised learning.](Images/SalinasA_Images/SalinasA_GT-crop.pdf "fig:"){width="\textwidth"} For this dataset, the proposed DLSS method performs best, with the only error made in splitting the bottom right cluster into two pieces, an error made by all algorithms. The simpler DL method also performs well, as does the benchmark spectral clustering algorithm. Comparing the labeling for DL and DLSS, the small regions of mislabeled pixels in DL are correctly labeled in DLSS, because these pixels are likely of low empirical density, and hence benefit from being labeled based on both spectral and spatial similarity, not spectral similarity alone. However, some pixels correctly classified by DL were labeled incorrectly by DLSS, indicating that the spatial proximity condition enforced in DLSS may not lead to improved results for every pixel. Details on this, and how to tune the size of the neighborhood with which spatial consensus labels are computed, are given in Section \[subsec:SpaceParameter\]. ### Kennedy Space Center Data Set {#subsubsec:KSC} The Kennedy Space Center dataset used for experiments consists of a subset of the original dataset, and contains four classes. Figure \[fig:KSC\] illustrates the first principal component of the data, as well as the labeled ground truth, which consists of the examples of four vegetation types which dominate the scene. Results appear in Table \[tab:Summary\]. ![\[fig:KSC\]The Kennedy Space Center data is a $250\times 100$ subset of the full Kennedy Space Center dataset. It contains 4 classes, some of which are not well-localized spatially. The scene was captured with the NASA AVIRIS instrument over the Kennedy Space Center (KSC), Florida, USA. The spatial resolution is 18 m. After removing low signal-to-noise-ratio and water-absorption bands, the dataset consists of 176 bands. Left: projection onto the first principal component of the data; right: ground truth (GT).](Images/KSC_Images/KSC_FirstPC-crop.pdf "fig:"){width=".24\textwidth"} ![\[fig:KSC\]The Kennedy Space Center data is a $250\times 100$ subset of the full Kennedy Space Center dataset. It contains 4 classes, some of which are not well-localized spatially. The scene was captured with the NASA AVIRIS instrument over the Kennedy Space Center (KSC), Florida, USA. The spatial resolution is 18 m. After removing low signal-to-noise-ratio and water-absorption bands, the dataset consists of 176 bands. Left: projection onto the first principal component of the data; right: ground truth (GT).](Images/KSC_Images/KSC_GT-crop.pdf "fig:"){width=".24\textwidth"} The proposed methods yield the best results, noting that the FMS method also performs well. Most linear methods, such as $K$-means with linear dimension reduction or NMF, perform poorly, suggesting that nonlinear methods are needed for this data. Spectral clustering performs much better than the linear methods. We note that spatial information for this dataset is less helpful than for the Indian Pines and Pavia datasets. ### Overall Comments on Clustering Quantitative results for the clustering experiments appear in Table \[tab:Summary\]. We see that the DLSS method performs best among all metrics for all datasets. The DL method generally performs second best, though DBSCAN, spectral clustering, and SMCE occasionally perform comparably to DL. It is notable that DL outperforms FSFDPC, which uses a similar labeling scheme, but computes modes with Euclidean distances, rather than diffusion distances. This provides empirical evidence for the need to use nonlinear methods of measuring distances for HSI. [max width=]{} Method OA I.P. AA I.P. $\kappa$ I.P. OA P. AA P. $\kappa$ P. OA S.A. AA S.A. $\kappa$ S.A. OA K.S.C. AA K.S.C. $\kappa$ K.S.C. --------------- ---------- ---------- --------------- ---------- ---------- ------------- ---------- ---------- --------------- ----------- ----------- ----------------- $K$-means 0.43 0.38 0.09 0.78 0.62 0.72 0.63 0.66 0.52 0.36 0.25 0.01 PCA+$K$-means 0.43 0.38 0.10 0.78 0.62 0.72 0.63 0.66 0.52 0.36 0.25 0.01 ICA+$K$-means 0.41 0.36 0.06 0.67 0.55 0.58 0.57 0.56 0.44 0.36 0.25 0.01 RP+$K$-means 0.51 0.51 0.26 0.76 0.61 0.70 0.63 0.66 0.53 0.60 0.50 0.43 DBSCAN 0.63 0.43 0.73 0.72 0.66 0.71 0.71 0.63 0.36 0.25 0.01 SC 0.54 0.45 0.24 0.82 0.76 0.77 0.62 0.52 0.44 GMM 0.44 0.35 0.02 0.64 0.59 0.55 0.64 0.61 0.55 0.42 0.31 0.10 SMCE 0.52 0.45 0.22 0.83 0.77 0.79 0.47 0.42 0.30 0.36 0.26 0.01 HNMF 0.41 0.32 -0.02 0.72 0.74 0.66 0.63 0.66 0.53 0.36 0.25 0.00 FMS 0.57 0.50 0.27 0.77 0.64 0.69 0.70 0.81 0.65 0.74 0.70 0.65 FSFDPC 0.58 0.51 0.26 0.78 0.75 0.73 0.63 0.61 0.54 0.36 0.25 0.00 DL 0.79 DLSS **0.85** **0.82** **0.75** **0.94** **0.83** **0.93** **0.85** **0.90** **0.81** **0.83** **0.73** **0.76** Active Learning {#subsec:ActiveLearning} --------------- To evaluate our proposed active learning method, Algorithm \[alg:ActiveLearning\], the same $4$ labeled HSI datasets were clustered with increasing the percentage $\alpha$ of labeled points, chosen as in Algorithm \[alg:ActiveLearning\]. Note that $\alpha=0$ corresponds the the unsupervised DLSS algorithm. The empirical results for this active learning scheme appear in Figure \[fig:ActiveLearningAnalysis\]. We also consider selecting the labeled points uniformly at random; we hypothesize our principled approach will be superior to random sampling. [ .24]{} ![\[fig:ActiveLearningAnalysis\] Active learning parameter analysis. The $x$-axis denotes the parameter $\alpha$. As $\alpha$ increases, more labeled pixels are introduced. All measures of accuracy are monotonic increasing in $\alpha$, and a small increase can lead to a huge jump in accuracy, as seen in the Indian Pines and Salinas A datasets. We see that randomly selecting points has a more incremental impact on improving accuracy than the principled approach, and may require a very large number of labels to achieve the performance achieved by active learning with a small number of labels. Many iterations of randomly selected points were used and averaged to produce the plots.](Images/ParameterAnalysis/ActiveParameter/IP_ActiveParameter-crop.pdf "fig:"){width="\textwidth"} [ .24]{} ![\[fig:ActiveLearningAnalysis\] Active learning parameter analysis. The $x$-axis denotes the parameter $\alpha$. As $\alpha$ increases, more labeled pixels are introduced. All measures of accuracy are monotonic increasing in $\alpha$, and a small increase can lead to a huge jump in accuracy, as seen in the Indian Pines and Salinas A datasets. We see that randomly selecting points has a more incremental impact on improving accuracy than the principled approach, and may require a very large number of labels to achieve the performance achieved by active learning with a small number of labels. Many iterations of randomly selected points were used and averaged to produce the plots.](Images/ParameterAnalysis/ActiveParameter/Pavia_ActiveParameter-crop.pdf "fig:"){width="\textwidth"} [.24]{} ![\[fig:ActiveLearningAnalysis\] Active learning parameter analysis. The $x$-axis denotes the parameter $\alpha$. As $\alpha$ increases, more labeled pixels are introduced. All measures of accuracy are monotonic increasing in $\alpha$, and a small increase can lead to a huge jump in accuracy, as seen in the Indian Pines and Salinas A datasets. We see that randomly selecting points has a more incremental impact on improving accuracy than the principled approach, and may require a very large number of labels to achieve the performance achieved by active learning with a small number of labels. Many iterations of randomly selected points were used and averaged to produce the plots.](Images/ParameterAnalysis/ActiveParameter/SalinasA_ActiveParameter-crop.pdf "fig:"){width="\textwidth"} [.24]{} ![\[fig:ActiveLearningAnalysis\] Active learning parameter analysis. The $x$-axis denotes the parameter $\alpha$. As $\alpha$ increases, more labeled pixels are introduced. All measures of accuracy are monotonic increasing in $\alpha$, and a small increase can lead to a huge jump in accuracy, as seen in the Indian Pines and Salinas A datasets. We see that randomly selecting points has a more incremental impact on improving accuracy than the principled approach, and may require a very large number of labels to achieve the performance achieved by active learning with a small number of labels. Many iterations of randomly selected points were used and averaged to produce the plots.](Images/ParameterAnalysis/ActiveParameter/KSC_ActiveParameter-crop.pdf "fig:"){width="\textwidth"} The plots indicate that the proposed active learning can produce dramatic improvements in labeling with very few training labels. Indeed, an improvement in overall accuracy from $85\%$ to $87\%$ for Indian Pines can be achieved with only 3 labels. Even more dramatic is the Salinas A dataset, in which 3 labeled points improves the overall accuracy from $84\%$ to $99.5\%$. The Pavia dataset enjoys some improved performance, though the random labels do about as well as the principled labels, and Kennedy Space Center dataset labelings are not affected by the small collection of labeled points. In the case of Pavia, however, the overall accuracy was already very large, so active learning seems not needed for this data set. Note that our principled scheme is generally superior to using randomly selected labeled points, which leads to a more gradual improvement in accuracy, compared to the huge gains that can be seen with the proposed principled method. It is interesting to compare our active learning results to a state-of-the-art *supervised* method. We consider the supervised HSI classification with edge preserving filtering method (EPF) [@Kang2014] algorithm, which combines a support vector machine with an analysis of spectral-spatial probability maps to label points. Using a publicly available implementation [^7], we ran this algorithm using $1\%$ and $5\%$ of points as training data, generated as a uniformly random sample over all labeled points. 10 experiments were ran on each of the four datasets considered, with results averaged. Quantitative results are shown in Table \[tab:SupervisedComparison\]. The supervised results are generally superior to the results achieved by the unsupervised DL and DLSS method. The proposed active learning, however, is able to achieve the same performance on the Salinas A dataset using two orders of magnitude fewer points. This is because the proposed active learning method only uses training points for pixels that are considered especially important, whereas the EPF algorithm trains on a random subset of points. Moreover, when only $1\%$ of training points are used, our active learning DLSS method with $.2\%$ of training points used outperforms the EPF method on the Indian Pines, Salinas A, and Kennedy Space Center datasets. This indicates the promise of the proposed active learning method, as it is able to outperform a state-of-the-art supervised method in the regime in which a low proportion of training points is available. [max width=]{} Method OA I.P. AA I.P. $\kappa$ I.P. OA P. AA P. $\kappa$ P. OA S.A. AA S.A. $\kappa$ S.A. OA K.S.C. AA K.S.C. $\kappa$ K.S.C. ------------------------------- --------- --------- --------------- ------- ------- ------------- --------- --------- --------------- ----------- ----------- ----------------- DLSS (unsupervised) 0.85 0.82 0.75 0.95 0.83 0.93 0.85 0.90 0.81 0.83 0.73 0.76 Active learning, .2% training 0.87 0.84 0.79 0.95 0.83 0.93 1.00 1.00 1.00 0.83 0.73 0.76 EPF, 1% training 0.49 0.33 0.16 0.99 0.99 0.99 0.97 0.97 0.96 0.51 0.37 0.31 EPF, 5% training 0.82 0.86 0.72 0.99 0.99 0.99 1.00 1.00 1.00 0.98 0.98 0.98 Parameter Analysis {#subsec:ParameterAnalysis} ------------------ We now discuss the parameters used in all methods, starting with those used for all comparison methods, and then discussing the two key parameters for the proposed method: diffusion time $t$ and radius size $r_{s}$ for the computation of the spatial consensus label in Algorithm \[alg:labels\]. For experimental parameters except these, a range of parameters were considered, and those with best empirical results were used. All instances of the $K$-means algorithm are run with $100$ iterations, with $10$ random initializations each time, and number of clusters $K$ equal to the known number of classes in the ground truth. Each of the linear dimension reduction techniques, PCA, ICA, and random projection, embeds the data into $\mathbb{R}^{K}$, where $K$ is the number of clusters. DBSCAN is highly dependent on several parameters, and a grid search was used on each dataset to select optimal parameters. Note that this means DBSCAN was optimized specifically for each dataset, while other methods used a fixed set of parameters across all experiments. Spectral clustering is run by computing a weight matrix as in $(\ref{eqn:W})$, with $k=100$ and $\sigma=1$. The top $K$ eigenvectors are then normalized to have Euclidean norm $1$, then used as features with $K$-means. Among the state-of-the-art methods, HNMF uses the recommended settings listed in the available online toolbox[^8]. For FSFDPC, the empirical density estimate is computed as described in Section \[subsec:AlgorithmDescription\], with a Gaussian kernel and 20 nearest neighbors. For SMCE, the sparsity parameter was set to be 10, as suggested in the online toolbox[^9]. The FMS algorithm depends on several key parameters; grid search was implemented, and empirically optimal parameters with respect to a given dataset were used. Note that this means FMS was, like DBSCAN, optimized specifically for each dataset, while other methods used a fixed set of parameters across all experiments. For the proposed algorithm, the same parameters for the density estimator as described above are used, in order to make a fair comparison with FSFDPC. Moreover, in the construction of the graph used to compute diffusion distances, we use the same construction as in spectral clustering and SMCE, again to make fair comparisons. The remaining parameters, diffusion time and spatial radius, were set to $30$ and $3$, respectively, for all experiments. We justify these choices and analyze their robustness in the following subsections. ### Diffusion Time $t$ {#subsec:TimeParameter} The most important parameter when using diffusion distances $d_t(x,y)$ is the time parameter $t\ge0$, see eqn. $(\ref{eqn:DD_eigen})$. The larger $t$ is, the smaller the contribution of the smaller eigenvalues in the spectral computation of $d_{t}$. Allowing $t$ to vary, connections in the dataset are explored by allowing the diffusion process to evolve. For small values of $t$, all points appear far apart because the diffusion process has not reached far, while for large values of $t$, all points appear close together because the diffusion process has run for so long that it has dissipated over the entire state space. In general, the interesting choice of $t$ is moderate, which allows for the data geometry to be discovered, but not washed out in long-term. In Figure \[fig:TimeParameterAnalysis\], all the accuracy measures for $t$ in $[0,100]$ are displayed. [ .24]{} ![\[fig:TimeParameterAnalysis\] Time parameter analysis for the four real datasets. In general, the time parameter has little impact on performance of the proposed algorithm. As suggested by these plots, $t=30$ is used for all ex periments.](Images/ParameterAnalysis/TimeParameter/IP_TimeParameter-crop.pdf "fig:"){width="\textwidth"} [ .24]{} ![\[fig:TimeParameterAnalysis\] Time parameter analysis for the four real datasets. In general, the time parameter has little impact on performance of the proposed algorithm. As suggested by these plots, $t=30$ is used for all ex periments.](Images/ParameterAnalysis/TimeParameter/Pavia_TimeParameter-crop.pdf "fig:"){width="\textwidth"} [.24]{} ![\[fig:TimeParameterAnalysis\] Time parameter analysis for the four real datasets. In general, the time parameter has little impact on performance of the proposed algorithm. As suggested by these plots, $t=30$ is used for all ex periments.](Images/ParameterAnalysis/TimeParameter/SalinasA_TimeParameter-crop.pdf "fig:"){width="\textwidth"} [.24]{} ![\[fig:TimeParameterAnalysis\] Time parameter analysis for the four real datasets. In general, the time parameter has little impact on performance of the proposed algorithm. As suggested by these plots, $t=30$ is used for all ex periments.](Images/ParameterAnalysis/TimeParameter/KSC_TimeParameter-crop.pdf "fig:"){width="\textwidth"} The behavior is robust with respect to time. For Indian Pines, performance is largely constant, except for a dip from time $t=15$ to $t=25$. For the Pavia, Salinas A, and Kennedy Space Center examples, the performance is invariant with respect to the diffusion time. We conclude that a large range of $25\le t\le 65$ or $t\ge 75$ would have led to the same empirical results as our choice $t=30$. ### Spatial Diffusion Radius {#subsec:SpaceParameter} The spatial consensus radius $r_s$ can also impact the performance of the proposed DLSS algorithm. Recall that this is the distance in the spatial domain used to compute the spatial consensus label (see Section \[subsec:AlgorithmDescription\] and definition ). If $r_{s}$ is too small, insufficient spatial information is incorporated; if $r_{s}$ is too large, the spectral information becomes drowned out. All measures of accuracy for each dataset for $r_{s}$ in $[0,10]$ appear in Figure \[fig:SpaceParameterAnalysis\]. [ .24]{} ![\[fig:SpaceParameterAnalysis\] Space parameter analysis for DLSS. For each curve, increasing the radius of the neighborhood in which the spatial consensus label is computed improves quantitative performance up until a certain point, after which performance decays. The point at which the decay sets in differs for each example. These plots suggest a spectral-spatial tradeoff: spectral and spatial information must be balanced to achieve empirically optimal clustering.](Images/ParameterAnalysis/SpaceParameter/IP_SpatialParameter-crop.pdf "fig:"){width="\textwidth"} [ .24]{} ![\[fig:SpaceParameterAnalysis\] Space parameter analysis for DLSS. For each curve, increasing the radius of the neighborhood in which the spatial consensus label is computed improves quantitative performance up until a certain point, after which performance decays. The point at which the decay sets in differs for each example. These plots suggest a spectral-spatial tradeoff: spectral and spatial information must be balanced to achieve empirically optimal clustering.](Images/ParameterAnalysis/SpaceParameter/Pavia_SpatialParameter-crop.pdf "fig:"){width="\textwidth"} [.24]{} ![\[fig:SpaceParameterAnalysis\] Space parameter analysis for DLSS. For each curve, increasing the radius of the neighborhood in which the spatial consensus label is computed improves quantitative performance up until a certain point, after which performance decays. The point at which the decay sets in differs for each example. These plots suggest a spectral-spatial tradeoff: spectral and spatial information must be balanced to achieve empirically optimal clustering.](Images/ParameterAnalysis/SpaceParameter/SalinasA_SpatialParameter-crop.pdf "fig:"){width="\textwidth"} [.24]{} ![\[fig:SpaceParameterAnalysis\] Space parameter analysis for DLSS. For each curve, increasing the radius of the neighborhood in which the spatial consensus label is computed improves quantitative performance up until a certain point, after which performance decays. The point at which the decay sets in differs for each example. These plots suggest a spectral-spatial tradeoff: spectral and spatial information must be balanced to achieve empirically optimal clustering.](Images/ParameterAnalysis/SpaceParameter/KSC_SpatialParameter-crop.pdf "fig:"){width="\textwidth"} We see a trade-off between spectral and spatial information, suggesting that $r_{s}$ should take a moderate value sufficiently greater than 0 but less than 10. This trade-off can be interpreted in the following way: empirically optimal results are achieved when both spectral and spatial information contribute harmoniously, and results deteriorate when one or the other dominates. We choose $r_{s}=3$ for all experiments, though other choices would give comparable (or sometimes better) quantitative results for the datasets considered. We note that the role of the spatial radius is analogous to the role of a regularization parameter for a general statistical learning problem. Taken too small, the problem is insufficiently regularized, leading to poor results. Taken too large, the regularization term dominates the fidelity term, leading also to poor results. In particular, the geometric regularity of the clusters in the spatial domain determine how large $r_{s}$ may be taken while still preserving the spectral information. If the clusters are convex and not too elongated, then taking $r_{s}$ large is reasonable. On the other hand, if the classes are very irregular spatially, for example highly non-convex or elongated, choosing $r_{s}$ too large will wash out the spectral information which is generally more discriminative than the spatial information, resulting in inaccurate clustering. Large Scale Experiments {#sec:LargeScaleExperiments} ----------------------- The results of Section \[subsec:HyperspectralClustering\] analyzed subsets of larger images, in order to reduce the number of classes to allow for effective unsupervised learning [@Zhu2017]. In order to evaluate the robustness of these results, we performed experiments in which the full HSI scenes were subdivided into small patches with fewer classes, then each patch—with a smaller number of classes than the total scene—were clustered. The results on individual patches may be used as the basis for a statistical evaluation of the performance of each clustering method. For the Indian Pines, Pavia, and Kennedy Space Center datasets, experiments for the entire dataset, suitably partitioned into smaller patches, were performed, with DLSS again performing best among all studied methods. Note that Salinas A had only 6 classes, and was considered in its entirety. The Indian Pines data set was partitioned into 24 rectangular patches of equal size; Pavia into 50 rectangular patches of equal size, and Kennedy Space Center into 25 patches of equal size. On each piece that contained some non-trivial ground truth, all clustering algorithms were ran. A series of statistical tests on the differences in performance were then executed as follows. For a pair of methods—denoted method $i$ and $j$— let $OA_{k}^{i}, OA_{k}^{j}$ be the overall accuracy of methods $i$ and $j$ on patch $k$, and let $\Delta_{k}^{i,j}=OA_{k}^{i}-OA_{k}^{j}$. The sample mean difference in error between methods $i$ and $j$ across the different patches is $\overline{\Delta^{i,j}}=\sum_{k=1}^{{N_{\text{patches}}}}\Delta_{k}^{i,j}/{N_{\text{patches}}}$, where ${N_{\text{patches}}}$ is the total number of patches with ground truth. It is of interest to investigate whether $\overline{\Delta_{i,j}}$ can be inferred to be different from 0. In order to perform a statistical test, the sample standard deviation of difference between methods $i,j$ is computed as $\sigma^{i,j}=\sqrt{\sum_{k}(\Delta_{k}^{i,j}-\overline{\Delta^{i,j}})^2/({N_{\text{patches}}}-1)}$. Then, the null hypothesis that $\overline{\Delta_{i,j}}=0$ may be tested against the alternative hypothesis that $\overline{\Delta_{i,j}}\neq0$ by performing a two-sided $t$-test [@Wasserman2013_All] with ${N_{\text{patches}}}-1=72$ degrees of freedom. The normalized $t$-scores for the $j$ corresponding to the DLSS method and $i$ running through all other methods are reported in Table \[tab:StatisticalAnalysis\]. The test confirms that for all methods $i$, the hypothesis that DLSS $(j=13)$ does not significantly differ from method $i$ ($\overline{\Delta_{i,13}}=0$) is rejected in favor of the alternative hypothesis that DLSS significantly differs from method $i$ $(\overline{\Delta_{i,13}}\neq 0$) at the $95\%$ level. This provides evidence that DLSS performs competetively with benchmark and state-of-the-art HSI clustering algorithms across HSI with a variety of land cover types and complexity. Note that the values are $\Delta_{i,j}^{k}$ are not independent for different $k$, due to correlations across images. However, the $t$-test still provides a powerful method for inferring statistical significance in this case, despite this theoretical assumption not being satisfied. [.09]{} ![Results of clustering individual patches of the Indian Pines data, without synchronizing the labels.\[fig:IP\_Patches\]](Images/Combined_Images/IP/IP_Raw_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Indian Pines data, without synchronizing the labels.\[fig:IP\_Patches\]](Images/Combined_Images/IP/IP_PCA_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Indian Pines data, without synchronizing the labels.\[fig:IP\_Patches\]](Images/Combined_Images/IP/IP_ICA_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Indian Pines data, without synchronizing the labels.\[fig:IP\_Patches\]](Images/Combined_Images/IP/IP_RandProj_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of the Indian Pines data, without synchronizing the labels.\[fig:IP\_Patches\]](Images/Combined_Images/IP/IP_DBSCAN_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Indian Pines data, without synchronizing the labels.\[fig:IP\_Patches\]](Images/Combined_Images/IP/IP_SC_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Indian Pines data, without synchronizing the labels.\[fig:IP\_Patches\]](Images/Combined_Images/IP/IP_GMM_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Indian Pines data, without synchronizing the labels.\[fig:IP\_Patches\]](Images/Combined_Images/IP/IP_SMCE_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Indian Pines data, without synchronizing the labels.\[fig:IP\_Patches\]](Images/Combined_Images/IP/IP_NMF_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of the Indian Pines data, without synchronizing the labels.\[fig:IP\_Patches\]](Images/Combined_Images/IP/IP_FMS_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of the Indian Pines data, without synchronizing the labels.\[fig:IP\_Patches\]](Images/Combined_Images/IP/IP_FSFDPC_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of the Indian Pines data, without synchronizing the labels.\[fig:IP\_Patches\]](Images/Combined_Images/IP/IP_DL_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of the Indian Pines data, without synchronizing the labels.\[fig:IP\_Patches\]](Images/Combined_Images/IP/IP_DLSS_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Pavia data, without synchronizing the labels.\[fig:Pavia\_Patches\]](Images/Combined_Images/Pavia/Pavia_Raw_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Pavia data, without synchronizing the labels.\[fig:Pavia\_Patches\]](Images/Combined_Images/Pavia/Pavia_PCA_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Pavia data, without synchronizing the labels.\[fig:Pavia\_Patches\]](Images/Combined_Images/Pavia/Pavia_ICA_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Pavia data, without synchronizing the labels.\[fig:Pavia\_Patches\]](Images/Combined_Images/Pavia/Pavia_RandProj_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of the Pavia data, without synchronizing the labels.\[fig:Pavia\_Patches\]](Images/Combined_Images/Pavia/Pavia_DBSCAN_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Pavia data, without synchronizing the labels.\[fig:Pavia\_Patches\]](Images/Combined_Images/Pavia/Pavia_SC_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Pavia data, without synchronizing the labels.\[fig:Pavia\_Patches\]](Images/Combined_Images/Pavia/Pavia_GMM_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Pavia data, without synchronizing the labels.\[fig:Pavia\_Patches\]](Images/Combined_Images/Pavia/Pavia_SMCE_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of the Pavia data, without synchronizing the labels.\[fig:Pavia\_Patches\]](Images/Combined_Images/Pavia/Pavia_NMF_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of the Pavia data, without synchronizing the labels.\[fig:Pavia\_Patches\]](Images/Combined_Images/Pavia/Pavia_FMS_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of the Pavia data, without synchronizing the labels.\[fig:Pavia\_Patches\]](Images/Combined_Images/Pavia/Pavia_FSFDPC_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of the Pavia data, without synchronizing the labels.\[fig:Pavia\_Patches\]](Images/Combined_Images/Pavia/Pavia_DL_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of the Pavia data, without synchronizing the labels.\[fig:Pavia\_Patches\]](Images/Combined_Images/Pavia/Pavia_DLSS_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of Kennedy Space Center data, without synchronizing the labels.\[fig:KSC\_Patches\]](Images/Combined_Images/KSC/KSC_Raw_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of Kennedy Space Center data, without synchronizing the labels.\[fig:KSC\_Patches\]](Images/Combined_Images/KSC/KSC_PCA_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of Kennedy Space Center data, without synchronizing the labels.\[fig:KSC\_Patches\]](Images/Combined_Images/KSC/KSC_ICA_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of Kennedy Space Center data, without synchronizing the labels.\[fig:KSC\_Patches\]](Images/Combined_Images/KSC/KSC_RandProj_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of Kennedy Space Center data, without synchronizing the labels.\[fig:KSC\_Patches\]](Images/Combined_Images/KSC/KSC_DBSCAN_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of Kennedy Space Center data, without synchronizing the labels.\[fig:KSC\_Patches\]](Images/Combined_Images/KSC/KSC_SC_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of Kennedy Space Center data, without synchronizing the labels.\[fig:KSC\_Patches\]](Images/Combined_Images/KSC/KSC_GMM_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of Kennedy Space Center data, without synchronizing the labels.\[fig:KSC\_Patches\]](Images/Combined_Images/KSC/KSC_SMCE_Combined-crop.pdf "fig:"){width="\textwidth"} [.09]{} ![Results of clustering individual patches of Kennedy Space Center data, without synchronizing the labels.\[fig:KSC\_Patches\]](Images/Combined_Images/KSC/KSC_NMF_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of Kennedy Space Center data, without synchronizing the labels.\[fig:KSC\_Patches\]](Images/Combined_Images/KSC/KSC_FMS_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of Kennedy Space Center data, without synchronizing the labels.\[fig:KSC\_Patches\]](Images/Combined_Images/KSC/KSC_FSFDPC_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of Kennedy Space Center data, without synchronizing the labels.\[fig:KSC\_Patches\]](Images/Combined_Images/KSC/KSC_DL_Combined-crop.pdf "fig:"){width="\textwidth"} [ .09]{} ![Results of clustering individual patches of Kennedy Space Center data, without synchronizing the labels.\[fig:KSC\_Patches\]](Images/Combined_Images/KSC/KSC_DLSS_Combined-crop.pdf "fig:"){width="\textwidth"} In addition to providing the basis for a statistical evaluation of the proposed algorithm, splitting large, complicated HSI into patches for clustering allows to over-segment the image. Examples of the oversegmented maps, where we do not attempt to synchronize the labels across patches, appear in Figures \[fig:IP\_Patches\], \[fig:Pavia\_Patches\], \[fig:KSC\_Patches\]. It is a topic of future research to combine these patches using the DLSS framework. [max width=]{} Method $K$-means PCA+$K$-means ICA+$K$-means RP+$K$-means DBSCAN SC GMM SMCE HNMF FMS FSFDPC DL DLSS --------------- ----------- --------------- --------------- -------------- --------- -------- -------- -------- -------- -------- -------- -------- ------ $t$-statistic 2.0163 2.1039 2.7710 3.1630 6.44774 2.6301 3.4461 4.1093 2.2160 2.1810 2.1219 2.1357 - Overall Comments on the Experiments and Conclusion {#subsec:CC} ================================================== We proposed a novel unsupervised method for clustering HSI, using data-dependent diffusion distances to learn modes of clusters, followed by a spectral-spatial labeling scheme based on diffusion in both the spectral and spatial domains. We demonstrated on various data sets that the proposed DLSS algorithm performs well compared to state-of-the-art techniques, and that the DLSS algorithm outperforms DL thanks to the incorporation of spatial information. We remark that the methods which employ linear dimension reduction, including PCA, ICA, and random projections, generally outperform methods that use no dimension reduction, but do not perform as well as those which used nonlinear dimension reduction, including spectral clustering, SMCE, DL, and DLSS. This indicates that while HSI data does exhibit intrinsically low-dimensional structure, the data lies close not to subspaces, but manifolds, i.e. nonlinear sets of low dimensionality. The proposed DL method, consisting of the geometric learning of modes but only spectral assignment of labels, largely outperforms all comparison methods (see Table \[tab:Summary\]). In particular, it outperforms in all examples considered the very popular and recent FSFDPC algorithm. This indicates that Euclidean distance is insufficient for learning the modes of complex HSI data. Moreover, the joint spectral-spatial labeling scheme DLSS improves over DL in all instances. In fact, DLSS gives the overall best performance for all datasets and all performance metrics. The incorporation of active learning in the DLSS algorithm dramatically improves the accuracy of labeling of the Indian Pines, Pavia, and Salinas A datasets with very few label queries. This parsimonious use of training labels has the potential to greatly improve the efficiency of machine learning tasks for HSI, in which the number of labels necessary to label a significant proportion of the image is very high. The proposed active learning method can perform competitively with the state-of-the-art supervised EPF spectral-spatial classification algorithm, using a fraction of the number of labeled pixels. Computational Complexity and Runtime {#subsec:CC} ------------------------------------ Let the data be $X=\{x_{n}\}_{n=1}^{N}\subset\mathbb{R}^{D}$. For the Indian Pines dataset, $N=1250, D=200$; for the Pavia dataset, $N=13500,D=102$; for the Salinas A dataset, $N=7138, D=224$; for the Kennedy Space Center dataset, $N=25000, D=176$. The most expensive step in DLSS is the construction of the nearest neighbor graph: we achieve near-linear scaling in $N$, $O(C_{d}DN\log N+k_{1}DN)$, using the cover trees algorithm [@LangfordICML06-CoverTree] with $C_{d}$ a constant that depends exponentially on the intrinsic dimension $d$ of the data, which is quite small in all the data sets considered. Once the nearest neighbors are found, the kernel density estimator, the random walk, and its eigenvectors can all be quickly constructed in time $O(N\log N)$, assuming that the number of nearest neighbors used in the density estimator is $O(\log N)$ and that the number of eigenvectors needed is $O(1)$. Computing the nearest spectral neighbor of higher empirical density, computing the spatial consensus labels, and active learning respectively have negligible computational complexity. We show empirical runtimes in Table \[tab:RunTime\], which demonstrates that the proposed methods have superior runtimes to spectral clustering and DBSCAN, and are substantially faster than SMCE. [max width=.49]{} --------------- ------- -------- ----------- --------- Method IP Pavia Salinas A KSC $K$-means 0.44 3.89 1.26 5.97 PCA+$K$-means 0.01 0.01 0.01 0.16 ICA+$K$-means 0.22 0.87 0.30 1.29 RP+$K$-means 0.11 0.79 0.13 0.58 DBSCAN 0.58 56.10 13.53 112.43 SC 0.77 101.20 13.87 483.56 GMM 0.40 3.07 1.91 2.16 SMCE 11.74 466.90 222.40 1315.21 HNMF 0.52 0.93 0.74 1.41 FMS 0.15 0.73 0.29 0.89 FSFDPC 1.48 33.64 10.05 69.46 DL 0.80 36.46 8.73 80.28 DLSS 1.40 73.77 12.24 106.05 --------------- ------- -------- ----------- --------- : \[tab:RunTime\]Run times for each method and each dataset, measured in seconds. The linear dimension reduction methods are extremely fast, as are NMF and GMM. The spectral clustering and FSFDPC algorithms are slower than DL, and DLSS is slightly slower is slightly slower than DL. The SMCE algorithm is substantially slower. Note that although FMS is quite fast, it is implemented in parallelized C++ code and ran on a machine with 24 cores and 48 threads. Future Research Directions {#sec:Future} ========================== A drawback of many clustering algorithms, including the ones presented in this paper, is the assumption that the number of clusters, $K$, is known a priori. While unsupervised clustering experiments typically assume $K$ is known, it is of interest to develop methods that allow efficient and accurate estimation of $K$, in order to make a truly unsupervised clustering method. Initial investigations suggest that looking for the “kink" in the sorted plot of $\mathcal{D}_{t}(x_i)$ could be used to detect $K$ automatically. More precisely, we check if there is a prominent peak in the value $\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})$, where where the points $\{x_{i}^{\text{sort}}\}_{i=1}^{n}$ are the data, sorted in decreasing order of their $\mathcal{D}_{t}$ values. This is a discrete version of the gradient, so we are looking for a sharp drop-off in the sorted $\mathcal{D}_{t}$ curve. If there is a prominent such maxima in $\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})$, precisely defined as a local maximum that is greater in magnitude than double the previous value, and also at least half the magnitude of the global maximum, we estimate $\hat{K}$ as this peak. If there is no such prominent peak, then we proceed to examine the second order information $(\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i}))/(\mathcal{D}_{t}(x^{\text{sort}}_{i+2})-\mathcal{D}_{t}(x^{\text{sort}}_{i+1}))$. This is a discrete approximation to the second derivative of $\mathcal{D}_{t}$, to find when $\mathcal{D}_{t}$ begins to flatten. Initial analysis on the Indian Pines, Pavia, Salinas A and Kennedy Space Center datasets used in this article confirm the promise of analyzing the decay of $\mathcal{D}_{t}(x_i)$; results showing plots of $\{\mathcal{D}_{t}(x_{i}^{\text{sort}})\}$ values appear in Figure \[fig:KinkAnalysis\], while the corresponding statistics $\{\mathcal{D}_{t}(x_{i+1})-\mathcal{D}_{t}(x_{i})\}$ and $\{({\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})})/({\mathcal{D}_{t}(x^{\text{sort}}_{i+2})-\mathcal{D}_{t}(x^{\text{sort}}_{i+1})})\}$ are shown in Figure \[fig:DiffRatioAnalysis\]. The estimated number of clusters $\hat K$ appear in Table \[tab:EstimatedK\]. ![\[fig:KinkAnalysis\]The sorted $\mathcal{D}_{t}(x_{i}^{\text{sorted}})$ values for each of the four datasets. From left to right: Indian Pines, Pavia, Salinas A, KSC. The estimate $\hat{K}$ is shown as a red star.](Images/IP_Images/IP_Kink-crop.pdf "fig:"){width="11.00000%"} ![\[fig:KinkAnalysis\]The sorted $\mathcal{D}_{t}(x_{i}^{\text{sorted}})$ values for each of the four datasets. From left to right: Indian Pines, Pavia, Salinas A, KSC. The estimate $\hat{K}$ is shown as a red star.](Images/Pavia_Images/Pavia_Kink-crop.pdf "fig:"){width="11.00000%"} ![\[fig:KinkAnalysis\]The sorted $\mathcal{D}_{t}(x_{i}^{\text{sorted}})$ values for each of the four datasets. From left to right: Indian Pines, Pavia, Salinas A, KSC. The estimate $\hat{K}$ is shown as a red star.](Images/SalinasA_Images/SalinasA_Kink-crop.pdf "fig:"){width="11.00000%"} ![\[fig:KinkAnalysis\]The sorted $\mathcal{D}_{t}(x_{i}^{\text{sorted}})$ values for each of the four datasets. From left to right: Indian Pines, Pavia, Salinas A, KSC. The estimate $\hat{K}$ is shown as a red star.](Images/KSC_Images/KSC_Kink-crop.pdf "fig:"){width="11.00000%"} ![\[fig:DiffRatioAnalysis\]We show the plots of $\{\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})\}$ (top row) and the ratios $\{({\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})})/({\mathcal{D}_{t}(x^{\text{sort}}_{i+2})-\mathcal{D}_{t}(x^{\text{sort}}_{i+1})}\})$ (bottom row) for each of the four HSI datasets considered in this article. The true number of classes is shown with a red star: the proposed method of estimating $K$ is accurate for all the datasets except Indian Pines. We see that the first order information correctly determines that there are 6 clusters in Salinas A and 4 clusters in the KSC HSI, owing to the prominent peaks. The first order information is ambiguous for Indian Pines and Pavia, since there are no prominent peaks. The second order information correctly estimates that there are 6 clusters in the Pavia data, and incorrectly estimates 4 clusters for Indian Pines.](Images/IP_Images/IP_Diff-crop.pdf "fig:"){width="11.00000%"} ![\[fig:DiffRatioAnalysis\]We show the plots of $\{\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})\}$ (top row) and the ratios $\{({\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})})/({\mathcal{D}_{t}(x^{\text{sort}}_{i+2})-\mathcal{D}_{t}(x^{\text{sort}}_{i+1})}\})$ (bottom row) for each of the four HSI datasets considered in this article. The true number of classes is shown with a red star: the proposed method of estimating $K$ is accurate for all the datasets except Indian Pines. We see that the first order information correctly determines that there are 6 clusters in Salinas A and 4 clusters in the KSC HSI, owing to the prominent peaks. The first order information is ambiguous for Indian Pines and Pavia, since there are no prominent peaks. The second order information correctly estimates that there are 6 clusters in the Pavia data, and incorrectly estimates 4 clusters for Indian Pines.](Images/Pavia_Images/Pavia_Diff-crop.pdf "fig:"){width="11.00000%"} ![\[fig:DiffRatioAnalysis\]We show the plots of $\{\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})\}$ (top row) and the ratios $\{({\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})})/({\mathcal{D}_{t}(x^{\text{sort}}_{i+2})-\mathcal{D}_{t}(x^{\text{sort}}_{i+1})}\})$ (bottom row) for each of the four HSI datasets considered in this article. The true number of classes is shown with a red star: the proposed method of estimating $K$ is accurate for all the datasets except Indian Pines. We see that the first order information correctly determines that there are 6 clusters in Salinas A and 4 clusters in the KSC HSI, owing to the prominent peaks. The first order information is ambiguous for Indian Pines and Pavia, since there are no prominent peaks. The second order information correctly estimates that there are 6 clusters in the Pavia data, and incorrectly estimates 4 clusters for Indian Pines.](Images/SalinasA_Images/SalinasA_Diff-crop.pdf "fig:"){width="11.00000%"} ![\[fig:DiffRatioAnalysis\]We show the plots of $\{\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})\}$ (top row) and the ratios $\{({\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})})/({\mathcal{D}_{t}(x^{\text{sort}}_{i+2})-\mathcal{D}_{t}(x^{\text{sort}}_{i+1})}\})$ (bottom row) for each of the four HSI datasets considered in this article. The true number of classes is shown with a red star: the proposed method of estimating $K$ is accurate for all the datasets except Indian Pines. We see that the first order information correctly determines that there are 6 clusters in Salinas A and 4 clusters in the KSC HSI, owing to the prominent peaks. The first order information is ambiguous for Indian Pines and Pavia, since there are no prominent peaks. The second order information correctly estimates that there are 6 clusters in the Pavia data, and incorrectly estimates 4 clusters for Indian Pines.](Images/KSC_Images/KSC_Diff-crop.pdf "fig:"){width="11.00000%"}\ [ .11]{} ![\[fig:DiffRatioAnalysis\]We show the plots of $\{\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})\}$ (top row) and the ratios $\{({\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})})/({\mathcal{D}_{t}(x^{\text{sort}}_{i+2})-\mathcal{D}_{t}(x^{\text{sort}}_{i+1})}\})$ (bottom row) for each of the four HSI datasets considered in this article. The true number of classes is shown with a red star: the proposed method of estimating $K$ is accurate for all the datasets except Indian Pines. We see that the first order information correctly determines that there are 6 clusters in Salinas A and 4 clusters in the KSC HSI, owing to the prominent peaks. The first order information is ambiguous for Indian Pines and Pavia, since there are no prominent peaks. The second order information correctly estimates that there are 6 clusters in the Pavia data, and incorrectly estimates 4 clusters for Indian Pines.](Images/IP_Images/IP_DiffRatio-crop.pdf "fig:"){width="\textwidth"} [ .11]{} ![\[fig:DiffRatioAnalysis\]We show the plots of $\{\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})\}$ (top row) and the ratios $\{({\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})})/({\mathcal{D}_{t}(x^{\text{sort}}_{i+2})-\mathcal{D}_{t}(x^{\text{sort}}_{i+1})}\})$ (bottom row) for each of the four HSI datasets considered in this article. The true number of classes is shown with a red star: the proposed method of estimating $K$ is accurate for all the datasets except Indian Pines. We see that the first order information correctly determines that there are 6 clusters in Salinas A and 4 clusters in the KSC HSI, owing to the prominent peaks. The first order information is ambiguous for Indian Pines and Pavia, since there are no prominent peaks. The second order information correctly estimates that there are 6 clusters in the Pavia data, and incorrectly estimates 4 clusters for Indian Pines.](Images/Pavia_Images/Pavia_DiffRatio-crop.pdf "fig:"){width="\textwidth"} [.11]{} ![\[fig:DiffRatioAnalysis\]We show the plots of $\{\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})\}$ (top row) and the ratios $\{({\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})})/({\mathcal{D}_{t}(x^{\text{sort}}_{i+2})-\mathcal{D}_{t}(x^{\text{sort}}_{i+1})}\})$ (bottom row) for each of the four HSI datasets considered in this article. The true number of classes is shown with a red star: the proposed method of estimating $K$ is accurate for all the datasets except Indian Pines. We see that the first order information correctly determines that there are 6 clusters in Salinas A and 4 clusters in the KSC HSI, owing to the prominent peaks. The first order information is ambiguous for Indian Pines and Pavia, since there are no prominent peaks. The second order information correctly estimates that there are 6 clusters in the Pavia data, and incorrectly estimates 4 clusters for Indian Pines.](Images/SalinasA_Images/SalinasA_DiffRatio-crop.pdf "fig:"){width="\textwidth"} [.11]{} ![\[fig:DiffRatioAnalysis\]We show the plots of $\{\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})\}$ (top row) and the ratios $\{({\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})})/({\mathcal{D}_{t}(x^{\text{sort}}_{i+2})-\mathcal{D}_{t}(x^{\text{sort}}_{i+1})}\})$ (bottom row) for each of the four HSI datasets considered in this article. The true number of classes is shown with a red star: the proposed method of estimating $K$ is accurate for all the datasets except Indian Pines. We see that the first order information correctly determines that there are 6 clusters in Salinas A and 4 clusters in the KSC HSI, owing to the prominent peaks. The first order information is ambiguous for Indian Pines and Pavia, since there are no prominent peaks. The second order information correctly estimates that there are 6 clusters in the Pavia data, and incorrectly estimates 4 clusters for Indian Pines.](Images/KSC_Images/KSC_DiffRatio-crop.pdf "fig:"){width="\textwidth"} Dataset IP Pavia Salinas A KSC --------------------------------------- ---- ------- ----------- ----- -- Estimated Number of Classes $\hat{K}$ 4 6 6 4 Number of Labeled GT Classes $K$ 3 6 6 4 : \[tab:EstimatedK\]We show the number of classes estimated by looking for the “kink" in the sorted plot of $\mathcal{D}_{t}(x_n)$. We see that for the Salinas A and Kennedy Space Center datasets, estimating based on $\{\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})\}$ correctly estimates the number of labeled classes, while this statistic is inconclusive for Indian Pines and Salinas A. For these data, we move to second order information, namely estimating $K$ by maximizing $({\mathcal{D}_{t}(x^{\text{sort}}_{i+1})-\mathcal{D}_{t}(x^{\text{sort}}_{i})})/({\mathcal{D}_{t}(x^{\text{sort}}_{i+2})-\mathcal{D}_{t}(x^{\text{sort}}_{i+1})})$. This estimator overestimates the number of classes for Indian Pines, estimating 4 instead of 3, while it correctly estimates the number of classes in Pavia. The Indian Pines dataset is the most challenging of the four labeled datasets analyzed, which suggests that the the proposed method for estimating $K$ may be insufficient for challenging HSI data. It is of interest to prove under what assumptions on the distributions and mixture model the plot $\mathcal{D}_{t}(x_{n})$ correctly determines $K$. Moreover, in the case that one cluster is noticeably smaller or harder to detect than others, as in the case of the Indian Pines dataset, it may be advantageous to use a different statistic on the finite difference curve, rather than the proposed derivative conditions on $\mathcal{D}_{t}$. Initial mathematical results and more subtle conjectures are proposed in an upcoming article [@Murphy2018]. Moreover, all unsupervised algorithms considered in this paper struggle with very large HSI scenes consisting of many classes. This is due to the large variation within clusters compared to the differences between clusters, which leads to genuine classes being split incorrectly; this is a well-known challenge for unsupervised clustering of HSI [@Zhu2017]. In Section \[sec:LargeScaleExperiments\] it is shown that DLSS is very effective at clustering on different patches of a large HSI. It remains an open question how to combine the results on these patches into a global clustering, which amounts to determining when to merge clusters learned in distinct patches. Automatically implementing such mergers with the DLSS framework is a direction of future research. The present work is essentially empirical: it is not known mathematically under what constraints on the mixture model the method proposed for learning modes succeeds with high probability. Besides being of mathematical interest, this would be useful for understanding the limitations of the proposed method for HSI. To understand this phenomenon rigorously, a careful analysis of diffusion distances for data drawn from a non-parametric mixture model is required, which is related to investigating performance guarantees for spectral clustering and mode detection [@Genovese2016; @Schiebinger2015]. It is also of interest to explicitly incorporate spectral-spatial features into the diffusion construction. It is known that use of spectral-spatial features is beneficial for supervised learning of HSI [@Kang2014; @Ghamisi2016; @Kang2017], and their use in unsupervised learning is an exciting research direction. Indeed, incorporating the spatial properties of the scene into the graph from which diffusion distances are generated may render the explicit spatial regularization step of the proposed algorithm redundant, thus improving runtime. Acknowledgments {#acknowledgments .unnumbered} =============== The authors would like to thank Ed Bosch for his helpful comments on a preliminary version of this paper. This research was partially supported by NSF-ATD-1222567, NSF-ATD-1737984, AFOSR FA9550-14-1-0033, AFOSR FA9550-17-1-0280, NSF-IIS-1546392, and ARO subcontract to W911NF-17-P-0039. We would also like to thank the anonymous reviewers of this article, whose valuable comments significantly improved the presentation and content of this article. [^1]: <http://www.ehu.eus/ccwintco/index.php?title=Hyperspectral_Remote_Sensing_Scenes> [^2]: <https://www.cs.helsinki.fi/u/ahyvarin/papers/fastica.shtml> [^3]: <http://vision.jhu.edu/code/> [^4]: <https://sites.google.com/site/nicolasgillis/code> [^5]: <http://www.ipol.im/pub/art/2017/204/?utm_source=doi> [^6]: <http://www.math.jhu.edu/~jmurphy/> [^7]: <http://xudongkang.weebly.com/> [^8]: <https://sites.google.com/site/nicolasgillis/code> [^9]: <http://www.vision.jhu.edu/code/>

ArXiv

--- abstract: 'The effect of full-duplex cooperative relaying in a random access multiuser network is investigated here. First, we model the self-interference incurred due to full-duplex operation, assuming multi-packet reception capabilities for both the relay and the destination node. Traffic at the source nodes is considered saturated and the cooperative relay, which does not have packets of its own, stores a source packet that it receives successfully in its queue when the transmission to the destination has failed. We obtain analytical expressions for key performance metrics at the relay, such as arrival and service rates, stability conditions, and average queue length, as functions of the transmission probabilities, the self interference coefficient, and the links’ outage probabilities. Furthermore, we study the impact of the relay node and the self-interference coefficient on the per-user and aggregate throughput, and the average delay per packet. We show that perfect self-interference cancelation plays a crucial role when the SINR threshold is small, since it may result to worse performance in throughput and delay comparing with the half-duplex case. This is because perfect self-interference cancelation can cause an unstable queue at the relay under some conditions.' author: - bibliography: - 'thesis.bib' title: 'Relay-assisted Multiple Access with Full-duplex Multi-Packet Reception' --- Full-duplex, relay, cooperative communications, network-level cooperation, multiple access, stability, random access networks. Introduction {#sec:intro} ============ Driven by the exponential traffic growth and the ever-increasing demands for wider spectrum, the quest for higher spectral efficiency and enhanced reliability and coverage is creating a new impetus for cooperative communication systems. Cooperative communication aims at increasing the link data rates and the reliability of time-varying links, by overcoming fading and interference in wireless networks. Among the various cooperation techniques to increase throughput, full-duplex relaying has recently gained significant attention. The vast majority of research papers have considered half-duplex or out-of-band full-duplex systems, in which terminals cannot transmit and receive at the same time, or over the same frequency band. However, the use of nodes with in-band full-duplex capability, i.e. terminals that transmit and receive simultaneously over the same frequency band, is constantly increasing in current wireless networks as they can potentially double the network spectral efficiency. Moreover, full-duplex relay systems open a whole new spectrum of capabilities, such as collision detection in contention-based networks. In this work, we focus on a relay-assisted random access network and we analyze the effect of full-duplex cooperative relaying in the network performance, namely arrival and service rates, stability conditions, and average queue length at the relay. Related Work ------------ The classical relay channel was originally introduced by van der Meulen [@b:Muelen], and earlier work on the relay channel was based on information-theoretic formulations, e.g.  [@b:CoverGamal]. Most cooperative techniques that have been studied so far focus on the benefits of physical layer cooperation [@b:Yates-NOW]. Nevertheless, there is evidence that the same gains can be achieved with network layer cooperation, which is plain relaying without any physical layer considerations [@b:Sadek; @b:Rong1]. Recently several works have investigated relaying performance at the MAC layer [@b:Sadek; @b:Rong1; @b:Rong2; @b:Simeone; @b:Pappas; @b:PappasITW2013Relay; @b:PappasGlobalsip2013; @b:PappasTIT2015]. More specifically, in [@b:Sadek], the authors have studied the impact of cooperative communication at the medium access control layer with TDMA. They introduced a new cognitive multiple access protocol in the presence of a relay in the network. In [@b:Pappas-ISIT] the notion of partial network level cooperation is introduced by adding a flow controller at the relay, which regulates the amount of provided cooperation depending on the conditions of the network. The classical analysis of random multiple access schemes, like slotted ALOHA [@b:Bertsekas], has focused on the so-called collision model. Random access with multi-packet reception (MPR) has attracted attention recently [@b:AlohaVerdu; @b:Angel; @b:Naware; @b:PappasMPR]. All the above approaches come together in the model that we consider. In wireless networks, when a wireless node transmits and receives simultaneously in the same frequency, the problem of self-interference arises. Self-interference mitigation is a key challenge in in-band full-duplex systems. Information-theoretic aspects of this problem can be found in the pioneering work of Shannon [@b:Shannon2way], although the capacity region of the two-way channel is not known for the general case [@b:Cover]. The information-theoretic limits of in-band full-duplex relaying have been studied focusing on the idealistic case of perfect self-interference cancelation [@Somekh07; @Jafar09]. There exist several techniques that allow the possibility of perfect self-interference cancelation [@b:Cover]. However, in practice, there are several technological limitations and challenges [@b:selflimit1; @b:selflimit2], which may limit the accuracy and the effectiveness of self-interference cancelation. Various methods for performing self-interference cancelation at the receivers can be found in [@b:selfcancel1] and [@b:selfcancel2]. The main result therein is that there is a tradeoff between transceiver complexity and self-interference cancelation accuracy. In  [@b:Tsubouchi93; @b:Chen98], it was demonstrated in practice real implementations of simultaneous transceivers, where the self-interference problem has been mitigated through RF isolators and echo cancellers, coupled with base-band digital filtering. Furthermore, some recent results have also shown that full duplex is possible, proposing specific designs, e.g. [@b:Choi-Mobicom2010; @b:Jain-Mobicom2011], which mainly focus on the physical and the medium access control (MAC) layer design. Choi et al. in [@b:Choi-Mobicom2010] designed a practical single-channel full-duplex wireless system, combining three self-interference cancellation schemes, as well as RF and digital interference cancellation. Jain et al. [@b:Jain-Mobicom2011] presented a full-duplex radio design using signal inversion and adaptive cancellation. Unlike [@b:Choi-Mobicom2010], the authors in [@b:Jain-Mobicom2011] consider wideband and high power systems. In theory, this new design has no limitation in terms of bandwidth or power. Therefore, building full-duplex wireless networks (such as full-duplex 802.11n wireless networks) has started becoming feasible. Fang et al. [@b:Pathbook] proposed a collision-free full-duplex broadcast MAC and studied cross-layer optimization of MAC and routing in full-duplex wireless networks under various resource and social constraints. In [@b:KwonTVT12] the comparison of performance of half and full-duplex relay is studied at the physical layer, in [@b:KimTVT13] is investigated the effect of channel estimation errors on the ergodic capacities for bidirectional full-duplex transmission. An information theoretic study in [@b:ErkipCISS13] compares multi-antenna half and full-duplex relaying from the perspective of achievable rates. Contribution ------------ In this work, we complement and extend the work in [@b:Pappas-ITW2]. We study the operation of a cooperative node relaying packets from a number of users/sources to a destination node as shown in Fig. \[fig:netmodel\]. We assume MPR capability for both the relay and the destination node. The relay node can transmit and receive at the same time over the same frequency band (in-band full duplex). We assume random medium access, slotted time, and that each packet transmission takes one timeslot. The wireless channel is modeled as Rayleigh flat-fading channel with additive white Gaussian noise. A user transmission is successful if the received signal-to-interference-plus-noise ratio (SINR) is above a certain threshold $\gamma$. We also assume that acknowledgements (ACKs) are instantaneous and error free. The relay does not have packets by itself and the source nodes are considered saturated with unlimited amount of traffic. The self-interference cancellation at the relay is modeled as a variable power gain, mainly because we are studying the impact on the network layer[^1]. Studying in detail the physical layer implementation of self-interference mitigation and considering specific self-interference cancelation mechanisms is beyond the scope of this paper. We obtain analytical expressions for key performance characteristics of the relay queue, such as arrival and service rates, and we derive conditions for stability and the average queue length as functions of the transmission probabilities, the self-interference coefficient, and the links’ outage probabilities. In particular, we study the impact of the relay node and the self-interference coefficient on the per-user and the network-wide throughput, as well as the average delay per packet. Furthermore, we derive expressions for both the per-user and aggregate throughput when the queue at the relay is unstable, for which case we do not have though any guarantees for bounded delay. The remainder of the paper is organized as follows: Section \[sec:sysmod\] describes the system model and in Section \[sec:analysis\] we present the main characteristics of the relay queue, such as the average arrival and service rates. In Section \[sec:thr\_analysis\], we provide expressions for the per-user and the aggregate throughput. The average delay per packet is obtained in Section \[sec:delay\_analysis\]. Numerical results are presented in Section \[sec:results\], and finally Section \[sec:conclusions\] concludes the paper. System Model {#sec:sysmod} ============ Network Model ------------- We consider a network with $n$ sources, one relay node, and a single destination node. The sources transmit packets to the destination using a cooperative relay; the case of $n=2$ is depicted in Fig. \[fig:netmodel\]. We assume that the queues of both sources are saturated, i.e., no external arrivals and unlimited buffer size, and that the relay does not have packets of its own but only forwards the packets it has received from the two users. The relay node stores a source packet that it receives successfully in its queue when the direct transmission to the destination node has failed. We assume a random access channel where $q_{0}$ is the transmit probability of the relay given that it has packets in its queue, and $q_{i}$ for $i \neq 0$ is the transmit probability for the $i$-th user. The receivers at the relay and the destination nodes are equipped with multiuser detectors, hence they can decode packets from more than one transmitter at a time. Furthermore, the relay can simultaneously transmit and receive packets (full duplex).  \[fig:netmodel\] Physical Layer Model -------------------- The MPR channel model used in this paper is a generalized form of the packet erasure model. We assume that a packet transmitted by node $i$ is successfully received by node $j$ if and only if ${\rm SINR}(i,j)\geq \gamma_{j}$, where $\gamma_{j}$ is a threshold characteristic of node $j$. The wireless channel is subject to fading; let $P_{tx}(i)$ be the transmit power at node $i$ and $r(i,j)$ be the distance between $i$ and $j$. The received power at $j$ when $i$ transmits is $P_{rx}(i,j)=A(i,j)h(i,j)$ where $A(i,j)$ is a random variable representing small-scale fading. Under Rayleigh fading, $A(i,j)$ is exponentially distributed [@b:Tse]. The received power factor $h(i,j)$ is given by $h(i,j) = P_{tx}(i)(r(i,j))^{-\alpha}$ where $\alpha$ is the path loss exponent with typical values between $2$ and $6$. We model the self-interference by a scalar $g \in [0,1]$ as in [@b:CodreanuWiOpt10] and [@b:CodreanuITW10]. We refer to the $g$ as the self-interference coefficient. When $g = 1$, no self-interference cancelation technique is used, while $g = 0$ models perfect self-interference cancelation. The success probability in the link $ij$ is given by $$\begin{aligned} \label{eq:succprob} P_{i/\mathcal{T}}^{j}=\exp\left(-\frac{\gamma_{j}\eta_{j}}{v(i,j)h(i,j)}\right) \left(1+\gamma_{j}(r(i,j))^{\alpha}g \right)^{-m} \times \\ \times \prod_{k\in \mathcal{T}\backslash \left\{i,j\right\}}{\left(1+\gamma_{j}\frac{v(k,j)h(k,j)}{v(i,j)h(i,j)}\right)}^{-1}, \end{aligned}$$ where $\mathcal{T}$ is the set of transmitting nodes at the same time, $v(i,j)$ is the parameter of the Rayleigh fading random variable, $\eta_{j}$ is the receiver noise power at $j$ and $m=1$ when $j \in \mathcal{T}$ and $m=0$ else. The analytical derivation for this success probability can be found in [@b:Tse]. *Note:* The self-interference is modeled through $g$ and it affects the success probability when the relay transmits and receives simultaneously. The value of $g$ captures the accuracy of the self-interference cancelation. As $g$ approaches $0$ it is closer to the pure full duplex operation. When $g$ is $1$ the operation is the half duplex operation since the success probabilities for the users in this case are very close to $0$. Queue Stability --------------- We adopt the definition of queue stability used in [@Szpankowski:stability]. Denote by $Q_i^t$ the length of queue $i$ at the beginning of timeslot $t$. The queue is said to be *stable* if $$\label{eqn:definition_stability} \lim_{t \rightarrow \infty} {Pr}[Q_i^t < {x}] = F(x) \text{ and } \lim_{ {x} \rightarrow \infty} F(x) = 1.$$ If $\lim_{x \rightarrow \infty} \lim_{t \rightarrow \infty} \inf {Pr}[Q_i^t < {x}] = 1$, the queue is *substable*. If a queue is stable, then it is also substable. If a queue is not substable, then we say it is unstable. Loynes’ theorem [@b:Loynes] states that if the arrival and service processes of a queue are strictly jointly stationary and the average arrival rate is less than the average service rate, then the queue is stable. Performance Analysis for the Relay Queue {#sec:analysis} ======================================== In this section, we derive expressions for key performance metrics for the relay queue, namely arrival and service rates, stability conditions, and average queue length. The analysis is provided for two cases: (i) when the network consists of two non-symmetric in general users, (ii) for $n>2$ symmetric users. This section is an intermediate step before investigating the impact of the relay node in the per-user throughput, the aggregate throughput, and the average per packet delay. In order to study those quantities, we need to first compute the average arrival and service rate of the relay, the average queue length, and the stability conditions. The stability of a queue is translated to bounded queue size, which implies finite queuing delay. Two-user Case ------------- We study first the relay queue characteristics for the two-user case. In this network, at each timeslot, the relay can receive at most two packets (one per user) and to transmit at most one. The probability that the relay receives $i$ packets in a given timeslot when its queue is empty is denoted by $r_{i}^{0}$, and $r_{i}^{1}$ otherwise (not empty). The expressions for the $r_{i}^{j}$ are rather lengthy and are presented in Appendix \[sec:app\_proof\_2\]. The average arrival rate at the relay when its queue is empty is denoted by $\lambda_{0}$, and by $\lambda_{1}$ when it is not (derived in Appendix \[sec:app\_proof\_2\]). The probability that the relay queue increases by $i$ packets when is empty is denoted by $p_{i}^{0}$, and $p_{i}^{1}$ when it is not; $p_{-1}^{i}$ is the probability that the queue decreases by one packet. Note that $p_{i}^{j}$ and $r_{i}^{j}$ are in general different quantities, however $p_{i}^{j}=r_{i}^{j}$ in half-duplex relay systems[^2]. The next theorem presents the main relay queue characteristics for the two-user case. \[thm:2users\] The key performance measures for the relay queue in a two-user network are provided below. - The average service rate is $$\label{eq:m2} \begin{aligned} \mu=q_{0}(1-q_{1})(1-q_{2})P_{0/0}^{d}+q_{0}q_{1}(1-q_{2})P_{0/0,1}^{d}+\\+q_{0}q_{2}(1-q_{1})P_{0/0,2}^{d}+q_{0}q_{1}q_{2}P_{0/0,1,2}^{d}. \end{aligned}$$ where $P_{0/0,i,j}^{d}$ is the success probability between the relay and the destination when the transmitting nodes are the relay and nodes $i$ and $j$. $P_{0/0,i,j}^{d}$ can be computed from (\[eq:succprob\]). - The probability that the queue at the relay is empty is $$\label{eq:probempty2} P\left(Q=0\right)=\frac{p_{-1}^{1}-p_{1}^{1}-2p_{2}^{1}}{p_{-1}^{1}-p_{1}^{1}-2p_{2}^{1}+\lambda_{0}}.$$ - The average arrival rate $\lambda$ is $$\label{eq:lambda2} \lambda=\frac{p_{-1}^{1}-p_{1}^{1}-2p_{2}^{1}}{p_{-1}^{1}-p_{1}^{1}-2p_{2}^{1}+\lambda_{0}}\lambda_{0}+\frac{\lambda_{0}}{p_{-1}^{1}-p_{1}^{1}-2p_{2}^{1}+\lambda_{0}}\lambda_{1}.$$ - The average relay queue size $Q$ is [$$\label{eq:avQ2} \overline{Q}=\frac{(p_{1}^{1}+2p_{2}^{1}-p_{-1}^{1})(4p_{1}^{0}+10p_{2}^{0})+\lambda_{0}(2p_{-1}^{1}-4p_{1}^{1}-10p_{2}^{1})}{2(p_{1}^{1}+2p_{2}^{1}-p_{-1}^{1})(p_{-1}^{1}-p_{1}^{1}-2p_{2}^{1}+\lambda_{0})}.$$ ]{} See Appendix \[sec:app\_proof\_2\]. *Note:* The values of $q_{0}$ for which the queue is stable are given by $q_{0min}<q_{0}<1$, where $q_{0min}$ is given in (\[eq:q0min2\]) in Appendix \[sec:app\_proof\_2\]. Queue stability is an important parameter of quality-of-service (QoS), as it implies finite queue delay (due to bounded queue size). The queueing delay is computed in Section \[sec:delay\_analysis\]. Symmetric $n$-user Case {#sec:relay-results_n} ----------------------- We now investigate the case of a symmetric $n$-user network[^3]. Each user attempts to transmit in a slot with probability $q$; the success probability to the relay and the destination when $i$ nodes transmit are given by $P_{0,i}$ and $P_{d,i}$, respectively. There are two cases for the $P_{d,i}$, i.e., $P_{d,i,0}$ and $P_{d,i,1}$, denoting the success probability when relay remains silent or transmits, respectively. Those success probabilities for the symmetric case are given by $P_{d,i,j}=P_{d}\left(\frac{1}{1+\gamma_{d}} \right)^{i-1} \left(\frac{1}{1+\beta\gamma_{0}} \right)^{j},\text{ } j=0,1$ and $\beta=\frac{v_{0d}h_{0d}}{v_{d}h_{d}}>1$. $P_{0d,i}=P_{0d}\left(\frac{1}{1+\frac{1}{\beta}\gamma_{d}}\right)^{i}$, $P_{0}=\exp\left(-\frac{\gamma_{0}\eta_{0}}{v_{0}h_{0}}\right)$, $P_{d}=\exp\left(-\frac{\gamma_{d}\eta_{d}}{v_{d}h_{d}}\right)$, $P_{0d}=\exp\left(-\frac{\gamma_{0}\eta_{0}}{v_{0}h_{0}}\right)$. There are two cases for the $P_{0,i}$, i.e., $P_{0,i,0}$ and $P_{0,i,1}$, denoting the success probability when the relay remains silent or transmits respectively. The success probabilities are given by $P_{0,i,0}=P_{0}\left(\frac{1}{1+\gamma_{0}} \right)^{i-1}$ and $P_{0,i,1}=P_{0}\left(1+\gamma_{0}r_{0}^{\alpha}g \right)^{-1}\left(\frac{1}{1+\gamma_{0}} \right)^{i-1}$, where $r_{0}$ is the distance between the users and the relay, $v_i$ is the parameter of the Rayleigh fading random variable at channel $i$, $\alpha$ is the path loss exponent and $g$ is the self-interference coefficient. The next theorem summarizes the results for the characteristics of the relay queue for the symmetric $n$-user case. \[thm:nusers\] The key performance measures for the relay queue in the $n$-symmetric user network are provided below. - The average service rate is $$\label{eq:mun} \mu=\sum_{k=0}^{n}{{n \choose k} {q_{0}q^{k}(1-q)^{n-k}}P_{0d,k}}.$$ - The probability that the queue at the relay is empty is $$\label{eq:probemptyn} P\left( Q=0 \right)=\frac{\displaystyle p_{-1}^{1}-\sum_{i=1}^{n}{ip_{i}^{1}}}{\displaystyle p_{-1}^{1}-\sum_{i=1}^{n}{ip_{i}^{1}}+\lambda_{0}}.$$ - The average arrival rate $\lambda$ is $$\label{eq:lamdan} \lambda=P\left(Q=0\right)\lambda_{0}+P\left(Q>0\right)\lambda_{1}.$$ The expressions for $\lambda_{0}$ and $\lambda_{1}$ are given in Appendix \[sec:app\_proof\_n\]. - The average relay queue size $Q$ is $${\scriptsize \label{eq:avQn} \begin{aligned} \overline{Q}=\frac{\displaystyle \left(\sum_{i=1}^{n}{ip_{i}^{1}}-p_{-1}^{1} \right)\sum_{i=1}^{n}{i(i+3)p_{i}^{0}}+\lambda_{0}\left(2p_{-1}^{1}-\sum_{i=1}^{n}{i(i+3)p_{i}^{1}} \right)}{\displaystyle 2\left(\sum_{i=1}^{n}{ip_{i}^{1}}-p_{-1}^{1} \right) \left(p_{-1}^{1}-\sum_{i=1}^{n}{ip_{i}^{1}}+\lambda_{0} \right)}. \end{aligned} }$$ See Appendix \[sec:app\_proof\_n\]. The values of $q_{0}$ for which the queue is stable are given by $q_{0min}<q_{0}<1$, where $q_{0min}$ is given in (\[eq:q0minn\]) in Appendix \[sec:app\_proof\_n\]. Throughput Analysis {#sec:thr_analysis} =================== In the previous section, we provided the main results on the relay queue characteristics, including the empty queue probability and the average queue length. Here, we derive the per-user throughput and the network aggregate throughput with one cooperative relay and $n$ users. The per-user throughput, $T_i$ for the $i$-th user is given by $T_i=T_{D,i}+T_{R,i}$, where $T_{D,i}$ denotes the direct throughput from user $i$ to the destination, i.e., the transmitted packet reaches the destination directly, without using the relay. When the transmission to the destination is not successful, and at the same time the relay node receives the packet correctly, then it stores it to its queue, and the contributed throughput by the relay for the user $i$ is denoted by $T_{R,i}$. When the queue at the relay is stable, $T_{R,i}$ is the arrival rate from user $i$ to the queue. The term $T_{D,i}$ can also be interpreted as the probability that a transmitted packet from user $i$ reaches the destination directly, and $T_{R,i}$ is the probability of unsuccessful transmission from user $i$ to the destination while the packet is received at the relay. The percentage of $i$-th user’s traffic that is being relayed is $\frac{T_{R,i}}{T_i}$. In the following subsection, we provide expressions for $T_{D,i}$ and $T_{R,i}$ for the two-user and the symmetric $n$-user cases. Per-user and Aggregate Throughput: Two-user Case ------------------------------------------------ The direct throughput to the destination for the $i$-th user, $T_{D,i}$, is given by $$\begin{aligned} T_{D,i} = q_{0}P\left(Q>0\right)q_{i} \left[ (1-q_{j}) P_{i/0,i}^{d}+q_{j} P_{i/0,i,j}^{d} \right]+ \\ +\left[1-q_{0}P\left(Q>0\right)\right]q_{i} \left[ (1-q_{j}) P_{i/i}^{d}+q_{j} P_{i/i,j}^{d} \right]. \end{aligned}$$ When the relay queue is stable, the contributed throughput to user $i$, $T_{R,i}$, is the arrival rate from user $i$ to the relay queue. Note that a packet from user $i$ enters the relay queue when the transmission to the destination is not successful and at the same time the relay is able to decode that packet. The relayed throughput $T_{R,i}$ of user $i$ is given by $${\scriptsize \begin{aligned} T_{R,i}=q_{0}P\left(Q>0\right)q_{i} \left[ (1-q_{j}) (1-P_{i/0,i}^{d})P_{i/0,i}^{0} +q_{j} (1-P_{i/0,i,j}^{d})P_{i/0,i,j}^{0} \right]+\\ +\left[1-q_{0}P\left(Q>0\right)\right]q_{i} \left[ (1-q_{j}) (1-P_{i/i}^{d})P_{i/i}^{0} +q_{j} (1-P_{i/i,j}^{d})P_{i/i,j}^{0}\right]. \end{aligned} }$$ The throughput $T_{i}$ for the $i$-th user is given by $$\label{eq:thr2_1} T_{i}= T_{D,i} + T_{R,i}.$$ In the above equations, the queue is assumed to be stable, hence the arrival rate from each user to the queue is a contribution to the overall throughput. The aggregate throughput is $T_{aggr} = T_{1} + T_{2}$. Notice that the per-user throughput is independent of $q_{0}$ as long as it is in the stability region. This is due to the fact that the product $q_{0}P\left(Q>0\right)$ is constant (does not depend on $q_0$). The proof is straightforward and thus is omitted. Per-user and Aggregate Throughput: Symmetric $n$-user Case ---------------------------------------------------------- In this subsection, we provide expressions for the direct and the relayed per-user and aggregate throughput. The notation used in Section \[sec:relay-results\_n\] applies here as well. Furthermore, the per-user throughput is denoted by $T$, the direct throughput to the destination by $T_D$, and the relayed throughput by $T_R$. The direct throughput $T_{D}$ is given by $$\begin{aligned} T_{D}=q_{0}P\left(Q>0\right)\sum_{k=0}^{n-1}{{n-1 \choose k}q^{k+1}(1-q)^{n-1-k} P_{d,k+1,1}}+\\ +\left[1-q_{0}P\left(Q>0\right)\right]\sum_{k=0}^{n-1}{{n-1 \choose k}q^{k+1}(1-q)^{n-1-k} P_{d,k+1,0}}. \end{aligned}$$ The throughput contributed by the relay (when the queue at the relay is stable), $T_{R}$, is given by $${\scriptsize \begin{aligned} T_{R}=q_{0}P\left(Q>0\right)\sum_{k=0}^{n-1}{{n-1 \choose k}q^{k+1}(1-q)^{n-1-k} (1-P_{d,k+1,1})P_{0,k+1,1}}+ \\ +\left[1-q_{0}P\left(Q>0\right)\right]\sum_{k=0}^{n-1}{{n-1 \choose k}q^{k+1}(1-q)^{n-1-k} \left(1-P_{d,k+1,0}\right)P_{0,k+1,0}}. \end{aligned} }$$ The per-user throughput $T$ for the cooperative relay network when the relay queue is stable is given by $$\label{eq:thrn} T = T_D + T_R.$$ The aggregate throughput is $T_{aggr} = n T$. When the queue is unstable, the aggregate throughput is the summation of the direct throughput among the users and the destination plus the service rate of the relay. However, when the queue is unstable, the queue size increases to infinity, thus there is no guarantee for finite queueing delay. Delay Analysis {#sec:delay_analysis} ============== In Section \[sec:analysis\], we studied the performance of the relay queue in terms of the probability of empty queue and the average queue length. That section was an intermediate step for our main goal, which is to study the impact of the relay node in the network in terms of throughput and the delay. In the previous section, we obtained the per-user and the aggregate throughput for a relay network with stable relay queue and commented on the case of unstable relay queue. In this section, we analyze an important network performance measure, the delay, and derive analytical expressions for the average delay required to deliver a packet from the source to the destination. \[thm:delay\] The average delay for a packet received at the destination when it is in the head of the user queue is given by $$\label{eq:delay} D_i=\frac{1+T_{R,i}\left(\frac{\overline{Q}}{\lambda} + \frac{1}{\mu} \right)}{T_i},$$ where $T_{R,i}$ and $T_i$ is the $i$-th user relayed and per-user throughput, respectively. $\lambda$ and $\mu$ is the average arrival and service rate of the relay, respectively, and $\overline{Q}$ is the average queue length of the relay. See Appendix \[sec:app\_delay\]. The expressions for $T_{R,i}$ and $T_i$ are given in Section \[sec:thr\_analysis\]. The expressions for $\lambda$, $\mu$, $\overline{Q}$ are summarized in Theorem \[thm:2users\] and \[thm:nusers\] for the two-user and the symmetric $n$-user case, respectively. Note that the term $\frac{\overline{Q}}{\lambda}$ in (\[eq:delay\]) is the queueing delay, which is the time a packet spends in queue, the time the packet is assigned to the queue for transmission and the time it starts being transmitted. In the meantime, the packet waits while other packets in the queue are transmitted. When the relay queue is unstable, the average queue length can be arbitrarily large, thus the average queueing delay tends to infinity. In (\[eq:delay\]), when the queue is unstable, then the average delay also tends to infinity. In the case of unstable queues, flow control policies could be applied for packet dropping, however this is beyond the scope of our paper. Numerical Results {#sec:results} ================= In this section, we provide numerical results to validate the above theoretical performance analysis. For exposition convenience, we consider the case where all users have the same link characteristics and transmission probabilities. The parameters used in the numerical results are as follows: distances are $r_{d}=130$, $r_{0}=60$, and $r_{0d}=80$ in meters, the path loss exponent is $\alpha=4$, and the receiver noise power $\eta=10^{-11}$. The transmit power for the relay is $P_{tx}(0)=10$ mW and for the $i$-th user is $P_{tx}(i)=1$ mW. Per user and Aggregate Throughput --------------------------------- Figs. \[fig:thr\_n\_02\] and \[fig:thr\_n\_06\] present the per-user throughput versus the number of users in the network for different values of $q$ and $g$, and for $\gamma=0.2$ and $\gamma=0.6$, respectively. Figs. \[fig:athr\_n\_02\] and \[fig:athr\_n\_06\] show the aggregate throughput versus the number of users. When $\gamma=0.2$, we observe that for $g=10^{-10}$ and $g=10^{-8}$ (almost perfect self-interference cancelation) the relay queue is unstable for relative small number of users. This is because for small values of $\gamma$, it is more likely to have more successful transmissions from the users to the relay, while at the same time the relay can transmit at most one packet per timeslot. For $\gamma=0.6$ the queue is never unstable for the selected set of parameters, while for $g=10^{-10}$ and $g=10^{-8}$, throughput gains are evident as compared to no self-interference cancelation. In Figs. \[fig:percent\_02\] and \[fig:percent\_06\], we plot the percentage of traffic that is being relayed in the network (cf. Section \[sec:thr\_analysis\]) for $\gamma=0.2$ and $\gamma=0.6$ respectively, for the case of a stable queue.  \[fig:percent\_02\]  \[fig:percent\_06\] Figs. \[fig:thr\_n\_12\] and \[fig:thr\_n\_25\] present the per-user throughput versus the number of users in the network for different values of $q$ and $g$, and for $\gamma=1.2$ and $\gamma=2.5$, respectively. Figs. \[fig:athr\_n\_12\] and \[fig:athr\_n\_25\] show the aggregate throughput versus the number of users. Finally, Figs. \[fig:percent\_12\] and \[fig:percent\_25\] show the percentage of traffic that is being relayed. Note that when the percentage tends to $1$ (or $100\%$), the contributed throughput by the relay tends to be the total network throughput. The gains from the relay are more pronounced for large $\gamma$, whilst in the case of $\gamma=0.2$ and quasi perfect self-interference cancelation, we tend to have an unstable queue, which affects the delay per packet as we will see in the next subsection.  \[fig:percent\_12\]  \[fig:percent\_25\] Average Queue Length and Average Delay per Packet ------------------------------------------------- In this subsection, we provide numerical results for two key performance metrics, namely the average relay queue size and the average delay per packet. Figs. \[fig:avQ\_n\_02\] and \[fig:avQ\_n\_06\] present the average queue length of the relay for $\gamma=0.2$ and $\gamma=0.6$. The average queue length is among the factors that affect the average delay per packet as presented in (\[eq:delay\]) of Theorem \[thm:delay\]. Figs. \[fig:Delay\_n\_02\] and \[fig:Delay\_n\_06\] illustrate the average delay per packet for $\gamma=0.2$ and $\gamma=0.6$. For $\gamma=0.2$ we included the per-packet delay for the network without the relay for comparison reasons. We observe that in that case the cooperative relay node does not provide any gains for increasing number of users, as the delay for the relay network is larger than the delay without using a relay. For $\gamma=0.6$, the delay for the network without the relay is much larger, e.g., it starts with $50$ timeslots for $1$ users and goes up to $400$ for $50$ users. In that case, the use of a relay is beneficial in terms of throughput and per-packet delay. Figs. \[fig:avQ\_n\_12\] and \[fig:avQ\_n\_25\] show the average queue length for $\gamma=1.2$ and $\gamma=2.5$ respectively. In Figs. \[fig:Delay\_n\_12\] and \[fig:Delay\_n\_25\] and \[fig:Delay\_n\_25\], we illustrate the average delay per packet. The delay for the network without the relay is significantly large, e.g., for $\gamma=1.2$ the delay is greater than $500$ timeslots and for $\gamma=2.5$ the delay is more than $10000$ timeslots. In those cases, the existence of the relay offers significant gains not only in terms of throughput but also in the delay performance. When we have almost perfect self-interference cancelation (except the case of $\gamma=0.2$), we observe significant gains in the delay performance compared to the case of the quasi half-duplex relay ($g \to 1$). Conclusions {#sec:conclusions} =========== In this paper, we explored full-duplex communication in which a cooperative node relays packets from a number of sources to a common destination node in a random access network with multi-packet reception capability for both the relay and the destination node. Considering a multiple capture model and the self-interference due to full-duplex relay operation, a transmission is successful if the received SINR is above a certain threshold $\gamma$. We provided analytical expressions for the performance of the relay queue, namely stability conditions, arrival and service rates, and average queue length. We studied the per-user and the aggregate throughput, and showed that the per-user throughput does not depend on the relay transmit probability under stability conditions. We also studied the impact of the self-interference coefficient $g$ on the per-user throughput, the network-wide throughput, and the average per-packet delay. We showed that the self-interference coefficient plays a crucial role when $\gamma$ is small (and $g$ tends to zero) since it may result in an unstable queue. However, for large $\gamma$ values and perfect self-interference cancelation, the gains in terms of throughput and delay are more pronounced. Future extensions of this work may include users with non-saturated queues, i.e. sources with external random arrivals, as well as scenarios where the cooperative relay node has packets on its own and different service priorities. Proof of Theorem \[thm:2users\] {#sec:app_proof_2} =============================== We provide here the proof of Theorem \[thm:2users\], which presents the main result for the relay queue characteristics for the two-user case. The average service rate $\mu$, is given by (\[eq:m2\]), where $q_{0}$ is the transmit probability of the relay given that it has packets in its queue, and $q_{i}$ for $i \neq 0$ is the transmit probability for the $i$-th user. The term $P_{i/i,k}^{j}$ is the success probability of link $ij$ when the transmitting nodes are $i$ and $k$ and can be calculated based on (\[eq:succprob\]). The average arrival rate $\lambda$ of the queue is given by $\lambda=P\left(Q=0\right)\lambda_{0}+P\left(Q>0\right)\lambda_{1}$, where $\lambda_{0}$ is the average arrival rate at the relay queue when the queue is empty and $\lambda_{1}$ when it is not. $\lambda_{0}=r_{1}^{0}+2r_{2}^{0}$, where $r_{i}^{0}$ is the probability of receiving $i$ packets given that the queue is empty. Accordingly, $\lambda_{1}=r_{1}^{1}+2r_{2}^{1}$, where $r_{i}^{1}$ is the probability of receiving $i$ packets when the queue is not empty. The expressions for $r_{i}^{0}$ are given by $$\label{eq:r10} {\scriptsize \begin{aligned} r_{1}^{0}=q_{1}(1-q_{2})(1-P_{1/1}^{d})P_{1/1}^{0}+q_{2}(1-q_{1})(1-P_{2/2}^{d})P_{2/2}^{0}+\\ +q_{1}q_{2}(1-P_{1/1,2}^{d})P_{1/1,2}^{0}P_{2/1,2}^{d}+q_{1}q_{2}(1-P_{2/1,2}^{d})P_{2/1,2}^{0}P_{1/1,2}^{d}+\\+q_{1}q_{2}(1-P_{1/1,2}^{d})P_{1/1,2}^{0}(1-P_{2/1,2}^{d})(1-P_{2/1,2}^{0})+\\ +q_{1}q_{2}(1-P_{2/1,2}^{d})P_{2/1,2}^{0}(1-P_{1/1,2}^{d})(1-P_{1/1,2}^{0}), \end{aligned} }$$ $$\label{eq:r20} \begin{aligned} r_{2}^{0}=q_{1}q_{2}(1-P_{1/1,2}^{d})(1-P_{2/1,2}^{d})P_{1/1,2}^{0}P_{2/1,2}^{0}. \end{aligned}$$ In order to compute for instance $r_{1}^{0}$ (i.e., the relay receives one packet), we have to take into account all the possible combinations, which are either the received packet is transmitted by the first or the second user (with all the possible combinations of active/idle users). When the relay queue is not empty, the expressions for the $r_{i}^{1}$ are given by $${\scriptsize \label{eq:r11} \begin{aligned} r_{1}^{1}=(1-q_{0})q_{1}(1-q_{2})(1-P_{1/1}^{d})P_{1/1}^{0}+q_{0}q_{1}(1-q_{2})(1-P_{1/0,1}^{d})P_{1/0,1}^{0}+\\ +(1-q_{0})q_{2}(1-q_{1})(1-P_{2/2}^{d})P_{2/2}^{0}+q_{0}q_{2}(1-q_{1})(1-P_{2/0,2}^{d})P_{2/0,2}^{0}+\\ +(1-q_{0})q_{1}q_{2}(1-P_{1/1,2}^{d})P_{1/1,2}^{0}(1-P_{2/1,2}^{d})(1-P_{2/1,2}^{0})+\\ +q_{0}q_{1}q_{2}(1-P_{1/0,1,2}^{d})P_{1/0,1,2}^{0}(1-P_{2/0,1,2}^{d})(1-P_{2/0,1,2}^{0})+\\ +(1-q_{0})q_{1}q_{2}(1-P_{1/1,2}^{d})P_{1/1,2}^{0}P_{2/1,2}^{d}+\\ +q_{0}q_{1}q_{2}(1-P_{1/0,1,2}^{d})P_{1/0,1,2}^{0}P_{2/0,1,2}^{d}+\\+(1-q_{0})q_{1}q_{2}(1-P_{2/1,2}^{d})P_{2/1,2}^{0}(1-P_{1/1,2}^{d})(1-P_{1/1,2}^{0})+\\ +q_{0}q_{1}q_{2}(1-P_{2/0,1,2}^{d})P_{2/0,1,2}^{0}(1-P_{1/0,1,2}^{d})(1-P_{1/0,1,2}^{0})+\\ +(1-q_{0})q_{1}q_{2}(1-P_{2/1,2}^{d})P_{2/1,2}^{0}P_{1/1,2}^{d}+\\+q_{0}q_{1}q_{2}(1-P_{2/0,1,2}^{d})P_{2/0,1,2}^{0}P_{1/0,1,2}^{d}, \end{aligned} }$$ $$\label{eq:r21} \begin{aligned} r_{2}^{1}=(1-q_{0})q_{1}q_{2}(1-P_{1/1,2}^{d})P_{1/1,2}^{0}(1-P_{2/1,2}^{d})P_{2/1,2}^{0}+\\ +q_{0}q_{1}q_{2}(1-P_{1/0,1,2}^{d})P_{1/0,1,2}^{0}(1-P_{2/0,1,2}^{d})P_{2/0,1,2}^{0}. \end{aligned}$$ ![The Markov Chain model for the two-user case.[]{data-label="fig:mc_rxtx"}](dtmc_rxtx.pdf) In order to fully characterize the average arrival rate at the relay, we have to compute the probability the queue is empty. We model the queue at the relay as a discrete time Markov Chain (DTMC), which describes the queue evolution and is presented in Fig. \[fig:mc\_rxtx\]. Each state is denoted by an integer and represents the queue size at the relay node. The transition matrix of the above DTMC is a lower Hessenberg matrix given by $$P=\left(\begin{array}{ccccc} a_{0} & b_{0} & 0 & 0 & \cdots \\ a_{1} & b_{1} & b_{0} & 0 & \cdots \\ a_{2} & b_{2} & b_{1} & b_{0} & \cdots \\ 0 & b_{3} & b_{2} & b_{1} & \cdots \\ 0 & 0 & b_{3} & b_{2} & \cdots\\ \vdots & \vdots & \vdots & \vdots & \ddots \end{array} \right) .$$ The elements of the matrix P, are $a_{0}=1-p_{1}^{0}-p_{2}^{0},a_{1}=p_{1}^{0},a_{2}=p_{2}^{0}$, $b_{0}=p_{-1}^{1}$ and $b_{i+1}=p_{i}^{1} \text{ } i=0,1,2,3$. The quantity $p_{i}^{0}$ ($p_{i}^{1}$) is the probability that the queue size increases by $i$ packets when the queue is empty (not empty). Note that $p_{i}^{0}=r_{i}^{0}$ because when the queue is empty, the probability of $i$ packets arriving is the same with the probability that the queue size increases by $i$ packets; when the queue is not empty however, this is not true. For example the probability of two packets arriving is not the same with the probability of doubling the queue size; this is because both arrivals and departures can occur at the same time. The expressions for the $p_{i}^{j}$ are given by $${\scriptsize \label{eq:p-11} \begin{aligned} p_{-1}^{1}=q_{0}(1-q_{1})(1-q_{2})P_{0/0}^{d}+q_{0}(1-q_{1})q_{2}P_{0/0,2}^{d}P_{2/0,2}^{d}+\\ +q_{0}(1-q_{1})q_{2}P_{0/0,2}^{d}(1-P_{2/0,2}^{d})(1-P_{2/0,2}^{0})+q_{0}q_{1}(1-q_{2})P_{0/0,1}^{d}P_{1/0,1}^{d}+\\ +q_{0}q_{1}(1-q_{2})P_{0/0,1}^{d}(1-P_{1/0,1}^{d})(1-P_{1/0,1}^{0})+\\+q_{0}q_{1}q_{2}P_{0/0,1,2}^{d}P_{1/0,1,2}^{d}P_{2/0,1,2}^{d}+\\ +q_{0}q_{1}q_{2}P_{0/0,1,2}^{d}(1-P_{1/0,1,2}^{d})(1-P_{1/0,1,2}^{0})(1-P_{2/0,1,2}^{d})(1-P_{2/0,1,2}^{0})+\\ +q_{0}q_{1}q_{2}P_{0/0,1,2}^{d}P_{1/0,1,2}^{d}(1-P_{2/0,1,2}^{d})(1-P_{2/0,1,2}^{0})+\\ +q_{0}q_{1}q_{2}P_{0/0,1,2}^{d}(1-P_{1/0,1,2}^{d})(1-P_{1/0,1,2}^{0})P_{2/0,1,2}^{d}, \end{aligned} }$$ $$\label{eq:p01} \begin{aligned} p_{0}^{1}=1-p_{-1}^{1}-p_{1}^{1}-p_{2}^{1}, \end{aligned}$$ $${\scriptsize \label{eq:p11} \begin{aligned} p_{1}^{1}=(1-q_{0})q_{1}(1-q_{2})(1-P_{1/1}^{d})P_{1/1}^{0}+(1-q_{0})q_{1}q_{2}(1-P_{1/1,2}^{d})P_{1/1,2}^{0}P_{2/1,2}^{d}+\\ +(1-q_{0})q_{1}q_{2}(1-P_{1/1,2}^{d})P_{1/1,2}^{0}(1-P_{2/1,2}^{d})(1-P_{2/1,2}^{0})+\\ +(1-q_{0})(1-q_{1})q_{2}(1-P_{2/2}^{d})P_{2/2}^{0}+(1-q_{0})q_{1}q_{2}(1-P_{2/1,2}^{d})P_{2/1,2}^{0}P_{1/1,2}^{d}+\\ +(1-q_{0})q_{1}q_{2}(1-P_{2/1,2}^{d})P_{2/1,2}^{0}(1-P_{1/1,2}^{d})(1-P_{1/1,2}^{0})+\\ +q_{0}q_{1}q_{2}P_{0/0,1,2}^{d}(1-P_{1/0,1,2}^{d})P_{1/0,1,2}^{0}(1-P_{2/0,1,2}^{d})P_{2/0,1,2}^{0}+\\ +q_{0}q_{1}(1-q_{2})(1-P_{0/0,1}^{d})(1-P_{1/0,1}^{d})P_{1/0,1}^{0}+\\ +q_{0}q_{1}q_{2}(1-P_{0/0,1,2}^{d})(1-P_{1/0,1,2}^{d})P_{1/0,1,2}^{0}P_{2/0,1,2}^{d}+\\ +q_{0}q_{1}q_{2}(1-P_{0/0,1,2}^{d})(1-P_{1/0,1,2}^{d})P_{1/0,1,2}^{0}(1-P_{2/0,1,2}^{d})(1-P_{2/0,1,2}^{0})+\\ +q_{0}q_{2}(1-q_{1})(1-P_{0/0,2}^{d})(1-P_{2/0,2}^{d})P_{2/0,2}^{0}+\\ +q_{0}q_{1}q_{2}(1-P_{0/0,1,2}^{d})(1-P_{2/0,1,2}^{d})P_{2/0,1,2}^{0}P_{1/0,1,2}^{d}+\\ +q_{0}q_{1}q_{2}(1-P_{0/0,1,2}^{d})(1-P_{2/0,1,2}^{d})P_{2/0,1,2}^{0}(1-P_{1/0,1,2}^{d})(1-P_{1/0,1,2}^{0}), \end{aligned} }$$ $$\label{eq:p21} \begin{aligned} p_{2}^{1}=(1-q_{0})q_{1}q_{2}(1-P_{1/1,2}^{d})P_{1/1,2}^{0}(1-P_{2/1,2}^{d})P_{2/1,2}^{0}+\\ +q_{0}q_{1}q_{2}(1-P_{0/0,1,2}^{d})(1-P_{1/0,1,2}^{d})P_{1/0,1,2}^{0}(1-P_{2/0,1,2}^{d})P_{2/0,1,2}^{0}. \end{aligned}$$ Note that $p_{i}^{j}$ and $r_{i}^{j}$ are in general different quantities, however in half-duplex relay systems we have $p_{i}^{j}=r_{i}^{j}$. The difference equations that govern the evolution of the states are given by $$Ps=s \Rightarrow s_{i}=a_{i}s_{0}+\sum_{j=1}^{i+1}{b_{i-j+1}s_{j}}.$$ We apply the z-transform technique to compute the steady state distribution, i.e., we let $$A(z)=\sum_{i=0}^{2}{a_{i}z^{-i}}, B(z)=\sum_{i=0}^{3}{b_{i}z^{-i}}, S(z)=\sum_{i=0}^{\infty}{s_{i}z^{-i}}.$$ We know that [@b:Gebali] $$S(z)=s_{0}\frac{z^{-1}A(z)-B(z)}{z^{-1}-B(z)}.$$ It is also known that the probability of the queue in the relay is empty is given by [@b:Gebali] $$\label{eq:P(Q=0)} P\left(Q=0\right)=\frac{1+B^{'}(1)}{1+B^{'}(1)-A^{'}(1)}.$$ The expressions of $A^{'}(1)$ and $B^{'}(1)$ are: $$\label{eq:A'(1)} \begin{aligned} A^{'}(z)=\left(\sum_{i=0}^{2}{a_{i}z^{-i}} \right)^{'}=-\sum_{i=1}^{2}{ia_{i}z^{-(i+1)}} \\ \Rightarrow A^{'}(1)=-\sum_{i=1}^{2}{ia_{i}}\Rightarrow A^{'}(1)=-\sum_{i=1}^{2}{ip_{i}^{0}}=-\lambda_{0}, \end{aligned}$$ $$\label{eq:B'(1)} \begin{aligned} B^{'}(z)=\left(\sum_{i=0}^{3}{b_{i}z^{-i}} \right)^{'}=-\sum_{i=0}^{3}{ib_{i}z^{-(i+1)}} \\ \Rightarrow B^{'}(1)=-\sum_{i=0}^{3}{ib_{i}}=-1+p_{-1}^{1}-p_{1}^{1}-2p_{2}^{1}. \end{aligned}$$ Then, the probability of the queue in the relay is empty is given by (\[eq:probempty2\]). Therefore, the average arrival rate $\lambda$ is given by (\[eq:lambda2\]). The average queue length is known to be $\overline{Q}=-S^{'}(1)$, where $S^{'}(1)=s_{0}\frac{K^{''}(1)}{L^{''}(1)}$ [@b:Gebali]. The expressions for $K(z)$ and $L(z)$ are given by $$\label{eq:K(z)} \begin{aligned} K(z)=\left(-z^{-2}A(z)+z^{-1}A^{'}(z)-B^{'}(z) \right) \left(z^{-1}-B(z)\right) - \\ - \left(z^{-1}A(z)-B(z) \right) \left(-z^{-2}-B^{'}(z) \right), \end{aligned}$$ $$\label{eq:L(z)} \begin{aligned} L(z)=\left(z^{-1}-B(z) \right)^{2}. \end{aligned}$$ Then $K^{''}(1)$ and $L^{''}(1)$ are given by $$\label{eq:K''(1)} {\scriptsize \begin{aligned} K^{''}(1)=\left(2A(1)-2A^{'}(1)+A^{''}(1)-B^{''}(1) \right) \left(-1-B^{'}(1) \right)-\\ - \left(2-B^{''}(1) \right) \left(-A(1)+A^{'}(1)-B^{'}(1) \right), \end{aligned} }$$ $$\label{eq:L''(1)} \begin{aligned} L^{''}(z)=\left[2\left(z^{-1}-B(z) \right) \left(-z^{-2}-B^{'}(z) \right) \right]^{'} \\ \Rightarrow L^{''}(1)=2\left(-1-B^{'}(1)\right)^{2}. \end{aligned}$$ The values of $A^{''}(1)$ and $B^{''}(1)$ are: $$\label{eq:A''(1)} \begin{aligned} A^{''}(z)=\left(-\sum_{i=1}^{2}{ia_{i}z^{-(i+1)}} \right)^{'}=\sum_{i=1}^{2}{i(i+1)a_{i}z^{-(i+2)}} \\ \Rightarrow A^{''}(1)=2p_{1}^{0}+6p_{2}^{0}, \end{aligned}$$ $$\label{eq:B''(1)} \begin{aligned} B^{''}(z)=\left(-\sum_{i=1}^{3}{ib_{i}z^{-(i+1)}} \right)^{'}=\sum_{i=1}^{3}{i(i+1)b_{i}z^{-(i+2)}} \\ \Rightarrow B^{''}(1)=2-2p_{-1}^{1}+4p_{1}^{1}+10p_{2}^{1}. \end{aligned}$$ The average queue length is given by (\[eq:avQ2\]). An important tool to determine stability is Loyne’s criterion [@b:Loynes], which states that if the arrival and service processes of a queue are jointly strictly stationary and ergodic, the queue is stable if and only if the average arrival rate is strictly less than the average service rate. If the queue is stable, the departure rate (throughput) is equal to the arrival rate, i.e., $\lambda_{1}<\mu\Leftrightarrow r_{1}^{1}+2r_{2}^{1}<\mu$ where $r_{1}^{1}=(1-q_{0})A_{1}+q_{0}B_{1}$, $r_{2}^{1}=(1-q_{0})A_{2}+q_{0}B_{2}$ and $\mu=q_{0}A$. The expressions for $A,A_{i},B_{i}$ are given by $$\label{eq:A1B1} \begin{aligned} A_{1}=q_{1}(1-q_{2})(1-P_{1/1}^{d})P_{1/1}^{0}+q_{2}(1-q_{1})(1-P_{2/2}^{d})P_{2/2}^{0}+\\ +q_{1}q_{2}(1-P_{1/1,2}^{d})P_{1/1,2}^{0}(1-P_{2/1,2}^{d})(1-P_{2/1,2}^{0})+\\+q_{1}q_{2}(1-P_{1/1,2}^{d})P_{1/1,2}^{0}P_{2/1,2}^{d}+\\ +q_{1}q_{2}(1-P_{2/1,2}^{d})P_{2/1,2}^{0}(1-P_{1/1,2}^{d})(1-P_{1/1,2}^{0})+\\+q_{1}q_{2}(1-P_{2/1,2}^{d})P_{2/1,2}^{0}P_{1/1,2}^{d}, \end{aligned}$$ $${\scriptsize \begin{aligned} B_{1}=q_{1}(1-q_{2})(1-P_{1/0,1}^{d})P_{1/0,1}^{0}+q_{2}(1-q_{1})(1-P_{2/0,2}^{d})P_{2/0,2}^{0}+\\ +q_{1}q_{2}(1-P_{1/0,1,2}^{d})P_{1/0,1,2}^{0}(1-P_{2/0,1,2}^{d})(1-P_{2/0,1,2}^{0})+\\+q_{1}q_{2}(1-P_{1/0,1,2}^{d})P_{1/0,1,2}^{0}P_{2/0,1,2}^{d}+\\ +q_{1}q_{2}(1-P_{2/0,1,2}^{d})P_{2/0,1,2}^{0}(1-P_{1/0,1,2}^{d})(1-P_{1/0,1,2}^{0})+\\+q_{1}q_{2}(1-P_{2/0,1,2}^{d})P_{2/0,1,2}^{0}P_{1/0,1,2}^{d}, \end{aligned} }$$ $$\label{eq:A2B2} \begin{aligned} A_{2}=q_{1}q_{2}(1-P_{1/1,2}^{d})P_{1/1,2}^{0}(1-P_{2/1,2}^{d})P_{2/1,2}^{0},\\ B_{2}=q_{1}q_{2}(1-P_{1/0,1,2}^{d})P_{1/0,1,2}^{0}(1-P_{2/0,1,2}^{d})P_{2/0,1,2}^{0}, \end{aligned}$$ $$\label{eq:A} \begin{aligned} A=(1-q_{1})(1-q_{2})P_{0/0}^{d}+q_{1}(1-q_{2})P_{0/0,1}^{d}+\\+q_{2}(1-q_{1})P_{0/0,2}^{d}+q_{1}q_{2}P_{0/0,1,2}^{d}. \end{aligned} $$ Then the values of $q_{0}$ for which the queue is stable are given by $q_{0min}<q_{0}<1$, where $$\label{eq:q0min2} q_{0min}=\frac{A_{1}+2A_{2}}{A+A_{1}+2A_{2}-B_{1}-2B_{2}}.$$ Proof of Theorem \[thm:nusers\] {#sec:app_proof_n} =============================== In this appendix, we provide the proof of Theorem \[thm:nusers\], which presents the relay’s queue characteristics for the symmetric $n$-user case. The service rate is given by (\[eq:mun\]). The average arrival rate $\lambda$ of the queue is $\lambda=P\left(Q=0\right)\lambda_{0}+P\left(Q>0\right)\lambda_{1}$, with $\lambda_{0}=\sum_{k=1}^{n}{kr_{k}^{0}}$, where the $r_{k}^{0}$ is the probability that the relay received $k$ packets when the queue is empty, given by $$\label{eq:rk0n} \begin{aligned} r_{k}^{0}=\sum_{i=k}^{n}{{n \choose i}{i \choose k} {q^{i}(1-q)^{n-i}}P_{0,i,0}^{k}} \times \\ \times \left(1-P_{d,i,0}\right)^{k}\left[1-P_{0,i,0}(1-P_{d,i,0})\right]^{i-k},\text{ }1 \leq k \leq n. \end{aligned}$$ $\lambda_{1}=\sum_{k=1}^{n}{kr_{k}^{1}}$, where the $r_{k}^{1}$ is the probability that the relay received $k$ packets when the queue is not empty and is given by $$\label{eq:rk1n} \begin{aligned} r_{k}^{1}=(1-q_{0})\sum_{i=k}^{n}{{n \choose i}{i \choose k} {q^{i}(1-q)^{n-i}}P_{0,i,0}^{k}} \times \\ \times \left(1-P_{d,i,0}\right)^{k}\left[1-P_{0,i,0}(1-P_{d,i,0})\right]^{i-k}+\\ +q_{0}\sum_{i=k}^{n}{{n \choose i}{i \choose k} {q^{i}(1-q)^{n-i}}P_{0,i,1}^{k}\left(1-P_{d,i,1}\right)^{k}} \times \\ \times \left[1-P_{0,i,1}(1-P_{d,i,1})\right]^{i-k},\text{ }1 \leq k \leq n. \end{aligned}$$ The elements of the transition matrix are given by $a_{k}=p_{k}^{0}$, $b_{0}=p_{-1}^{1}$, $b_{1}=p_{0}^{1}$ and $b_{k+1}=p_{k}^{1} \text{ } \forall k>0$ where $$\label{eq:pk0n} \begin{aligned} p_{k}^{0}=\sum_{i=k}^{n}{{n \choose i}{i \choose k} {q^{i}(1-q)^{n-i}}P_{0,i,0}^{k}\left(1-P_{d,i,0}\right)^{k}} \times \\ \times \left[1-P_{0,i,0}(1-P_{d,i,0})\right]^{i-k},\text{ }1 \leq k \leq n, \end{aligned}$$ $$\label{eq:p-11n} p_{-1}^{1}=q_{0}\sum_{k=0}^{n}{{n \choose k}q^{k}(1-q)^{n-k}P_{0d,k}\left[1-P_{0,k,1}(1-P_{d,k,1})\right]^{k}},$$ $$\label{eq:pk1n} \begin{aligned} p_{k}^{1}=(1-q_{0})\sum_{i=k}^{n}{{n \choose i}{i \choose k} {q^{i}(1-q)^{n-i}}P_{0,i,0}^{k}} \times \\ \times \left(1-P_{d,i,0}\right)^{k}\left[1-P_{0,i,0}(1-P_{d,i,0})\right]^{i-k}+\\ +q_{0}\sum_{i=k}^{n}{{n \choose i}{i \choose k} {q^{i}(1-q)^{n-i}}(1-P_{0d,i})P_{0,i,1}^{k}} \times \\ \times \left(1-P_{d,i,1}\right)^{k}\left[1-P_{0,i,1}(1-P_{d,i,1})\right]^{i-k}+\\ +q_{0}\sum_{i=k+1}^{n}{{n \choose i}{i \choose k+1} {q^{i}(1-q)^{n-i}}P_{0d,i}P_{0,i,1}^{k+1}} \times \\ \times \left(1-P_{d,i,1}\right)^{k+1}\left[1-P_{0,i,1}(1-P_{d,i,1})\right]^{i-k-1}, \end{aligned}$$ $$\label{eq:p01n} p_{0}^{1}=1-p_{-1}^{1}-\sum_{i=1}^{n}{p_{i}^{1}}.$$ The probability that the queue in the relay is empty is given by (\[eq:P(Q=0)\]), where the expressions for $A^{'}(1)$ and $B^{'}(1)$ are $$\label{eq:A'(1)} \begin{aligned} A^{'}(z)=\left(\sum_{i=0}^{n}{a_{i}z^{-i}} \right)^{'}=-\sum_{i=1}^{n}{ia_{i}z^{-(i+1)}} \\ \Rightarrow A^{'}(1)=-\sum_{i=1}^{n}{ia_{i}}\Rightarrow A^{'}(1)=-\sum_{i=1}^{n}{ip_{i}^{0}}=-\lambda_{0}, \end{aligned}$$ $$\label{eq:B'(1)} \begin{aligned} B^{'}(z)=\left(\sum_{i=0}^{n+1}{b_{i}z^{-i}} \right)^{'}=-\sum_{i=i}^{n+1}{ib_{i}z^{-(i+1)}} \\ \Rightarrow B^{'}(1)=-\sum_{i=i}^{n+1}{ib_{i}}=-b_{1}-\sum_{i=2}^{n+1}{ib_{i}}=-1+p_{-1}^{1}-\sum_{i=1}^{n}{ip_{i}^{1}}. \end{aligned}$$ Then the probability that the queue in the relay is empty is given by (\[eq:probemptyn\]). As we showed in Appendix \[sec:app\_proof\_2\], the average queue length is given by $\overline{Q}=-S^{'}(1)$, where $S^{'}(1)=s_{0}\frac{K^{''}(1)}{L^{''}(1)}$. The expressions for $K^{''}(1)$ and $L^{''}(1)$ are given by (\[eq:K”(1)\]) and (\[eq:L”(1)\]). The expressions for $A^{''}(1)$ and $B^{''}(1)$ are $$\label{eq:A''(1)} \begin{aligned} A^{''}(z)=\left(-\sum_{i=1}^{n}{ia_{i}z^{-(i+1)}} \right)^{'}=\sum_{i=1}^{n}{i(i+1)a_{i}z^{-(i+2)}} \\ \Rightarrow A^{''}(1)=\sum_{i=1}^{n}{i(i+1)a_{i}}=\sum_{i=1}^{n}{i(i+1)p_{i}^{0}}, \end{aligned}$$ $$\label{eq:B''(1)} \begin{aligned} B^{''}(z)=\left(-\sum_{i=i}^{n+1}{ib_{i}z^{-(i+1)}} \right)^{'}=\sum_{i=1}^{n+1}{i(i+1)b_{i}z^{-(i+2)}} \\ \Rightarrow B^{''}(1)=\sum_{i=1}^{n+1}{i(i+1)b_{i}}=2-2p_{-1}^{1}+\sum_{i=1}^{n}{i(i+3)p_{i}^{1}}. \end{aligned}$$ Following the same methodology as in Appendix \[sec:app\_proof\_2\], we obtain that the average queue length given by (\[eq:avQn\]). As in Appendix \[sec:app\_proof\_2\], the queue is stable if $\lambda_{1}<\mu\Leftrightarrow \sum_{k=1}^{n}{kr_{k}^{1}}<\mu$, where $r_{k}^{1}=(1-q_{0})A_{k}+q_{0}B_{k}$ and $\mu=q_{0}A$. The expressions for $A,A_{k},B_{k}$ are : [$$\label{eq:Akn} \begin{aligned} A_{k}=\sum_{i=k}^{n}{{n \choose i}{i \choose k} {q^{i}(1-q)^{n-i}}P_{0,i,0}^{k}\left(1-P_{d,i,0}\right)^{k}\left[1-P_{0,i,0}(1-P_{d,i,0})\right]^{i-k}}, \end{aligned}$$ $$\label{eq:Bkn} \begin{aligned} B_{k}=\sum_{i=k}^{n}{{n \choose i}{i \choose k} {q^{i}(1-q)^{n-i}}P_{0,i,1}^{k}\left(1-P_{d,i,1}\right)^{k}\left[1-P_{0,i,1}(1-P_{d,i,1})\right]^{i-k}}, \end{aligned}$$ ]{} $$\label{eq:An} \begin{aligned} A=\sum_{k=0}^{n}{{n \choose k} {q^{k}(1-q)^{n-k}}P_{0d,k}}. \end{aligned}$$ The values of $q_{0}$ for which the queue is stable are given by $q_{0min}<q_{0}<1$, where $$\label{eq:q0minn} q_{0min}=\frac{\displaystyle \sum_{k=1}^{n}{kA_{k}}}{\displaystyle A+\sum_{k=1}^{n}{kA_{k}}-\sum_{k=1}^{n}{kB_{k}}}.$$ Proof of Theorem \[thm:delay\] {#sec:app_delay} ============================== We first present the analysis for the average delay $D_i$ required to deliver a packet from source $i$ to the destination. This delay is the summation of the transmission delay from the source (to either the destination directly or the relay node), the queueing delay at the relay node, and the transmission delay from the relay to the destination. When a packet is transmitted from the $i$-th source, there is a probability that this packet reaches the destination directly, which is $T_{D,i}$. In the case that the transmission to the destination is not successful but is successful to the relay, the packet enters the relay queue, this is with probability $T_{R,i}$. The total time that the packet entering the relay queue reaches the destination is denoted by $D_Q$. If the transmission from the source to the destination is unsuccessful to both destination and relay nodes, then it remains at the source for future retransmission (with probability $1-T_{D,i}-T_{R,i}$). The average delay $D_i$ is given by $D_i=T_{D,i}+T_{R,i}\left(1+D_{R}\right)+(1-T_{D,i}-T_{R,i})\left(1+D_i \right)$, which after some simplifications results in $$\label{eq:Di} D_i=\frac{1+T_{R,i}D_{R}}{T_i}.$$ The expressions for $T_{R,i}$, $T_{D,i}$, and $T_i$, are given in Section \[sec:thr\_analysis\]. When the packet from source $i$ that enters the queue waits, while other packets in the queue are transmitted, this waiting time is the queue delay and is denoted by $D_Q$. When the packet that waits at the queue reaches the head of the queue, then it is transmitted from the relay with a probability (due to the random access assumption), the transmission delay from the relay to the destination is $\frac{1}{\mu}$, where $\mu$ is the service rate. The total delay in the relay node is denoted by $D_{R}$. The expression for $D_R$ is $D_{R}=D_{Q}+\mu+\left(1-\mu \right) \left(1+\frac{1}{\mu} \right)$, which is $$\begin{aligned} \label{eq:DR} D_{R}=D_{Q}+\frac{1}{\mu}.\end{aligned}$$ From Little’s law, we obtain that $D_{Q}=\frac{\overline{Q}}{\lambda}$, where $\overline{Q}$ is the average queue length for the relay and $\lambda$ is the average arrival rate. The expressions for $\overline{Q}$ and $\lambda$ are presented in Section \[sec:analysis\]. After substituting (\[eq:DR\]) into (\[eq:Di\]), we obtain (\[eq:delay\]) in Theorem \[thm:delay\]. Note that in our study we do not take into account the processing and the propagation delay. [^1]: The self-interference cancellation at the relay is modeled as a variable power gain that affects the success probability with which the relay will receive a packet and is described in Section \[sec:sysmod\]. [^2]: The case of half-duplex relay is studied in [@b:PappasMPR], for which the analysis is simpler compared to the full-duplex case. [^3]: Our work could be generalized to the asymmetric case; nevertheless the expressions will be significantly involved without providing any meaningful or crisp insights.

ArXiv

--- abstract: 'We consider an aspect of the open problem: Does every square-integrable function on $SU(2)$ have an almost everywhere convergent Fourier series? Let $0 < \alpha < 1$. We show that to each countable set $E$ in $SU(2)$ there corresponds an $\alpha$-Hölder continuous function on $SU(2)$ whose Fourier series diverges on $E$. We also show that the Fourier series of each $\alpha$-Hölder continuous function on $SU(2)$ converges almost everywhere.' address: | Missouri University of Science and Technology\ Department of Mathematics and Statistics\ Rolla, Missouri 65409-0020, USA. author: - David Grow and Donnie Myers date: '2020-03-24' title: 'Almost everywhere convergence of Fourier series on SU(2): the case of Hölder continuous functions ' --- Introduction ============ The Peter-Weyl theorem suggests the study of the formal Fourier series\ ${\sum d_{\lambda}(\chi_{\lambda}\ast f)}$ of a function $f$ on a compact, connected, semisimple Lie group $G$. Here the sum is over the equivalence classes of continuous irreducible unitary representations of $G$, $d_{\lambda}$ is the degree of the representation, and $\chi_{\lambda}$ is its character. The vast literature of Fourier analysis on $G$ is primarily concerned with mean convergence or divergence of the Fourier series of $f \in L^p(G)$ (e.g.[@GST; @GLZ; @S; @ST]), uniform or absolute convergence of the partial sums if $f$ is smooth (e.g.[@C; @GLZ; @Ma1; @DFM; @MG; @R1; @R2; @Sug; @T]), almost everywhere convergence or divergence of the partial sums if $f$ is a central function in $L^p(G)$ (e.g.[@CGTV; @GT]), and uniform, mean, or almost everywhere summability of the partial sums if $f$ belongs to various subspaces of $L^1(G)$ (e.g.[@CF; @CGT; @GLZ; @Z]). The aim of this work is to advance the study of almost everywhere convergence or divergence of Fourier partial sums of nonsmooth, possibly noncentral functions in $L^2(G)$. Let $SU(2)$ denote the two-dimensional special unitary group. We show that to each $\alpha$ in $(0,1)$ and to each countable subset $E$ of $SU(2)$ there corresponds an $\alpha$-Hölder continuous function on $SU(2)$ whose Fourier series diverges at each $x$ in $E$. Since it is possible to arrange that such a set $E$ is dense in $SU(2)$, the Fourier series of the corresponding function is divergent at infinitely many points in every nonempty open subset of $SU(2)$. Nevertheless, the Fourier series of each $\alpha$-Hölder continuous function on $SU(2)$ converges almost everywhere on $SU(2)$. In fact, relying on a general almost everywhere convergence result of Dai [@D] for Fourier-Laplace series on spheres, it follows that if $f$ in $L^2(SU(2))$ has an integral modulus of continuity $\Omega(f,t)$ satisfying $$\int_0^1\frac{\Omega^2(f,t)}{t}dt < \infty,$$ then the sequence of Fourier partial sums $\{S_{N}f(x)\}_{N=1}^{\infty}$ converges to $f(x)$ almost everywhere on $SU(2)$. In particular, if $f$ is an $\alpha$-Hölder continuous function on $SU(2)$ for some $\alpha$ in $(0,1)$, or more generally if $\Omega(f,t) = O(t^{\alpha})$ for some $\alpha$ in $(0,1)$, then the Fourier partial sums of $f$ converge to $f(x)$ almost everywhere on $SU(2)$. It is still an open problem whether $$\label{aeconvergence} \underset{N\rightarrow\infty}{\lim}S_{N}f(x) = f(x)$$ holds almost everywhere for every $f$ in $L^2(SU(2))$. It follows from a result of Pollard [@P] on Jacobi series that if $f$ is a central function in $L^p(SU(2))$ for some $p > 4/3$, then (\[aeconvergence\]) holds almost everywhere. On the other hand, a general theorem of Stanton and Tomas [@ST] for compact, connected, semisimple Lie groups shows that if $p < 2$ then there correspond an $f$ in $L^p(SU(2))$ and a subset $E$ of $SU(2)$ of full measure such that (\[aeconvergence\]) fails for all $x$ in $E$. Finally, the Peter-Weyl theorem implies that to each $f$ in $L^2(SU(2))$ there corresponds an increasing sequence $\{N_j\}$ of positive integers such that $$\underset{j\rightarrow\infty}{\lim}S_{N_j}f(x) = f(x)$$ for almost every $x$ in $SU(2)$. This lends some hope for a positive answer to the problem (\[aeconvergence\]), as do the results in this paper. Preliminaries ============= Let $SU(2)$ denote the two-dimensional special unitary group. General matrices $x,y \in SU(2)$ can be expressed as $$x=\left[\begin{array}{cc} \alpha_{1}+i\alpha_{2} & \beta_{1}+i\beta_{2}\\ -\left(\beta_{1}-i\beta_{2}\right) & \alpha_{1}-i\alpha_{2} \end{array}\right], \quad y=\left[\begin{array}{cc} \gamma_{1}+i\gamma_{2} & \delta_{1}+i\delta_{2}\\ -\left(\delta_{1}-i\delta_{2}\right) & \gamma_{1}-i\gamma_{2} \end{array}\right]$$ where $\alpha_{1},\alpha_{2},\beta_{1},\beta_{2},\gamma_{1},\gamma_{2},\delta_{1},\delta_{2}$ are real numbers satisfying $$\alpha_{1}^2+\alpha_{2}^2+\beta_{1}^2+\beta_{2}^2 = 1 = \gamma_{1}^2+\gamma_{2}^2+\delta_{1}^2+\delta_{2}^2.$$ Equip $SU(2)$ with the left and right translation invariant metric $d$ given by $$\begin{aligned} d(x,y)=& \sqrt{\frac{1}{2}\textrm{tr}((x-y)(x-y)^{*})} \\ =& \left((\alpha_1-\gamma_1)^2+(\alpha_2-\gamma_2)^2+(\beta_1-\delta_1)^2+(\beta_2-\delta_2)^2\right)^{1/2}. \notag\end{aligned}$$ Let $0 < \alpha < 1$ and let $f$ be a real function on $SU(2)$. If there exists a real number $M \geq 0$ such that $$\vert f(x)-f(y) \vert \leq Md^{\alpha}(x,y)$$ for all $x,y \in SU(2)$, then we say that $f$ is an $\alpha$-Hölder continuous function on $SU(2)$ and write $f\in\textrm{Lip}_{\alpha}(SU(2))$. Let $\mu$ denote normalized Haar measure on $SU(2)$. If $f$ and $g$ are Haar-integrable functions on $SU(2)$, then their convolution product is defined for all $x \in SU(2)$ by $$(f \ast g)(x)=\int_{SU(2)}f(xy^{-1})g(y)d\mu(y).$$ If $f \in \textrm{L}^1(SU(2))$ and $f \ast g = g \ast f$ for all $g \in \textrm{L}^1(SU(2))$, then we call $f$ a central function on $SU(2)$. This is equivalent to the property that for $\mu$-almost every $x \in SU(2)$, $f(yxy^{-1})=f(x)$ for all $y \in SU(2)$. In particular, since every $x \in SU(2)$ is diagonalizable via a similarity transformation: $$yxy^{-1} = \begin{bmatrix} &e^{i\theta} &0 \\ &0 &e^{-i\theta} \end{bmatrix} \equiv \omega(\theta)$$ where $y \in SU(2)$ and $e^{\pm i\theta}$ are the eigenvalues of $x$, it follows that if $f$ is central then for $\mu$-almost every $x \in SU(2)$, $$f(x) = f(\omega(\theta))$$ where $\theta \in [0,\pi]$. Furthermore, we have $$\label{CentralIntegral} \int_{SU(2)}f(x)d\mu(x)=\frac{2}{\pi}\int_{0}^{\pi}f(\omega(\theta))\sin^2(\theta)d\theta$$ when $f$ is central; this is a special case of equation (\[HaarIntegral\]) below. We denote the family of all (inequivalent) continuous, irreducible, unitary representations of $SU(2)$ by $\{\pi_n\}_{n=0}^{\infty}$ (cf. pp. 125-136 in vol. 2 of [@HR]). Observe that $\pi_n$ has dimension $n+1$ and its character $\chi_n$ is the continuous central function on $SU(2)$ given by $$\chi_n(x) = \textrm{trace}\left(\pi_n(x)\right).$$ If $e^{\pm i\theta}$ are the eigenvalues of $x$, then $$\chi_n(x) = \chi_n(\omega(\theta)) = \frac{\sin((n+1)\theta)}{\sin(\theta)};$$ the rightmost member this identity is the $n^{\text{th}}$ Chebyshev polynomial of the second kind on $[0,\pi]$. The Dirichlet kernel $\{\mathbf{D}_N\}_{N=0}^{\infty}$ on $SU(2)$ is the sequence of continuous central functions given by $$\mathbf{D}_N(x)=\sum_{n=0}^N (n+1)\chi_n(x).$$ The $N$th Fourier partial sum of a function $f \in \textrm{L}^1(SU(2)$ is the continuous function on $SU(2)$ given by $$S_Nf(x) = (f \ast \mathbf{D}_N)(x).$$ Divergence of Fourier partial sums on a countable subset ======================================================== \[divergence\] Let $\alpha\in(0,1)$ and let $\{x_{i}\}_{i=1}^{\infty}$ be any countable subset of $SU(2)$. Then there exists a function $f\in\textrm{Lip}_{\alpha}(SU(2))$ such that $$\underset{N\geq1}{\sup}|S_{N}f(x_{i})|=\infty$$ for all $i=1,2,3,\ldots$. Recall that for $\alpha \in (0,1)$, $\textrm{Lip}_{\alpha}(SU(2))$ is a Banach space with norm $$\Vert f\Vert_{\textrm{Lip}_{\alpha}(SU(2))}=\underset{x\in SU(2)}{\sup}|f(x)|+\underset{\underset{x\neq y}{x,y\in}SU(2)}{\sup}\frac{|f(x)-f(y)|}{d^{\alpha}(x,y)}.$$ Fix $x \in SU(2)$ and $n \in \mathbb{N}$, and set $\varLambda_{n}^{x}(f)=S_{n}f(x)$. Each $\varLambda_{n}^{x}$ is a bounded linear functional on $\textrm{Lip}_{\alpha}(SU(2))$ of norm $$\Vert\varLambda_{n}^{x}\Vert =\sup\left\{ |S_{n}f(x)|: f\in\textrm{Lip}_{\alpha}(SU(2)),\;\Vert f\Vert_{\textrm{Lip}_{\alpha}(SU(2))}\leq1\right\} \leq \Vert\boldsymbol{D}_{n}\Vert_{L^{1}(SU(2))}.$$ Specializing to the case when $x = e$, the identity matrix in $SU(2)$, we have $$\varLambda_{n}^{e}(f)=(f\ast\boldsymbol{D}_{n})(e)= \int_{SU(2)}f(y)\boldsymbol{D}_{n}(y)d\mu(y)= \frac{2}{\pi}\int_{0}^{\pi}f(\omega(\theta))\boldsymbol{D}_{n}(\omega(\theta))\sin^2(\theta)d\theta$$ by (\[CentralIntegral\]). Note that $$\boldsymbol{D}_{n}(\omega(\theta))=\frac{-1}{2\sin(\theta)}D_{n+1}^{\prime}(\theta)$$ where $$D_n(t)=\sum_{j=-n}^{n}e^{ijt}=1+2\sum_{j=1}^{n}\cos(jt)=\frac{\sin((2n+1)t/2)}{\sin(t/2)}$$ is the Dirichlet kernel on $[-\pi,\pi]$. It follows that $$\begin{aligned} \varLambda_{n}^{e}(f) =&\frac{1}{\pi}\int_{0}^{\pi}f(\omega(\theta))\cos^{2}\left(\frac{\theta}{2}\right)D_{n+1}(\theta)d\theta\\ &\quad - \frac{(2n+3)}{\pi}\int_{0}^{\pi}f(\omega(\theta))\cos\left(\left(n+\frac{3}{2}\right)\theta\right)\cos\left(\frac{\theta}{2}\right)d\theta.\end{aligned}$$ The absolute maxima and minima of the function $h_n(\theta)=\cos\left(\left(n+\frac{3}{2}\right)\theta\right)$ on $[0,\pi]$ occur at the endpoints of the intervals $I_{k}=\left[\frac{2k\pi}{2n+3},\frac{2(k+1)\pi}{2n+3}\right]$ where $k \in \{0,1,2,...,n\}$. Let $g_n$ be the sawtooth function on $[0,\pi]$ determined by $g_n\left(\frac{2k\pi}{2n+3}\right) = (-1)^k$ for $0\leq k \leq n+1$, $g_n(\pi)=0$, and $g_n$ is piecewise linear between these points. Define a central function $f_n$ on $SU(2)$ by $f_n(\omega(\theta)) = g_n(\theta)$ for $\theta \in [0,\pi]$. It is easy to see that $f_n$ belongs to $\textrm{Lip}_{\alpha}(SU(2))$; in fact, $$\frac{\vert f_n(x)-f_n(y) \vert}{d^{\alpha}(x,y)} \leq \left(\frac{\pi}{2}\right)^{\alpha}\left(\frac{2\pi}{2n+3}\right)^{1-\alpha} \leq \pi$$ for all distinct matrices $x$ and $y$ in $SU(2)$. Since $g_n(\theta)\cos\left(\left(n+\frac{3}{2}\right)\theta\right) \geq g_n^2(\theta) \geq 0$ on each interval $I_k$ and on $\left[\frac{2(n+1)\pi}{2n+3},\pi\right]$, and since the function $\theta \mapsto \cos(\theta/2)$ is positive and decreasing on $[0,\pi)$, it follows that $$\begin{aligned} \int_{I_k}f_n(\omega(\theta))\cos\left(\left(n+\frac{3}{2}\right)\theta\right)\cos(\theta/2)d\theta &\geq \int_{I_k}g_n^2(\theta)\cos(\theta)d\theta \\ &\geq \cos\left(\frac{(k+1)\pi}{2n+3}\right)\int_{I_k}g_n^2(\theta)d\theta \\ &= \frac{2\pi}{3(2n+3)}\cos\left(\frac{(k+1)\pi}{2n+3}\right)\end{aligned}$$ for all $k \in \{0,1,2,...,n\}$ and $$\int_{\frac{2(n+1)\pi}{2n+3}}^{\pi}f_n(\omega(\theta))\cos\left(\left(n+\frac{3}{2}\right)\theta\right)\cos(\theta/2)d\theta \geq 0.$$ Adding these inequalities we obtain $$\begin{aligned} \intop_{0}^{\pi}f_n(\omega(\theta))\cos\left(\left(n+\frac{3}{2}\right)\theta\right)\cos\left(\frac{\theta}{2}\right)d\theta & \geq \frac{2}{3}\left(\frac{\pi}{2n+3}\right)\sum_{k=1}^{n+1}\cos\left(\frac{k\pi}{2n+3}\right)\\ & = \frac{2}{3}\left(\frac{\pi}{2n+3}\right)\left\{ D_{n+1}\left(\frac{\pi}{2n+3}\right)-1\right\},\end{aligned}$$ and hence $$\biggl|\frac{(2n+3)}{\pi}\int_{0}^{\pi}f_n(\omega(\theta))\cos\left(\left(n+\frac{3}{2}\right)\theta\right)\cos\left(\frac{\theta}{2}\right)d\theta\biggr| \geq\frac{2}{3}\left\{ D_{n+1}\left(\frac{\pi}{2n+3}\right)-1\right\} .$$ Since the function $\theta\mapsto f_n(\omega(\theta))\cos^{2}(\theta/2)$ is uniformly bounded by $1$ on $[0,\pi]$, $$\begin{aligned} \biggl |\frac{1}{\pi}\int_{0}^{\pi}f_n(\omega(\theta))\cos^{2}(\theta/2)D_{n+1}(\theta)d\theta \biggr| & \leq \frac{1}{\pi}\intop_{0}^{\pi}|D_{n+1}(\theta)|d\theta \\ & = \frac{4}{\pi^{2}}\log(n+1)+o(1)\end{aligned}$$ as $n\rightarrow\infty$. Consequently $$\frac{|\varLambda_{n}^{e}(f_n)|}{\Vert f_n\Vert_{\textrm{Lip}_{\alpha}(SU(2))}} \geq \frac{\frac{2}{3}\left\{ D_{n+1}\left(\frac{\pi}{2n+3}\right)-1\right\} -\left(\frac{4}{\pi^{2}}\log(n+1)+o(1)\right)}{1+\pi}.$$ But $D_{n+1}\left(\frac{\pi}{2n+3}\right) = \left(\sin\left(\frac{\pi}{2(2n+3)}\right)\right)^{-1} \geq \frac{2(2n+3)}{\pi}$ and hence $$\Vert\varLambda_{n}^{e}\Vert = \sup\left\{\frac{|\varLambda_{n}^{e}(f)|}{\Vert f\Vert_{\textrm{Lip}_{\alpha}(SU(2))}}:\quad f\in\textrm{Lip}_{\alpha}(SU(2)), \quad f\neq 0 \right\}$$ is asymptotically bounded below by $$\frac{\frac{2}{\pi}(2n+3)-\frac{4}{\pi^{2}}\log(n+1)}{1+\pi}$$ as $n\rightarrow\infty$. Consequently, the sequence of bounded linear functionals $$\varLambda_{n}^{e}(f)=S_{n}f(e)$$ is not uniformly bounded on the Banach space $\textrm{Lip}_{\alpha}(SU(2))$ as $n\rightarrow\infty$. By the uniform boundedness principle $$\underset{n\geq1}{\sup}|S_{n}f(e)| = \infty$$ for all $f$ belonging to some dense $G_{\delta}$ set in $\textrm{Lip}_{\alpha}(SU(2))$. If $z \in SU(2)$, define the left translation operator $L_z$ on $\textrm{Lip}_{\alpha}(SU(2))$ by $L_{z}f(y) = f(zy)$ for all $y \in SU(2)$. For each element of the countable subset $\{x_{i}\}_{i=1}^{\infty}$ of $SU(2)$, observe that $$\frac{|\varLambda_{n}^{x_i}\left(L_{{x_i}^{-1}}f_{n}\right)|}{\Vert L_{{x_i}^{-1}}f_{n}\Vert_{\textrm{Lip}_{\alpha}(SU(2))}} =\frac{|\varLambda_{n}^{e}\left(f_{n}\right)|}{\Vert f_{n}\Vert_{\textrm{Lip}_{\alpha}(SU(2))}},$$ so there corresponds a dense $G_{\delta}$ subset $E_{x_i}$ of $\textrm{Lip}_{\alpha}(SU(2))$ such that $$\underset{n\geq1}{\sup}|S_{n}f(x_i)|=\infty$$ for all $f \in E_{x_i}$. By the Baire category theorem $E=\bigcap_{n=1}^{\infty}E_{x_{i}}$ is dense in $\textrm{Lip}_{\alpha}(SU(2))$. In particular, $E$ is nonempty and any $f \in E$ gives the desired conclusion. The problem of pointwise convergence for the Fourier series of central functions in $\textrm{Lip}_{\alpha}(SU(2))$ for some $\alpha \in (0,1)$ is related to an analogous problem for Fourier-Jacobi series of functions $f:[-1,1] \rightarrow \mathbb{R}$. Recall that the Chebyshev polynomials of the second kind $U_n (n=0,1,2,...)$ are a special case of the Jacobi polynomials $P_n^{(\alpha,\beta)}$ with $\alpha = \beta = 1/2$; i.e. $$U_n(\cos(\theta))= P_n^{(1/2,1/2)}(\cos(\theta)) = \frac{\sin((n+1)\theta)}{\sin(\theta)} \quad (n=0,1,2,...;\theta \in [0,\pi]).$$ Clearly $\{U_n\}_{n=0}^{\infty}$ is an orthogonal set of functions with respect to the inner product $$\langle F,G \rangle = \int_{-1}^{1}F(t)G(t)\sqrt{1-t^2}dt.$$ If $F \in L^2([-1,1],\sqrt{1-t^2}dt)$, let $s_N(F;t)$ denote the $N$th partial sum of the Fourier series of $F$ with respect to $\{U_n\}_{n=0}^{\infty}$. If $F$ has a modulus of continuity $\omega(F,h)$ satisfying the Dini-Lipschitz condition $$\underset{h\rightarrow 0^{+}}{\lim}\left(\omega(F,h)\log(h^{-1})\right)= 0,$$ then a classical theorem [@Sue] assures $s_N(F;t) \rightarrow F(t)$ uniformly on any interval $[-1+\delta, 1-\delta]$ with $\delta \in (0,1)$. Furthermore, a general theorem of Belen’kii [@B] guarantees that if $F$ satisfies a Dini-Lipschitz condition and $\{s_N^{(\alpha,\beta)}(F;\pm 1)\}_{N=0}^{\infty}$ converge for some $\alpha > -1$ and $\beta > -1$, then the Fourier-Jacobi series of $F$ converges uniformly to $F$ on $[-1,1]$. Observe that a central function $f \in L^2(SU(2))$ corresponds to a function $F \in L^2([-1,1],\sqrt{1-t^2}dt)$ via $F(\cos(\theta)) = f(x)$ where $e^{i\theta},e^{-i\theta}$ are the eigenvalues of $x \in SU(2)$. It is easy to check that the Fourier partial sums of $F$ and $f$ satisfy the identity $$s_N(F;\cos(\theta)) = (S_Nf)(x) \quad (N=0,1,2,...),$$ and if $f \in \textrm{Lip}_{\alpha}(SU(2))$ for some $\alpha \in (0,1)$ then $F$ satisfies a Dini-Lipschitz condition. This leads to the following result. \[central\] Let $f$ be a central function in $\textrm{Lip}_{\alpha}(SU(2))$ for some $\alpha \in (0,1)$. Then: 1. $S_Nf(x) \rightarrow f(x)$ uniformly outside any open set containing $\{e,-e\}$; 2. $S_Nf(x) \rightarrow f(x)$ uniformly on $SU(2)$ if $\{S_Nf(\pm e)\}_{N=0}^{\infty}$ converge. The authors of this paper originally wondered if they could delete the word “central” from the hypothesis of Theorem \[central\] and still obtain conclusions (a) and (b). Theorem \[divergence\] shows that there is no possibility of such an analogue of Theorem \[central\] for general functions in $\textrm{Lip}_{\alpha}(SU(2))$ for $0 < \alpha < 1$. This is so because the points of divergence for the Fourier partial sums of such a noncentral function need no longer be at the “poles” $\pm e$. According to Theorem \[divergence\], points of divergence for $\textrm{Lip}_{\alpha}(SU(2))$ functions can be dense in $SU(2)$. Almost everywhere convergence of Fourier partial sums ===================================================== Note that $SU(2)$ is isometrically homeomorphic to the unit sphere $S^3$ in $\mathbb{R}^4$ via the isometry $\eta$ from $SU(2)$ onto $S^3$ given by $$\left[\begin{array}{cc} \alpha_{1}+i\alpha_{2} & \beta_{1}+i\beta_{2}\\ -\beta_{1}+i\beta_{2} & \alpha_{1}-i\alpha_{2} \end{array}\right] \mapsto \left(\alpha_1,\alpha_2,\beta_1,\beta_2\right).$$ Furthermore, the spherical coordinate system on $S^3$: $$\alpha_{1}= \cos(\theta),\alpha_{2}= \sin(\theta)\cos(\phi),\beta_{1}= \sin(\theta)\sin(\phi)\cos(\psi), \beta_{2}= \sin(\theta)\sin(\phi)\sin(\psi),$$ where $\phi \in [0,\pi]$, $\theta \in [0,\pi]$, and $\psi \in [0,2\pi]$, can be transferred to $SU(2)$ and forms a convenient parametrization thereof: $$x(\phi,\theta,\psi)=\left[\begin{array}{cc} \cos(\theta)+i\sin(\theta)\cos(\phi) & \sin(\theta)\sin(\phi)e^{i\psi}\\ -\sin(\theta)\sin(\phi)e^{-i\psi} & \cos(\theta)-i\sin(\theta)\cos(\phi) \end{array}\right].$$ We denote by $\Phi$ the mapping from $[0,\pi]\times[0,\pi]\times[0,2\pi]$ onto $SU(2)$ given by $(\phi,\theta,\psi) \mapsto x(\phi,\theta,\psi)$. In spherical coordinates, normalized Haar measure $\mu$ on $SU(2)$ satisfies $$\label{HaarIntegral} \int_{SU(2)}f(x)d\mu(x) = \frac{1}{2\pi^2}\int_0^{\pi}\int_0^{\pi}\int_0^{2\pi} (f\circ\Phi)(\phi,\theta,\psi))\sin^2(\theta)\sin(\phi)d\psi d\phi d\theta$$ for every $f \in \textrm{L}^1(SU(2),\mu) = \textrm{L}^1(SU(2))$ (cf. pp. 133-134 in vol. 2 of [@HR]). Using spherical coordinates, the $N$th Fourier partial sum of a function $f \in \textrm{L}^1(SU(2))$ can be written as $$\begin{aligned} (&S_{N}f)(x(\phi_{0},\theta_{0},\psi_{0})) =\int_{SU(2)}\mathbf{D}_{N}(y)f(x(\phi_{0},\theta_{0},\psi_{0})y^{-1}(\phi,\theta,\psi))d\mu(y) \\ & \nonumber =\frac{-1}{4\pi^{2}}\int_{0}^{\pi}\int_{0}^{\pi}\int_{0}^{2\pi}D_{N+1}^{\prime}(\theta) f(x(\phi_{0},\theta_{0},\psi_{0})y^{-1}(\phi,\theta,\psi))\sin(\theta)\sin(\phi)d\psi d\phi d\theta\\ & \nonumber =\frac{-1}{\pi}\intop_{0}^{\pi}D_{N+1}^{\prime}(\theta)\sin(\theta)[Q_{x}f](\theta)d\theta,\end{aligned}$$ where $$[Q_{x}f](\theta)=\frac{1}{4\pi}\int_{0}^{\pi}\int_{0}^{2\pi} f(x(\phi_{0},\theta_{0},\psi_{0})y^{-1}(\phi,\theta,\psi))\sin(\phi)d\psi d\phi.$$ The above identities for the Fourier partial sums of $f \in L^p(SU(2))$ reduce convergence problems at a point $x$ in the three-dimensional manifold $SU(2)$ to the behavior of the real function $[Q_{x}f]$ on the interval $[0,\pi]$. See [@DFM] and [@MG] for applications of this principle. Because $SU(2)$ is isometrically homeomorphic to the unit sphere $S^3$, the theory of Fourier series of functions on $SU(2)$ is closely connected with that of Fourier-Laplace series on $S^3$. Specifically, if $\sigma_3$ denotes normalized surface measure on $S^3$, then the $N$th partial sum of the Fourier-Laplace series of $F \in L^2(S^3,\sigma_3)$ is equal to the $N$th partial sum of the Fourier series of $f = F\circ\eta \in L^2(SU(2))$ [@DFM pp. 93-94]. Moreover, it follows from (\[HaarIntegral\]) that $\sigma_{3}(E) = \mu(\eta^{-1}(E))$ for all spherical boxes $$E = \{\overrightarrow{r}(\phi,\theta,\psi) \in S^3 : (\phi,\theta,\psi) \in [\alpha,\beta]\times[\gamma,\delta]\times[\epsilon,\nu]\}$$ in $S^3$; in particular, sets of zero Haar measure in $SU(2)$ correspond to sets of zero surface measure in $S^3$. More generally, let $S^n$ denote the unit sphere in $\mathbb{R}^{n+1}$. For $F \in L^2(S^n,\sigma_n) = L^2(S^n)$ and $\theta \in (0,\pi)$, define the spherical translation operator $T_{\theta}$ by $$(T_{\theta}F)(\overrightarrow{x})=\frac{1}{\vert{S^{n-1}}\vert}\int_{\{\overrightarrow{y}\in S^n : \overrightarrow{x}^T\overrightarrow{y} = 0 \}}F(\overrightarrow{x}\cos(\theta)+\overrightarrow{y}\sin(\theta))d\nu(\overrightarrow{y})$$ where $\nu$ denotes surface measure on the $n-1$ dimensional sphere $\{\overrightarrow{y} \in S^n : \overrightarrow{x}^T\overrightarrow{y} = 0\}$. Let $$(\Delta_{\theta}F)(\overrightarrow{x})= F(\overrightarrow{x})-(T_{\theta}F)(\overrightarrow{x})$$ denote the difference operator acting on $F \in L^2(S^n)$, and for $t>0$ let $$\omega(F,t)=\sup\{\Vert\Delta_{\theta}F\Vert_{L^2(S^n)}:0 < \theta \leq t\}$$ denote the integral modulus of continuity of $F \in L^2(S^n)$. The following result is the $r=2$ case of Theorem 2 obtained by Feng Dai in [@D]. \[Dai\] Let $F \in L^2(S^n)$. If $$\int_0^1\frac{\omega^2(F,t)}{t}dt < \infty$$ then the Fourier-Laplace partial sums of $F$ at $\overrightarrow{x}$ converge to $F(\overrightarrow{x})$ $\sigma_n$-almost everywhere. Let us specialize to the case $n=3$ of the previous theorem. If $\overrightarrow{x}=\eta(x(\phi_0,\theta_0,\psi_0))$ then $$\begin{aligned} (T_{\theta}F)(\overrightarrow{x}) & =\frac{1}{4\pi}\int_{0}^{\pi}\int_{0}^{2\pi}F(O(\cos(\theta)\overrightarrow{e_1}+\sin(\theta)\overrightarrow{v}))\sin(\phi)d\psi d\phi \\ & = \frac{1}{4\pi}\int_{0}^{\pi}\int_{0}^{2\pi}F(\cos(\theta)\overrightarrow{x}+\sin(\theta)O\overrightarrow{v})\sin(\phi)d\psi d\phi\end{aligned}$$ where $\overrightarrow{e_1} = [1,0,0,0]^T$, $\overrightarrow{v}=[0,\cos(\phi),\sin(\phi)\cos(\psi),\sin(\phi)\sin(\psi)]^T$, and $$O = \cos(\theta_0)\operatorname{diag}(1,-1,-1,-1) + \sin(\theta_0)S$$ with $$S = \begin{bmatrix} 0 & \cos(\phi_0) & \sin(\phi_0)\cos(\psi_0) & \sin(\phi_0)\sin(\psi_0) \\ \cos(\phi_0) & 0 & \sin(\phi_0)\sin(\psi_0) & -\sin(\phi_0)\cos(\psi_0) \\ \sin(\phi_0)\cos(\psi_0) & -\sin(\phi_0)\sin(\psi_0) & 0 & \cos(\phi_0) \\ \sin(\phi_0)\sin(\psi_0) & \sin(\phi_0)\cos(\psi_0) & -\cos(\phi_0) & 0 \end{bmatrix}.$$ If $F = f\circ\eta^{-1}$ then a routine computation yields $$(T_{\theta}F)(\overrightarrow{x})=\frac{1}{4\pi}\int_{0}^{\pi}\int_{0}^{2\pi}f(x(\phi_0,\theta_0,\psi_0)y^{-1}(\phi,\theta,\psi)) \sin(\phi)d\psi d\phi$$ and hence it follows that $$\label{SphericalTranslationIdentity} (T_{\theta}F)(\overrightarrow{x}) = [Q_{x}f](\theta)$$ for all $0 < \theta < \pi$. Let $\theta \in [0,\pi]$. Define the difference operator $\delta_{\theta}$ on $L^2(SU(2))$ by $$\delta_{\theta}(f)(x) = f(x) - [Q_{x}f](\theta)$$ and let $$\Omega(f,t)=\sup\{\Vert\delta_{\theta}f\Vert_{L^2(SU(2))}:0 < \theta \leq t\}$$ denote the integral modulus of continuity of $f \in L^2(SU(2))$. Using Theorem \[Dai\] and identity (\[SphericalTranslationIdentity\]) yields the following almost everywhere convergence result on $SU(2)$. Let $f \in L^2(SU(2))$. If $$\int_0^1\frac{\Omega^2(f,t)}{t}dt < \infty$$ then the Fourier partial sums of $f$ at $x$ converge to $f(x)$ $\mu$-almost everywhere on $SU(2)$. The next three corollaries are immediate consequences. Let $f \in L^2(SU(2))$. If $\Omega(f,t)=O(\frac{1}{\log^{\beta}(1/t)})$ for some $\beta > 1/2$ then $S_{N}f(x) \rightarrow f(x)$ for almost every $x \in SU(2)$. Let $f \in L^2(SU(2))$. If $\Omega(f,t)=O\left(t^{\alpha}\right)$ for some $\alpha \in (0,1)$ then $S_{N}f(x) \rightarrow f(x)$ for almost every $x \in SU(2)$. Let $f \in \textrm{Lip}_{\alpha}(SU(2))$ for some $\alpha \in (0,1)$. Then $S_{N}f(x) \rightarrow f(x)$ for almost every $x \in SU(2)$. [7]{} Belen’kii, A. M., Uniform convergence of Fourier-Jacobi series on the orthogonality segment, Mathematical notes of the Academy of Sciences of the USSR, 46 (1989), 901-906. Chen, X. and D. Fan, On almost everywhere divergence of Bochner-Riesz means on compact Lie groups, Math. Z. 289 (2018), 961-981. Clerc, J.L., Sommes de Riesz et multiplicateurs sur un groupe de Lie, Ann. Inst. Fourier (Grenoble) 24 (1974), 149-172. Colzani, L., S. Giulini, and G. Travaglini, Sharp results for the mean summability of Fourier series on compact Lie groups, Math. Ann. 285 (1989), 75-84. Colzani, L., S. Giulini, G. Travaglini, and M. Vignati, Pointwise convergence of Fourier series on compact Lie groups, Coll. Math. 61 (1990), 379-386. Dai, F., A note on a.e. convergence of Fourier-Laplace series in $L^2$, J. Beijing Normal Univ. (Natur. Sci.), 35 (1999), 6-9. Giulini, S., P.M. Soardi, and G. Travaglini, Norms of characters and Fourier series on compact groups, J. Funct. Anal. 46 (1982), 88-101. Giulini, S. and G. Travaglini, Central Fourier analysis for Lorentz spaces on compact Lie groups, Monatsh. Math. 107 (1989), 207-215. Gong, S., S.X. Li, and X.A. Zheng, Harmonic analysis on classical groups, Proceedings of the Analysis Conference, Singapore 1986, S.T.L. Choy et al (ed.), Elsevier, 1988. Hewitt, E. and K.A. Ross, *Abstract Harmonic Analyis* (two volumes), Springer, 1963 and 1970. Mayer, R.A., Fourier series of differentiable functions on SU(2), Duke Math. J. 3 (1967), 549-554. Myers, D., *Pointwise and Uniform Convergence of Fourier Series* (PhD dissertation), Missouri University of Science and Technology, 2016. Myers, D. and D. Grow, Lipschitz functions on SU(2) have uniformly convergent Fourier series, J. Math. Anal. Appl., 458 (2018), 730-741. Pollard, H., The convergence almost everywhere of Legendre series, Proc. Amer. Math. Soc. 35 (1972), 442-444. Ragozin, D.L., Polynomial approximation on compact manifolds and homogeneous spaces, Trans. Amer. Math. Soc. 150 (1970), 41-53. Ragozin, D.L., Approximation theory, absolute convergence, and smoothness of random Fourier series on compact Lie groups, Math. Ann. 219 (1976), 1-11. Stanton, R.J., Mean convergence of Fourier series on compact Lie groups, Trans. Amer. Math. Soc., 218 (1976), 61-87. Stanton, R.J. and P.A. Tomas, Convergence of Fourier series on compact Lie groups, Bull. Amer. Math. Soc. 82 (1976), 61-62. Suetin, P. K., *Classical Orthogonal Polynomials* \[in Russian\], pp. 146, 300. Nauka, Moscow. Sugiura, M., Fourier series of smooth functions on compact Lie groups, Osaka J. Math. 8 (1971), 33-47. Taylor, M.E., Fourier series on compact Lie groups, Proc. Amer. Math. Soc. 19 (1968), 1103-1105. Založnik, A., Function spaces generated by blocks associated with spheres, Lie groups, and spaces of homogeneous type, Trans. Amer. Math. Soc., 309 (1988), 139-164.

ArXiv

--- abstract: 'We present deep single-dish  observations of a sample of six nearby E+A galaxies ($0.05<z<0.1$). A non-negligible fraction of a local sample of E+As are detected in HI. In four galaxies, we have detected up to a few times $10^9$ [M$_\sun$]{} of neutral gas. These E+A galaxies are almost as gas-rich as spiral galaxies with comparable luminosities. There appears to exist no direct correlation between the amount of  present in an E+A galaxy and its star-formation rate as traced by radio continuum emission. Moreover, the end of the starburst does not necessarily require the complete exhaustion of the neutral gas reservoir. Most likely, an intense burst of star formation consumed the dense molecular clouds, which are the sites of massive star formation. This effectively stops star formation, even though copious amounts of diffuse neutral gas remain. The remaining reservoir may eventually lead to further episodes of star formation. This may indicate that some E+As are observed in the inactive phase of the star-formation duty cycle.' author: - 'P. Buyle, D. Michielsen, S. De Rijcke, D.J. Pisano, H. Dejonghe, K. Freeman' title: The  content of E+A galaxies --- Introduction {#intro} ============ At $z \approx 0.5$, galaxy clusters contain a population of blue, distorted galaxies that is missing in local clusters: the so-called Butcher-Oemler effect [@bo78]. In their spectroscopic study of blue galaxies in three clusters at z$\sim$0.31 @cs87 found that sixty percent of these blue galaxies (classified as Type 3 galaxies) have optical spectra characterised by strong Balmer absorption lines, typical for a very young stellar population, but with weak, if any, emission lines, such as [\[\]]{} $\lambda$3727[Å]{}. Since these spectra are a superposition of an old stellar population, resembling that of an elliptical ’or E’ galaxy, and a young stellar population, dominated by A stars, these galaxies are called E+A galaxies. In the notation of @dr99, these galaxies would be classified as k+a/a+k, satisfying the criteria EW([\[\]]{})$< 5$[Å]{} and EW(H$\delta)>$3[Å]{}. While at $z \approx 0.4$, about 20% of all cluster galaxies are classified as E+As [@belloni95], they constitute less than 1% of the present-day clusterpopulation [@fabricant91]. Apparently, these galaxies are observed during a quiescent phase, which explains the lack of emission lines, soon after a vigorous starburst, which explains the strong Balmer absorption lines. Due to the short lifetimes of the stars causing the Balmer absorption, the starburst must have ended no more than $\sim 1$ Gyr ago [@dg83; @p99]. E+As often have disturbed morphologies, e.g. tidal tails, suggestive of recent merger or interaction events. They span the whole morphological range, from bulge-dominated with underlying disks to disk-dominated [@tran03; @yang04]. Their high surface brightness sets them apart from the elliptical and lenticular galaxies in the Fundamental Plane. Over time, fading of the stellar population will drive them towards the locus of the E/S0s [@yang04; @p99]. Internal velocity dispersions of galaxies classified as E+As appear to increase as a function of redshift, going from $\sigma \simeq 150$ [kms$^{-1}$]{} at $z = 0.3$ to $\sigma \simeq 250$ [kms$^{-1}$]{} at $z=0.83$ [@tran03]. This trend suggests that massive galaxies undergo an E+A phase, i.e. are observed in a post-starburst phase, at earlier cosmic times than less massive ones. This is reminiscent of the “down-sizing” phenomenon in star-forming galaxies [@c96], according to which the masses of galaxies hosting star formation decrease as the Universe ages. ------------------------------ ------------ -------------------------------- ------- ------------------ ------------------ ------------------ ----------------------- Galaxy RA (J2000) $\delta$ (J2000) $M_B$ $v_{\rm hel}$ $\Delta v$ $\int S(v)\,dv$ H[i]{} mass (h,m,s) ($^\circ$,$\arcmin$,$\arcsec$) (mag) ([kms$^{-1}$]{}) ([kms$^{-1}$]{}) (Jy km s$^{-1}$) ($10^9$ [M$_\sun$]{}) [SDSS J210258.87+103300.6]{} 21 02 58.9 +10 33 01 -21.7 27821 440 $0.18 \pm 0.02$ $6.5 \pm 0.8$ [SDSS J230743.41+152558.4]{} 23 07 43.4 +15 25 58 -20.4 20894 240 $0.04 \pm 0.01$ $0.9 \pm 0.3$ [SDSS J233453.20+145048.7]{} 23 34 53.2 +14 50 49 -20.2 19388 380 $0.15 \pm 0.02$ $2.7 \pm 0.3$c LCRS B101120.1-024053 (EA17) 10 13 52.4 -02 55 48 -18.2 18258 $<0.18$ $< 2.9 $ LCRS B002018.8-415015 (EA18) 00 22 47.1 -41 33 37 -18.9 17941 660 $0.15 \pm 0.02$ $2.3 \pm 0.3$ LCRS B020551.6-453502 (EA19) 02 07 49.7 -45 20 50 -18.9 19186 $<0.07$ $< 1.2 $ ------------------------------ ------------ -------------------------------- ------- ------------------ ------------------ ------------------ ----------------------- \ ------------------------------ ------------- ----------- ------- -------------- --------------------------- Galaxy \[O[ii]{}\] H$\delta$ $r_e$ Sérsic index SFR ([Å]{}) ([Å]{}) (kpc) ([M$_\sun$]{}/yr) [SDSS J210258.87+103300.6]{} -0.68 5.13 4.6 3.2 $10.2^{^{+15.7}_{-6.5}}$ [SDSS J230743.41+152558.4]{} -0.92 6.78 2.0 3.1 $1.2^{^{+1.8}_{-0.8}}$ [SDSS J233453.20+145048.7]{} -1.27 5.18 2.1 3.3 $5.6^{^{+10.1}_{-3.7}}$ LCRS B101120.1-024053 (EA17) 1.68 6.92 1.5 2.7/1.3 $< 8.2^{^{+18.1}_{-5.9}}$ LCRS B002018.8-415015 (EA18) 1.75 5.96 1.7 2.2 $5.3^{^{+10.4}_{-3.7}}$ LCRS B020551.6-453502 (EA19) 0.98 6.08 2.1 1.1 $<1.8^{^{+2.7}_{-1.1}}$ ------------------------------ ------------- ----------- ------- -------------- --------------------------- Although the majority of low-$z$ E+A galaxies have a smooth light distribution, many of them also show slightly disturbed morphologies (e.g. warps and dust lanes). Based on this, @zab96 argued that the E+A phase is the aftermath of a vigorous starburst, triggered by a major merger or interaction. This is corroborated by a study of E+A galaxies drawn from the 2dFGRS catalogue [@blake04]. About three quarters of all E+A galaxies are found in the field (i.e. outside clusters), simply because most of the galaxies in the universe do not reside in clusters. However, the *fraction* of E+As is four times higher in clusters than in the field [@zab96; @tran04]. The spatial distribution of E+A galaxies in clusters is more extended than that of quiescent galaxies, but less extended than that of emission-line galaxies [@dr99], suggesting that processes such as galaxy harassment or ram-pressure stripping, which are specific to clusters, can also cause the E+A phenomenon. Based on the 2dFGRS E+A sample, @blake04 find that the distribution of [*local*]{} environments of E+A galaxies closely traces that of the ensemble of 2dFGRS galaxies and conclude that whatever causes the E+A phenomenon, must be a very local mechanism, such as encounters of galaxy pairs. This is corroborated by a recent analysis of the environments of E+As selected from the SDSS [@goto05b]. Most E+As have E/S0-like morphologies, with a small fraction of ongoing interactions. Their luminosity distribution is more similar to the distribution of spectroscopically defined elliptical galaxies than to the luminosity distribution of the ensemble of 2dFGRS galaxies [@blake04]. However, not all E+As can be associated with mergers and, obviously, more than one evolutionary pathway can lead to a post-starburst galaxy [@tran03; @dr99]. Numerical simulations show that E+As can indeed be formed via a major merger of two gas-rich spiral galaxies [@bekki05]. A disk-disk merger event then triggers a starburst, which consumes, or, by feedback, expels most of the available gas and then subsides. The young stars then dominate the optical spectrum for the following few hundred Myr while emission lines are absent. During this time-span, a galaxy would be classified as an E+A. In this case, one expects star-formation to be centrally concentrated, leading to radially decreasing Balmer line strengths [@pracy05]. Alternatively, star formation can be truncated more or less instantaneously over the whole disk of the galaxy without a starburst, e.g. by the gas being swept away by ram pressure stripping. In this case, as the young star population fades, the older bulge population causes the strengths of the Balmer lines to be radially increasing [@pracy05]. The red colours of some H$\delta$-strong E+As cannot be explained by any plausible starburst model [@cs87; @blake04], leaving heavy dust extinction as the only viable explanation. This hypothesis can be tested by using dust insensitive tracers of star formation. Since radio continuum emission is synchrotron radiation from electrons accelerated in supernova remnants, it is an indirect tracer of star formation. @mo01 observed part of the @zab96 sample and detected radio-continuum emission in only two out of fifteen galaxies. @smg99 detect five out of eight post-starburst galaxies at radio wavelengths. Radio-continuum observations of a sample of 36 E+As drawn from the SDSS yielded no detections [@goto05]. If this apparent lack of ongoing star-formation in E+As is true, then no dust obscuration, hiding the star-formation sites, needs to be invoked. Near-infrared studies [@gal00; @bal05] have shown that the u$-$g and r$-$k colours, and the H$\delta$ line-strengths of E+As can be well explained by dust-free models in which more than 5% of the stellar mass has recently been produced in a starburst. Hence, the presence of dust is still uncertain because of these contradictory observations. Up to now only one search for  in E+As was conducted [@chang01]. VLA observations of five E+As from the sample of @zab96 resulted in the detection of only one field E+A galaxy, EA1, with a total H[i]{} mass of $7.1 \pm 0.4 \times 10^9\,M_\odot$ (assuming $H_0 = 70$ [kms$^{-1}$]{}Mpc$^{-1}$). For the four other galaxies, upper limits of order $10^9\,M_\odot$ could be derived. EA1 consists of two components, and is most likely a merger remnant. However, other galaxies in this sample are also optical mergers but they do not contain detectable amounts of gas. We started an  study of a sample of E+A galaxies in order to constrain the amount of neutral gas present in these systems. In section \[obs\] we describe the sample and the observations. The results are presented in section \[results\]. We discuss these results in section \[discussion\] and summarise our conclusions in section \[concl\].  \  \   \[HIspectra\] Observations {#obs} ============ The sample ---------- Since [*(i)*]{} the occurrence of the E+A phenomenon seems to be determined predominantly by the local environment, [*(ii)*]{} the mechanisms triggering the E+A phenomenon, i.e. mergers, interactions, ram-pressure stripping, act both at high and low redshift, and [*(iii)*]{} for a given flux, the mass scales with distance squared, we opted to use distance as our main selection criterion to maximise our chances of a detection. We selected the three closest E+A galaxies from the large catalogue compiled by @goto03 from the SDSS, and the three closest E+A galaxies from the compilation of @zab96. The E+A sample selected from the SDSS by @goto03 satisfies very strict criteria and contains only galaxies with EW(H$\delta)>4{\rm\AA}+\Delta{\rm EW}({\rm H}\delta)$, with $\Delta{\rm EW}({\rm H}\delta)$ the 1$\sigma$ error on the H$\delta$ equivalent width, and no detectable [\[\]]{} and H$\alpha$ emission, quantified by the constraints EW([\[\]]{}$)<\Delta$EW([\[\]]{}) and EW(H$\alpha)<\Delta{\rm EW}({\rm H}\alpha)$, respectively. The Zabludoff sample satisfies the following criteria: EW(H$\beta\gamma\delta)>5.5{\rm\AA}$ and EW([\[\]]{}$)<2.5$[Å]{}, with EW(H$\beta\gamma\delta$) the mean equivalent width of the H$\beta$, H$\gamma$, and H$\delta$ absorption lines. Table \[EAproperties\] summarises the properties of the galaxies in our sample. Arecibo observations {#arecibo} -------------------- We observed the galaxies SDSS J210258.87+103300.6, [[SDSS J230743.41+152558.4]{}]{}, and [[SDSS J233453.20+145048.7]{}]{} for 7.5 h each, including overhead, with the 305m Arecibo Radio Telescope[^1] in Puerto Rico. The observations were scheduled on the nights of 23$-$25 June, 13$-$15 July and 28$-$30 July 2005. Each galaxy was observed for 2.5 hours per day during night-time to minimise solar interference. We used the L-wide receiver which has an average system temperature of $\approx 27$ K (depending on the elevation of the source). We selected the interim correlator in both linear polarisations to process the data. This resulted in final spectra with a total bandwidth of 25 MHz and 12.5 MHz divided over 1024 channels, resulting in a velocity resolution of 6.3 [[kms$^{-1}$]{}]{} and 3.15 [[kms$^{-1}$]{}]{} respectively. For [[SDSS J230743.41+152558.4]{}]{} and [[SDSS J233453.20+145048.7]{}]{}, we used the radar blanker to minimise the effect of the FAA Airport radar at 1330 MHz and 1350 MHz. The beam size of the L-wide receiver is $3.1'\times3.5'$. We applied the standard position-switching algorithm. Each galaxy was observed for 5 minutes, followed by a 5 minute offset by 5$'$ in right ascension to blank sky, such that we tracked the same azimuth and zenith angle as the on-source scan. This mode was used for all galaxies. We reduced the data by means of the standard Arecibo [IDL]{} routines, written by P. Perrilat. We calibrated each bandwidth individually and the polarisations were averaged together before a second-order baseline was fit across the interference-free part of the spectrum. We checked, with NED, the recession velocities of all galaxies inside the Arecibo beam to avoid confusion with other objects. The results are presented in section \[results\]. Parkes observations {#parkes} ------------------- We observed the galaxies LCRS B101120.1-024053 (EA17), LCRS B002018.8-415015 (EA18), and LCRS B020551.6-453502 (EA19) with the 64m ATNF Parkes Radio Telescope[^2] in Australia from dusk till dawn from 11 till 15 October 2005. EA17 was observed during sunrise. We used the Multibeam Correlator in the MB13 configuration (beam-switching mode) which enabled us to observe with 7 beams simultaneously, with one beam on the source while the other six were pointed adjacently on the sky. The beams were switched in position each 5 minutes. This way, we derived spectra with a bandwidth of 64 MHz divided over 1024 channels which yields a spectral resolution of 13.19 [[kms$^{-1}$]{}]{} and a beam width of 14.1$'$. The integration times, including overhead, were 10 h for EA17, 21 h for EA18, and 10 h for EA19. This resulted in one clear 3$\sigma$-detection of EA18. All observational quantities are listed in table \[EAproperties\]. The data were reduced by means of the [Livedata]{} data reduction pipeline, which is especially developed for the Multibeam Correlator. No obvious radio interference could be observed and a second order polynomial was used to fit the spectral baseline, after masking out a region around the optical velocity. Afterwards, all data were combined with the help of the [Gridzilla]{} software package using the median of the weighted values as an estimator. Finally, residual baseline ripples were removed by using the [mbspect]{} fitting algorithm in the [MIRIAD]{} [@sault95] software package. Again we checked the recession velocities of all other objects within the Parkes beam to avoid confusion. The results are presented in section \[results\]. {width="7cm"} {width="7cm"}\ {width="7cm"} {width="7cm"}\ {width="7cm"} {width="7cm"} Results =======  masses ------- In Figure \[HIspectra\], we show the  spectra of our six E+As. Three of them have a clear 3$\sigma$ detection: SDSS J210258.87+103300.6, [[SDSS J233453.20+145048.7]{}]{}, and EA18. In the case of [SDSS J230743.41+152558.4]{}, there is a peak at the correct velocity, over a velocity width of 240 [kms$^{-1}$]{}, which could be a tentative 2.5$\sigma$ detection of this galaxy. We did not detect EA17 and EA19. The observations of EA17 were badly affected by solar interference though. We calculate  masses for the detected objects and 3$\sigma$ upper limits for the undetected galaxies using the formula $$M_{\rm H\,I} = 2.36 \times 10^5\,M_\odot\, D^2 \int S(v)\, dv$$ with the distance $D$ in Mpc and $\int S(v) \,dv$ the total flux density in Jy [kms$^{-1}$]{}. The distance $D$ was calculated as the Hubble distance $D = v/H_0$, using $H_0 = 70$ [kms$^{-1}$]{}Mpc$^{-1}$. We find masses of $6.5 \pm 0.8 \times 10^9\,M_\odot$ for [[SDSS J210258.87+103300.6]{}]{}, $0.9 \pm 0.3 \times 10^9\,M_\odot$ for [[SDSS J230743.41+152558.4]{}]{}, $2.7 \pm 0.3 \times 10^9\,M_\odot$ for [[SDSS J233453.20+145048.7]{}]{}, and $2.3 \pm 0.3 \times 10^9\,M_\odot$ for EA18, as listed in Table \[EAproperties\]. To estimate the error on the mass of a galaxy, we generated 50000 statistically equivalent renditions of its radio spectrum by adding Gaussian noise to the original spectrum, using the measured 1$\sigma$ noise on the original datapoints. The mass error is taken to be the rms of the 50000 masses measured from these spectra. The $3\sigma$ upperlimits for the gas content of EA17 and EA19 are $2.9 \times 10^9\,M_\odot$ and $1.2 \times 10^9\,M_\odot$ respectively, assuming a total velocity width of 450 [[kms$^{-1}$]{}]{} (the average velocity width of the detected galaxies). Dust content ------------ Given the small angular sizes of these galaxies, their dust content can only be studied accurately with the supreme spatial resolution of the Hubble Space Telescope (HST). The HST archive contains images of EA17, EA18, and EA19, obtained with the Advanced Camera for Surveys (ACS) through the F435W and F625W filters. In Fig. \[ima\_EA\], we show the F625W images and F425W$-$F625W color maps of these galaxies. From the images it is obvious that the E+A class of galaxies comprises very different morphological types. EA17 has a smooth appearance, but it also contains a dust lane, which is clearly visible in both the colour-map and plain image. The only galaxy of this sample of three, that was detected in , is EA18, which is an irregular-looking galaxy with lots of red and blue patches along the major axis, pointing to dust and/or young stars. EA18 appears also to be warped. Finally, EA19 is a spiral (Sab) galaxy, with two spiral arms nicely visible in both images. The spiral arms are slightly bluer than the rest of the disk and the centre contains blue and red patches, pointing again to dust and/or young stars. From just this small sample, it is already clear that dust is present near the centres of some E+A galaxies. Along with the presence of neutral gas this might indicate that on-going star formation could in fact be hidden by dust. Environment ----------- Although E+A galaxies are predominately located in the field (i.e. outside clusters), their fraction is four times higher in clusters [@zab96; @tran04]. By investigating their spatial distribution inside clusters, @dr99 showed that an E+A phenomenon can be caused by environmental processes such as galaxy harassment or ram-pressure stripping. Such processes act mostly on the neutral hydrogen content of galaxies and can therefore be investigated by means of our observations. We adopted the determination of cluster membership of @zab96 and also checked the number of neighbours within a radius of  0.5Mpc of each galaxy. Those E+As that were listed in @zab96 as being cluster members (EA4 and EA11) have more than 10 known neighbours. Others, such as EA1, EA3, EA18 have 1 or 2 known neighbours and are clearly field galaxies, while the remaining ones have 3-7 known neighbours. For the SDSS galaxies from @goto03, cluster membership is not given. Based on the number of known near neighbours, only [SDSS J230743.41+152558.4]{} is a possible cluster galaxy. Hence, of the five E+As detected at 21 cm up to now, one is a cluster member ( SDSS J230743.41+152558.4) and four are not. Of those not detected, two are cluster members and four are not. From this result, it is clear that it is premature to draw conclusions on the environment by means of neutral hydrogen observations. In order to do so, a larger sample is required. Optical emission line strengths ------------------------------- As E+A galaxies are defined by means of their optical spectra, and more precisely by the equivalent widths of primarily the emission lines [\[\]]{}and H$\delta$ [@zab96; @goto03], one can investigate any trend between the neutral hydrogen content and the equivalents widths. For our sample we list these values in Table \[EAproperties\]. We restate that our E+A galaxies are compiled from two different samples with each different constraints concerning the equivalent widths of [\[\]]{} and H$\delta$. In Table \[othersample\] we list the EW([\[\]]{}) and EW(H$\delta$) for the undetected galaxies from @chang01. Both groups of detected and undetected E+A galaxies at 21 cm contain a mixture of [\[\]]{} absorption and emission lines. Similarly, E+A galaxies with low and high EW(H$\delta$) are detected. We conclude that again a larger sample is needed in order to investigate any trend between the optical emission lines and their neutral hydrogen content. -------- -------------- ----------- ----------------------- -- -- -- -- -- -- Galaxy [[\[\]]{}]{} H$\delta$ H[i]{} mass ([Å]{}) ([Å]{}) ($10^9$ [M$_\sun$]{}) EA01 1.80 8.98 7.1 EA02 1.25 7.98 $<3.1$ EA03 -0.29 8.13 $<3.9$ EA04 1.37 9.82 $<2.0$ EA11 2.16 6.96 $<4.7$ -------- -------------- ----------- ----------------------- -- -- -- -- -- -- : Properties of the E+A galaxies of @chang01.[]{data-label="othersample"} Star-formation rate ------------------- We estimate the star-formation rate (SFR) associated with the observed H[i]{} masses using the relation $$\Sigma_{\rm SFR} \approx 2.5^{\pm 0.7} \times 10^{-10} \left( \frac{\Sigma_{\rm gas}}{M_\odot \, {\rm pc}^{-2}} \right)^{1.40^{\pm 0.15}} M_\odot\,{\rm pc}^{-2}\,{\rm yr}^{-1}, \label{sfr}$$ [@k98]. We substituted the total H[i]{} mass divided by $\pi R_{\rm e}^2$, with $R_{\rm e}$ the half-light radius, for the gas surface density $\Sigma_{\rm gas}$. From the archival F625W HST/ACS images of EA17, EA18 and EA19, we derived surface brightness profiles as a function of radius by integrating the light in circular apertures centered on the galaxy. A similar method was applied to derive the half-light radii for the SDSS galaxies from Sloan r-band images. We fitted seeing or PSF convolved Sérsic profiles to the surface brightness profiles of all galaxies. The seeing and PSF profiles are determined from about 10 stars in each image. We used the surface brightness profile of a galaxy, extrapolated beyond the last data point by the best fitting Sérsic profile, to measure the half-light radii of these E+As. The SFR surface density, $\Sigma_{\rm SFR}$, was then converted into a global SFR by multiplying with $\pi R_{\rm e}^2$. In all cases but SDSS J230743.41+152558.4 and EA19, we found the E+As detected at 21 cm to have SFRs in the range $5-10\,M_\odot\,{\rm yr}^{-1}$ (see Table \[EAproperties\]), which is higher than expected for post-starburst galaxies. Hence, these gas-rich E+As could be actively forming stars at quite high rates but the star-formation sites are obscured by dust. Alternatively, although much gas is present, almost no stars are being formed. The radio continuum observations of [@mo01] rule out star formation at a rate higher than $1.0-1.5\,M_\odot\,{\rm yr}^{-1}$ in the case of EA17 and EA18. These authors measure a star formation rate of $1.5\,M_\odot\,{\rm yr}^{-1}$ in EA19. While the star formation rate of EA19 agrees roughly with our upper limit, the high star formation rates of EA17 and EA18 derived from their H[i]{} content are in clear contradiction with the very low rates derived from radio continuum observations. Unfortunately, no radio continuum observations have been performed of [[SDSS J210258.87+103300.6]{}]{}, [[SDSS J230743.41+152558.4]{}]{}, and [[SDSS J233453.20+145048.7]{}]{} so far. Discussion ========== Three out of the six galaxies observed by us contain detectable amounts of neutral gas and one is a border case. Both the SDSS and the Zabludoff samples satisfy very strict selection criteria and surely do not show any optical evidence for star formation: these appear to be true post-starburst galaxies. They do, however, contain significant amounts of neutral gas. EA19 is not detected by us in 21 cm line emission, with a 3$\sigma$ upperlimit on the  mass of $1.2 \times 10^9\,M_\odot$. EA18 was detected by us, with a 21 cm flux consistent with an  mass of $2.3 \pm 0.3 \times 10^9\,M_\odot$ of . ![$\log(M_{\rm HI}/L_B)$ versus $\log(L_B)$. Spiral galaxies [@h04] are depicted by spiral symbols, elliptical galaxies [@g01] by blue circles (the shade of blue is an indication of spectroscopic age and the vertical black line shows the luminosity limit of log($L_B$)=9.5 that was used by @g01), and E+A galaxies by red circles (this work) and orange circles [@chang01]. The predicted trend for late-type galaxies by the semi-analytical models of @ny04 is plotted in green. \[logLM\]](f3.eps){width="8.5cm"} In Fig. \[logLM\], we plotted $\log(M_{\rm HI}/L_B)$ versus $\log(L_B)$, $L_B$ the B-band luminosity expressed in solar B-band luminosities, for a sample of spiral galaxies [@h04], elliptical galaxies [@g01], and E+A galaxies [this work and @chang01] along with the predicted $\log(M_{\rm HI}/L_B)$ versus $\log(L_B)$ relation for late-type galaxies [@ny04]. For the E+As from the SDSS sample, we converted the SDSS g and r magnitudes to the B-band magnitude using the conversion formulae of [@j05]. The b$_J$ magnitudes of the LCRS E+As were converted into B-band magnitudes using the relation $m_B = m_J + 0.28 ({\rm B}-{\rm V})$ from [@mes90]. Using the mean B$-$V colour of the SDSS E+As, $\langle ({\rm B}-{\rm V}) \rangle = 0.78$ mag, in this equation, we find $m_B \approx m_J + 0.22$ mag. The absolute B-band magnitudes found by applying these conversion formulae are also listed in Table 1. Ellipticals with young spectroscopic ages ($\lesssim 1-2$ Gyr) are generally more gas-rich than older elliptical galaxies, which @g01 interpret as evidence for a merger origin for elliptical galaxies: the starburst following the merger rapidly consumes the gas reservoir and subsequent quiescent star formation consumes the gas at a much slower rate. It should be noted that @g01 include only ellipticals brighter than $M_B = -18.5$ mag, or $\log(L_B) = 9.5$, in their sample because fainter galaxies are likely to have experienced different evolutionary histories. The resulting sample of ellipticals spans about the same luminosity range as the E+A samples. As is clear from Fig. \[logLM\], some E+As are more gas-rich than most young elliptical galaxies. This is most likely not a selection effect, since the E+A data set was assembled based on optical, spectral properties and the elliptical galaxies were selected according to optical morphological considerations, not on H[i]{} mass. One would, moreover, expect the @g01 data set to be complete for galaxies with large H[i]{} masses. We performed a Kolmogorov-Smirnov (KS) and Wilcoxon (W) test on the $\log(M_{\rm HI}/L_B)$ distribution (within the same magnitude limits, both in $L_B$ and $\log(M_{\rm HI}/L_B)$, i.e. in the region defined by the detected E+As) and found that the distribution of E+A galaxies (i) corresponds with that of spiral and young elliptical galaxies with a significance of respectively 99% (KS) or 94% (W) and 65% (KS) or 79% (W) and (ii) differs from that of old elliptical galaxies with a confidence level of 79% (KS) or 78% (W). This statistical test and Fig. \[logLM\] might suggest a gas depletion time sequence, with E+As being observed less than $\sim 500$ Myr after the termination of the starburst that followed the putative merger [@yang04], young ellipticals after $\lesssim 1-2$ Gyr, and old ellipticals at later times. However, one should note that the difference between the distributions of E+As and elliptical galaxies has just a significance of 1$\sigma$. In order to derive conclusive evidence for this suggestion, a larger statistical sample of H[i]{} masses and spectroscopic ages of E+As is essential. Nevertheless, if correct, this ties together with the morphological study of 5 E+As from the Zabludoff (1996) sample with HST by @yang04, who suggest that E+As are likely to evolve to elliptical galaxies with power-law density profiles. One should also note that the end of a starburst does not necessarily require the complete exhaustion of the neutral gas reservoir. This could be explained by the fact that neutral gas itself is not the raw material for star formation: stars form in the dense cores of molecular clouds. @ko02 have obtained CO($1\!\rightarrow\! 0$) and HCN($1\!\rightarrow\! 0$) observations of the nearby post-starburst galaxy NGC5195, which, interestingly, forms an interacting pair with NGC5194. These authors note a central decrease of the mass fraction of dense molecular cores, leaving only diffuse molecular gas, evidenced by a very low central $L_{\rm HCN}/L_{\rm CO}$ value. Most likely, an intense central burst of star formation $\sim 1$ Gyr ago, responsible for the observed population of A stars, evaporated the dense molecular clouds, which are the sites where massive stars form. This effectively stopped further star formation, although large amounts of diffuse neutral and molecular gas remain. A similar mechanism may be responsible for switching off the starburst in E+A galaxies without the necessity of consuming the complete gas reservoir. The remaining  reservoir in some E+As may eventually lead, after gas has been allowed to condense into molecular cores, to further episodes of star formation. Eq. (\[sfr\]), which is based on observations of normal spiral galaxies that host a balanced mix of neutral and molecular gas, may therefore not be applicable to post-starburst galaxies. Conclusions {#concl} =========== We present deep single-dish  observations of a sample of six nearby E+A galaxies ($0.05<z<0.1$). We find masses of $6.5 \pm 0.8 \times 10^9\,M_\odot$ for [[SDSS J210258.87+103300.6]{}]{}, $0.9 \pm 0.3 \times 10^9\,M_\odot$ for [[SDSS J230743.41+152558.4]{}]{}, $2.7 \pm 0.3 \times 10^9\,M_\odot$ for [[SDSS J233453.20+145048.7]{}]{}, and $2.3 \pm 0.3 \times 10^9\,M_\odot$ for EA18. The $3\sigma$ upperlimits for the gas content of EA17 and EA19 are $2.9 \times 10^9\,M_\odot$ and $1.2 \times 10^9\,M_\odot$ respectively, assuming a total velocity width of 450 [[kms$^{-1}$]{}]{} (the average velocity width of the detected galaxies). The three galaxies from the SDSS sample satisfy very strict selection criteria and, for all practical purposes, can be considered to be truly post-starburst galaxies. The E+A galaxies detected in 21 cm line emission are almost as gas-rich as spiral galaxies with comparable luminosities. By plotting these E+As, spiral galaxies, and elliptical galaxies in a $\log(M_{\rm HI}/L_B)$ versus $\log(L_B)$ diagram, and performing a Kolmogorov-Smirnov and Wilcoxon test we suggest that existence of a gas depletion time sequence, with E+As being observed very shortly after the termination of the starburst that ensued from the putative merger, young ellipticals after $\lesssim 1-2$ Gyr, and old ellipticals at even later times. This would tie together with the morphological study of 5 E+As from the Zabludoff (1996) sample with HST by @yang04, who suggest that E+As are likely to evolve to elliptical galaxies with power-law density profiles. However, the conclusions drawn here are just based on 1$\sigma$ trends. A larger sample of  detections and spectroscopic ages of E+As is required to investigate this hypothesis. The presence of  in an E+A galaxy can be explained in two ways.\ (i) We can interpret the lack of on-going star-formation in E+A galaxies, suggested by previous radio continuum observations [@mo01] and indirectly by their selection criteria, and the fact that the end of the starburst does not necessarily require the complete exhaustion of the neutral gas reservoir, as being due to the effect the starburst has on the dense molecular cores which are responsible for the massive star formation. An intense burst of star formation can evaporate the dense molecular clouds, effectively stopping further star formation, even though copious amounts of diffuse neutral and molecular gas remain. The remaining  reservoir in some E+As may eventually lead, after the gas has again condensed into molecular cores, to further episodes of star formation. This may indicate that E+As are observed in the inactive phase of the star-formation duty cycle.\ (ii) A second possibility is what previously has been proposed by Couch & Sharples (1987) and Blake et al. (2004). There might still be on-going star-formation associated with the presence of gas, which is hinted from the high SFR (see Table \[EAproperties\]) of our E+A galaxies; however it cannot be observed since it is obscured by dust [@smg99], which is suggested by the ACS images in Fig. \[ima\_EA\], and/or optical emission lines might be a poor way of isolating true post-starbust systems. In this case E+A galaxies are in the active star-formation phase and will presumably exhaust all their gas content. PB acknowledges financial support from the Bijzonder OnderzoeksFonds (BOF). DM is supported by the MAGPOP EU Marie Curie Training and Research Network. SD acknowledges financial support from the Fonds voor Wetenschappelijk Onderzoek – Vlaanderen (FWO). This research was performed while D. J. P. held a National Research Council Research Associateship Award at the Naval Research Laboratory. Basic research in astronomy at the Naval Research Laboratory is funded by the Office of Naval Research. This work has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. Based on observations made with the NASA/ESA Hubble Space Telescope, obtained from the data archive at the Space Telescope Institute. STScI is operated by the association of Universities for Research in Astronomy, Inc. under the NASA contract NAS 5-26555. Balogh, M. L., Miller, C., Nichol, R., Zabludoff, A., & Goto, T. 2005, , 360, 587 Bekki, K., Couch, W. J., Shioya, Y., & Vazdekis, A. 2005, , 359, 949 Belloni, P., Bruzual, A. G., Thimm, G. J., Roser, H.-J., 1995, A&A, 297, 61 Blake, C., Pracy, M. B., Couch, W. J., Kenji, B., Lewis, I., Glazebrook, K., Baldry, I. K., Baugh, b. M., Bland-Hawthorn, J., et al., 2004, , 355, 713 Butcher, H., Oemler, A., Jr., 1978, ApJ, 219, 18 Chang, T. C., van Gorkum, J. H., Zabludoff, A. I., Zaritsky, D., & Mihos, J. C. 2001, , 121, 1965 Couch, W. J., Sharples, R. M., 1987, MNRAS, 229, 423 Cowie, L. L., Songaila, A., Hu, Ester M., Cohen, J. G., 1996, AJ, 112, 839 Dressler, A., Gunn, J. E., 1983, ApJ, 270, 7 Dressler, A., Smail, I., Poggianti, B. M., Butcher, H., Couch, W. J., Ellis, R. S., & Oemler, A., Jr., 1999, ApJS, 122, 51 Fabricant, D.G., McClintock, J. E., Bautz, M.W., 1991, , 381, 33 Galaz, G. 2000, , 119, 2118 Georgakakis, A., Hopkins, A. M., Caulton, A., Wiklind, T., Terlevich, A. I., Forbes, D., A., 2001, , 326, 1431 Goto, T., Nichol, R. C., Okamura, S., Sekiguchi, M., Miller, C. J., Bernardi, M., Hopkins, A., Tremonti, C., Connolly, A., Castander, F. J., et al. 2003, , 55, 771 Goto, T. 2004, A&A, 427, 125 Goto, T. 2005, MNRAS, 357, 937 Helmboldt, J. F., Walterbos, R. A. M., bothun, G. D., O’Neil, K., de Blok, W. J. G., 2004, , 613, 914 Jester, S., Schneider, D. P., Richards, G. T., Green, R. F., Schmidt, M., Hall, P. B., Strauss, M. A., Vanden Berk, D. E., et al., 2005,, 130, 873 Kennicutt, R. C., Jr., 1998, ARA&A, 36, 189 Kohno, K., Tosaki, T., Matsushita, S., Vila-Vilaó, B., Shibatsuka, T., Kawabe, R., 2002, PASJ, 54, 541 Maddox, S. J., Efstathiou, G., Sutherland, W. J., 1990, , 246, 433 Miller, N. A., & Owen, F. N. 2001, , 554, 125 Nagashima, M. & Yoshii, Y., 2004, , 610, 23 Poggianti, B. M., Smail, I., Dressler, A., Couch, W. J., Barger, J., Butcher, H., Ellis, E. S., & Oemler, A., Jr., 1999, ApJ, 518, 576 Pracy, M. B., Couch, W. J., Blake, C., Bekki, K., Harrison, C., Colless, M., Kuntschner, H., & de Propris, R.  2005, , 359, 1421 Sault, R.J., Teuben, P.J. & Wright, M.C.H., 1995, in Shaw R., Payne H.E., Hayes, J.J.E., eds, ASP Conf Ser. Vol. 77, Astronomical Data Analysis Software and Systems IV. Astron. Soc. Pac., San Fransisco, p. 433 2005, , 359, 1421 Smail, I., Morrison, G., Gray, M. E., Owen, F. N., Ivison, R. J., Kneib, J.-P., Ellis, R. S. Gray, M. E., et al. 1999, ApJ, 525, 609 Tran, K.-V. H., Franx, M., Illingworth, G., Kelson, D. D., & van Dokkum, P. 2003, , 599, 865 Tran, K.-V. H., Franx, M., Illingworth, G. D., van Dokkum, P., Kelson, D. D., & Magee, D. 2004, , 609, 683 Yang, Y., Zabludoff, A. I., Zaritsky, D., Lauer, T. R., & Mihos, J. C. 2004, , 607, 258 Zabludoff, A. I., Zaritsky, D., Lin, H., Tucker, D., Hashimoto, Y., Shectman, S. A., Oemler, A., & Kirshner, R. P. 1996, , 466, 104 [^1]: The Arecibo Observatory is part of the National Astronomy and Ionosphere Center, which is operated by Cornell University under a cooperative agreement with the National Science Foundation. [^2]: The Parkes telescope is part of the Australia Telescope which is funded by the Commonwealth of Australia for operation as a National Facility managed by CSIRO.

ArXiv

--- abstract: | Let $A$ be a positive semidefinite $m\times m$ block matrix with each block $n$-square, then the following determinantal inequality for partial traces holds $$(\tr A)^{mn} - \det(\tr_2 A)^n \ge \bigl| \det A - \det(\tr_1 A)^m \bigr|,$$ where $\tr_1$ and $\tr_2$ stand for the first and second partial trace, respectively. This result improves a recent result of Lin [@Lin16]. author: - | Yongtao Li$^a$, Lihua Feng$^b$, Weijun Liu$^b$, Yang Huang$^{\dag, b}$\ [${}^a$School of Mathematics, Hunan University]{}\ [Changsha, Hunan, 410082, P.R. China ]{}\ [$^b$School of Mathematics and Statistics, Central South University]{}\ [New Campus, Changsha, Hunan, 410083, P.R. China. ]{} title: ' [Another determinantal inequality involving partial traces ]{} [^1]' --- \[section\] \[theorem\][Corollary]{} \[theorem\][Definition]{} \[theorem\][Conjecture]{} \[theorem\][Question]{} \[theorem\][Lemma]{} \[theorem\][Proposition]{} \[theorem\][Question]{} \[theorem\][Example]{} [[**Key words:**]{} Partial traces; Block matrices; Determinantal inequality; Numerical range in a sector. ]{}\ [2010 Mathematics Subject Classication. 15A45, 15A60, 47B65.]{} Introduction {#sec1} ============ Throughout the paper, we use the following standard notation. The set of $n\times n$ complex matrices is denoted by $\mathbb{M}_n(\mathbb{C})$, or simply by $\mathbb{M}_n$, and the identity matrix of order $n$ by $I_n$, or $I$ for short. If $A=[a_{ij}]$ is of order $m\times n$ and $B$ is of order $s\times t$, the tensor product of $A$ with $B$, denoted by $A\otimes B$, is an $ms\times nt$ matrix, partitioned into $m\times n$ block matrix with the $(i,j)$-block the $s\times t$ matrix $a_{ij}B$. In this paper, we are interested in complex block matrices. Let $\mathbb{M}_m(\mathbb{M}_n)$ be the set of complex matrices partitioned into $m\times m$ blocks with each block being a $n\times n$ matrix. The element of $\mathbb{M}_m(\mathbb{M}_n)$ is usually written as ${ A}=[A_{i,j}]_{i,j=1}^m$, where $A_{i,j}\in \mathbb{M}_n$ for all $i,j$. By convention, if $X\in \mathbb{M}_n$ is positive semidefinite, we write $X\ge 0$. For two Hermitian matrices $A$ and $B$ of the same size, $A\ge B$ means $A-B\ge 0$. It is easy to see that $\ge$ is a partial ordering on the set of Hermitian matrices, referred to as [*Löuner ordering*]{}. Now we introduce the definition of partial traces, which comes from Quantum Information Theory [@Petz08 p. 12]. For $A\in \mathbb{M}_m(\mathbb{M}_n)$, the first partial trace (map) $A \mapsto \mathrm{tr}_1 A \in \mathbb{M}_n$ is defined as the adjoint map of the imbedding map $X \mapsto I_m\otimes X\in \mathbb{M}_m\otimes \mathbb{M}_n$. Correspondingly, the second partial trace (map) $A \mapsto \mathrm{tr}_2 A\in \mathbb{M}_m$ is defined as the adjoint map of the imbedding map $Y\mapsto Y\otimes I_n \in \mathbb{M}_m\otimes \mathbb{M}_n$. Therefore, we have $$\label{eqdef} \langle I_m\otimes X, A \rangle =\langle X, \mathrm{tr}_1A \rangle , \quad \forall X\in \mathbb{M}_n,$$ and $$\langle Y\otimes I_n, A \rangle =\langle Y,\mathrm{tr}_2 A \rangle, \quad \forall Y\in \mathbb{M}_m.$$ Assume that $A=[A_{i,j}]_{i,j=1}^m$ with $A_{i,j}\in \mathbb{M}_n$, then the visualized forms of the partial traces are actually given in [@Bh07 pp. 120–123] as $$\label{eqdef2} \mathrm{tr}_1 { A}=\sum\limits_{i=1}^m A_{i,i},\quad \mathrm{tr}_2{ A}=\bigl[ \mathrm{tr}A_{i,j}\bigr]_{i,j=1}^m.$$ If ${ A}=[A_{i,j}]_{i,j=1}^m \in \mathbb{M}_m(\mathbb{M}_n)$ is positive semidefinite, it is easy to see that both $\tr_1 A$ and $\tr_2 A$ are positive semidefinite; see, e.g., [@Zha12]. To some extent, these two partial traces are closely related. For instance, Audenaert [@Aud07] proved an inequality for Schatten $p$-norms, $$\label{eqaud} \tr A + \lVert A \lVert_q \ge \lVert \tr_1 A \rVert_q + \lVert \tr_2 A \rVert_q .$$ Inequality (\[eqaud\]) was used to prove the subadditivity of Tsallis entropies. Moreover, Ando (see [@Ando14]) established the following, $$\label{eqando} (\tr A)I_m\otimes I_n+ A \ge I_m\otimes (\mathrm{tr}_1 A) + (\mathrm{tr}_2 A) \otimes I_n,$$ where $\ge $ means the Löuner ordering. Furthermore, Motivated by inequalities (\[eqaud\]) and (\[eqando\]), Lin [@Lin16] proved an analogous result for determinant, which states that $$\label{eqlin} (\tr A)^{mn} +\det A \ge \det (\tr_1 A)^m +\det (\tr_2 A)^n.$$ In this paper, we improve Lin’s result (\[eqlin\]) as follows. \[thm34\] Let $A\in \mathbb{M}_m(\mathbb{M}_n)$ be positive semidefinite. Then $$(\tr A)^{mn} - \det(\tr_2 A)^n \ge \bigl| \det A - \det(\tr_1 A)^m \bigr|.$$ The paper is organized as follows. We first present some auxiliary results, and then we show our proof of Theorem \[thm34\]. Finally, we extend our result to a larger class of matrices, namely, matrices whose numerical ranges are contained in a sector (Theorem \[thm36\]). Auxiliary results and proofs {#sec2} ============================ For $A=[A_{i,j}]_{i,j=1}^m\in \mathbb{M}_m(\mathbb{M}_n)$, we define the partial tranpose of $A$ by $A^{\tau}=[A_{j,i}]_{i,j=1}^m$. It is clear that $A\ge 0$ does not necessarily imply $A^{\tau}\ge 0$. If both $A$ and $A^{\tau}$ are positive semidefinite, then $A$ is called to be [*positive partial tranpose*]{} (or PPT for short). Recall that a linear map $\Phi: \mathbb{M}_n\to \mathbb{M}_k$ is called positive if it maps positive matrices to positive matrices. A linear map $\Phi: \mathbb{M}_n\to \mathbb{M}_k$ is said to be $m$-positive if for $[A_{i,j}]_{i,j=1}^m\in \mathbb{M}_m(\mathbb{M}_n)$, $$\label{eq1} [A_{i,j}]_{i,j=1}^m \ge 0 \Rightarrow [\Phi (A_{i,j})]_{i,j=1}^m\ge 0.$$ It is said to be [*completely positive*]{} if (\[eq1\]) holds for any integer $m\ge 1$. It is well known that both the trace map and determinant map are completely positive; see, e.g., [@Zhang11 p. 221, p. 237]. On the other hand, a linear map $\Phi $ is said to be $m$-copositive if for $[A_{i,j}]_{i,j=1}^m\in \mathbb{M}_m(\mathbb{M}_n)$, $$\label{eq2} [A_{i,j}]_{i,j=1}^m \ge 0 \Rightarrow [\Phi (A_{j,i})]_{i,j=1}^m\ge 0,$$ and $\Phi$ is said to be [*completely copositive*]{} if (\[eq2\]) holds for any positive integer $m\ge 1$. Furthermore, $\Phi$ is called [*a completely PPT map*]{} if it is completely positive and completely copositive. A comprehensive survey of the standard results on completely positive maps can be found in [@Bh07 Chapter 3] or [@Paulsen02]. We need the following lemma, which is the main result in [@Lin14]; see, e.g., [@Li20laa]. (see [@Lin14]) \[lem21\] The map $\Phi(X)=(\tr X)I +X$ is a completely PPT map. In the proof of the next proposition, we only employ the fact that $\Psi (X)=(\tr X)I+X$ is $2$-copositive. Proposition \[thm22\], first proved by the authors [@HuangLi20] recently, which is a complement of Ando’s result (\[eqando\]) and play a vital role in our derivation of Theorem \[thm34\]. We here provide an alternative proof for convenience of readers. Our proof is slightly more transparent than the original proof in [@HuangLi20]. \[thm22\] Let $A=[A_{i,j}]_{i,j=1}^m\in \mathbb{M}_m(\mathbb{M}_n)$ be positive semidefinite. Then $$\label{eqmain} (\tr A)I_m\otimes I_n - (\tr_2 A) \otimes I_n \ge A -I_m\otimes (\tr_1 A) .$$ The proof is by induction on $m$. When $m=1$, there is nothing to prove. We now prove the base case $m=2$. In this case, the required inequality is $$\begin{aligned} & \begin{bmatrix} (\tr A )I_n& 0\\ 0& (\tr A)I_n \end{bmatrix} - \begin{bmatrix} (\tr A_{1,1})I_n& (\tr A_{1,2})I_n\\ (\tr A_{2,1})I_n & (\tr A_{2,2})I_n \end{bmatrix} \\ &\quad \ge \begin{bmatrix} A_{1,1}& A_{1,2}\\ A_{2,1} & A_{2,2} \end{bmatrix} - \begin{bmatrix} A_{1,1}+A_{2,2}& 0\\ 0& A_{1,1}+A_{2,2}\end{bmatrix}, \end{aligned}$$ or equivalently (note that $\tr A=\tr A_{1,1} +\tr A_{2,2}$), $$\begin{aligned} \label{proofe1} H:=\begin{bmatrix} (\tr A_{2,2})I_n+ A_{2,2}& -A_{1,2}-(\tr A_{1,2})I_n\\ -A_{2,1}-(\tr A_{2,1})I_n& (\tr A_{1,1})I_n+ A_{1,1} \end{bmatrix}\ge 0. \end{aligned}$$ By Lemma \[lem21\], we have $$\begin{aligned} \begin{bmatrix} (\tr A_{1,1})I_n+ A_{1,1}& (\tr A_{2,1})I_n+ A_{2,1}\\ (\tr A_{1,2})I_n+ A_{1,2}& (\tr A_{2,2})I_n+ A_{2,2} \end{bmatrix}\ge 0, \end{aligned}$$ and so $$\begin{aligned} H=\begin{bmatrix}0 & -I_n\\ I_n& 0\end{bmatrix} \begin{bmatrix} (\tr A_{1,1})I_n+ A_{1,1}& (\tr A_{2,1})I_n+ A_{2,1}\\ (\tr A_{1,2})I_n+ A_{1,2}& (\tr A_{2,2})I_n+ A_{2,2} \end{bmatrix} \begin{bmatrix}0 & I_n\\ -I_n& 0\end{bmatrix} \ge 0, \end{aligned}$$ which confirms the desired (\[proofe1\]). Suppose the result (\[eqmain\]) is true for $m=k-1>1$, and then we consider the case $m=k$, $$\begin{aligned} \Gamma &:=(\tr A)I_k\otimes I_n+ I_k \otimes(\tr_1 A)-A-(\tr_2 A) \otimes I_n \\ &= \left(\tr \sum_{i=1}^{k}A_{i,i}\right)I_k\otimes I_n +I_k\otimes \left(\sum_{j=1}^k A_{j,j}\right)-A-([\tr A_{i,j}]_{i,j=1}^k) \otimes I_n \\ &= \begin{bmatrix} \sum_{i=1}^{k-1}(\tr A_{i,i})I_n & && \\ &\ddots & \\& & \sum_{i=1}^{k-1}(\tr A_{i,i})I_n& \\ & & & 0 \end{bmatrix} \\ &\quad + \begin{bmatrix} (\tr A_{k,k})I_n & && \\ &\ddots & \\& & (\tr A_{k,k})I_n & \\ & & & \sum_{i=1}^k(\tr A_{i,i})I_n \end{bmatrix} \\ &\quad + \begin{bmatrix} \sum_{i=1}^{k-1} A_{i,i} & && \\ &\ddots & \\& & \sum_{i=1}^{k-1} A_{i,i}& \\ & & & 0 \end{bmatrix}+ \begin{bmatrix} A_{k,k} & && \\ &\ddots & \\& & A_{k,k} & \\ & & & \sum_{i=1}^k A_{i,i} \end{bmatrix} \\ & \quad - \begin{bmatrix} A_{1,1} & \cdots & A_{1, k-1}& 0 \\ \vdots & & \vdots & \vdots \\ A_{k-1,1}& \cdots & A_{k-1, k-1} & 0 \\ 0 & \cdots & 0& 0 \end{bmatrix}-\begin{bmatrix} 0 & \cdots & 0& A_{1, k} \\ \vdots & & \vdots & \vdots \\0& \cdots & 0 & A_{k-1,k} \\ A_{k,1} & \cdots & A_{k,k-1} & A_{k,k} \end{bmatrix} \\ &\quad - \begin{bmatrix} (\tr A_{1,1})I_n & \cdots & (\tr A_{1, k-1})I_n& 0 \\ \vdots & & \vdots & \vdots \\ (\tr A_{k-1,1})I_n& \cdots & (\tr A_{k-1, k-1})I_n & 0 \\ 0 & \cdots & 0& 0 \end{bmatrix} \\ &\quad -\begin{bmatrix} 0 & \cdots & 0& (\tr A_{1, k})I_n \\ \vdots & & \vdots & \vdots \\0& \cdots & 0 & (\tr A_{k-1,k})I_n \\ (\tr A_{k,1})I_n & \cdots & (\tr A_{k,k-1})I_n & (\tr A_{k,k})I_n \end{bmatrix}.\end{aligned}$$ After some rearrangement, we have $$\Gamma =\Gamma_1 +\Gamma_2,$$ where $$\begin{aligned} \Gamma_1&:=\begin{bmatrix} \sum_{i=1}^{k-1}(\tr A_{i,i})I_n & && \\ &\ddots & \\& & \sum_{i=1}^{k-1}(\tr A_{i,i})I_n& \\ & & & 0 \end{bmatrix} + \begin{bmatrix} \sum_{i=1}^{k-1} A_{i,i} & && \\ &\ddots & \\& & \sum_{i=1}^{k-1} A_{i,i}& \\ & & & 0 \end{bmatrix} \\ & \phantom{:}- \begin{bmatrix} A_{1,1} & \cdots & A_{1, k-1}& 0 \\ \vdots & & \vdots & \vdots \\ A_{k-1,1}& \cdots & A_{k-1, k-1} & 0 \\ 0 & \cdots & 0& 0 \end{bmatrix} - \begin{bmatrix} (\tr A_{1,1})I_n & \cdots & (\tr A_{1, k-1})I_n& 0 \\ \vdots & & \vdots & \vdots \\ (\tr A_{k-1,1})I_n& \cdots & (\tr A_{k-1, k-1})I_n & 0 \\ 0 & \cdots & 0& 0 \end{bmatrix}, \end{aligned}$$ and $$\begin{aligned} \Gamma_2&:= \begin{bmatrix} (\tr A_{k,k})I_n & && \\ &\ddots & \\& & (\tr A_{k,k})I_n & \\ & & & \sum_{i=1}^k(\tr A_{i,i})I_n \end{bmatrix} + \begin{bmatrix} A_{k,k} & && \\ &\ddots & \\& & A_{k,k} & \\ & & & \sum_{i=1}^k A_{i,i} \end{bmatrix} \\ & \phantom{:} - \begin{bmatrix} 0 & \cdots & 0& A_{1, k} \\ \vdots & & \vdots & \vdots \\0& \cdots & 0 & A_{k-1,k} \\ A_{k,1} & \cdots & A_{k,k-1} & A_{k,k} \end{bmatrix} - \begin{bmatrix} 0 & \cdots & 0& (\tr A_{1, k})I_n \\ \vdots & & \vdots & \vdots \\ 0& \cdots & 0 & (\tr A_{k-1,k})I_n \\ (\tr A_{k,1})I_n & \cdots & (\tr A_{k,k-1})I_n & (\tr A_{k,k})I_n \end{bmatrix} \\ &\phantom{:}= \begin{bmatrix} (\tr A_{k,k})I_n+A_{k,k} & & & -A_{1,k}-(\tr A_{1, k})I_n \\ & \ddots & & \vdots \\ & & (\tr A_{k,k})I_n+A_{k,k} & -A_{k-1,k}-(\tr A_{k-1,k})I_n \\ -A_{k,1}-(\tr A_{k,1})I_n & \cdots & -A_{k,k-1}-(\tr A_{k,k-1})I_n & \sum_{i=1}^{k-1}\big((\tr A_{i,i})I_n+A_{i,i}\big) \end{bmatrix}.\end{aligned}$$ Now by induction hypothesis, we get that $\Gamma_1$ is positive semidefinite. It remains to show that $\Gamma_2$ is positive semidefinite. Observing that $\Gamma_2$ can be written as a sum of $k-1$ matrices, in which each summand is $*$-congruent to $$H_i:=\begin{bmatrix} (\tr A_{k,k})I_n+A_{k,k} & -A_{i,k}-(\tr A_{i, k}))I_n \\ - A_{k,i}-(\tr A_{k,i})I_n & (\tr A_{i,i})I_n+A_{i,i} \end{bmatrix}, \quad i=1, 2,\ldots, k-1.$$ Just like the proof of the base case, we infer from Lemma \[lem21\] that $H_i\ge 0$ for all $i=1,2, \ldots, k-1$. Therefore, $\Gamma_2 \ge 0$, thus the proof of induction step is complete. Before showing our proof of Theorem \[thm34\], we need one more lemma for our purpose. (see [@Lin16])\[lem31\] Let $X, Y, W, Z\in \mathbb{M}_\ell$ be positive semidefinite. If $X+Y\ge W+Z$, $X\ge W$ and $X\ge Z$, then $$\label{lin} \det X+\det Y\ge \det W+\det Z.$$ We remark that Lemma \[lem31\] implies the determinantal inequality: $$\det (A+B+C) +\det C \ge \det (A+C) +\det (B+C),$$ where $A,B$ and $C$ are positive semidefinite; see [@Lin14b] and [@Li20] for more details. We are now in a position to present the proof of Theorem \[thm34\]. [**Proof of Theorem \[thm34\]**]{}  In view of (\[eqlin\]), it suffices to show $$\label{eq111} (\tr A)^{mn} +\det (\tr_1 A)^m \ge \det A +\det (\tr_2 A)^n.$$ Let $X=(\tr A)I_m\otimes I_n, Y=I_m\otimes (\tr_1 A),W=A,Z=(\tr_2 A)\otimes I_n$, respectively. It is easy to see that $$(\tr A)I_m =\sum_{i=1}^m (\tr A_{i,i})I_m = \bigl( \tr (\tr_2 A)\bigr)I_m \ge \tr_2 A,$$ which implies that $X\ge Z\ge 0$, and clearly $X\ge W\ge 0$. Moreover, by Proposition \[thm22\], $X+Y\ge W+Z$. That is, all conditions in Lemma \[lem31\] are satisfied. Therefore, $$\begin{aligned} (\tr A)^{mn} + \det \bigl( I_m\otimes (\tr_1 A) \bigr) \ge \det A +\det \bigl( (\tr_2 A) \otimes I_n \bigr).\end{aligned}$$ Since $\det (X\otimes Y) =(\det X)^n(\det Y)^m$ for every $X\in \mathbb{M}_m$ and $Y\in \mathbb{M}_n$, this completes the proof. Using the same idea in previous proof and combining [@HuangLi20 Proposition 2.3], one could also get the following Proposition \[prop35\]. We leave the details for the interested reader. \[prop35\] Let $A\in \mathbb{M}_m(\mathbb{M}_n)$ be PPT. Then $$(\tr A)^{mn} + \det(\tr_2 A)^n \ge \det A + \det(\tr_1 A)^m .$$ At the end of the paper, we extend the determinantal inequality (\[eq111\]) to a larger class of matrices whose numerical ranges are contained in a sector. The same extension of (\[eqlin\]) can be found in [@YLC19]. Before showing our extension, we first introduce some standard notations. For $A\in \mathbb{M}_n$, the Cartesian (Toeptliz) decomposition $A=\Re A+i\Im A$, where $\Re A=\frac{1}{2}(A+A^*)$ and $\Im A=\frac{1}{2i}(A-A^*)$. Let $|A|$ denote the positive square root of $A^*A$, i.e., $|A|=(A^*A)^{1/2}$. We denote the $i$-th largest singular value of $A$ by $s_i(A)$, then $s_i(A)=\lambda_i(|A|)$, the $i$-th largest eigenvalue of $|A|$. Recall that the numerical range of $A\in \mathbb{M}_n$ is defined by $$W(A)=\{x^*Ax : x\in \mathbb{C}^*,x^*x=1\}.$$ For $\alpha \in [0,\frac{\pi}{2})$, let $S_{\alpha}$ be the sector on the complex plane given by $$S_{\alpha}=\{z\in \mathbb{C}: \Re z>0,|\Im z|\le (\Re z)\tan \alpha \} =\{re^{i\theta } : r>0,|\theta |\le \alpha \}.$$ Obviously, if $W(A)\subseteq S_{\alpha}$ for $\alpha \in [0,\frac{\pi}{2})$, then $\Re (A)$ is positive definite and if $W(A)\subseteq S_0$, then $A$ is positive definite. Such class of matrices whose numerical ranges are contained in a sector is called the [*sector matrices class*]{}. Clearly, the concept of sector matrices is an extension of that of positive definite matrices. Over the past years, various studies on sector matrices have been obtained in the literature; see, e.g., [@Choi19; @Jiang19; @Kua17; @Lin15; @YLC19; @Zhang15]. First, we list two lemmas which are useful to establish our extension (Theorem \[thm36\]). (see [@Lin15]) \[lem34\] Let $0\le \alpha <\frac{\pi}{2}$ and $A\in \mathbb{M}_n$ with $W(A)\subseteq S_{\alpha}$. Then $$|\det A| \le (\sec \alpha)^n \det (\Re A).$$ (see [[@HJ13 p. 510]]{}) \[lem35\] Let $A$ be an $n$-square complex matrix. Then $$\lambda_i(\Re A) \le s_i(A),\quad i=1,2,\ldots ,n.$$ Moreover, if $X$ has positive definite real part, then $$\det \Re A + |\det \Im A| \le |\det A|.$$ Now, we provide the extension of (\[eq111\]). \[thm36\] Let $H\in \mathbb{M}_m(\mathbb{M}_n)$ be such that $W(H)\subseteq S_{\alpha}$. Then $$(\tr |A|)^{mn} + |\det (\tr_1 A)|^m \ge (\cos \alpha )^{mn} \det |A| + (\cos \alpha)^{mn} |\det (\tr_2 A)|^n.$$ By Lemma \[lem35\], we have $$\label{eq12} \tr |A| =\sum_{i=1}^{mn} s_i(A) \ge \sum_{i=1}^{mn} \lambda_i(\Re A) = \tr (\Re A) \ge 0.$$ Since $W(A)\subseteq S_{\alpha}$, it is noteworthy that $W(\tr_1 A) \subseteq S_{\alpha}$ and $W(\tr_2 A) \subseteq S_{\alpha}$; see, e.g., [@Kua17]. Observe that $\Re (\tr_1 A) = \tr_1( \Re A)$ and $\Re( \tr_2 A) = \tr_2( \Re A)$. Therefore, $$\begin{aligned} (\tr |A|)^{mn} + |\det (\tr_1 A)|^m &\ge (\tr \Re A)^{mn} + \det \bigl( \Re (\tr_1 A) \bigr)^m \\ & = (\tr \Re A)^{mn} + \det \bigl( \tr_1( \Re A) \bigr)^m \\ & \ge \det (\Re A) + \det \bigl( \tr_2( \Re A) \bigr)^n \\ &= \det (\Re A) + \det \bigl( \Re( \tr_2 A) \bigr)^n \\ &\ge (\cos \alpha)^{mn} |\det A| + (\cos \alpha)^{mn} |\det (\tr_2 A)^n |, \end{aligned}$$ where the first inequality follows from (\[eq12\]) and Lemma \[lem35\], the second one follows by applying (\[eq111\]) to $\Re A$, the last one is by Lemma \[lem34\]. Acknowledgments {#acknowledgments .unnumbered} =============== The author would like to thank Dr. Minghua Lin for bringing the question to his attention and for naming the title of the manuscript, which can be regarded as a continuation and development of his result [@Lin16]. All authors are grateful for valuable comments from the referee, which considerably improve the presentation of our manuscript. This work was supported by NSFC (Grant Nos. 11671402, 11871479), Hunan Provincial Natural Science Foundation (Grant Nos. 2016JJ2138, 2018JJ2479) and Mathematics and Interdisciplinary Sciences Project of CSU. [30]{} K. M. R. Audenaert, Subadditivity of $q$-entropies for $q>1$, J. Math. Phys. 48 (2007), no. 8, 083507. T. Ando, Matrix inequalities involving partial traces, ILAS Conference, 2014. R. Bhatia, Positive Definite Matrices, Princeton University Press, Princeton, 2007. D. Choi, T. Y. Tam, P. Zhang, Extension of Fischer’s inequality, Linear Algebra Appl. 569 (2019) 311–322. Y. Huang, W. Liu, L. Feng, Y. Li, A complement of Ando’s inequality involving partial traces, arXiv. R. A. Horn, C. R. Johnson, Matrix Analysis, 2nd ed., Cambridge University Press, Cambridge, 2013. X. Jiang, Y. Zheng, X. Chen, Extending a refinement of Kotelianskii’s inequality, Linear Algebra Appl. 574 (2019) 252–261. L. Kuai, An extension of the Fiedler-Markham determinant inequality, Linear Multilinear Algebra 66 (2018) 547–553. Y. Li, L. Feng, Z. Huang, W. Liu, Inequalities regarding partial trace and partial determinant, Math. Ineq. Appl. 23 (2020) 477–485. Y. Li, Y. Huang, L. Feng, W. Liu, Some applications of two completely copositive maps, Linear Algebra Appl. 590 (2020) 124–132. M. Lin, A completely PPT map, Linear Algebra Appl. 459 (2014) 404–410. M. Lin, An Oppenheim type inequality for a block Hadamard product, Linear Algebra Appl. 452 (2014) 1–6. M. Lin, Extension of a result of Haynsworth and Hartfiel, Arch. Math. 104 (2015) 93–100. M. Lin, A determinantal inequality involving partial traces, Canad. Math. Bull. 59 (2016) 585–591. V. Paulsen, Completely Bounded Maps and Operator Algebras, Cambridge University Press, Cambridge, 2002. D. Petz, Quantum Information Theory and Quantum Statistics. Theoretical and Mathematical Physics, Springer, Berlin, 2008. J. Yang, L. Lu, Z. Chen, Schatten $q$-norms and determinantal inequalities for matrices with numerical ranges in a sector, Linear Multilinear Algebra 67 (2019) 221–227. F. Zhang, Positivity of matrices with generalized matrix functions. Acta Math. Sin. (Engl. Ser.) 28 (2012) 1779–1786. P. Zhang, Extension of Matic’s results, Linear Algebra Appl. 486 (2015) 328–334. F. Zhang, Matrix Theory: Basic Results and Techniques, 2nd ed., Springer, New York, 2011. [^1]: This work was supported by NSFC (Grant Nos. 11671402, 11871479), Hunan Provincial Natural Science Foundation (Grant Nos. 2016JJ2138, 2018JJ2479) and Mathematics and Interdisciplinary Sciences Project of CSU. E-mail addresses: ytli0921@hnu.edu.cn(Y. Li), fenglh@163.com (L. Feng), wjliu6210@126.com (W. Liu), FairyHuang@csu.edu.cn (Y. Huang, corresponding author).

ArXiv

--- abstract: 'The ionizing spectral energy distribution of quasars exhibits a steepening of the distribution shortward of $\sim 1200\,$Å. The change of the power-law index from approximately $-1$ (near-UV) to $-2$ (far-UV) has so far been interpreted as being intrinsic to quasars. We consider the possibility that the steepening may result from a tenuous absorption component that is anticorrelated with large mass overdensities. UV sensitive satellites, whose detectors can extend down to 1000 Å, can set a useful limit to such an absorption component through the search of a flux increase in the window 1050–1190Å (observer frame) with respect to an extrapolation of the continuum above 1230 Å. Since the recent FUSE or HST-STIS data do not show any obvious discontinuity in this region, this effectively rules out the possibility that intergalactic  absorption is very important, and it is concluded that most if not all of the steepening is intrinsic to quasars. A smaller flux discontinuity of order 1% cannot, however, be ruled out yet and would still be consistent with the warm-hot intergalactic component if it amounts to 30% of the baryonic mass, as predicted by some models of large scale structure formation, provided its temperature lies around $10^{5.3}$ K.' author: - 'Luc Binette, Mario Rodríguez-Martínez, Sinhue Haro-Corzo and Isidro Ballinas' title: | Technique for detecting warm-hot intergalactic gas\ in quasar UV spectra --- Introduction {#sec:intro} ============ The ionizing spectral energy distribution (hereafter ISED) of nearby active galactic nuclei cannot be observed directly, due to the large Galactic absorption beyond the Lyman limit. Owing to the redshift effect, however, we can get a glimpse of the ISED from the spectra of distant quasars. The pioneering work of Zheng et al. (1997, ZK97), using HST-FOS archived data, identified a noticeable steepening in quasar SEDs near 1200 Å. The power-law index $\alpha$ ($F_{\nu}\propto \nu^{\alpha}$) in quasars steepens from $ -1$ for $\lambda > 1200\,$Å to $\simeq -2$ at shorter wavelengths[^1], down to 350Å (). In a more recent study, Telfer et al. 2002 (hereafter TZ02) found similar results with a composite spectrum characterized by a mean near-UV index of $-0.7$ steepening to $\simeq -1.7$ in the far-UV. The above authors favor the interpretation that the steepening is intrinsic to quasars and that it is the signature of a comptonized accretion disk. On the other hand, certain distributions of intergalactic absorption gas can also cause an (apparent) steepening of the SED, starting at 1200Å (), as shown in this Paper, that is, located in the same position as the break encountered by TZ02. As long as ISED studies are based on detectors that do not extend beyond 1200Å (), one cannot readily distinguish between the contribution to the observed steepening from an intervening absorption model and that of a purely intrinsic break in the quasar SED as proposed by ZK97 or TZ02. FUSE or HST-STIS spectra, however, extend much farther into the UV and will be shown here to provide us with a compelling test to discriminate between the two interpretations. Large scale structure formation models predict that a substantial fraction of baryons should reside in a warm-hot phase at current epoch, possibly up to 30–40% of $\Omega_{bar.}$ [@davea]. Depending on its temperature and distribution with redshift, this warm-hot gas may contribute to a reduced fraction of the observed steepening of the ISED. In this Paper, we show that this component may give rise to a slight flux increase (that is, a discontinuity) in the region 1050–1190Å () with respect to longer wavelengths. The calibration of the level of this discontinuous excess flux can be used as a technique to set useful limits for the baryonic mass contribution from the warm-hot intergalactic component. The objective (and results) of this Paper is threefold: 1. Find the simplest and yet physically meaningful absorption gas distribution which can mimic the observed steepening (§\[sec:steep\]), 2. Devise a technique to falsify the previous proposition that the break is due to absorption, thereby confirming the idea that the steepening is intrinsic to quasars (§\[sec:false\]), 3. Propose a dependable technique to reveal traces of a warm-hot baryonic component in the local Universe using UV spectra (§\[sec:detect\]). Procedure and calculations {#sec:cal} ========================== Generalized Gunn-Peterson effect {#sec:gp} -------------------------------- The Gunn-Peterson (GP) effect [@gunn] sets stringent limits for the presence of neutral diffuse gas at high redshifts. In the simplest form of the GP test, the absorption gas produces a flux decrement between  and . The decrement is measured against a continuum level, usually taken to be a power-law extrapolated from the region redward of the  emission line. With high resolution spectra, this technique is sensitive to  columns otherwise too small to produce resolvable  lines. @songa in this manner could set a limit of the GP opacity $\tau_{GP} < 0.1$ at $z=4.7$. Another manifestation of the GP effect is illustrated by the staircase transmission curve of @moeller, which is the result of unresolved  forest lines in intermediate resolution spectra. This transmission curve is also characterized by a broad trough (partly due to photoelectric absorption), the Lyman valley, whose minimum occurs at a rest-frame wavelength $\sim 650$Å. Before averaging their quasar spectra, ZK97 and TZ02 statistically corrected each quasar spectrum for the presence of such a Lyman valley by calculating the appropriate transmission curve, using the scheme developed by @moeller. Finally, the GP effect can be generalized to include the broadened wings of faint  forest lines as a result of the Hubble flow in underdense regions. Our contention is that these manifestations of the GP effect are sensitive to certain classes of  distributions as a function of redshift. Other  distributions might instead be associated to large scale voids. In that case, the amount of absorption gas would be decreasing towards the background quasar, against which we are trying to detect it. Such opacity behavior might be harder to pinpoint, since it would not necessarily produce a recognizable absorption break at the rest-frame . Another factor that could hinder the detection of intergalactic absorption is that the absorption component evolves in the opposite direction as the  clouds, becoming in other words progressively more abundant towards lower redshifts. For instance, hierarchical structure formation is expected to generate a hot gas phase, whose mass is predicted to increase with time [e.g. @phillips], that is, with a temporal evolution which is opposite to that of the  forest. Any residual absorption it may cause would occur at much shorter wavelengths, in a wavelength domain too far from the  emission line () to rely on an extrapolation of the continuum observed redward of this line. With these considerations in mind, we have explored various distributions, whose absorption behavior is not as evident as that expected from the traditional GP effect. The hypothesis being tested is this Paper is whether all or just a small fraction of the far-UV steepening in $F_{\nu}$ is caused by absorption. We will demonstrate in §\[sec:res\] how this hypothesis is observationally falsifiable. Even if most of the SED break turns out to be intrinsic to quasars, the technique can still be used to set a lower limit on the temperature of the warm-hot intergalactic medium (WHIM). Derivation of the transmission curve {#sec:tra} ------------------------------------ The technique used to generate a simulated composite spectrum to compare with the observed spectrum of ZK97 is described in @binb. Briefly, we assume a typical spectrograph sensitivity window for FOS of 1300–3000Å. We then simply redshift this window in locked steps in order to simulate quasars of different redshifts in the range $0.33 \le z_Q \le 3.6$ (the same range as ZK97). Before averaging, each redshifted quasar SED is multiplied by the redshift integrated transmission curve $T_{\lambda}$ (defined below). We assume the concordance $\Lambda$CDM cosmology with $\Omega_{\Lambda} = 0.7$, $\Omega_{M}=0.3$ and $h = 0.67$ with $h = H_0/100$. To be definite, we adopt a uniform gas distribution [see @bina]. Equivalent calculations assuming a clumpy medium were also carried out for comparison. We found that the ill-defined additional parameters required then do not reveal anything fundamental or even interesting. Furthermore, the two formalisms are equivalent for the assumed opacity regime in which the diffuse absorption component lies in the linear part of the curve of growth, that is for  columns $N_{H^0} \ll 10^{12}\,$. The simulated quasar spectra are divided in energy bins, and for each quasar rest-frame wavelength bin $j$, we calculate the transmitted intensity $I_{\lambda_j}^{tr} = I_{\lambda_j} T_{\lambda_j} = I_{\lambda_j} e^{-\tau(\lambda_j)}$ by integrating the opacity along the line-of-sight to quasar redshift $z_Q$ $$\begin{array}{cc} \tau(\lambda_j) = \sum_{i=0}^{10}{\int_{0}^{z_Q} \sigma_i(\frac{\lambda_j}{1+z})\; n_{H^0}(z) \frac{dl}{dz} \; dz} \label{eq:tau} \end{array}$$ where $\lambda_j$ is the quasar rest-frame wavelength for bin $j$ and $n_{H^0}(z)$ the intergalactic neutral hydrogen density, which consists of one of the three density distributions discussed below \[eqs (1–3)\]. The summation is carried out over the following opacity sources: photoionization ($i=0$) and line absorption from the Lyman series of hydrogen ($1\le i \le 10$). Although our code could include up to 40 levels, we found that considering only the 10 lowest proved to be adequate. We adopted a fiducial velocity dispersion $b$ of 30 km/s and assumed a Gaussian profile for the $\sigma_i$ of the lines. Heuristic  gas distributions {#sec:hi} ---------------------------- ### Three  distributions to test {#sec:three} In calculating the transmission function of a quasar at redshift $z_Q$, we will consider three different distributions of absorption gas density, , with redshift, which we will then compare with the composite SED of either ZK97 or TZ02. The first is given by $$\label{eq:nha} \begin{array}{cc} n_{H^0}(z) = n^0_{H^0} \,(1+z)^3 \, {(1+z)^{\gamma}} \, {(z_Q^{\prime})^{\beta}} \end{array}$$ where $z$ is the absorbing gas redshift, $z_Q$ the quasar redshift,  the quasar redshift as seen from the absorbing gas at $z$ \[that is $z^{\prime}_Q = (1+z_Q)/(1+z)\,-1$\], and  the neutral gas density at zero redshift. $\gamma$ and $\beta$ are adjustable parameters. The $(1+z)^3$ factor represents the cosmological expansion of the Universe. The factor $(z_Q^{\prime})^{\beta}$ with $\beta > 0$ implies that the density decreases in the neighborhood of individual quasars. The second distribution which will be considered is $$\label{eq:nhb} \begin{array}{cc} n_{H^0}(z) = n^0_{H^0} \,(1+z)^3\, {\frac{ {(1+z)^{\gamma}} }{1+({{z_P}/{z_Q^{\prime}}})^{\beta} }} \end{array}$$ where  is the size of the region (near each quasar) within which the density decreases as $({z_P}/{z_Q^{\prime}})^{\beta}$ towards the quasar \[note that $\beta > 0$ as in eq. (\[eq:nha\])\]. It is possible to substitute the denominator above by the expression ${1+({{r_P}/{r_Q^{\prime}}})^{\beta}}$. For each value of discussed below, we we have searched for the value of  which would give us an equivalent fit of the composite ISED. The parameter  represents the zone of influence (or cavity radius) of the quasars in spatial units. Although the density decreases smoothly towards each quasar, in this text we will refer to this zone, where ${\hbox{$r_Q^{\prime}$}}< {\hbox{$r_P$}}$ near each quasar, as a ‘cavity’. The third distribution is given by $$\label{eq:nhc} \begin{array}{cc} n_{H^0}(z) = n^0_{H^0} \,(1+z)^3 \, {\frac{ {\rm exp}[-(z/1.6)^{1.4}] }{1+({{z_P}/{z_Q^{\prime}}})^{\beta} }} \end{array}$$ where the exponential function in the numerator is a parametric fit to the increase of the warm-hot intergalactic medium (WHIM) towards the current epoch as calculated by @davea (their Model D2). We will again discuss which value of  produces a fit equivalent to that provided by our favored  values. ### Three conditions to satisfy {#sec:cond} The above distributions lie in a sequence of increasing plausibility, as discussed below. We emphasize that our exploration included many more functions than the above three, the vast majority of which could not reproduce a smooth steepening of the spectrum near 1100 Å() (failing condition [*a*]{} below). Nevertheless, these three distributions will suffice for the purpose of unequivocally testing the absorption hypothesis. In our search, the following procedure was adopted: we explored a large set of functions with the following three priorities in mind: [*a)*]{} the  distribution must produce an acceptable fit of the the composite ISED without a discontinuity near , [*b)*]{} it must be a physically meaningful function, and [*c)*]{} it must be consistent with other information we have about the Universe such as, for instance, the volume density of quasars. We were able to satisfy conditions [*a*]{} and [*b*]{}, although [*c*]{} could not be strictly respected for our favored test case Model C, as discussed below. Model results {#sec:res} ============= Reproducing the observed steepening {#sec:steep} ----------------------------------- If we adopt the first distribution of  corresponding to eq. (\[eq:nha\]) and the parameters listed in Table \[tbl\_1\] for Model A, that is, a zero redshift density ${\hbox{$n^0_{H^0}$}}= 4.7 \times 10^{-12} $, $\gamma=-1.5$ and $\beta = 0.8$, we can reproduce the observed smooth steepening of the composite spectrum of ZK97 remarkably well as shown in Fig. \[fig1\] by the thick solid line. The intrinsic SED considered in Model A is a single power-law of index $-1$ (short-long dashed line), which we extend to all wavelengths. This index provides a good fit at the longer wavelengths (1300–2300Å) of the underlying continuum in the ZK97 composite spectrum. This first distribution has also the property that the far-UV index, , defined in the range 350–900 Å (), increases very little with quasar redshift [@bina]. This agrees with the findings of TF02 (c.f. their Fig. 12). However, although the simplest, this distribution is not physically meaningful in the following aspect: the limitless range of the dependence of eq. (\[eq:nha\]) on implies that the local  density depends on the quasar redshift. To clarify this point: two independent quasars, one nearby and the other at redshift 3, but lying along very nearby line-of-sights on the sky, would be characterized by very different local  densities according to eq. (\[eq:nha\]). Such behavior is unphysical, because the density in the local Universe should not depend at what distance the background quasar is located. We therefore conclude that condition [*b*]{} (‘physically meaningful’, see §\[sec:cond\]) is not satisfied and that the  distribution of eq. (\[eq:nha\]) must be rejected. In the text, we will refer to this undesirable property as the “nearby line-of-sight problem”. A distribution without any dependence on  would not suffer from this problem. However, in [*all*]{} the distributions that succeeded in reproducing the smooth roll-over near 1200 Å, some kind of smooth gas cavity near each quasar had to be taken in, otherwise, a deep and sharp absorption discontinuity[^2] unavoidably appears at 1216 Å () in the synthesized composite SED, a feature that is obviously not observed. The thin long-dashed line in Fig. \[fig1\], for instance, illustrates the magnitude of this sharp discontinuity present near  when no cavity is considered (here with $\beta$ set to zero in eq. (\[eq:nha\])). The solution to this problem is to have the dependence on  limited to a reduced zone of influence near each quasar. This is achieved with the expression in the denominator in eqs (\[eq:nhb\]) and (\[eq:nhc\]), where  limits the distance scale within which the density is decreasing (towards the quasar) as a power-law of index $\beta$. We therefore have two competing effects present in our subsequent distributions \[of eq. (\[eq:nhb\]) or eq. (\[eq:nhc\])\]: one factor describes the  density as a function of redshift (seen from us) $z$, which has no other QSO close to the selected line-of-sight, and the other describes the density within the neighborhood of QSOs (within a zone of influence $\simeq {\hbox{$r_P$}}$). Two different line-of-sights lying at a projected distance on the sky greater than  will be described by a unique and self-consistent distribution of  with redshift. As a consistency check, the volume occupied by all the QSO cavities must be smaller than the total volume sampled. A possible interpretation of the decreasing density towards individual quasars is that our putative  distribution is associated to the largest scale voids. The environment of quasars might also be hotter and transparent as a result of (protocluster) stellar winds from the associated large scale and most massive structures. Our aim will be to make  as small as possible, but without producing any discontinuity or dip near 1200 Å. The short-dashed line in Fig. \[fig1\] shows that the distribution corresponding to eq. (\[eq:nhb\]) (Model B) results in as good a fit to the ZK97 spectrum as Model A, when using $\gamma = -3.0$, ${\hbox{$z_P$}}= 0.7$ and $\beta = 1.5$ (see Table \[tbl\_1\]). Nevertheless, this Model B is not satisfactory. In effect, a cavity with ${\hbox{$z_P$}}= 0.7$ requires a cavity of size ${\hbox{$r_P$}}\sim 2$ Gpc, which is more than half the path-length to the highest-$z$ quasar in the sample (i.e. 3.75 Gpc for $z_Q=3.6$). Hence most line-of-sights would cross many overlapping cavities (hence condition [*b*]{} is not satisfied) and such a situation not properly described by eq. (\[eq:nhb\]). Condition [*c*]{} (‘consistency with other reliable information’, see §\[sec:cond\]) is not respected either. First, the cumulative volume of quasar influence on exceeeds 15% of the total volume sampled. Second, the exponent of $(1+z)$, $\gamma = -3.0$, is much too steep if we were to compare our Model B to model-predictions of the evolution of the WHIM. For instance, most models in @davea are characterized by redshift-averaged values of $\gamma$ in the range $-1.1$ to $-1.5$. In order to ensure a physically more meaningful  distribution, we replaced the ($\gamma$) power-law dependence on redshift by an exponential fit to the WHIM Model D2 of @davea, the description of which lies in the numerator of eq. (\[eq:nhc\]). Replacing the previous best fit distribution with $\gamma =-3$ by a distribution that matches the behavior of the WHIM with redshift better \[by using either eq. (\[eq:nhc\]) or $\gamma \simeq -1.2$ and eq. (\[eq:nhb\])\] comes at a price, however. In effect, such distributions always result in a smooth rise in transmission towards the short wavelength extremity of the SED, beyond 500 Å. Interestingly, the more recent composite spectrum from TZ02 presents such a rise in the far UV as displayed in Fig. \[fig2\]. TZ02 have questioned the reality of this rise. The latter might be caused by a tendency of high redshift quasars to have a harder ISED. Nonetheless, because the TZ02 work contains more objects than the ZK97, and because this feature is seen in both radio-loud as well as radio-quiet quasars, we will take their results at face values and compare their radio-quiet composite SED with models which used eq. (\[eq:nhc\]). The thick solid line in Fig. \[fig2\] represents our Model C, which is characterized by a significantly smaller ${\hbox{$z_P$}}= 0.3$. Although the fit is imperfect, it matches the overall trends present in the TZ02 composite spectrum. The adopted intrinsic SED used here is somewhat harder, corresponding to an index of $-0.72$ (which is the value inferred by TZ02). For Model C, with ${\hbox{$z_P$}}= 0.3$ we find that ${\hbox{$r_P$}}= 800$ Mpc. A single cavity occupies therefore 1% of the total volume sampled. Using the known luminosity function of bright QSOs \[i.e. a quasar density of order $\xi = 3 \times 10^{-7} \, {\rm Mpc}^{-3}$ [@Boyle]\], the volume[^3] corresponding to this value of  is too large by a factor $\sim 650$ and condition [*c*]{} is therefore not fulfilled on the ground of the known density of QSOs. With this caveat in mind, we proceed to study the behavior of the absorption of our putative intergalactic component at wavelengths shorter than provided by HST-FOS. This will allow us to calibrate a new method to detect the WHIM and, indirectly, to determine to what extent condition [*b*]{} is satisfied in Model C. Prediction of a possible discontinuity in the far UV {#sec:false} ---------------------------------------------------- ### Analysis of the transmission curves {#sec:ana} The derivation of the composite SED of Model C required using eq. (\[eq:nhc\]) and calculating the transmission function at evenly spaced redshift values. We now turn to the particularities of these transmission curves. Two such transmission curves are presented in Fig. \[fig3\] for illustrative purposes for quasars of redshifts 2.76 and 0.33. Clearly, the main opacity source is the  line. The contribution of photoelectric absorption to the total opacity is relatively small as illustrated in the Figure while the contribution of higher series lines can be appreciated by the reduced size of the jumps that are visible in the transmission curves. The main result to emphasize is that a discontinuity appears to the blue of  () (that is beyond the UV limit of the HST-FOS detector) in all the calculated transmission curves, not only when assuming the parameters of Model C, but in all the  distributions that satisfied condition [*a*]{} (‘good fit’, see §\[sec:cond\]). By comparing the transmission curves in Fig. \[fig3\] with those resulting from  forest absorbers, a clear difference emerges. The transmission functions derived by [@moeller] (or ZK97) are characterized by a stair-case behavior, in which the opacity from each physical process (either a Lyman series line or photoelectric absorption) [*decreases*]{} towards shorter wavelengths, due to the strong evolutionary nature of the  forest, in which the density of absorbers increases steeply with redshift. In the case of the transmission resulting from equations (1)–(3), the opacity (along a given step of the staircase transmission curve) tends to either remain high or increase towards shorter wavelengths (but, of course, only up to the threshold wavelength of the absorption process involved). The obvious reason for this marked difference is that the distributions used here are characterized by a density that increases towards smaller redshifts $z$. An inspection of Fig. \[fig3\] suggests that, rather than looking for a sharp absorption edge near  in the quasar rest-frame, the most obvious demarcation produced by  distributions (which satisfy condition [*a*]{} of a ‘good fit’) is at the shorter wavelength end, in the 1190–1070 Å observer-frame region where  rather than  dominates the absorption. In order to detect such a discontinuity near 1216 Å (), one requires detectors with a sensitivity, which extends to wavelengths shorter than the earlier HST-FOS window, such as provided by FUSE or HST-STIS. ### Operative measurement of the jump: {#sec:def} In order to test our absorption hypothesis against individual quasar spectra observed by STIS or FUSE, we must first calibrate the depth of the expected discontinuity (on the blue side of local ) as a function of  for our different models. Since the intrinsic far-UV spectral index, , is not precisely known and, as suggested by the substantial scatter found by TZ02, may even vary from quasar to quasar, it makes sense to measure the depth of the jump in a way that does not depend on an a priori knowledge of the intrinsic ISED. The technique proposed here is to evaluate the discontinuity’s depth at 1160Å () by comparing the flux there to the extrapolated value from a power-law fit redward of  (), within the narrow window 1260–1360Å. This window is meant to exclude the Galactic  absorption trough as well as the geocoronal . We define the quantity ${\hbox{$\tau_{1160}$}}= {\rm log_e} \, [F^{1160}_{obs.}/F_{extr.}^{1160}]$, which can be shown to be insensitive to the power-law index of the intrinsic ISED assumed. In Fig. \[fig3\], we illustrate the procedure by taking as example the transmitted SED of a $z_Q=1.0$ quasar and Model C (the calculation assumes an ideal detector without wavelength coverage limitations). The two squares illustrate the position (albeit here in the quasar rest-frame), at which it is proposed to define . For each model discussed (A–D), we list in Table \[tbl\_1\] the value of  evaluated at $z_Q =1$, while in Fig. \[fig4\] we plot the behavior of  as a function of $z_Q$. ### Condition b and the behavior of {#sec:phy} As discussed in §\[sec:steep\] concerning the “nearby line-of-sight problem”, it is unphysical to have the local density vary when the background quasar, for instance, lies at redshift 2 rather than 3. We can easily find out when a distribution suffers from this problem by checking whether or not  is constant at moderate and high values of quasar redshift. It is apparent in Fig. \[fig4\] that Models A and B suffer severely from the above-mentioned problem. In the case of Model C, however,  is flat, at least beyond $z_Q \ga 0.6$. However, within a radius around us given by the large value of ${\hbox{$r_P$}}\sim 800$ Mpc, we may reasonably expect to lie within the zone of influence of a (dominant) quasar. In this case, rather than the smoothly increasing function depicted in Fig. \[fig4\], a single value of  may instead apply ( reflects the local density). Its value would depend on our distance from this dominant quasar. Furthermore,  would not necessarily be isotropic. In conclusion, the smooth initial rise of  in Model C is at best an idealization. Given the likelihood of being positioned inside the cavity of a single quasar, the predicted  constitutes an upper limit (a more probable value may lie a factor of a few lower). ### Negative results using two EUV quasar spectra {#sec:neg} We now use Model C as a test case, since it is the most physical at hand under the assumption that the 500 Å rise in the composite ISED of TZ02 is real and that the break is entirely due to intergalactic absorption. For any quasar $z_Q > 0.5$, one finds that  is quite significant. Typically, ${\hbox{$\tau_{1160}$}}\ga 0.25$ as seen in Fig.\[fig4\]. If that much absorption were present, it would therefore be a striking feature in the far-UV spectra \[provided the quasar could be observed blueward of 1200 Å ()\]. It turns out, however, that the observations of the quasar HS1543+5921 by [@bowen $z_Q=0.807$] with STIS and HE2347–4342 by [@kriss $z_Q=2.885$] with FUSE, show no indication at all of a discontinuity at the expected level of ${\hbox{$\tau_{1160}$}}=0.27$ and 0.28, respectively, in the 1160 Å region, as compared to the extrapolation from the 1270 Å region. We consider that a discontinuity as small as ${\hbox{$\tau_{1160}$}}= 0.05$, if present in the two spectra, would have been apparent. We conclude that intergalactic absorption cannot be the main cause of the observed steepening in the composite SED. Model C fails not only because of the observed absence of a discontinuity in the far-UV but also because it is inconsistent with the known density of quasars (condition [*c*]{}). No other  distribution could be found that would solve alll three conditions. Our study therefore confirms that the change of slope near 1100 Å must be in origin intrinsic to quasars, as proposed by ZK97 and TZ02. This general conclusion is not dependent on the particular distribution adopted, since in all the distributions we have explored (which satisfied condition [*a*]{}),  was always $\ge 0.2$ when $z_Q > 0.5$. Detecting the WHIM in absorption {#sec:detect} -------------------------------- Having recognized that most of the break is intrinsic to quasars, we now turn to the problem of determining how much flux discontinuity near 1160 Å (i.e. ) can be expected if we assume the  density predicted by WHIM models and a fiducial temperature of $10^{5.3}$ K. For this purpose, we adopt an intrinsic SED, which already incorporates the intrinsic far-UV steepening. It consists of a broken power-law, which has the same index of $-0.72$ in the near-UV as in Fig. \[fig2\], but that sharply turns over at 1200 Å () into a steeper index of $-1.57$ in the far-UV, as represented by the short-long dashed line in Fig. \[fig5\]. These two indices correspond to the values characterizing the radio-quiet quasar composite of TZ02 while the turn-over wavelength is within the range of values encountered by TZ02 ($\approx $ 1200–1300 Å). ### Model D using a broken power-law SED Adopting eq. (\[eq:nhc\]) and the above broken power-law, we first determine which values of  and  best reproduce the observed composite SED. The result is Model D, plotted as the solid line in Fig. \[fig5\]. The same $\beta=1.5$ is used as for Model C and could now be set to the small value of 0.08 without producing any significant discontinuity[^4] on the blue side of the quasar  emission line. Model D succeeds rather well in fitting the TZ02 composite. It does not constitute, however, a unique solution because of the uncertainties regarding the intrinsic SED. If for instance, we shifted the break from 1200 to 1100 Å, the assumed ISED would lie higher above the composite of TZ02 and a higher density WHIM would therefore be required for the model to overlap the data. With a predicted value of  as small as $0.05$ (see Table \[tbl\_1\]), Model D is characterized by a much smaller discontinuity than Models A–C, simply because the assumed ISED is much closer to the observed composite of TZ02. Furthermore, does not depend on redshift anymore, as shown in Fig. \[fig4\], which indicates that the cavity near the individual quasars is not affecting our determination of . Inspection of the published spectra of HS1543+5921 by [@bowen] and HE2347–4342 by [@kriss] suggests, however, that  is likely to be smaller than 0.05. A thorough analysis aiming at setting stringent upper limits on  would clarify the matter. ### Deriving  from the expected WHIM {#sec:wh} Let us now turn to the inverse problem of determining the height of the jump  expected from a WHIM component that would contribute 30% of the total baryonic mass in the local Universe [as in Model D2 of @davea]. One advantage of the operative definition of  proposed in §\[sec:def\] is that the choice of a given intrinsic SED is not critical to the purpose of determining . This is verified from Table \[tbl\_1\] where the quotient ${\hbox{$n^0_{H^0}$}}/{\hbox{$\tau_{1160}$}}$ evaluated at $z_Q=1$ in Models C and D are $1.89 \times 10^{-11} $ and $1.80 \times 10^{-11} \,$ , respectively. They differ by only 5%, despite the use of very different intrinsic SEDs. Because  scales linearly with , all the information needed to derive  can be extracted from Model D independently of whether or not the density inferred turns out very different than in Table \[tbl\_1\] or whether the ISED assumed is somewhat incorrect. For any arbitrarily small [*observed*]{} discontinuity , the inferred  density is given by ${\hbox{$n^0_{H^0}$}}= 1.8 \times 10^{-11} \, {\hbox{$\tau_{1160}^{obs.}$}}$ which does not depend on $z_Q$ because we have ${\hbox{$z_P$}}\ll 0.4$. In terms of the critical density $3H_0^2/8\pi G m_H$, this density becomes $$\begin{array}{cc} \Omega_{WHIM} = {{1.6 \times 10^{-6}}\over{x \, h^2}} \; {\hbox{$\tau_{1160}^{obs.}$}}\label{eq:om} \end{array}$$ where $x$ is the neutral fraction of the putative WHIM absorption gas component. We can express $\Omega_{WHIM}$ as a fraction $F_b$ of the (current epoch) total baryonic mass $\Omega_{bar.} = 0.021 h^{-2}$ $$\begin{array}{cc} F_b = \Omega_{WHIM}/\Omega_{bar.} = {{7.6 \times 10^{-5} }\over{x}} \; {\hbox{$\tau_{1160}^{obs.}$}}\label{eq:fb} \end{array}$$ With $F_b=0.3$, we obtain that ${\hbox{$\tau_{1160}^{obs.}$}}\simeq x/2.5 \times 10^{-4}$. Since the WHIM is likely to be highly ionized, we are lead to expect very small values of . For example, for a collisionally ionized WHIM at a temperature $10^{5.3}$K, we derive $x \sim 3.1 \times 10^{-6}$ and therefore ${\hbox{$\tau_{1160}^{obs.}$}}= 0.012$. Clearly, a high S/N spectrum is required in order to be able to detect a 1% excess flux at 1160Å. It remains nevertheless a worthwhile exercise to carry out, since any limit that can be put on  would directly translate into a lower limit on the neutral fraction (hence on the temperature) of the WHIM component. Note that, with that small , we have the liberty to set  to zero in eq. (\[eq:nhc\]) and condition [*c*]{} (§\[sec:cond\]) is therefore implicitly satisfied. By the same token, we do not have to worry about whether we possibly lie in the zone of influence of a single quasar anymore. In order to optimize the detection sensitivity, the strategy to follow would be to add together as many EUV quasar spectra as possible (ideally with redshift in the range $0.5 \la z_Q \la 0.8$) and then extract  from the summed stack. Conclusions {#sec:conc} =========== In all the distributions of ${\hbox{$n_{H^0}$}}(z)$ used for reproducing the steepening observed in the energy distribution of quasars, the dominant atomic process contributing to the opacity is   scattering. An important implication of these calculations is that the transmission curve shows a significant discontinuity to the blue of  () where the transmission rises sharply. The size of the discontinuity, , for any distribution of intergalactic gas that can successfully reproduce the SED steepening reported by TZ02 (assuming a single power-law SED), is predicted to be typically $> 0.25$ in spectra of quasar redshift $z_Q > 0.5$. Since the two quasars HS1543+5921 and HE2347–4342, which were observed by HST-STIS and FUSE, respectively, do not show any sign of a discontinuity anywhere near this level, we consider the case proven that most if not all of the 1200 Åcontinuum steepening is intrinsic to quasars rather than due to absorption. On the other hand, a small flux increase by as little as 1% percent is still possible and would be consistent with a WHIM temperature $\sim 10^{5.3}$K contributing up to 30% of the current baryonic matter. This work was supported by the Mexican Science Funding Agency CONACyT under grant 32139-E. We are indebted to Wei Zheng and Randal Telfer for sharing their published composite spectrum which are used in Fig. \[fig1\] and \[fig2\], respectively. We are grateful to V. Avila-Reese for constructive discussions about the manuscript. Binette, L., Rodríguez-Martínez, M., & Ballinas, I. 2002, in ASP Conf. Series, Galaxies: the third dimension, ed. Rosado, M., Binette, L. & Arias, L. (San Francisco: ASP), 282, 317 Binette, L., Rodríguez-Martínez, M., & Ballinas, I. 2003, in Revista Mexicana de Astronomía y Astrofísica, Conf. Series, Galaxy Evolution: Theory and Observations, ed. V. Avila-Reese, C. Firmani, C. Frenk & C. Allen (Mexico City: Mexico), in press (astro-ph/0205139) Boyle, B. J. 2001, Advanced Lectures on the Starburst-AGN, Edited by Itziar Aretxaga, Daniel Kunth, and Raúl Mújica. Singapore: World Scientific, p. 325 Bowen,D.V., Tripp, T. M., & Jenkins, E. B. 2001, , 121, 1456 Davé, R., et al. 2001, , 552, 473 Eastman, R. G., MacAlpine, G. M., & Richstone, D. O. 1983, , 275, 53 Gunn, J.E., & Peterson, B. A. 1965, , 142, 1633 Kriss, G. A., et al. 2001, Science, 293, 1112 Møller, P., & Jakobsen, P. 1990, , 228, 299 Phillips, A. A., Ostriker, J. P., & Cen, R. 2001, , 554, L9 Songaila, A., Hu, E. M., & Cowie, L. L. 1999, 525, L5 Telfer, R. C., Zheng, W., Kriss, G. A., & Davidsen, A. F. 2002, , 565, 773 (TZ02) Zheng, W., Kriss, G. A., Telfer, R. C., Grimes, J. P., & Davidsen, A. F. 1997, , 475, 469 (ZK97) [ccllcccc]{} A & (1) & $\nu^{-1}$ & $ 4.7 \times 10^{-12} $ & $-1.5$ & 0.8 & — & 0.306\ B & (2) & $\nu^{-1}$ & $ 1.5 \times 10^{-11} $ & $-3.0$ & 1.5 & 0.70 & 0.628\ C & (3) & $\nu^{-0.72}$ & $ 5.2 \times 10^{-12} $ & – & 1.5 & 0.30 & 0.275\ D & (3) & $\nu^{-1.57}$ & $ 9.0 \times 10^{-13} $ & – & 1.5 & 0.08 & 0.050\ [^1]: Throughout the text,  and  will denote wavelengths in the quasar rest-frame or in the observer-frame, respectively \[${\hbox{$\lambda_{rest}$}}= (1+z_Q)^{-1} {\hbox{$\lambda_{obs.}$}}$\]. [^2]: In @binb, such a discontinuity at 1216 Å () was smoothed out by invoking photoionization by the background quasar (in analogy with the “proximity effect”). However, the metagalactic background radiation (relative to a single quasar) is much too strong for this proposition to be sustained. @eastman avoided the discontinuity by invoking absorption cloudlets being progressively accelerated up to $0.8 c$ within a region $< 10$kpc near the background quasar. The main problem was that the cloudlets themselves emitted more UV than the background quasar. [^3]: That is, one should realize the condition $V_Q = \frac{4}{3} \pi {{\hbox{$r_P$}}}^{3} \la \xi^{-1}$. [^4]: The value of  which can provide an equivalent fit is 200 Mpc. Interestingly, this value is only a factor of two too large with respect to the density of quasars at high redshifts. This can be resolved if we use a function  of redshift such as $300 \,(1+z)^{-1}$ Mpc. Although  is then larger in the local Universe, at redshifts $\sim 2$ we get ${\hbox{$r_P$}}\ \simeq 100$ Mpc, which is the epoch when the density of quasar is known to peak. Condition [*c*]{} is marginally satisfied in Model D.

ArXiv

--- abstract: 'We give estimates of the $L^p$ norm of the Bergman projection on a strongly pseudoconvex domain in $\mathbb{C}^n$. We show that this norm is comparable to $\frac{p^2}{p - 1}$ for $1 <p< \infty$.' address: 'Department of Mathematics and Statistics, University of Toledo, Toledo, Ohio, 43606 zcuckovi@math.utoledo.edu' author: - 'Željko Čučkovi'' c' title: Estimates of the $L^p$ norms of the Bergman projection on strongly pseudoconvex domains --- Introduction ============ Let $\Omega\subset \mathbb{C}^n$ be a bounded domain and let $H(\Omega)$ denote the holomorphic functions on $\Omega$. Let $L^2(\Omega)$ be the standard space of square-integrable functions with respect to Lebesgue measure with the usual $L^2$ inner product. The Bergman space is defined as $ A^2 (\Omega) =:H(\Omega)\cap L^2(\Omega)$. The Bergman projection is the orthogonal projection operator $P:L^2(\Omega)\longrightarrow A^2(\Omega).$ This operator is one of the most fundamental operators in complex analysis. The Bergman kernel function $K(z,w)$ defined on $\Omega \times \Omega$ represents the Bergman projection as an integral operator $$Pf(z)=\int_\Omega K(z,w)f(w)\, dw,\qquad f\in L^2(\Omega)$$ where $dw$ denotes integration in the $w$ variables, with respect to the euclidean volume form. In this paper we are interested in estimating the $L^p$ norm of the Bergman projection on strongly pseudoconvex domains. Suppose that $\Omega$ is smoothly bounded, that is: there exists a $C^\infty$, real-valued function $r:\,\,\text{nbhd} \left(\overline\Omega\right)\longrightarrow\mathbb{R}$ such that $\Omega=\{ z: r(z) <0\}$ and $dr\neq 0$ when $r=0$. If $\Omega$ is strongly pseudoconvex, i.e. $i\partial\bar\partial r(p)\left(\xi,\bar\xi\right) >0$ for all $p\in b\Omega$ and all vectors $\xi\in\mathbb{C}^n$ satisfying $\partial r(p)\left(\xi\right)=0$, the boundary behavior of the Bergman kernel function associated to $\Omega$ is understood quite precisely. For this class of domains, the mapping properties of $P$ in many classical Banach spaces have been established. For our purposes, we mention the classical result that $P$ is a bounded operator on $L^p(\Omega)$ for $1 < p < \infty$ [@Pho-Ste]. On strongly pseudoconvex domains, Fefferman [@Fef] has established a complete asymptotic expansion of $K(z,w)$, in terms of $r(z), r(w)$ and a pseudo-distance between $z$ and $w$, as $z,w\to b\Omega$ (see also [@BoM-Sjo]). A crucial feature of the Bergman kernel on a strongly pseudoconvex domain is that its singularities occur only on the boundary diagonal, instead of on the full diagonal in $\Omega\times\Omega$. The simplest example of a strongly pseudoconvex domain in $ \mathbb{C}^n$ is an open unit ball $\mathbb{B}_n$. An important precursor to our work is the result by Zhu who established a sharp estimate of the $L^p$ norm of $\|P\|$ [@Zhu]. We state the unweighted version of his main theorem. \[TheoremA\] For all $1 < p < \infty$ there exists a constant $C > 0$, depending on $n$ but not on $p$, such that the norm of $P:L^p(\mathbb{B}_n)\longrightarrow A^p(\mathbb{B}_n)$ satisfies the estimate $$C^{-1} \csc \frac{\pi}{p} \leq \|P\|_p \leq C \csc \frac{\pi}{p}.$$ Zhu has also restated his results in the following way: there exists a constant $C > 0$ independent of $p$ such that $$C^{-1} \frac{p^2}{p-1} \leq \|P\|_p \leq C \frac{p^2}{p-1}.$$ We will prove estimates of $\|P\|_p$ for $P:L^p(\Omega)\longrightarrow A^p(\Omega),$ in case $\Omega$ is strongly pseudoconvex. The first theorem gives the upper estimate that is analogous to the estimate of Zhu. \[T:Theorem1\] Let $\Omega$ be a smoothly bounded strongly pseudoconvex domain. For all $1 < p < \infty$ there exists a constant $C' > 0$, depending on $n$ but not on $p$, such that the norm of $P:L^p(\Omega)\longrightarrow A^p(\Omega)$ satisfies the estimate $$\|P\|_p \leq C' \frac{p^2}{p-1}.$$ The next theorem gives lower estimates for the norm of $P$ for more general domains than strongly pseudoconvex domains. Recall that a domain $\Omega$ satisfies the Condition R if $P$ maps the space $C^{\infty} (\overline\Omega)$ into itself. \[T:Theorem2\] Let $\Omega\subset \mathbb{C}^n$ be a smoothly bounded domain and assume that the Condition R holds on $\Omega$. Then there exists $2 < p_0 < \infty$ and a constant $C' > 0$, depending on $n$ but not on $p$, such that $$C' \frac{p^2}{p-1} \leq \|P\|_p$$ for all $p \in (1, q_0) \cup (p_0, \infty)$, with $\frac 1p_0 +\frac 1q_0 =1.$ Since our domain $\Omega$ is strongly pseudoconvex, it does satisfy the Condition R. The following corollary follows immediately. \[T:Corollary 3\] Let $\Omega\subset \mathbb{C}^n$ be a smoothly bounded strongly pseudoconvex domain. Then there exists $2 < p_0 < \infty$ and a constant $C' > 0$, depending on $n$ but not on $p$, such that $$C' \frac{p^2}{p-1} \leq \|P\|_p$$ for all $p \in (1, q_0) \cup (p_0, \infty)$, with $\frac 1p_0 +\frac 1q_0 =1.$ Notice that the lower estimates are obtained for $p$ sufficiently large (or sufficiently close to 1). We do not know if there is an underlying reason for this restriction. The advantage of working on the open unit ball is that an explicit formula for the Bergman kernel is known. On strongly pseudoconvex domains we only have estimates for the Bergman kernel and it seems that the type of results in Theorem \[T:Theorem1\] and \[T:Theorem2\] has not been considered before. The paper is laid out as follows. In Section 2 we prove the upper estimate on $\|P\|_p$, while in Section 3 we prove the lower estimate. [**Acknowledgment.**]{} The author would like to thank Sonmez Sahutoglu for several discussions during the preparation of this work and to Akaki Tikaradze for pointing out a geometric argument in the lower estimate. Upper estimate ============== The work in this section was motivated by the work [@McN-Ste] and the related work on Toeplitz operators in [@Cuc-McN]. The proof of our upper estimate depends on the Schur lemma applied to the Bergman kernel. \[P:Prop2.1\] Suppose $\mu$ is a positive measure on a space $X$ and $K(x, y)$ is a nonnegative measurable function on $X \times X$. Let $1< p <\infty$ be given and let $q$ be the conjugate exponent of $p$, $\frac 1p +\frac 1q =1$. Suppose there exists $C>0$ and a positive function $h(x)$ on $X$ such that $$\label{E:2.2} \int_X K(x,y) h(y)^q d \mu(y) \leq C h(x)^q%\tag 2.8$$$$ for almost all $x$ in $X$ and $$\label{E:2.3} \int_X K(x,y) h(x)^p d \mu(x) \leq Ch(y)^p%\tag 2.8$$$$ for almost all $y$ in $X$. Then the integral operator $$\label{E:2.4} Tf(x)=\int_X K(x,y)f(y)\, d \mu(y)%\tag 2.9$$$$ is bounded on $L^p(X, \mu)$ with norm not exceeding $C$. If $\Omega$ is smoothly bounded strongly pseudoconvex domain, the earlier work in [@Hor] contained only estimates on the Bergman kernel restricted to the diagonal. As we have already mentioned Fefferman [@Fef] has established a complete asymptotic expansion of $K(z,w)$ on $\Omega$. For our purposes we now recall only the upper bounds on the Bergman kernel. \[P:Prop2.5\] Let $\Omega=\{ r<0\}$ be a smooth, bounded, strongly pseudoconvex domain in $\mathbb{C}^n$. Let $K(z, w)$ denote the Bergman kernel of $\Omega.$ For each $p\in b\Omega$, there exists a neighborhood $U$ of $p$, holomorphic coordinates $(\zeta_1,\dots ,\zeta_N)$ and a constant $C>0$, such that for $z,w\in U\cap\Omega$ $$\label{E:2.6} |K(z,w)|\leq C\left( |r(z)|+|r(w)|+ |z_1- w_1| +\sum_{k=2}^n |z_k -w_k|^2\right)^{-(n+1)}.%\tag 3.1$$$$ Here $z=(z_1,\dots , z_n)$ in the $\zeta$-coordinates, and similarly for $w$. Inequality (\[E:2.6\]) can be extracted from the results in [@Fef]; it may also be obtained by simpler methods as shown by [@McN] and [@N-R-S-W], see the remark (5.3) in [@McN]. In what follows we use the notation $f(z)\lesssim g(z)$ to denote that there exists a constant $C$, independent of $z$ and $\epsilon$, such that $f(z)\leq Cg(z)$. In order to apply the Schur lemma, we need the following proposition. \[P:Prop2.7\] Let $\Omega=\{ r<0\}$ be a smooth, bounded, strongly pseudoconvex domain in $\mathbb{C}^n$, and let $K(z,w)$ be the Bergman kernel associated to $\Omega$. Then $$\label{E:2.8} \int_\Omega \left| K(z,w)\right||r(w)|^{-\epsilon}\, dw \lesssim \frac{1}{\epsilon (1 - \epsilon)} |r(z)|^{-\epsilon},%\tag 3.3$$$$ for all $0 <\epsilon <1$. The proof follows the lines of Lemma 1 in [@McN-Ste]. We will skip certain steps to avoid the repetition. Let $\Delta_b=\{ (z,z): z\in b\Omega\}$ be the boundary diagonal of $\overline\Omega\times\overline\Omega$. It is well known that $$\label{E:2.9} K(z,w)\in C^\infty\left( \overline\Omega\times\overline\Omega\setminus \Delta_b\right)%\tag 3.4$$$$ see [@Ker]. Cover $b\Omega$ by neighborhoods $U_1,\dots U_M$ given by Proposition \[P:Prop2.5\]; we may assume that the neighborhoods are so small that the quantity in parenthesis on the right hand side of (\[E:2.6\]) is less than 1. Now consider an arbitrary $U_j$, $1\leq j\leq M$ and let $z,w\in \overline\Omega\cap U_j$. Assume $z\in U_j$ is temporarily fixed. Then Proposition \[P:Prop2.5\] gives $$\begin{split} I_j&=\int_{U_j}\left| K(z,w)\right||r(w)|^{-\epsilon}\, dw \\ &\leq C\int_{\mathbb{C}^n} \left( |r(z)|+|r(w)|+ |z_1- w_1| +\sum_{k=2}^n |z_k -w_k|^2\right)^{-(n+1)} |r(w)|^{-\epsilon}\, dw. \end{split}$$ As in [@McN-Ste], we change the coordinates: $\widetilde w_k = w_k- z_k$, $k=2,\dots , n$, $\text{Re } \widetilde w_1 = r(w)$, $\text{Im }\widetilde w_1 = \text{Im } w_1$. Now let $x=\text{Re }\widetilde w_1$ and $y=\text{Im }z_1 -\text{Im } w_1$. We then obtain $$\label{E:2.10} I_j\leq C\int_{\mathbb{C}^n}\left( |r(z)|+|x|+ |y| +\sum_{k=2}^n |\widetilde w_k|^2\right)^{-(n+1)} |x|^{-\epsilon}\, d\widetilde w_2\dots d\widetilde w_n\, dx\, dy.%\tag 3.6$$$$ We start evaluating this iterated integral by performing the $\widetilde w_2$ integration first. Define $$\begin{aligned} R_1 &=&\left\{ \widetilde w_2: |\widetilde w_2|^2 > |r(z)|+|x|+ |y| +\sum_{k=3}^n |\widetilde w_k|^2\right\} \\ R_2 &= &\left\{ \widetilde w_2: |\widetilde w_2|^2 < |r(z)|+|x|+ |y| +\sum_{k=3}^n |\widetilde w_k|^2\right\}.\end{aligned}$$ Using polar coordinates on the region $R_1$ we have $$\begin{split} \int_{R_1} &\left( |r(z)|+|x|+ |y| +\sum_{k=2}^n |\widetilde w_k|^2\right)^{-(n+1)} |x|^{-\epsilon}\, d\widetilde w_2 \\ &\leq \int_{R_1} \left(|\widetilde w_2|^2\right)^{-(n+1)} |x|^{-\epsilon}\, d\widetilde w_2 \\ &\lesssim \left(|r(z)|+|x|+ |y| +\sum_{k=3}^n |\widetilde w_k|^2\right)^{-n}|x|^{-\epsilon}. \end{split}$$ On the region $R_2$ we obtain the same upper bound by using the estimate $$\begin{split} \int_{R_2} &\left( |r(z)|+|x|+ |y| +\sum_{k=2}^n |\widetilde w_k|^2\right)^{-(n+1)} |x|^{-\epsilon}\, d\widetilde w_2 \\ &\leq\left(|r(z)|+|x|+ |y| +\sum_{k=3}^n |\widetilde w_k|^2\right)^{-(n+1)}|x|^{-\epsilon}\,\text{vol}(R_2). \\ \end{split}$$ We continue in the same way to perform the integration on $d\widetilde w_3, \dots d\widetilde w_n$ and $dy$ integrals, reducing one negative power of the integrand at each step, to obtain $$I_j\lesssim \int_\mathbb{ R}\left(|r(z)| +|x|\right)^{-1}|x|^{-\epsilon}\, dx.$$ We now estimate this final integral: $$\begin{split} 2 \int_ 0^{\infty}&\left( |r(z)|+x \right)^{-1} x^{-\epsilon}\, dx\\ &=2 \int_{|r(z)|}^{\infty} \left( |r(z)|+x \right)^{-1} x^{-\epsilon}\, dx + 2 \int_{0}^{|r(z)|} \left( |r(z)|+x \right)^{-1} x^{-\epsilon}\, dx \\ &\leq 2 \int_{|r(z)|}^{\infty} x^{-\epsilon - 1}\, dx + 2 \int_{0}^{|r(z)|} |r(z)|^{-1} x^{-\epsilon}\, dx \\ & = \frac{1}{\epsilon} |r(z)|^{-\epsilon} + \frac{1}{1 - \epsilon} |r(z)|^{-\epsilon} \\ &= \frac{1}{\epsilon (1 - \epsilon)} |r(z)|^{-\epsilon}. \\ \end{split}$$ Thus we obtain $$I_j\lesssim \frac{1}{\epsilon (1 - \epsilon)} |r(z)|^{-\epsilon}.$$ If $U_0=\overline\Omega\setminus\cup_{j=1}^M U_j$ then, it follows from (\[E:2.9\]) that there exists $M > 0$ so that $|K(z, w)| \leq M$ for all $z, w \in U_0$. Similarly for $w \in U_0$, there exists $\eta_1$ and $\eta_2$ such that $ \eta_1 \leq |r(w)| \leq \eta_2$ for all $w \in U_0$. Hence $$\begin{split} \int_{U_0} \left| K(z,w)\right||r(w)|^{-\epsilon}\, dw &\lesssim \eta_1^{-\epsilon} \\ & \lesssim (\frac{\eta_2}{\eta_1})^{\epsilon} |r(z)|^{-\epsilon}\\ & \lesssim |r(z)|^{-\epsilon}\\ \end{split}$$ since $\frac{\eta_2}{\eta_1} > 1$ and $0 < \epsilon < 1$. Thus we obtained the same bounds on each of the open sets $U_0, \dots , U_M$, we have shown (\[E:2.8\]). We are now prepared to give the proof of the upper estimate. We apply the Schur lemma to the function $h(z) = |r(z)|^{-\frac{1}{pq}}.$ Then (\[E:2.2\]) becomes $$\int_\Omega |K(z,w)|h(w)^q dw = \int_\Omega |K(z,w)||r(z)|^{-\frac{1}{p}} dw.\\$$ Now apply Proposition 2.7 with $\epsilon = \frac{1}{p}$ to obtain for all $z \in {\Omega}$ $$\begin{split} \int_\Omega |K(z,w)|h(w)^q dw &\lesssim pq |r(z)|^{-\frac{1}{p}}\\ &= \frac{p^2}{p - 1}h(z)^q.\\ \end{split}$$ By the symmetry of the kernel and the estimate $pq$ on the right hand side, (\[E:2.3\]) is satisfied for all $w \in {\Omega}$ with the same upper bound. We conclude that $$\|P\|_p \lesssim \frac{p^2}{p-1}.\qedhere$$ Lower estimate ============== After doing a holomorphic change of coordinates, for simplicity, we can assume that $0$ is in the boundary of $\Omega$ and $y_n = \text{Im}z_n > 0$ is the outward real normal. Let us consider the case $p >2.$ Now for each $p$ choose $z_p$ to be the point $z_p = (0, 0, \dots, -ie^{-p}).$ Then $z_p \in \Omega$ for $p$ large enough. Now we follow the idea from [@Zhu]. Let $$f(z) = \log (iz_n) - \overline{\log(iz_n)}.$$ Clearly $f(z) = 2i \arg(iz_n)$ is a bounded function on $\Omega$ and hence $\|f\|_p \leq 2\pi.$ It is not difficult to see that $\log (iz_n) \in A^2(\Omega)$. Hence $Pf(z) = \log (iz_n) - P(\overline{\log(iz_n)}).$ Now we recall a recent result from [@HMS], Corollary 1.12. \[T:Theorem3.1\] Let $\Omega\subset \mathbb{C}^n$ be a smoothly bounded domain and assume that the Condition R holds on $\Omega$. Then for any $f \in A^2(\Omega)$, $P \overline f$ is smooth up to the boundary. We now use this theorem to conclude that there exists a constant $M > 0$ so that $|P(\overline{\log(iz_n)})| < M$ on $\Omega$. It is well known that point evaluations are bounded on $A^p(\Omega)$, see for ex. [@Kr], Lemma 1.4.1 modified for arbitrary $p > 1$ or [@Dur-Sch] in case of planar domains. More concretely, for a $z\in {\Omega}$ we have $$|f(z)| \leq C(n) d(z)^{-\frac{2n}{p}} \|f\|_p$$ for all $f \in A^p(\Omega)$, where $d(z)$ denotes the distance of $z$ to the boundary. We will apply this inequality to $f(z)$ with the point $z_p$ chosen above. Hence we obtain $$\begin{aligned} \label{E:3.4} \nonumber \|Pf\|_p &\geq C(n) d(z_p)^{\frac{2n}{p}} |Pf(z_p)|\\ & \gtrsim (e^{-2n}) |Pf(z_p)|\\ \nonumber & \gtrsim \left| |\log(e^{-p})| - |P(\overline{\log(iz_n))(e^{-p})}|\right| \\ \nonumber & > p - M \\ \nonumber & \gtrsim p\end{aligned}$$ for $p$ large enough. Hence we conclude that there exists a $p_0 > 2$ such that $\|Pf\|_p \gtrsim p$ for all $p > p_0$. This shows that $$\frac{\|Pf\|_p}{\|f\|_p} \gtrsim p$$ for $p > p_0$ which shows that $\|Pf\|_p > Cp$, with $C$ depending on $n$ but not on $p > p_0$. Since $p > 2$, we have $p > \frac{1}{2} \frac{p^2}{p-1}$ which gives $$\|Pf\|_p > C'\frac{p^2}{p-1},$$ with $C'$ depending on $n$ but not on $p > p_0$. This gives the lower estimate in our theorem. By duality and the symmetry of $pq$ we get the estimate for the range $p \in (1, q_0)$. [99]{} L. Boutet de Monvel & J. Sjöstrand, *Sur la singularité des noyaux de Bergman et de Szegö*, Soc. Math. France Astérisque **34-35** (1976), 123-164 Ž. Čučkovi' c & J. D. McNeal, *Special Toeplitz operators on strongly pseudoconvex domains*, Rev. Mat. Iberoamericana **22** (2006), 851-866 P. Duren & A. Schuster, *Bergman spaces*, Math. Surveys and Monographs **100**, Amer. Math. Soc. (2004) C. Fefferman, *The Bergman kernel and biholomorphic mappings of pseudoconvex domains*, Inv. Math. **26** (1974), 1-65 A.-K.Herbig, J.D. McNeal & E.J. Straube, *Duality of holomorphic function spaces and smoothing properties of the Bergman projection*, Trans. Amer. Math. Soc. **366** (2014), 647-665 L. Hörmander, *$L^2$ estimates and existence theorems for the $\overline\partial$-operator*, Acta Math. **113** (1965), 89-152 N. Kerzman, *The Bergman kernel function. Differentiability at the boundary*, Math. Ann. **195** (1972), 149-158 S. G. Krantz *Function theory of several complex variables*, Sec. Ed., AMS Chelsea Publishing 2001. J.D. McNeal, *Boundary behavior of the Bergman kernel function in $\mathbb{C}^2$*, Duke Math. J. **58** (1989), 499-512 J.D. McNeal & E.M. Stein, *Mapping properties of the Bergman projection on convex domains of finite type*, Duke Math. J. **73** (1994), 177-199 A. Nagel, J.P. Rosay, E.M. Stein, & S. Wainger, *Estimates for the Bergman and Szegö kernels in $\mathbb{C}^2$*, Ann. of Math. **129** (1989), 113-149 D.H. Phong & E.M. Stein, *Estimates for the Bergman and Szeg" o projections on strongly pseudoconvex domains*, Duke Math. J. **44** (1977), 695-704 Kehe Zhu, *A sharp norm estimate for the Bergman projection on $L^p$ spaces*, Bergman spaces and related topics in complex analysis, Contemp. Math **404** (2006), 199-205

ArXiv

--- abstract: 'When the sizes of photonic nanoparticles are much smaller than the excitation wavelength, their optical response can be efficiently described with a series of polarizability tensors. Here, we propose a universal method to extract the different components of the response tensors associated with small plasmonic or dielectric particles. We demonstrate that the optical response can be faithfully approximated, as long as the effective dipole is not induced by retardation effects, hence do not depend on the phase of the illumination. We show that the conventional approximation breaks down for a phase-driven dipolar response, such as optical magnetic resonances in dielectric nanostructures. To describe such retardation induced dipole resonances in intermediate-size dielectric nanostructures, we introduce “pseudo-polarizabilities” including first-order phase effects, which we demonstrate at the example of magnetic dipole resonances in dielectric spheres and ellipsoids. Our method paves the way for fast simulations of large and inhomogeneous meta-surfaces.' author: - - - - - - - bibliography: - 'extrac-pola-plasmonic.bib' title: Polarizabilities of complex individual dielectric or plasmonic nanostructures --- Introduction ============ In a multitude of topical areas in contemporary physics and chemistry, the concept of the polarizability has proven to be extremely useful. In particular, in the physics of gases and surfaces, the dynamic polarizability tensor of molecules appears explicitly in the description, for example, of the Van der Waals dispersion energy, or in the description of the Raman scattering process.[@girard_effective_1987; @girard_effective_1987-1; @buckingham_polarizability_1956; @a._d._buckingham_permanent_1967; @barron_rayleigh_1975; @buckingham_intermolecular_1975] During the 1970s, A. D. Buckingham wrote a founding article on this subject in which an exhaustive list of linear polarizabilities is proposed.[@buckingham_polarizability_1979] Although this work was restricted to atomic and molecular systems, it represents a valuable stand of the various possible contributions as well as their ranking, in terms of electric and magnetic multipolar polarizabilities. The theoretical study of the linear optical response of small metallic or dielectric particles has also been extensively investigated in the last decades. In particular, in the context of plasmonics the concept of polarizability is often applied to the description of both far-field and near–field plasmon spectra of sub-wavelength sized noble metal particles.[@evlyukhin_optical_2010; @sersic_magnetoelectric_2011; @bowen_using_2012; @arango_polarizability_2013; @bernal_arango_underpinning_2014; @asadchy_bianisotropic_2018; @bertrand_global_2019] In many situations, single metal particles can be schematized by a sphere of radius $a$, in which case their optical response can be described by a scalar, frequency-dependent polarizability $\alpha(\omega_{0})$. Then, the polarizability tensor is diagonal and all tensor elements are identical. In cgs units, it reads:[@draine_discrete-dipole_1988] $$\alpha_{ij}(\omega_{0})=a^{3}\left(\frac{\epsilon_{2}(\omega_{0})-\epsilon_{1}} {\epsilon_{2}(\omega_{0})+2\epsilon_{1}}\right) \; , \label{ALPHA1}$$ where $\epsilon_{1}$ (respectively $\epsilon_{2}$) is the dielectric constant of the medium (respectively the nano-sphere). From relation (\[ALPHA1\]), we can extract the extinction spectrum via the imaginary part of $\alpha(\omega_{0})$. Consequently, the extinction spectra of a sample containing a large number $N$ of such non–interacting nanoparticles $\alpha_i(\omega_{0})$ is given by:[@draine_discrete-dipole_1988; @girard_shaping_2008] $$I_{\text{ext}}(\lambda_{0})=\frac{8 \pi^{2}}{\lambda_{0}} \sum\limits_i^N {\text{Im}}\Big( \alpha_i(\omega_{0}) \Big) \; , \label{EXTINC}$$ where $\lambda_{0}$ represents the incident wavelength and “${\text{Im}}$” the imaginary part. The sphere represents the highest symmetry, belonging to the [*isotropic*]{} symmetry group. As stated above, in this case, all the diagonal elements of the polarizability are identical, and the system displays a scalar response defined by $\alpha_{ij}=\alpha \delta_{ij}$ (see equations (\[ALPHA1\]) and (\[EXTINC\])). When transforming the sphere into an ellipsoid of symmetry group $D_{\infty h}$, the polarizability must be defined with two independent components, and for even lower symmetry, all components $\alpha_{ij}$ of the polarizability tensor must be calculated. This situation corresponds to high anisotropy induced by a complex shape of particles (or nano-cavities). Note that other kinds of anisotropy can come from the intrinsic anisotropy of the dielectric constant of the particle but also from the surface of another object.[@sihvola_electromagnetic_1994] In the latter case, the concept of effective polarizability is generally introduced, and the final symmetry of the particle is dressed by the symmetry of the surface (i.e. $D_{\infty h}$, for a perfectly planar surface). As illustrated by these examples, the design of nanostructure polarizabilities starts with the conception of a reference geometry by intuitive considerations. Such an approach, however, is limited to rather simple problems. In case of complex structures or complicated phenomena, the intuitive method often fails, as unexpected effects such as polarization conversion occur in the polarizability tensors. In this work we propose a numerical method to extract the polarizability tensors for complex shaped metallic and dielectric nanostructures through a volume discretization technique, which uses the concept of a [*generalized propagator*]{}. Furthermore, in order to faithfully describe also magnetic optical effects in dielectric nanostructures, where the conventional dipolar polarizability approximation fails, we introduce “pseudo-polarizabilities” that include phase-induced magnetic dipole resonances, similar to some homogenization approaches for metamaterials,[@alu_first-principles_2011; @ciattoni_nonlocal_2015] but at the level of a single, isolated structure. Our pseudo-polarizabilities might then be used to construct aperiodic or random metasurface-like assemblies without periodicity. A Generalized Electromagnetic Propagator for Arbitrary Shaped Particles or Cavities =================================================================================== The concept of the [*generalized electric field propagator*]{} previously described in reference  can be easily extended to the general case of meta-systems displaying both an electric and a magnetic linear response. In this case, the source zone as depicted in figure \[fig:zone\] is characterized by the following susceptibility tensor, where $\fakebold{\mathbb{I}}$ is the identity tensor: ![ Sketch used to implement the concept of generalized electromagnetic propagator. (a) transparent reference medium with $\epsilon_{1}(\omega_0)=n_{1}^{2}$ and $\mu_{1}=1$; (b) material system of arbitrary shape, also called the [*source zone*]{}, embedded in the reference medium (permittivity $\epsilon_{s}(\omega_0)$ and permeability $\mu_{s}(\omega_0)$). []{data-label="fig:zone"}](fig_1_pola){width="\columnwidth"} $$\bm{\chi}(\omega_0)= \left( \begin{matrix} \chi_{e}(\omega_0)\, \fakebold{\mathbb{I}} & 0 \\[12pt] 0 & \chi_{m}(\omega_0) \fakebold{\mathbb{I}} \end{matrix}\right) \; ; \label{CHI}$$ where $\chi_{e}(\omega_0)$ and $\chi_{m}(\omega_0)$ are related to the permittivity $\epsilon_{s}(\omega_0)$, respectively the permeability $\mu_{s}(\omega_0)$ of the source zone: $$\chi_{e}(\omega_0)=\frac{\epsilon_{s}(\omega_0)-\epsilon_{1}(\omega_0)}{4\pi} \; ; \label{CHI-E}$$ and $$\chi_{m}(\omega_0)=\frac{\mu_{s}(\omega_0)-\mu_{1}(\omega_0)}{4\pi} \; . \label{CHI-M}$$ Introducing two super vectors $\mathbf{F}_{0}(\mathbf{r},\omega_0) = (\mathbf{E}_{0}(\mathbf{r},\omega_0),\, \mathbf{H}_{0}(\mathbf{r},\omega_0))$ and $\mathbf{F}(\mathbf{r},\omega_0) = (\mathbf{E}(\mathbf{r},\omega_0),\, \mathbf{H}(\mathbf{r},\omega_0))$ (where $\mathbf{E}$ and $\mathbf{H}$ refer to electric and magnetic fields, respectively) to describe the incident and total electromagnetic fields, we can define a unique $(6\times 6)$ dyadic tensor $\protect\fakebold{\mathbb{K}}(\mathbf{r},\mathbf{r}',\omega_0)$ operating in the volume $V$ of the source zone and establishing the link between $\mathbf{F}_{0}(\mathbf{r},\omega_0)$ and $\mathbf{F}(\mathbf{r},\omega_0)$: $$\mathbf{F}(\mathbf{r},\omega_0)=\int_{V}\protect\fakebold{\mathbb{K}}(\mathbf{r},\mathbf{r}',\omega_0) \cdot\mathbf{F}_{0}(\mathbf{r}',\omega_0)\mathrm{d}\mathbf{r}'\; . \label{EQ-INT-K}$$ Actually, the $(6\times 6)$ superpropagator $\protect\fakebold{\mathbb{K}}(\mathbf{r},\mathbf{r}',\omega_0)$ is composed of four mixed $(3\times 3)$ dyadic tensors: $$\protect\fakebold{\mathbb{K}}(\mathbf{r},\,\mathbf{r}',\omega_0)= \left( \begin{matrix} K^{EE}(\mathbf{r},\mathbf{r}',\omega_0) & K^{EH}(\mathbf{r},\mathbf{r}',\omega_0) \\[12pt] K^{HE}(\mathbf{r},\mathbf{r}',\omega_0) & K^{HH}(\mathbf{r},\mathbf{r}',\omega_0) \end{matrix}\right) \label{K-KMIX}$$ in which the first one, $K^{EE}(\mathbf{r},\mathbf{r}',\omega_0)$ that describes the [*electric–electric field couplings*]{} was introduced in the early beginning of [*near–field optics*]{} [@martin_generalized_1995]. The three other contributions, i. e. $K^{EH}(\mathbf{r},\mathbf{r}',\omega_0)$, $K^{HE}(\mathbf{r},\mathbf{r}',\omega_0)$ and $K^{HH}(\mathbf{r},\mathbf{r}',\omega_0)$, account for coupling with the magnetic field. All these propagators are related to the corresponding [*mixed field–susceptibilities*]{} $S^{EE}$, $S^{EH}$, $S^{HE}$, and $S^{HH}$, [@sersic_magnetoelectric_2011; @wiecha_decay_2018] associated with the [*source zone*]{}: $$\begin{aligned} &\mathbf{K}^{EE}(\mathbf{r},\mathbf{r}',\omega_0)=\delta(\mathbf{r}-\mathbf{r}')\protect \fakebold{\mathbb{I}}+\chi_{e}(\omega_0)\cdot S^{EE}(\mathbf{r},\mathbf{r}',\omega_0)\\[2pt] &\mathbf{K}^{EH}(\mathbf{r},\mathbf{r}',\omega_0)=\chi_{m}(\omega_0)\cdot S^{EH}(\mathbf{r},\mathbf{r}',\omega_0)\\[2pt] &\mathbf{K}^{HE}(\mathbf{r},\mathbf{r}',\omega_0)=\chi_{e}(\omega_0)\cdot S^{HE}(\mathbf{r},\mathbf{r}',\omega_0)\\[2pt] &\mathbf{K}^{HH}(\mathbf{r},\mathbf{r}',\omega_0)=\delta(\mathbf{r}-\mathbf{r}')\protect \fakebold{\mathbb{I}}+\chi_{m}(\omega_0)\cdot S^{HH}(\mathbf{r},\mathbf{r}',\omega_0) \end{aligned} \label{MIXED-S}$$ As explained in references  and , these dyadic tensors can be numerically computed by performing a volume discretization of the [*source zone*]{} together with a [*Dyson sequence procedure*]{}[@martin_generalized_1995] or other numerical inversion techniques, to extract the various field-susceptibilities in the [*source zone*]{}. Extraction of polarizabilities of small nanostructures ====================================================== The volume discretization of the [*source region*]{} leads to a mesh of $N$ identical elementary volumes $\Delta v$. Such a procedure converts integrals over the source volume $V$ into discrete summations. In reference [@girard_shaping_2008] we have gathered the expressions of the discretization volume elements $\Delta v$ for both cubic and hexagonal compact discretization grids together with the corresponding Green’s function renormalization terms. The electric polarization at the [*i*]{}th cell in the source region can be written as follows: $$\label{Pola_elec} \begin{aligned} \mathcal{P}(\mathbf{r}_{i},\omega_{0})= \Delta v^{2} & \chi_e(\omega_0) \\ \times \sum_{j=1}^{N} \Big( & \mathbf{K}^{EE}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0}) \cdot\mathbf{E}_{0}(\mathbf{r}_{j},\omega_{0}) \\ & + \mathbf{K}^{EH}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0}) \cdot\mathbf{H}_{0}(\mathbf{r}_{j},\omega_{0})\Big) \, . \end{aligned}$$ Concerning the magnetic polarization $\mathcal{M}(\mathbf{r}_{i},\omega_{0})$ induced in the [*source region*]{}, it may be split into two contributions related to $\chi_{e}(\omega_{0})$ and $\chi_{m}(\omega_{0})$, respectively: $$\label{Pola_mag} \mathcal{M}(\mathbf{r}_{i},\omega_{0})=\mathcal{M}_{e}(\mathbf{r}_{i},\omega_{0}) +\mathcal{M}_{m}(\mathbf{r}_{i},\omega_{0})$$ with $$\label{Pola_mage} \begin{aligned} \mathcal{M}_{e}(\mathbf{r}_{i},\omega_{0}) = &-\dfrac{ik_{0}}{2}\Delta v^{2} \chi_{e}(\omega_{0}) \\ \times \sum_{j=1}^{N} & \Big(\mathbf{r}_{i}\wedge\mathbf{K}^{EE} (\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0})\cdot\mathbf{E}_{0}(\mathbf{r}_{j},\omega_{0}) \\ & + \mathbf{r}_{i}\wedge\mathbf{K}^{EH}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0})\cdot\mathbf{H}_{0}(\mathbf{r}_{j},\omega_{0})\Big) \end{aligned}$$ and $$\label{Pola_MAGm} \begin{aligned} \mathcal{M}_{m}(\mathbf{r}_{i},\omega_{0}) = & \Delta v^{2}\chi_{m}(\omega_{0}) \\ \times \sum_{j=1}^{N} & \Big(\mathbf{K}^{HE}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0})\cdot\mathbf{E}_{0}(\mathbf{r}_{j},\omega_{0}) \\ & + \mathbf{K}^{HH}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0})\cdot\mathbf{H}_{0}(\mathbf{r}_{j},\omega_{0})\Big) \; , \end{aligned}$$ where the first contribution in Eq. (\[Pola\_mag\]), proportional to $k_{0}$ = $\omega_{0}/c$, originates from polarization vortices induced by phase changes inside the source region. These magnetic polarization effects have been extensively studied recently in the case of high index dielectric nano-structures.[@evlyukhin_optical_2010; @kuznetsov_magnetic_2012; @albella_low-loss_2013; @albella_electric_2014; @decker_resonant_2016; @barreda_recent_2019] Note that the choice of the center of the coordinate system is important, as it has an impact on the magnetic polarization $\mathcal{M}_e$. Usually, it is convenient to use the center of mass $\mathbf{r}_{c}$ of the nanostructure[@evlyukhin_multipole_2011] and we will adopt this choice for the following examples where we set $\mathbf{r}_{c}$ as the center of the coordinate system. The total electric polarization $\mathcal{P}(\omega_{0})$ (respectively magnetic polarization $\mathcal{M}(\omega_{0})$) is obtained by adding the local electric polarizations Eq. (\[Pola\_elec\]) (respectively the magnetic polarizations Eq. (\[Pola\_mag\])) of all the elementary cells of the volume discretization. These polarizations are related to the super vector $\mathbf{F}_{0}(\mathbf{r}_{c},\omega)$ at the center of mass $\mathbf{r}_{c}$ of the nanostructure by the $(6\times6)$ [*super polarizability*]{} $\bm{\alpha}(\omega_{0})$: ![ Spectral variation of the imaginary part of the dipolar polarizability of a spherical gold particle of radius 5nm. (a) nano-sphere suspended in vacuum ($n_{1}$=1). Comparison of the analytical Clausius-Mossotti polarizability (blue) with the numerical calculation (red). (b) numerically calculated in-plane (red) and out-of-plane (green) polarizability tensor elements for a nano-sphere deposited on a silica substrate ($n_{2}$=1.48). Insets show sketches of the particle environment. For the numerical calculations, we discretized the spheres using 6337 identical mesh cells on a hexagonal compact grid. []{data-label="fig:sphere_polarizability"}](plot_particule_spherique_2){width="0.7\columnwidth"} $$\left( \begin{matrix} \mathcal{P}(\omega_{0}) \\[12pt] \mathcal{M}(\omega_{0}) \end{matrix}\right)= \overbrace{ \left( \begin{matrix} \alpha^{EE}(\omega_{0}) & \alpha^{EH}(\omega_{0}) \\[12pt] \alpha^{HE}(\omega_{0}) & \alpha^{HH}(\omega_{0}) \end{matrix}\right) }^{\bm{\alpha}(\omega_{0})} \cdot \overbrace{ \left( \begin{matrix} \mathbf{E}_{0}(\mathbf{r}_{c},\omega_{0}) \\[12pt] \mathbf{H}_{0}(\mathbf{r}_{c},\omega_{0}) \end{matrix}\right)}^{\mathbf{F}_0(\mathbf{r}_c, \omega_{0})}$$ where the polarizabilities $\alpha^{EE}(\omega_{0})$, $\alpha^{EH}(\omega_{0})$, $\alpha^{HE}(\omega_{0})$ and $\alpha^{HH}(\omega_{0}) $ are four (3$\times$3) dyadic tensors, defined by \[eq:different\_alphas\] $$\label{alpha_EE} \alpha^{EE}(\omega_{0})=\Delta v^{2}\chi_{e}(\omega_{0})\sum_{i,j}^{N}\mathbf{K}^{EE}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0})e^{{\text{i}}\mathbf{k}\cdot\mathbf{r}_{j}}$$ $$\alpha^{EH}(\omega_{0})=\Delta v^{2}\chi_{e}(\omega_{0})\sum_{i,j}^{N}\mathbf{K}^{EH}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0})e^{{\text{i}}\mathbf{k}\cdot\mathbf{r}_{j}}$$ $$\begin{aligned} \alpha^{HE}&(\omega_{0})=\Delta v^{2}\sum_{i,j}^{N}\Big\lbrace \chi_{m}(\omega_{0})\mathbf{K}^{HE}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0})\\[1pt] &-\dfrac{ik_{0}}{2}\chi_{e}(\omega_{0})\mathbf{r}_{i}\wedge\mathbf{K}^{EE}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0})\Big\rbrace e^{{\text{i}}\mathbf{k}\cdot\mathbf{r}_{j}}\\[2pt] \end{aligned}\label{eq:alpha_HE}$$ $$\begin{aligned} \alpha^{HH}&(\omega_{0})=\Delta v^{2}\sum_{i,j}^{N}\Big\lbrace \chi_{m}(\omega_{0})\mathbf{K}^{HH}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0})\\[1pt] &-\dfrac{ik_{0}}{2}\chi_{e}(\omega_{0})\mathbf{r}_{i}\wedge\mathbf{K}^{EH}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0})\Big\rbrace e^{{\text{i}}\mathbf{k}\cdot\mathbf{r}_{j}}\!. \end{aligned}$$ To be more precise, these are [*pseudo-polarizabilities*]{} since they depend on the direction of illumination due to the phase term $\exp({\text{i}}\mathbf{k}\cdot\mathbf{r}_{j})$. Conventional polarizabilities depend only on the geometry and the material of the nanostructure.[@arango_polarizability_2013; @bernal_arango_underpinning_2014] This phase term is the direct cause of the emergence of polarization vortices, which are responsible for the existence of magnetic multipole moments in dielectric nanostructures.[@kuznetsov_magnetic_2012; @kuznetsov_optically_2016] In order to be able to describe the magnetic polarization due to the mixed field susceptibility, we keep the phase term in the expression of the [*pseudo polarizabilities*]{}. We note that this approximation requires that the wave vector of the incident field is known already during the calculation of $\bm \alpha (\omega_0)$. However, we will show later, that a further approximation can be used to generalize these pseudo-polarizabilites to any oblique illumination without prior knowledge of the angle of incidence. For the calculation of the polarizabilities we used our own python implementation “pyGDM” of the volume discretization procedure described above.[@wiecha_pygdmpython_2018] {width="\textwidth"} Results ======= Electric-electric polarizability for structures of arbitrary shape ------------------------------------------------------------------ In a first step, we compare the spectral variation of the imaginary part of the dipolar polarizability ${\text{Im}}[\alpha^{EE}(\omega_{0})]$ at the example of an isolated spherical gold particle (radius $r=5\,$nm). Fig. \[fig:sphere\_polarizability\]a shows a comparison of the first diagonal term $\alpha_{xx}$, calculated analytically (Eq. (\[ALPHA1\]), blue line) or numerically (using Eq. (\[alpha\_EE\]), red line). For the sphere suspended in vacuum, the diagonal terms of $\alpha^{EE}$ are identical, and off-diagonal terms vanish. Our numerical discretization approach reproduces the well-known plasmon resonance for gold nano-particles around $\lambda_{0}=520$nm.[@amendola_surface_2017] The slight quantitative difference between the two representations is due to the inaccuracy of the analytical formula on non-atomic size scales. If we add a silica substrate in the calculation (see inset in Fig. \[fig:sphere\_polarizability\]b), the symmetry is reduced from spherical to a cylindrical. In consequence, the polarizability tensor is no longer diagonal and $\alpha^{EE}_{xx}=\alpha^{EE}_{yy} \neq \alpha^{EE}_{zz}$, which is depicted in Fig. \[fig:sphere\_polarizability\]b. The volume discretization allows us to treat nanostructures of arbitrary shape.[@girard_near_2005; @wiecha_pygdmpython_2018] Therefore, in a next step we study the evolution of the different terms of the [*electric-electric pseudo-polarizability*]{} tensor $\alpha^{EE}(\omega_{0})$, while gradually increasing the structure complexity, as illustrated in Fig. \[fig:complex\_structures\](a-d). Note that the polarizability tensor is symmetric (see Eqs.  and ), so in Fig. \[fig:complex\_structures\](e-l) we plot only the upper triangular elements. First, we calculate the spectral variation of the polarizability matrix of a gold pad of size $(50\text{nm}\times50\text{nm}\times25\text{nm})$, discretized with cubic cells of side length $d=2.5$nm (cf. Fig. \[fig:complex\_structures\]a). The real and imaginary part of each tensor component are shown in Fig. \[fig:complex\_structures\]e, respectively \[fig:complex\_structures\]i. Due to the symmetry of the structure the off diagonal terms of $\alpha^{EE}$ are zero (cyan lines). Moreover, we observe that $\alpha^{EE}_{xx}= \alpha^{EE}_{yy}$ (blue lines) which is a result of the rectangular footprint of the structure. Because the height is only half of the structure’s width, $\alpha^{EE}_{zz}$ is significantly smaller (green line). Despite the small dimensions of the pad, localized plasmon resonances arise slightly red-shifted at around $550 - 600\,$nm. Now if we increase the size of the pad along $Ox$ by a factor of two, the $\alpha^{EE}_{xx}$ and $\alpha^{EE}_{yy}$ terms are not equal anymore, due to the aspect ratio of the elongated pad. In this case, the resonance for excitation along the long edge is even more red-shifted to around $650$nm, which reflects the effective wavelength scaling of the localized plasmon resonance.[@novotny_effective_2007] Next, we calculate the polarizability tensor for a symmetric L-shaped gold structure (illustrated in Fig. \[fig:complex\_structures\]c). In this structure, coupling between the horizontal and the vertical arm leads to a non-zero off-diagonal term $\alpha^{EE}_{xy}$, as can be seen in Fig. \[fig:complex\_structures\]g and \[fig:complex\_structures\]k (magenta lines). Due to this off-diagonal term, two additional resonances emerge around 690nm and 1170nm at which polarization conversion between the $Y$-arm and the $X$-arm of the antenna occurs.[@kats_giant_2012; @wiecha_polarization_2017] The two peaks at $690\,$nm and $1170\,$nm correspond to the anti-bonding, respectively bonding modes between the two arms.[@panaro_dark_2014; @black_optimal_2014] We note, that the opposite phase of the bonding and the anti-bonding mode is correctly reflected also in the spectrum of the polarizability off-diagonal element. Polarization conversion is only occurring between $X$ and $Y$, hence the other off-diagonal elements remain zero (cyan lines). Moreover, both arms are of the same length which leads to $\alpha^{EE}_{xx}= \alpha^{EE}_{yy}$ (blue lines). Finally, we construct a three-dimensional structure which introduces interactions between each Cartesian direction, as depicted in figure \[fig:complex\_structures\]d. In this case, the each matrix element shows a unique spectral behavior, representing the complex interaction mechanisms between the antenna arms in different directions (Fig. \[fig:complex\_structures\]h and \[fig:complex\_structures\]l). We note at this point, that the approach is also capable to deal with nano-cavities carved into a bulk medium, by using a non-unitary permittivity for the environment and $\chi_e=\chi_m=0$ in the hollow source region. ![ Dielectric nanosphere ($n=4$) of radius $r=100\,$nm in vacuum, illuminated by a plane wave of linear polarization. (a) Extinction spectrum (blue line) and electric (ED) as well as magnetic dipole (MD) contributions to the extinction (orange, respectively green line), calculated as described in Ref. . (b) Electric dipole extinction via the polarizability tensor $\alpha^{EE}$ with (solid orange line) and without the phase term (dashed orange line). The ED extinction from (b) is shown as dashed black line for comparison. (c) Magnetic dipole extinction via the polarizability tensor $\alpha^{HE}$ with (solid green line) and without the phase term (dashed green line). The MD extinction from (b) is shown as dashed black line for comparison. At the top, the internal electric field distribution (real parts) is shown at the ED (left) and MD resonance (right). []{data-label="fig:alpha_EH"}](fig_ED_MD){width="\columnwidth"} {width="\textwidth"} Magnetic-electric polarizability of a dielectric sphere ------------------------------------------------------- We now want to assess the role of the magnetic terms in the super polarizability. Since in nature no material with a significant direct magnetic optical response is known, we will assume $\chi_m = 0$, hence the magnetic field of light cannot directly interact with the nanostructure. In consequence the polarizability tensors Eqs. (\[eq:different\_alphas\]) drastically simplify. The mixed terms involving $\mathbf{K}^{EH}$ and $\mathbf{K}^{HE}$ all disappear, since they include the product $\chi_e \chi_m$ (see also Eq. ). In fact only the two terms that depend on $\mathbf{K}^{EE}$ remain. Hence, for media with $\chi_m=0$, the electric polarization is fully described by $\alpha^{EE}$ and the magnetic polarization is entirely governed by $\alpha^{HE}$. As an example we show in figure \[fig:alpha\_EH\]a the extinction cross section of a dielectric nano-sphere ($n=4$) of radius $r=100\,$nm in vacuum, calculated from the discretized electric polarization density (blue line).[@girard_near_2005] We show additionally the decomposition of the extinction into an effective electric and magnetic dipole moment at the sphere’s center of mass (orange, respectively green lines).[@evlyukhin_multipole_2011; @terekhov_multipolar_2017] The dielectric sphere has an electric dipole (ED) resonance at $600\,$nm and a magnetic dipole (MD) resonance at $790\,$nm, which are indicated by black vertical dashed lines. The real part of the electric field inside the nano-sphere at these resonances is qualitatively shown in 3D vector plots above figure \[fig:alpha\_EH\]a. In figure \[fig:alpha\_EH\]b and \[fig:alpha\_EH\]c we show the extinction cross section obtained from the effective polarizabilities $\alpha^{EE}$, respectively $\alpha^{HE}$. We compare the “static” effective polarizabilities without phase term (dashed colored lines) and the above introduced pseudo-polarizabilities including the phase term $\exp(i\,\mathbf{k}\cdot \mathbf{r}_j)$ (solid colored lines, see Eqs. ). The dotted black lines show the ED and MD response from the full internal fields. While the ED resonance in figure \[fig:alpha\_EH\]b is very well reproduced by both, the static and the phase-sensitive electric-electric pseudo-polarizability, the MD resonance cannot be reproduced if the phase term in Eq.  is omitted (dashed green line in figure \[fig:alpha\_EH\]c). Only if the phase term is taken into account, the extinction calculated from the pseudo-polarizability matches the magnetic dipole resonance in the dielectric sphere (solid green line in figure \[fig:alpha\_EH\]c). This is because the magnetic dipole is induced by the vortex formed by the electric displacement current (see illustration of the MD above Fig. \[fig:alpha\_EH\]a, right), which is a direct consequence of the phase difference of the incident field across the relatively large nano-sphere. Approximation of $\alpha^{HE}$ for arbitrary angles of incidence ---------------------------------------------------------------- In contrast to “classical” static polarizabilities,[@sersic_magnetoelectric_2011; @arango_polarizability_2013] the here introduced *pseudo polarizabilities* depend on the illumination wave-vector $\mathbf{k}_0$ as a result of the above discussed phase term. In consequence, to solve the general problem, the pseudo polarizability needs to be separately calculated for every incident field which limits the usefulness of the approximation. However, we can approximate arbitrary incident angles through a first order expansion of the phase term. While we keep the phase-term in the definition of the polarizabilities, we assume that the first order term of its Taylor expansion is sufficient to describe the magnetic dipolar response. Thus, while allowing retardation effects to a certain extent, we still stick with the assumption that the wavelength is large with respect to the nanostructure (i.e. $\lambda_0 \gg |\mathbf{r}|$). Since the optical interaction is still modelled as a point-response, the wave vector of the illumination is assumed to be constant across the nanostructure. Furthermore, the approximation requires that the location of the effective dipole is independent of the wave vector. We assume here that the effective electric and magnetic dipole moments $\mathcal{P}(\omega_{0})$, respectively $\mathcal{M}(\omega_{0})$ lie at the particle’s center of mass $\mathbf{r}_c$ for any angle of incidence and polarization of the illumination. Without loss of generality we now consider an incident wave vector in the $XZ$ plane, were we get: \[eq:polarizations\_sum\_wavevector\] $$\begin{split} \mathcal{P}\ \approx\ \Bigg( \left(\dfrac{k_x}{|\mathbf{k_0}|}\right)^2 \alpha^{EE}_{k_x} +\ \left(\dfrac{k_z}{|\mathbf{k_0}|}\right)^2 \alpha^{EE}_{k_z} \Bigg) \cdot \mathbf{E}_0 \end{split}$$ $$\begin{split} \mathcal{M}\ \approx\ \bigg( \dfrac{k_x}{|\mathbf{k_0}|} \alpha^{HE}_{k_x} +\ \dfrac{k_z}{|\mathbf{k_0}|} \alpha^{HE}_{k_z} \bigg) \cdot \mathbf{E}_0 \, . \end{split}$$ For a derivation of these approximations based on a first order expansion of the phase term $\exp(\text{i}\mathbf{k}\cdot\mathbf{r}_j)$ in Eqs. (\[eq:different\_alphas\]), see appendices \[appendix:alpha\_HE\] and \[appendix:alpha\_EE\]. The dependence on $\omega_0$ and $\mathbf{r}_c$ has been omitted for the sake of readability. $k_i$ is the wave vector component and $\alpha^{XX}_{k_i}$ the pseudo polarizability for the Cartesian direction $i\in \{x, z\}$. Both are evaluated at the position $\mathbf{r}_c$ of the effective dipole (here the center of mass). Using this superposition scheme, the response of the nanostructure to any oblique plane wave illumination is described by three “pseudo super-polariability tensors” $\bm{\alpha}_{k_i}$ (one for every Cartesian coordinate axis $i$). Once the approximations Eqs.  for the effective dipole moments are calculated, the extinction cross sections due to the induced electric and magnetic polarizations can be calculated as[@evlyukhin_multipole_2011] \[eq:extinction\_arbitrary\_incidence\] $$I_{\text{ext},\mathcal{P}}(\omega_0) = \frac{8 \pi^{2}}{\lambda_{0}} \ {\text{Im}}\Big( \mathbf{E}_0(\mathbf{r}_c, \omega_0)^* \cdot \mathcal{P}(\omega_0) \Big)$$ $$I_{\text{ext},\mathcal{M}}(\omega_0) = \frac{8 \pi^{2}}{\lambda_{0}} \ {\text{Im}}\Big( \mathbf{H}_0(\mathbf{r}_c, \omega_0)^* \cdot \mathcal{M}(\omega_0) \Big) \, ,$$ where the superscript asterisk ($^*$) indicates complex conjugation. In figure \[fig:interpolation\_k\] we show spectra of the extinction cross section of a dielectric spheroid (refractive index $n=4$) in vacuum, with a diameter of $D_1=250\,$nm along the $OX$-oriented long axis and two identical short axes with diameters ($D_2=120\,$nm), as illustrated at the left of figure \[fig:interpolation\_k\]. The extinction is shown for different incident angles for $s$-polarization (top row) and $p$-polarization (bottom row). Clearly, the pseudo polarizability superposition approximation (solid lines; $\alpha^{EE}$: orange, $\alpha^{HE}$: green) yields excellent agreement with the ED and MD decomposition of the extinction from full-field simulations (dashed orange and green lines, respectively). Once again, the static polarizability approximation breaks down in case of the magnetic dipole resonance $\alpha^{HE}_{\text{static}}$ (green dotted lines). In case of the electric dipole response, the static polarizability $\alpha^{EE}_{\text{static}}$ gives a reasonable approximation. However, if the incidence direction is along the long axis of the ellipsoid, phase effects start to play a non-negligible role, and significant deviations occur in the static polarizability approximation. Conclusion and perspectives =========================== In conclusion we introduced a mathematical scheme for a generalized description of light-matter interaction in nanostructures through both, optical electric and magnetic fields. We showed how the optical response of nanostructures can be approximated through a universal “super polarizability” tensor, which combines the optical response through electric and magnetic dipole moments. Using a volume discretization, the super polarizability can be numerically calculated for nano-structures of arbitrary shape and material. We demonstrated that our pseudo polarizability, which includes phase effects, is capable to faithfully describe also magnetic dipole resonances in dielectric nanostructures of important size, where a conventional, static point-response model is breaking down. In contrast to similar, computationally more complex multi-dipole methods,[@bertrand_global_2019] our approach of effective electric and magnetic polarizabilities is capable to capture the optical response of complex nanostructures in a single quantity, which strongly facilitates the further evaluation of the optical behavior, for instance under changing illumination conditions. We foresee that our framework can be used to calculate large assemblies of different and / or randomly positioned nanostructures. Our work paves the way to the development of powerful design methods for highly heterogeneous plasmonic and dielectric or hybrid metasurfaces. [**Acknowledgments**]{}: The authors thank G. Colas des Francs and A. Mlayah for fruitful discussions on the physics of small plasmonic particles. PRW acknowledges support by the German Research Foundation (DFG) through a research fellowship (WI 5261/1-1). This work was supported by the computing center CALMIP in Toulouse. All data supporting this study are openly available from the University of Southampton repository (DOI: 10.5258/SOTON/XXXXXXXXX). Appendix {#appendix .unnumbered} ======== Interpolation of magnetic-electric polarizabilities {#appendix:alpha_HE} --------------------------------------------------- For a non-magnetic nanostructure the electric-magnetic polarizability writes (see also main paper): $$\begin{aligned} \alpha^{HE} (\omega_{0}) = & \sum_{i,j}^{N} -\dfrac{ik_{0}}{2}\chi_{e} \Delta v^{2} \Big\lbrace \mathbf{r}_{i}\wedge\mathbf{K}^{EE}(\mathbf{r}_{i},\mathbf{r}_{j}) \Big\rbrace e^{{\text{i}}\mathbf{k}\cdot\mathbf{r}_{j}} \\ = & \sum_{i,j}^{N} A_{i,j}^{HE} e^{{\text{i}}\mathbf{k}\cdot\mathbf{r}_{j}} \, , \end{aligned}$$ where we neglected the dependence on $\omega_0$ for the sake of readability. Due to the phase term $\exp({\text{i}}\mathbf{k}\mathbf{r})$, the polarizability $\alpha^{HE}$ is dependent on the incident angle and writes for a wave vector $\mathbf{k}$ of arbitrary angle $\vartheta$: $$\alpha_{\vartheta}^{HE}=\sum_{i,j}^{N} A_{i,j}^{HE} e^{{\text{i}}\mathbf{k}\cdot\mathbf{r}_{j}} =\sum_{i,j}^{N} A_{i,j}^{HE} e^{{\text{i}}(k_x x_{j} + k_y y_{j} + k_z z_{j})}\, .$$ Now we consider three $\alpha^{HE}_{r_i}$ corresponding to plane wave incidence along each of the three Cartesian directions: $$\begin{aligned} \alpha_{x}^{HE} &= \sum_{i,j}^{N} A_{i,j}^{HE} e^{{\text{i}}k x_{j}}\\ \alpha_{y}^{HE} &= \sum_{i,j}^{N} A_{i,j}^{HE} e^{{\text{i}}k y_{j}}\\ \alpha_{z}^{HE} &= \sum_{i,j}^{N} A_{i,j}^{HE} e^{{\text{i}}k z_{j}}; \end{aligned}$$ where $k = \dfrac{2\pi n}{\lambda_0}$ and $n$ is the medium index. We now develop the sum of the polarizabilities for plane wave incidence along the Cartesian coordinate axis. We define also three parameters allowing to describe an arbitrary illumination direction : $$\label{eq:beta_definition} \beta_x = \dfrac{k_x}{ k}\, , \quad \beta_y = \dfrac{k_y}{ k}\, , \quad \beta_z = \dfrac{k_z}{ k}\, ,$$ In addition, we assumed that $k_x^2+k_y^2+k_z^2 = k^2$. We can now write $$\begin{gathered} \beta_x\alpha_x^{HE} + \beta_y\alpha_y^{HE} + \beta_z\alpha_z^{HE} = \\ \sum_{i,j}^{N} A_{i,j}^{HE} \Big[\beta_x e^{{\text{i}}k x_{j}} + \beta_y e^{{\text{i}}k y_{j}} + \beta_z e^{{\text{i}}k z_{j}}\Big]\, .\end{gathered}$$ Assuming that $\lambda_0 \gg |\mathbf{r}| $, we can approximate the exponentials by their first order Taylor series: $$\begin{aligned} & \beta_x\alpha_x^{HE} + \beta_y\alpha_y^{HE} + \beta_z\alpha_z^{HE} \approx \\ & \sum_{i,j}^{N} A_{i,j}^{HE}\Big[ \beta_x(1 + i k x_{j}) + \beta_y(1 + i k y_{j}) + \beta_z(1 + i k z_{j})\Big] = \\ & \sum_{i,j}^{N} A_{i,j}^{HE}\Big[\beta_x + \beta_y + \beta_z + {\text{i}}(\beta_x k x_{j} + \beta_y k y_{j} + \beta_z k z_{j})\Big]\, . \end{aligned}$$ By adding “$1-1$”, we can write $$\label{eq:development_aHE_with_constants} \begin{aligned} &\beta_x\alpha_x^{HE} + \beta_y\alpha_y^{HE} + \beta_z\alpha_z^{HE} \approx \\ &\sum_{i,j}^{N} A_{i,j}^{HE}[\beta_x + \beta_y + \beta_z - 1 + \underbrace{1 + i(k_x x_{j} + k_y y_{j} + k_z z_{j})}_{\approx\ e^{{\text{i}}(k_x x_{j} + k_y y_{j} + k_z z_{j})}}] \approx \\ & \sum_{i,j}^{N} A_{i,j}^{HE}[\beta_x + \beta_y + \beta_z - 1 + e^{{\text{i}}(k_x x_{j} + k_y y_{j} + k_z z_{j})}]\, . \end{aligned}$$ The constant terms in Eq.  are proportional to the static magnetic-electric polarizability, which, as we have shown in the main paper, is negligible compared to usual dipolar polarizabilities, since the vortices that generate the magnetic dipolar response in non-magnetic nanostructures cannot be described without the phase term $\exp({\text{i}}\mathbf{k}\mathbf{r})$: $$\label{eq:magnetic_0_pola_constant} \sum_{i,j}^{N} \Big( \text{const.} \times \, A_{i,j}^{HE} \Big) \approx 0\, .$$ Hence we find: $$\alpha_{\vartheta}^{HE} \approx \beta_x\alpha_x^{HE} + \beta_y\alpha_y^{HE} + \beta_z\alpha_z^{HE}$$ Interpolation of electric-electric polarizabilities {#appendix:alpha_EE} --------------------------------------------------- The electric-electric polarizability writes (see also main paper): $$\begin{aligned} \alpha^{EE}(\omega_{0}) = & \Delta v^{2}\chi_{e}(\omega_{0}) \sum_{i,j}^{N} \mathbf{K}^{EE}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0}) e^{i\mathbf{k}\cdot\mathbf{r}_{j}} \\ = & \sum_{i,j}^{N} A_{i,j}^{EE} e^{{\text{i}}\mathbf{k}\cdot\mathbf{r}_{j}} \, , \end{aligned}$$ where we neglected the dependence on $\omega_0$ for the sake of readability. Due to the phase term $\exp({\text{i}}\mathbf{k}\mathbf{r})$, the polarizability $\alpha^{EE}$ is dependent on the incident angle and writes for a wave vector $\mathbf{k}$ of arbitrary angle $\vartheta$: $$\label{eq:alpha_teta_EE_general_k} \alpha_{\vartheta}^{EE}=\sum_{i,j}^{N} A_{i,j}^{EE} e^{{\text{i}}\mathbf{k}\cdot\mathbf{r}_{j}} =\sum_{i,j}^{N} A_{i,j}^{EE} e^{{\text{i}}(k_x x_{j} + k_y y_{j} + k_z z_{j})}\, .$$ Now we consider three $\alpha^{EE}_{r_i}$ corresponding to plane wave incidence along each of the three Cartesian directions: $$\begin{aligned} \alpha_{x}^{EE} &= \sum_{i,j}^{N} A_{i,j}^{EE} e^{{\text{i}}k x_{j}}\\ \alpha_{y}^{EE} &= \sum_{i,j}^{N} A_{i,j}^{EE} e^{{\text{i}}k y_{j}}\\ \alpha_{z}^{EE} &= \sum_{i,j}^{N} A_{i,j}^{EE} e^{{\text{i}}k z_{j}} \end{aligned}$$ We use the same definition of $\beta_x$, $\beta_y$ and $\beta_z$ as in Eq. . We can now write: $$\begin{gathered} \beta_x^2\alpha_x^{EE} + \beta_y^2\alpha_y^{EE} + \beta_z^2\alpha_z^{EE} = \\ \sum_{i,j}^{N} A_{i,j}^{EE} \Big[\beta_x^2 e^{{\text{i}}k x_{j}} + \beta_y^2 e^{{\text{i}}k y_{j}} + \beta_z^2 e^{{\text{i}}k z_{j}}\Big]\, .\end{gathered}$$ Assuming that $\lambda_0 \gg |\mathbf{r}| $, we can approximate the exponentials by their first order Taylor series: $$\label{eq:apprixmate_sum_alphaEE} \begin{aligned} & \beta_x^2\alpha_x^{EE} + \beta_y^2\alpha_y^{EE} + \beta_z^2\alpha_z^{EE} \approx \\ & \sum_{i,j}^{N} A_{i,j}^{EE}\Big[ \beta_x^2(1 + i k x_{j}) + \beta_y^2(1 + i k y_{j}) + \beta_z^2(1 + i k z_{j})\Big] = \\ & \sum_{i,j}^{N} A_{i,j}^{EE}\Big[\underbrace{\beta_x^2 + \beta_y^2 + \beta_z^2}_{=1} + {\text{i}}(\beta_x^2 k x_{j} + \beta_y^2 k y_{j} + \beta_z^2 k z_{j})\Big] = \\ & \sum_{i,j}^{N} A_{i,j}^{EE}\Big[1 + {\text{i}}(\beta_x^2 k x_{j} + \beta_y^2 k y_{j} + \beta_z^2 k z_{j})\Big]\, . \end{aligned}$$ Now we subtract the first order Taylor expansion of equation  $$\alpha_{\vartheta}^{EE} \approx \sum_{i,j}^{N} A_{i,j}^{EE}\Big[1 + {\text{i}}(\beta_x k x_{j} + \beta_y k y_{j}+ \beta_z k z_{j})\Big]$$ from both sides of equation , which yields: $$\label{eq:final_difference_expression_aEE} \begin{aligned} & \beta_x^2\alpha_x^{EE} + \beta_y^2\alpha_y^{EE} + \beta_z^2\alpha_z^{EE} -\alpha_{\vartheta}^{EE} \approx \\ & \sum_{i,j}^{N} {\text{i}}A_{i,j}^{EE} \Big(\beta_x(\beta_x - 1) k x_{j} + \beta_y(\beta_y -1) k y_{j} + \beta_z(\beta_z-1) k z_{j}\Big) = \\ & \qquad \quad {\text{i}}\beta_x(\beta_x - 1) k \sum_{i,j}^{N} A_{i,j}^{EE} x_{j} \\ & \qquad + \quad {\text{i}}\beta_y(\beta_y - 1) k \sum_{i,j}^{N} A_{i,j}^{EE} y_{j} \\ & \qquad + \quad {\text{i}}\beta_z(\beta_z - 1) k \sum_{i,j}^{N} A_{i,j}^{EE} z_{j} \end{aligned}$$ To demonstrate that the expression on the right hand side in Eq.  is negligible, we use Eq. , which states that $$\begin{gathered} \sum_{i,j}^{N} \Big(A_{i,j}^{HE} \Big) = \\ - \Delta v^{2}\dfrac{ik_{0}}{2}\chi_{e}(\omega_{0})\sum_{i,j}^{N}\Big\lbrace\mathbf{r}_{i}\wedge\mathbf{K}^{EE}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0})\Big\rbrace \approx 0\, .\end{gathered}$$ Properly speaking, with “approximately zero” we mean that the term is negligible within the small particle approximation. Using now the symmetry of $\mathbf{K}^{EE}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0})$ and the antisymmetry of the cross product “$\mathbf{r}_{i}\, \wedge$”, we can anti-commute those two terms: $$\begin{aligned} \sum_{i,j}^{N} \Big(A_{i,j}^{HE} \Big) &= \Delta v^{2}\dfrac{ik_{0}}{2}\chi_{e}(\omega_{0})\sum_{i,j}^{N}\mathbf{K}^{EE}(\mathbf{r}_{i},\mathbf{r}_{j},\omega_{0})\cdot\Big\lbrace\mathbf{r}_{i}\, \wedge \Big\rbrace\\ &= \dfrac{ik_{0}}{2}\sum_{i,j}^{N} A_{i,j}^{EE}\cdot\Big\lbrace\mathbf{r}_{i}\, \wedge \Big\rbrace \approx 0 \end{aligned}$$ where $$\Big\lbrace\mathbf{r}_{i}\, \wedge\Big\rbrace = \begin{pmatrix} 0 & -z_i & y_i \\ z_i & 0 & -x_i \\ -y_i & x_i & 0 \\ \end{pmatrix} \, .$$ Using the symmetry of $\mathbf{K}^{EE}$ $\left(A_{i,j}^{EE} = A_{j,i}^{EE}\right)$, we get $$\label{eq:decomposition_approx_terms_Aij_xj_0} \begin{aligned} \dfrac{ik_{0}}{2}\sum_{i,j}^{N} A_{i,j}^{EE} x_{j} & \approx 0\\ \dfrac{ik_{0}}{2}\sum_{i,j}^{N }A_{i,j}^{EE} y_{j} & \approx 0\\ \dfrac{ik_{0}}{2}\sum_{i,j}^{N} A_{i,j}^{EE} z_{j} & \approx 0 \end{aligned}$$ Comparing Eq.  and Eq.  we find $$(\beta_x^2\alpha_x^{EE} + \beta_y^2\alpha_y^{EE} + \beta_z^2\alpha_z^{EE}) - \alpha_{\theta}^{EE} \approx 0 \, ,$$ hence $$\alpha_{\vartheta}^{EE} \approx \beta_x^2\alpha_x^{EE} + \beta_y^2\alpha_y^{EE} + \beta_z^2\alpha_z^{EE}\, .$$

ArXiv

--- abstract: 'We report on the observation of a radiation helicity sensitive photocurrent excited by terahertz (THz) radiation in dual-grating-gate (DGG) InAlAs/InGaAs/InAlAs/InP high electron mobility transistors (HEMT). For a circular polarization the current measured between source and drain contacts changes its sign with the inversion of the radiation helicity. For elliptically polarized radiation the total current is described by superposition of the Stokes parameters with different weights. Moreover, by variation of gate voltages applied to individual gratings the photocurrent can be defined either by the Stokes parameter defining the radiation helicity or those for linear polarization. We show that artificial non-centrosymmetric microperiodic structures with a two-dimensional electron system excited by THz radiation exhibit a $dc$ photocurrent caused by the combined action of a spatially periodic in-plane potential and spatially modulated light. The results provide a proof of principle for the application of DGG HEMT for all-electric detection of the radiation’s polarization state.' author: - 'P. Faltermeier,$^1$ P. Olbrich,$^1$ W. Probst,$^1$ L. Schell,$^1$ T. Watanabe$^2$, S. A. Boubanga-Tombet$^2$, T. Otsuji$^2$, and S. D.Ganichev$^{1}$' title: | Helicity sensitive terahertz radiation detection by\ dual-grating-gate high electron mobility transistors --- I. Introduction =============== Field-effect-transistors (FETs) have emerged as promising devices for sensitive and fast room temperature detection of terahertz (THz) radiation [@knaprev2013; @knaptredicuccirev2013]. They are considered as a good candidate for real-time THz imaging and spectroscopic analysis [@Boppel2012; @Muravev2012] as well as future THz wireless communications [@Tonouchi2007]. Devices employing plasmonic effects in FETs have already been applied for room temperature detection of radiation with frequencies from tens of GHz up to several THz and enable the combination of individual detectors in a matrix. They are characterized by high responsivity (up to a few kV/W), low noise equivalent power (down to 10 pW/$\sqrt{\rm Hz}$), fast response time (tens of picoseconds) and large dynamic range (linear power response up to 10 kW/cm$^2$), see e.g. Ref.[@knaptredicuccirev2013; @9; @14Roj; @11; @Watanabe; @But2014; @Knap2014]. The operation principle of FET THz detectors used so far is based on the nonlinear properties of the two-dimensional (2D) plasma in the transistor channel. The standard Dyakonov-Shur model [@Dyakonov] assumes that radiation is coupled to the transistor by an effective antenna, which generates an $ac$ voltage predominantly on one side of the transistor. Both resonant [@98] and non-resonant [@99] regimes of THz detection have been studied. While research aimed to development of THz FET detectors is focused on single gate structures recently several groups have shown that higher sensitivities are expected for structures with periodic symmetric and asymmetric metal stripes or gates [@Watanabe; @19w; @116; @20w; @16w; @18w; @31; @115; @35; @31Roj]. In particular, dual-grating-gate FET are considered as a good candidate for sensitive THz detection. The first data obtained on dual-gated-structures demonstrated a substantial enhancement of the photoelectric response and an ability to control detector parameters by variation of individual gate bias voltage [@Watanabe]. At the same time, THz electric field applied to FETs with asymmetric periodic dual gate structure is expected to give rise to electronic ratchet effects [@28; @30; @29; @31Roj; @33Roj] (for review see [@30]) and plasmonic ratchet effects [@36]. Besides improving the figure of merits of FET detectors, ratchet effects may also result in new functionalities. In particularly, they may induce photocurrents driven solely by the radiation helicity. Here, we report on the observation of a radiation helicity sensitive photocurrent excited by THz radiation in dual-grating-gate InAlAs/InGaAs/InAlAs/InP high electron mobility transistors (HEMT). We show that artificial non-centrosymmetric microperiodic structures with a two-dimensional electron system excited by THz radiation exhibit a $dc$ photocurrent caused by the lateral asymmetry of the applied static potential and terahertz electric field. We demonstrate that depending on gate voltages applied to the individual gratings of the dual-grating-gate the response can be proportional to either the Stokes parameters [@Saleh] defining the radiation helicity or those for linear polarization. As an important result, for a wide range of gate voltages we observed a photocurrent $j_{\rm C}$ being proportional to the radiation helicity $P_{\rm circ} = (I_{\sigma^+} - I_{\sigma^-})/(I_{\sigma^+} + I_{\sigma^-})$, where $I_{\sigma^+}$ and $I_{\sigma^-}$ are intensities of right- and left-handed circularly polarized light. For the circular photocurrent $j_{\rm C}$ measured between source and drain contacts changes its sign with the inversion of the radiation helicity. This observation is of particular importance for a basic understanding of plasmon-photogalvanic and quantum ratchet effects. It also has a large potential for the development of an all-electric detector of the radiation’s polarization state, which was so far realized applying less sensitive photogalvanic effects only [@122; @123; @124]. The observed phenomena is discussed in the framework of electronic ratchet [@35; @30; @29; @33Roj; @31Roj] and plasmonic ratchet effects excited in a 2D electron system with a spatially periodic *dc* in-plane potential [@Watanabe; @35; @36]. II. Experimental Technique ========================== The device structure is based on an InAlAs/InGaAs/InAlAs/InP high-electron mobility transistor (HEMT) and incorporates doubly inter-digitated grating gates (DGG) G$_1$ and G$_2$. A sketch and a photograph of the gates are shown in Fig. \[fig01\](a) and inset in Fig. \[fig01\](b). The 2D electron channel is formed in a quantum well (QW) at the heterointerface between a 16nm-thick undoped InGaAs composite channel layer and a 23nm-thick, Si-doped InGaAs carrier-supplying layer. The electron density of the 2DEG is about $3 \times 10^{12}$cm$^2$, electron effective mass normalized on free-electron mass $m_0$ and room temperature mobility are $m/m_0\,=\,0.04$ and $\mu_0$=11000cm$^2$/(Vs), respectively. The DGG gate is formed with 65nm-thick Ti/Au/Ti by a standard lift-off process. The footprint of the narrower gate fingers G$_1$ was defined by an E-beam lithography, whereas that of the wider gate fingers G$_2$ was defined by a photolithography. In all studied structures, the metal fingers of the grating gates G$_{1}$ and G$_{2}$ have the same length, being $d_{\rm G1}$=200nm and $d_{\rm G2}$=800nm. The spacing between narrow and wide DGG fingers is asymmetric with $a_{\rm G1}$=200nm and $a_{\rm G2}$=400nm, see Fig. \[fig01\]. The size of the active area, covered with the grating is about 20$\mu$m$\times$20$\mu$m. Ohmic contacts, forming source and drain of HEMTs, were fabricated by highly doped 15nm thick InAlAs and InGaAs layers. The axis along the gate’s fingers is denoted as $x$ and that along source and drain as $y$. The characteristic source/drain current - gate voltage dependence obtained by transport measurement is shown for sample\#A in Fig. \[fig01\](b). All experiments are performed at room temperature. The HEMT structures were illuminated with polarized THz and microwave (MW) radiation at normal incidence. For optical excitation we used low power $cw$ optically pumped CH$_3$OH THz laser [@3aa; @karch2010] and Gunn diodes providing monochromatic radiation with frequencies $f$=2.54THz and 95.5GHz, respectively. The radiation peak power $P$, being of the order of several milliwats at the sample’s position, has been controlled by pyroelectric detectors and focused onto samples by parabolic mirrors (THz laser) or horn antenna (Gunn diode). The spatial beam distribution of THz radiation had an almost Gaussian profile, checked with a pyroelectric camera [@edge; @Glazov2014]. THz laser radiation peak intensity, $I$, for laser spot being of about 1.2 mm diameter on the sample, was $I \approx 8$ W/cm$^2$. The profile of the microwave radiation and, in particular, the efficiency of the radiation coupling to the sample couldn’t be determined with satisfactory accuracy. Thus, all microwave data are given in arbitrary units. The polarization state of THz radiation has been varied applying crystal quartz $\lambda$/4- or $\lambda$/2-plates [@book]. To obtain circular and elliptically polarized light the quarter-wave plate was rotated by the angle, $\varphi$, between the initial polarization plane and the optical axis of the plate. The radiation polarization states for several angles $\varphi$ are illustrated on top of Fig. \[fig02\]. Orientation of the linearly polarized radiation is defined by the azimuth angle $\alpha$, with $\alpha\,=\,\varphi\,=\,0$ chosen in such a way that the electric field of incident linearly polarized light is directed along $x$-direction. Different orientation of linearly polarized MW radiation were obtained by rotation of a metal wire grid polarizer. The photocurrent excited between source and drain is measured across a 50 $\Omega$ load resistor applying the standard lock-in technique. ![(a) Sketch of the dual-grating-gate HEMT. Cross-section of the structure shows the layer sequence and indicates the width of the fingers ($d\rm_{1/2}$) and the fingers spacings ($a\rm_{1/2}$). THz radiation at 2.54 THz is applied at normal incidence. (b) Drain-to-source current as a function of the gate voltage $U_{\rm G1}$ measured at $U_{\rm G2} =0$ V. Inset shows the photograph of the structure. Here G$\rm_{1}$/G$\rm_{2}$, S and D denote first/second gate, source and drain, respectively. Part of G$\rm_{1}$/G$\rm_{2}$ structure is highlighted by yellow lines for visualization. []{data-label="fig01"}](fig01.eps){width="0.80\linewidth"} ![THz radiation induced normalized photocurrent $j_y /I$ as a function of the angle $\varphi$ defining the radiation helicity. The current is measured for different voltages applied to the first and second gates. (a) shows the data for [$U_{\rm G1}\,=\,-1.06$V]{} at gate 1 and zero gate voltage at gate 2. (b) shows the photocurrent measured for zero gate voltage at gate 1 and $U_{\rm G2}\,=\,-0.9$V. Full lines show fits to the total current calculated after Eq. (\[FITphi\]). The ellipses on top illustrate the polarization states for various $\varphi$. Insets show amplitudes of photocurrent contributions $j_{\rm C}/I$, driven by the light helicity, and $j_1/I$ ($j_2/I$), induced by linear polarization, as a function of the gate voltages $U_{\rm G1}$ or $U_{\rm G2}$. Second set of the insets schematically show corresponding gate potentials. Dashed lines are guide for the eye indicating the potential asymmetry in $y$-direction. Note that presence of the metal gates results in a nonzero potential even for $U_{\rm \rm G}\, =\, 0$. []{data-label="fig02"}](fig02.eps){width="0.85\linewidth"} ![(a) THz radiation induced normalized photocurrent $j_y /I$ as a function of the angle $\varphi$ defining the radiation helicity. The current is measured for comparable voltages applied to the first ($U_{\rm G2}\,=\,-1.1$V) and the second ($U_{\rm G2}\,=\,-0.92$V) gates. Full line shows fit to the total current calculated after Eq. (\[FITphi\]). The ellipses on top illustrate the polarization states for various $\varphi$. Right inset shows amplitudes of photocurrent contributions $j_{\rm C}/I$, driven by the light helicity, and $j_1/I$ ($j_2/I$), induced by linear polarization, as a function of the gate voltage $U_{\rm G1}$ or $U_{\rm G2}$. Upper inset schematically shows corresponding gate potentials. Dashed lines are guide for the eye indicating the potential asymmetry in $y$-direction. (b) shows amplitudes of the photocurrent contributions $j_{\rm C}/I$, driven by the light helicity, as a function of the gate voltage $U_{\rm G1}$ ($U_{\rm G2}$=0) measured for three different structures \#A, \#B, and \#C. The inset shows photovoltage measured in sample \#D across 50 $\Omega$ load resistance ($R_L \ll R_s$) and directly from the sample over the lock-in amplifiers input resistance being much larger than the sample resistance $R_s$. Note that the former signal is multiplied by factor 25. []{data-label="fig03"}](fig03.eps){width="0.85\linewidth"} ![(a) THz radiation induced normalized photocurrent $j_y/I$ excited by linearly polarized THz radiation in samples \#A and \#B as a function of the gate voltage $U_{\rm G1}$. The current is shown for $U_{\rm G2}\,=$0 and several in-plane orientations of the radiation electric field in respect to source-drain line defined by azimuth angles $\alpha$. Inset shows dependence of $j_y$ on the angle $\alpha$ obtained for $U_{\rm G1}\,=\,-1.08$ V and $U_{\rm G2}\,=\,0$. Full line shows fit to the total current calculated after Eq. (\[SCalpha\]). Arrows indicate electric field orientation for several angles $\alpha$. (b) Photocurrent $j_y/I$ excited by linearly polarized microwave radiation ($f =95.5$GHz) in samples \#A and \#B as a function of the gate voltage $U_{\rm G1}$ ($U_{\rm G2}= 0$). Inset shows dependence of $j_y/I$ on the azimuth angle $\alpha$ obtained in sample \#B for $U_{\rm G1}\,=\,-1.14$ V and $U_{\rm G2}\,=\,0$. Full line shows fit after $j_y \propto \cos^2(\alpha + \theta)$ with the phase angle $\theta$. []{data-label="fig04"}](fig04.eps){width="0.85\linewidth"} III. Photocurrent experiment ============================ Illuminating the structure with elliptically (circular) polarized radiation of terahertz laser operating at frequency $f$=2.54THz we observed a $dc$ current strongly depending on the radiation polarization. Figure \[fig02\](a) shows the photocurrent as a function of the phase angle $\varphi$ defining the radiation polarization state. The data are obtained for zero gate voltage at the gate2, $U_{\rm G2}$=0 and $U_{\rm G1}$=-1.06V. The principal observation is that for right- ($\sigma^+$) and left-handed ($\sigma^-$) polarizations, i.e., for $\varphi=45^\circ$ and $135^\circ$, the signs of the photocurrent $j_y$ are opposite. The overall dependence $j_y(\varphi)$ is well described by $$\label{FITphi} j_y(\varphi) = j_0 s_0 + j_1 s_1 (\varphi) + j_2 s_2(\varphi) + j_{\rm C} s_3(\varphi)\:,$$ and corresponds to the superposition of the Stokes parameters with different weights given by the coefficients $j_0$, $j_1$, $j_2$, and $j_{\rm C}$, which in the experimental geometry applying rotation of quarter-wave plate the Stokes parameters change after $$\begin{aligned} \label{S12} &&s_0 \equiv |E_x|^2+|E_y|^2, \\ &&s_1 \equiv |E_x|^2-|E_y|^2 = \frac{\cos{4\varphi} + 1}{2}, \\ &&s_2 \equiv E_xE_y^*+E_x^*E_y = \frac{\sin{4\varphi}}{2},\\ &&s_3 \equiv i(E_xE_y^*-E_x^*E_y) = -P_{\rm circ} = -\sin{2\varphi}\,\,\, , $$ Here $s_0$ determines the radiation intensity, $s_1$ and $s_2$ define the linear polarization of radiation in the $(xy)$ and rotated by $45^{\circ}$ coordinate frames, and $s_3$ describes the degree of circular polarization or helicity of radiation. Consequently individual photocurrent contributions in Eq. (\[FITphi\]) are induced by unpolarized, linearly or circularly polarized light components. While the polarization dependence given by Eq. (\[FITphi\]) has been detected for arbitrary relations between voltages applied to the first and second gates, the magnitude and even the sign of the individual contributions can be controlled by the gate voltages. The inset in Fig. \[fig02\](a) shows a gate dependence of the polarization dependent contributions to the total photocurrent [@footnote]. The dependence on the gate voltage $U_{\rm G1}$ is obtained for zero biased second gate. Photocurrent measured in the close circuit configuration with $R_L \ll R_s$ shows a maximum amplitude for $U_{\rm G1} = -1.1$V. For open circuit configuration the measured photovoltage increases at larger negative bias voltages and achieves maximum at the threshold voltage, $U_{\rm th} = -1.3$V. Corresponding data will be presented and discussed below. While the non-monotonic behavior of the signal for gate voltage variation is well known for FET detectors [@knaprev2013; @knaptredicuccirev2013; @30w] the signal sign inversion upon a change of the radiation polarization, see Fig. \[fig02\](a), is generally not expected for standard Dyakonov-Shur FET detectors indicating crucial role of the lateral superlattice in the photocurrent generation. To demonstrate that the observed effect indeed stems from the lateral asymmetry of the periodic potential we interchanged the voltages applied to the gates. Figure \[fig02\](b) shows the results obtained for zero gate voltage at the first gate and $U_{\rm G2}= -0.9$V at the second one. The figure reveals that changing the sign of the lateral potential asymmetry, see insets of Fig. \[fig02\](a) and (b), results in the sign inversion of all contributions besides the polarization independent offset. The situation holds for almost all values of $U_{\rm G2}$, see the insets in Fig. \[fig02\](a) and (b). Significantly, the proper choice of the relation between amplitudes of the individual gate potentials allows one to suppress completely one or the other photocurrent contribution. Figure \[fig03\](a) demonstrates that for close values of gate voltages the circular photocurrent vanishes (corresponding potential profile for $U_{\rm G1}$= -1.1V and $U_{\rm G2}$= -0.9V is shown in the inset in Fig. \[fig03\]). The interplay of the contributions upon variation of $U_{\rm G1}$ and for fixed $U_{\rm G2}$= -1.1V is shown in the inset in Fig. \[fig03\](a). It is seen that for nonzero second gate voltage the circular, $j_c$, and linear, $j_2$, photocurrent contributions change their direction with increasing $U_{\rm G1}$. Moreover, the inversions take place at different $U_{\rm G1}$ voltages. This fact can be used to switch on and off the circular photocurrent $j_C \propto P_{\rm circ}$ contribution. To support the conclusion that $j_{1}$ and $j_{2}$ photocurrent contribution are caused by the linear polarized light component we carried out additional measurements applying linearly polarized light. The gate dependence of the normalized photocurrent $j_{y}/I$ measured for samples \#A and \#B for several azimuth angles $\alpha$ are shown in Fig. \[fig04\](a). The inset in this figure presents the dependence of $j_{y}/I$ on the electric field orientation. The polarization dependence is well described by the Eq. (\[FITphi\]) taking into account that for linearly polarized light the last term vanishes and the Stokes parameters are given by $$\begin{aligned} \label{SCalpha} s_1(\alpha) = \cos{2 \alpha}\:, \quad s_2(\alpha) = \sin{2 \alpha} \:. \nonumber \end{aligned}$$ Here $\alpha=2\beta$ defines the orientation of the polarization plane and $\beta$ is the angle between the initial polarization plane and the optical axis of the half-wave plate. The magnitudes and signs of the coefficients $j_0$, $j_1$, and $j_2$ used for the fit coincide with that applied for fitting of $\varphi$-dependencies obtained at the same gate voltages. These results demonstrate that photocurrents $j_{1}$ and $j_{2}$ measured in set-up applying quarter-wave plate are indeed controlled by the degree of linear polarization of elliptically polarized radiation. The polarization sensitive photocurrent has been observed in all studied devices of similar design and arbitrary relation between second and first gate potentials. The photocurrent can always be well described by Eq. (\[FITphi\]). Figure \[fig03\](b) summarizes the data on the helicity driven photocurrent $j_{\rm C}/I$ detected in three HEMT structures upon change of $U_{\rm \rm G1}$ and for $U_{\rm \rm G2} = 0$. In all samples we detected similar dependencies of the photocurrent characterized by close maximum positions but different signal magnitudes. The data of Fig. \[fig03\](b) as well as circles in its inset are obtained in the close circuit configuration applying 50 $\Omega$ load resistance. The non-monotonic behavior of the photosignal measured in this geometry is caused by the interplay of the potential asymmetry, increasing with raising second gate voltage, and raising of the sample resistance for large gate voltages. For the open circuit geometry (signal is fed to the high input impedance of lock-in amplifier) the maximum of the signal is detected for gate voltages being equal to the threshold voltage, $U_{\rm th}$, see squares in the inset in Fig. \[fig03\](b). Following Ref. [@Watanabe] we estimate from the voltages measured in open circuit geometry the voltage responsivities for the signals corresponding to the photocurrents $j_2$ and $j_C$ as $R_v = U_s / P \times S / S_t \approx 0.3$ V/W and 0.15 V/W, respectively. Here $P$ the total power of the source at the detector plane, $S$ radiation beam spot area, and $S_t = 20 \times 20$ $\mu$m$^2$ transistor area. The voltage responsivities, being rather low as compared to that typically obtained for plasmonic FET detectors, indicates the necessity of further optimization of the structure design. Finally, we note that measurements applying microwave radiation show that for lower frequencies the polarization behavior changes qualitatively. Instead of the sign-alternating dependencies discussed above the signal now varies after $j_y \propto \cos^2 (\alpha + \theta)$, see inset in Fig. \[fig04\](b). This observation is in a good agreement with the Dyakonov-Shur theory [@Dyakonov] and was reported for many conventional plasmonic FET detectors, see e.g. [@knaprev2013; @knaptredicuccirev2013]. The gate voltage dependence of the response shown in Fig. \[fig04\] also reproduces well the results previously obtained for similar structures [@Watanabe; @Coquillat2014]. Even the fact that the maximum of the signal in various structures has been obtained for different directions of the electric field vector in respect to $y$-direction (source-drain) has already been reported for these transistors and attributed to the antenna coupling of MW radiation to transistor, see Ref. [@Coquillat2014]. IV. Discussion ============== The observation of the circular photocurrent and the sign-alternating linear photocurrent $j_2$ reveals that a microscopic process actuating these photocurrents goes beyond the plasmonic Dyakonov-Shur model typically applied to discuss operation of FETs THz detectors. Indeed, as addressed above, the latter implies an oscillating electric field along source-drain direction ($y$-direction) yielding sign conserving variation upon rotation of polarization plane, $j_y \propto \cos^2 \alpha $ [@footnote2]. As recently shown in Ref. [@117; @118], the Dyakonov-Shur model in fact may result in the circular photocurrent but only due to interference effects of two different channels and two interacting antennas in small size special design FETs - the model which can hardly be applied to the large DGG samples used in our experiments. At the same time, the observed polarization behavior is characteristic for the electronic ratchet effects excited in asymmetric periodic structures [@28; @30; @29; @33Roj] and linear/circular plasmonic ratchet effects [@35; @36]. The ratchet currents arise due to the phase shift between the periodic potential and the periodic light electric field resulting from near field diffraction in a system with broken symmetry. Microscopic theory developed in Ref. [@29] shows that the helicity dependent photocurrent appear because the carriers in the laterally modulated quantum wells move in two directions and are subjected to the action of the two-component electric field. Symmetry analysis of the photocurrent shows that in our DDG structures described by C$_1$ point group symmetry [@footnote3] it varies with radiation polarization after Eq. (\[FITphi\]), being in agreement with experimental observation shown in Figs. \[fig02\], \[fig03\] and \[fig04\](a). Moreover, as the ratchet photocurrents are proportional to the degree of the in-plane asymmetry, they reverse the sign upon inversion of static potential asymmetry. Exactly this behavior has been observed in experiment, see Fig. \[fig02\] (a) and (b). The proportionality to the degree of lateral asymmetry also explain the increase of the signal with raising voltage applied to one gate at constant voltage by the other. The interplay of the degree of lateral asymmetry and periodic modulation of THz electric field results in the complex gate-voltage dependence, in particular, for $U_{\rm G1} \approx U_{\rm G2}$. As the different individual contributions to the total current effect might imply different microscopic mechanisms of the photocurrent formation, their behavior upon change of external parameters can distinct from each other. This would result in a sign-alternating gate-voltage behavior, in particular for the range of comparable $U_{\rm G1}$ and $U_{\rm G2}$, like it is observed in experiment, see Fig. \[fig02\] (c). While all qualitative features of the observed phenomena can be rather good described in terms of ratchet effects we would like to address another possible effect, which might trigger the helicity-driven photocurrent. It could be the differential plasmonic drag effect in the two-dimensional structure with an asymmetric double-grating gate considered in Refs. [@35; @Popov2015]. As shown in Ref. [@35] for a periodic AlAs/InGaAs/InAlAs/InP structure and linearly polarized THz radiation, photon drag effect can be comparable in strength with the plasmonic ratchet effect at THz frequencies. As the circular photon drag effect has been observed in different low dimensional materials [@karch2010; @Shalygin2006] we can expect that modification of the theory developed in [@35] can also yield helicity driven plasmonic drag current compatible with the ratchet one. Finally, we note that the ratchet effects (either electronic or plasmonic) can be greatly increased due to the resonant enhancement of the near-field in two-dimensional electron system at the plasmon resonance excitation as it was shown for the plasmonic ratchet in Refs. [@31Roj; @36]. The resonant plasmon condition $\omega \tau >1$, see Ref. [@Dyakonov] can be well satisfied in our structure ($\omega \tau =4$ at 2.54 THz). As shown in Ref. [@31Roj], the fundamental plasmon resonance is excited in a similar structure at frequency around 2 THz. Therefore, the plasmon resonance excitation can contribute to the observed ratchet effects independently of particular microscopic mechanisms of the ratchet photocurrent formation. The measurements in a broader THz frequency range could elucidate the role of the plasmonic resonance excitation in the ratchet photocurrent enhancement. V. Summary ========== To summarize, our measurements demonstrate that dual-grating-gate InAlAs/InGaAs/InAlAs/InP excited by terahertz radiation can yield a helicity sensitive photocurrent response at THz frequencies. We show, that HEMTs with asymmetric lateral superlattice of gate fingers with unequal widths and spacing can be applied for generation of a photocurrent defined by linearly and circularly radiation polarization components. Moreover, one can obtain photoresponse being proportional to one of the Stokes parameters simply by variation of voltages applied to the individual gates. The photocurrent formations can be well described in terms of ratchet effects excited by terahertz radiation. By that the lateral grating induces a periodical lateral potential acting on the 2D electron gas in QW. This grating also modulates the incident radiation in the near field and hence in the plane of the 2DES, resulting in circular, linear and polarization-independent ratchet effects. While the responsivity of the polarization dependent response is lower than that reported for FET transistors it can be substantially improved by optimization of the structure design leading the resonant enhancement of the ratchet effects the plasmon resonance excitation. We thank V. Popov for helpful discussions. The financial support from the DFG (SFB 689) is gratefully acknowledged. [99]{} W. Knap and M. Dyakonov, *Plasma wave THz detectors and emitters* in Handbook of Terahertz Technology edited by D. Saeedkia (Woodhead Publishing, Waterloo, Canada, 2013), pp. 121-155. W. Knap, , S. Rumyantsev, M. S Pea Vitiello, D. Coquillat, S. Blin, N. Dyakonova, M. Shur, F. Teppe, A. Tredicucci, T. Nagatsuma, Nanotechnology **24**, 214002 (2013). S. Boppel, A. Lisauskas, A. Max, V. Krozer, and H. G. Roskos, Opt. Lett. **37**, 536 (2012). V. M. Muravev and I. V. Kukushkin, Appl. Phys. Lett. **100**, 082102 (2012). M. Tonouchi, Nature Photon. **1**, 97 (2007). F. Schuster, D. Coquillat, H. Videlier, M. Sakowicz, F. Teppe, L. Dussopt, B. Giffard, T. Skotnicki, and W. Knap, Opt. Express **19**, 7827 (2011). G. C. Dyer, S. Preu, G. R. Aizin, J. Mikalopas, A. D. Grine, J. L. Reno, J. M. Hensley, N. Q. Vinh, A. C. Gossard, M. S. Sherwin, S. J. Allen, and E. A. Shaner, Appl. Phys. Lett. **100**, 083506 (2012). S. Preu, M. Mittendorff, S. Winnerl, H. Lu, A. C. Gossard, and H. B. Weber, Opt. Express **21**, 17941 (2013). T. Watanabe, S. A. Boubanga-Tombet, Y. Tanimoto, D. Fateev, V. Popov, D. Coquillat, W. Knap, Y. M. Meziani, Y. Wang, H. Minamide, H. Ito, and T. Otsuji, IEEE Sensors Journal **13**, 89 (2013). D. B. But, C. Drexler, M. V. Sakhno, N. Dyakonova , O. Drachenko, F. F. Sizov, A. Gutin, S. D. Ganichev, W. Knap, J. Appl. Phys. **115**, 164514 (2014). W. Knap, D. B. But, N. Dyakonova, D. Coquillat, A. Gutin, O. Klimenko, S. Blin, F. Teppe, M.S. Shur, T. Nagatsuma, S.D. Ganichev, and T. Otsuji, *Recent Results on Broadband Nanotransistor Based THz Detectors* in NATO Science for Peace and Security Series B, Physics and Biophysics: THz and Security Applications, edited by C. Corsi, F. Sizov, (Springer, Dordrecht, Netherlands, 2014) pp.189 - 210. M. Dyakonov and M. S. Shur, IEEE-Trans-ED **43**(3), 380 (1996). W. Knap, Y. Deng, S. Rumyantsev, J.-Q. Lu, M. S. Shur, C. A. Saylor and L. C. Brunel, Appl. Phys. Lett. **80**, 3433 (2002). W. Knap, V. Kachorovskii, Y. Deng, S. Rumyantsev, J.-Q. Lu, R. Gaska, M. S. Shur, G. Simin, X. Hu, M. Asif Khan, C. A. Saylor, and L. C. Brunel, J. Appl. Phys. **91**, 9346 (2002). T. Otsuji, M. Hanabe, T. Nishimura, and E. Sano, Opt. Exp. **14**, 4815 (2006). S. Sassine, Yu. Krupko, J.-C. Portal, Z. D. Kvon, R. Murali, K. P. Martin, G. Hill, and A. D. Wieck, Phys. Rev. B **78**, 045431 (2008). D. Coquillat , S. Nadar, F. Teppe, N. Dyakonova, S. Boubanga-Tombet, W. Knap, T. Nishimura, T. Otsuji, Y. M. Meziani, G. M. Tsymbalov, and V. V. Popov, Opt. Exp. **18**, 6024 (2010). V. V. Popov, J. Infr. Millim. THz Waves **32**, 1178 (2011). G. C. Dyer, G. R. Aizin, J. L. Reno, E. A. Shaner, and S. J. Allen, IEEE J. Sel. Topics Quantum Electron. **17**, 85 (2011). V. V. Popov, D. V. Fateev, T. Otsuji, Y. M. Meziani, D. Coquillat, and W. Knap, Appl. Phys. Lett. **99**, 243504 (2011). E. S. Kannan, I. Bisotto, J.-C. Portal, T. J. Beck, and L. Jalabert, Appl. Phys. Lett. **101**, 143504 (2012). V. V. Popov, Appl. Phys. Lett. **102**, 253504 (2013). V.V. Popov, D.V. Fateev, T. Otsuji, Y.M. Meziani, D. Coquillat, and W. Knap, Appl. Phys. Lett. **99**, 243504 (2011). P. Olbrich, E.L. Ivchenko, T. Feil, R. Ravash, S.D. Danilov, J.Allerdings, D. Weiss, and S. D. Ganichev, Phys. Rev. Lett **103**, 090603 (2009). E. L. Ivchenko and S. D. Ganichev, JETP Lett. **93**, 752 (2011). P. Olbrich, J. Karch, E. L. Ivchenko, J. Kamann, B. Maerz, M. Fehrenbacher, D. Weiss, and S. D. Ganichev, Phys. Rev. B **83**, 165320 (2011). A. V. Nalitov, L. E. Golub, E. L. Ivchenko, Phys. Rev. B **86**, 115301 (2012). I. V. Rozhansky, V. Yu. Kachorovskii, and M. S. Shur, Phys. Rev. Lett. **114**, 246601 (2015). B. E. A. Saleh, M. C. Teich, *Fundamentals of Photonics* (John Wiley & Sons, New York, 2003). S. N. Danilov, B. Wittmann, P. Olbrich, W. Eder, W. Prettl, L. E. Golub, E. V. Beregulin, Z. D. Kvon, N. N. Mikhailov, S. A. Dvoretsky, V. A. Shalygin, N. Q. Vinh, A. F. G. van der Meer, B. Murdin, and S. D. Ganichev, J. Appl. Physics **105**, 013106 (2009). S. D. Ganichev, J. Kiermaier, W. Weber, S. N. Danilov, D. Schuh, Ch. Gerl, W. Wegscheider, D. Bougeard, G. Abstreiter, and W. Prettl, Appl. Phys. Lett. **91**, 091101 (2007). S. D. Ganichev, W. Weber, J. Kiermaier, S. N. Danilov, D. Schuh, W. Wegscheider, Ch. Gerl, D. Bougeard, G. Abstreiter and W. Prettl, J. Appl. Physics **103**, 114504 (2008). S.D. Ganichev, S.A. Tarasenko, V.V. Bel’kov, P. Olbrich, W. Eder,D.R. Yakovlev, V. Kolkovsky, W. Zaleszczyk, G. Karczewski, T. Wojtowicz, and D. Weiss, Phys. Rev. Lett. **102**, 156602 (2009). J. Karch, P. Olbrich, M. Schmalzbauer, C. Zoth, C. Brinsteiner, M. Fehrenbacher, U. Wurstbauer, M. M. Glazov, S. A. Tarasenko, E. L. Ivchenko, D. Weiss, J. Eroms, R. Yakimova, S. Lara-Avila, S. Kubatkin, S. D. Ganichev, Phys. Rev. Lett. **105**, 227402 (2010). J. Karch, C. Drexler, P. Olbrich, M. Fehrenbacher, M. Hirmer, M. M. Glazov, S. A. Tarasenko, E. L. Ivchenko, B. Birkner, J. Eroms, D. Weiss, R. Yakimova, S. Lara-Avila, S. Kubatkin, M. Ostler, T. Seyller, S. D. Ganichev, Phys. Rev. Lett. **107**, 276601 (2011). M.M.Glazov and S.D.Ganichev, Physics Reports **535**, 101 (2014). S. D. Ganichev and W. Prettl, *Intense Terahertz Excitation of Semiconductors* (Oxford University Press, Oxford, 2006). While being detected in all reported measurements a polarization independent offset given by the coefficient $j_0$ will not be discussed in details. Instead, hereafter we focus on helicity sensitive photocurrent, $j_{\rm C}$, and currents driven by linearly polarized light, $j_{1}$ and $j_{2}$. M. Sakowicz, M. B. Lifshits, O. A. Klimenko, F. Schuster, D. Coquillat, F. Teppe, and W. Knap, J. Appl. Phys. **110**, 054512 (2011). D. Coquillat, V. Nodjiadjim, A. Konczykowska, M. Riet, N. Dyakonova, C. Consejo, F. Teppe, J. Godin, W. Knap, Didgest of Int. Conf. on Infrared, Millimeter, and Terahertz Waves, Tucson, USA, (2014). Note that signal variation with polarization is, apart the offset, identical with that of $s_1$, therefore this polarization dependence can also be used to describe the $j_1$-related photocurrent behavior. C. Drexler, N. Dyakonova, P. Olbrich, J. Karch, M. Schafberger, K. Karpierz, Yu. Mityagin, M. B. Lifshits, F. Teppe, O. Klimenko, Y. M. Meziani, W. Knap, and S. D. Ganichev, J. Appl. Physics **111**, 124504 (2012). K. S. Romanov and M. I. Dyakonov, Appl. Phys. Lett. **102**, 153502 (2013). All previous works aimed to the radiation induced ratchet effects discuss the case of unconnected parallel metal stripes: a system belonging to C$_s$ point group symmetry consisting of the identity element and the reflection in the plane perpendicular to the stripes [@28; @30; @29; @33Roj; @35; @36]. For this symmetry circular photocurrent $j_C$ and the photocurrent $j_2$ can be generated along stripes only whereas polarization independent offset and $j_0$ and photocurrent $j_1$ are allowed in the perpendicular to that direction (source-drain). Design of our DDG structures with interconnected metal stripes in each of gates excludes reflection plane reducing the point group symmetry to C$_1$. As a result the symmetry does not imply any restrictions and the photocurrent includes all four individual contributions ($j_0$, $j_1$, $j_2$ and $j_C$) which are allowed in any in-plain direction. More details on the symmetry analysis of photocurrents in quantum wells of C$_1$ symmetry can be found in [@Wittmann; @Belkov2008]. B. Wittmann, S.N. Danilov, V.V. Bel’kov, S.A. Tarasenko, E.G. Novik, H. Buhmann, C. Brüne, L.W. Molenkamp, E.L. Ivchenko, Z.D. Kvon, N.N. Mikhailov, S.A. Dvoretsky, N.Q.Vinh, A.F.G. van der Meer, B. Murdin, and S.D. Ganichev Semicond. Sci. and Technology **25**, 095005 (2010). V.V. Bel’kov, and S.D. Ganichev, Semicond. Sci. Technol. **23**, 114003 (2008). V.V. Popov, D.V. Fateev, E.L. Ivchenko, and S.D. Ganichev, Phys. Rev. B **91**, 235436 (2015). V.A. Shalygin, H. Diehl, Ch. Hoffmann, S.N. Danilov, T. Herrle, S.A. Tarasenko, D. Schuh, Ch. Gerl, W. Wegscheider, W. Prettl and S.D. Ganichev, JETP Lett. **84**, 570 (2006).

ArXiv

--- abstract: 'Traditional classification algorithms assume that training and test data come from similar distributions. This assumption is violated in adversarial settings, where malicious actors modify instances to evade detection. A number of custom methods have been developed for both adversarial evasion attacks and robust learning. We propose the first systematic and general-purpose retraining framework which can: a) boost robustness of an *arbitrary* learning algorithm, in the face of b) a broader class of adversarial models than any prior methods. We show that, under natural conditions, the retraining framework minimizes an upper bound on optimal adversarial risk, and show how to extend this result to account for approximations of evasion attacks. Extensive experimental evaluation demonstrates that our retraining methods are nearly indistinguishable from state-of-the-art algorithms for optimizing adversarial risk, but are *more general and far more scalable*. The experiments also confirm that without retraining, our adversarial framework dramatically reduces the effectiveness of learning. In contrast, retraining significantly boosts robustness to evasion attacks without significantly compromising overall accuracy.' author: - | Bo Li\ `bo.li.2@vanderbilt.edu`\ Vanderbilt University\ Yevgeniy Vorobeychik\ `yevgeniy.vorobeychik@vanderbilt.edu`\ Vanderbilt University\ Xinyun Chen\ `jungyhuk@gmail.com`\ Shanghai Jiaotong University\ bibliography: - 'nips\_2016\_retraining.bib' title: A General Retraining Framework for Scalable Adversarial Classification --- Introduction ============ Experimental Results ==================== Supplement To: A General Retraining Framework for Scalable Adversarial Classification {#supplement-to-a-general-retraining-framework-for-scalable-adversarial-classification .unnumbered} =====================================================================================

ArXiv

--- abstract: 'Our ignorance of the period between the end of inflation and the beginning of Big Bang Nucleosynthesis limits our understanding of the origins and evolution of dark matter. One possibility is that the Universe’s energy density was dominated by a fast-rolling scalar field while the radiation bath was hot enough to thermally produce dark matter. We investigate the evolution of the dark matter density and derive analytic expressions for the dark matter relic abundance generated during such a period of kination. Kination scenarios in which dark matter does not reach thermal equilibrium require $\langle \sigma v \rangle < 2.7\times 10^{-38} {\,\mathrm{cm^3\,s^{-1}}}$ to generate the observed dark matter density while allowing the Universe to become radiation dominated by a temperature of $3 \, \mathrm{MeV}$. Kination scenarios in which dark matter does reach thermal equilibrium require $\langle \sigma v \rangle > 3\times 10^{-26} {\,\mathrm{cm^3\,s^{-1}}}$ in order to generate the observed dark matter abundance. We use observations of dwarf spheroidal galaxies by the Fermi Gamma-Ray Telescope and observations of the Galactic Center by the High Energy Stereoscopic System to constrain these kination scenarios. Combining the unitarity constraint on $\langle \sigma v \rangle$ with these observational constraints sets a lower limit on the temperature at which the Universe can become radiation dominated following a period of kination if ${\langle \sigma v \rangle > 3\times 10^{-31} {\,\mathrm{cm^3\,s^{-1}}}}$. This lower limit is between ${0.05 \, \mathrm{GeV}}$ and ${1 \, \mathrm{GeV}}$, depending on the dark matter annihilation channel.' author: - Kayla Redmond - 'Adrienne L. Erickcek' title: New Constraints on Dark Matter Production during Kination --- Introduction {#sec:Intro} ============ The expansion history of the Universe before Big Bang Nucleosynthesis (BBN) is uncertain. The fact that the primordial curvature perturbation spectrum is almost scale invariant strongly suggests that shortly after the Big Bang, the Universe experienced a period of inflation [@Guth:1980; @Linde:1981; @Albrecht:1982]. The energy scale of inflation is not known, but it is generally assumed to be greater than ${10^{10} \, \mathrm{GeV}}$. The successful BBN prediction of the abundances of light elements only requires that the Universe be radiation dominated at a temperature of ${3 \, \mathrm{MeV}}$ [@Kawasaki:1999; @Kawasaki:2000; @Hannestad:2004; @Ichikawa:2005; @Ichikawa:2006]. Thus, there is a gap in the cosmological record between the theorized energy scale of inflation and ${3 \, \mathrm{MeV}}$. In the simplest model, inflation is powered by a scalar field defined as the inflaton, and the Universe becomes radiation dominated when the inflaton decays into relativistic particles [@Kofman:1994; @Kofman:1997; @Allahverdi:2010]. Another possibility is that a different scalar field dominates the Universe after inflation [@Kane:2015]. If either of these scalar fields oscillates around the minimum of their potential, it behaves like a pressureless fluid and the Universe would be in an early-matter-dominated era [@Turner:1983; @Chung:1998; @Kamionkowski:1990; @Giudice:2000; @Fornengo:2002; @Pallis:2004; @Gondolo:2006; @Gelmini:2006; @Drewes:2014; @Kumar:2015; @Erickcek:2015; @Kane:2015]. An alternative scenario is that a fast-rolling scalar field (a kinaton) dominates the energy density of the Universe prior to the onset of radiation domination. When the kinaton’s energy density is dominant, the Universe is said to be in a period of kination [@Spokoiny:1993; @Joyce:1996; @Ferreira:1997]. Kination was initially proposed as an inflationary model that does not require the complete conversion of the false vacuum energy into radiation to initiate the onset of radiation domination [@Spokoiny:1993]. Kination also facilitates baryogenesis; if the electroweak phase transition occurs during kination, then baryogenesis is possible during a second-order phase transition [@Joyce:1996]. Finally, if the kinaton’s potential energy becomes dominant at very late times, it can accelerate the expansion of the Universe and mimic the effects of a cosmological constant [@Ferreira:1997; @Peebles:1998; @Dimopoulos:2001; @Dimopoulos:2002Curvaton; @Chung:2007]. The uncertainties in the thermal history of the Universe prior to BBN limit our understanding of the origins of dark matter [@Kamionkowski:1990; @Giudice:2000; @Gelmini:2006; @Drees:2006; @Grin:2007; @Watson:2009; @Kane:2015; @Erickcek:2015; @Beniwal:2017]. We study the effects of kination on the thermal production of dark matter. We derive analytic expressions for the dark matter relic abundance generated during kination and confirm that our analytic results match the numeric solutions to the Boltzmann equation. Our relic abundance expressions depend on the dark matter mass $m_\chi$, the velocity-averaged dark matter annihilation cross section $\langle \sigma v \rangle$, and the temperature at which the Universe becomes radiation dominated, $T_{\mathrm{RH}}$. Our analytic expressions allow us to solve for the $\langle \sigma v \rangle$ values that will generate the observed dark matter abundance. We determine that in order to achieve the observed dark matter abundance, kination models in which dark matter reaches thermal equilibrium require $\langle \sigma v \rangle$ values that would underproduce dark matter during radiation domination. In contrast, kination models in which dark matter does not reach thermal equilibrium require $\langle \sigma v \rangle$ values that would overproduce dark matter during radiation domination. Using the most recent constraints on $m_\chi$ and $\langle \sigma v \rangle $ from Fermi-LAT PASS-8 observations of dwarf spheroidal galaxies [@Fermi:Constraints] and High Energy Stereoscopic System (H.E.S.S.) observations of the Galactic Center [@HESS:Constraints], we constrain $T_{{\mathrm{RH}}}$ for kination scenarios where dark matter reaches thermal equilibrium. Prior investigations of dark matter production during kination have focused on specific kinaton potentials. References [@Profumo:2003; @Pallis:2005; @Pallis:2nd2005; @Pallis:2006; @Gomez:2008] investigated how the relic abundance of dark matter is affected if the kinaton has an exponential potential, while Refs. [@Lola:2009; @Pallis:2009] studied kination models where the kinaton has an inverse power-law potential. While these prior works did place constraints on dark matter parameters, those constraints were dependent on the specified kinaton potential. Our relic abundance expressions are independent of the kinaton potential, and our constraints on $m_\chi$, $T_{{\mathrm{RH}}}$, and $\langle \sigma v \rangle$ are applicable to all kination scenarios in which dark matter is a thermal relic. Furthermore, improvements in the observational constraints on $\langle \sigma v \rangle$ over the past six years allow us to place tighter constraints than previous works. We determine that kination scenarios in which dark matter reaches thermal equilibrium have a minimum allowed reheat temperature between ${0.05 \, \mathrm{GeV}}$ and ${1 \, \mathrm{GeV}}$, depending on the dark matter annihilation channel. In Section \[sec:Background\], we discuss the evolution equations that govern the thermal production of dark matter during kination. In Sections \[sec:FreezeOut\] and \[sec:FreezeIn\], we present analytic derivations of the dark matter relic abundance for dark matter that does and does not reach thermal equilibrium. In Section \[sec:Constraints\], we use observational data from Fermi-LAT and H.E.S.S. to constrain $m_\chi$, $T_{{\mathrm{RH}}}$, and $\langle \sigma v \rangle$. In Section \[sec:end\], we summarize our results. Natural units $(\hbar = c=k_B=1)$ are used throughout this work. Thermal Dark Matter During Kination {#sec:Kination} =================================== Kinaton Cosmology {#sec:Background} ----------------- The scenario we consider consists of a fast-rolling scalar field (the kinaton) that dominates the energy density of the Universe prior to BBN. The kinaton’s energy density is dominated by its kinetic energy, meaning that the kinaton’s energy density equals its pressure and that the equation of state parameter is $w=1$. Therefore, the kinaton’s energy density scales as $a^{-6}$, where $a$ is the scale factor, and will eventually become subdominant to radiation, whose energy density scales as $a^{-4}$. Reheating is defined as the point at which the radiation energy density becomes the dominant component of the Universe. It is important to note, however, that during kination, the temperature of the radiation bath is higher than the temperature at reheating. Therefore, it is possible to thermally produce dark matter prior to the onset of radiation domination. We consider three energy density components during kination: dark matter, radiation, and the kinaton. The evolution of these energy densities are governed by three free parameters: the dark matter mass $m_{\chi}$, the reheat temperature $T_{{\mathrm{RH}}}$, and the velocity-averaged dark matter annihilation cross section $\langle \sigma v \rangle$. Throughout this work, we assume *s*-wave dark matter annihilation. In kination scenarios, $T_{\mathrm{RH}}$ is the temperature at which the radiation energy density equals the kinaton energy density. We assume that the kinaton does not decay nor interact with dark matter or radiation (see Ref. [@Pallis:2nd2005] for an analysis of decaying kinaton cosmologies). Radiation and dark matter on the other hand are thermally coupled via pair production and annihilation. Therefore, the equations for the energy density of the scalar field $\rho_{\phi}$, the radiation energy density $\rho_r$, and the dark matter number density $n_\chi$ are $$\begin{aligned} & \frac{d}{dt} \rho_{\phi} = -6H\rho_{\phi}, \\ & \frac{d}{dt}n_{\chi} = -3Hn_{\chi} - \langle \sigma v \rangle (n_{\chi}^2 - n_{\chi,{\mathrm{eq}}}^2), \\ & \frac{d}{dt}\rho_r = -4H\rho_r + \langle \sigma v \rangle E_{\chi} (n_{\chi}^2 - n_{\chi,{\mathrm{eq}}}^2),\end{aligned}$$ \[eq:Boltz\] where ${ \langle E_{\chi} \rangle = \rho_{\chi}/n_{\chi}}$ is the average energy of a dark matter particle and $n_{\chi,{\mathrm{eq}}}$ is the number density of dark matter particles in thermal equilibrium.[^1] For a dark matter particle with mass $m_{\chi}$ and internal degrees of freedom $g_{\chi}$ within a thermal bath of temperature $T$, ![The density evolution of the kinaton, radiation, and dark matter. In this figure, ${m_\chi = 10^4 \, \mathrm{GeV}}$, and reheating occurs when ${a/a_I=2.7\times 10^4}$; the reheat temperature is ${2\,\mathrm{GeV}}$. The two solid curves show the evolution of $\rho_{\chi}$ for the two values of $\langle \sigma v \rangle$ that produce the observed dark matter density: ${\Omega_{\chi} h^2 = 0.12}$ [@Planck:2015]. The top solid curve corresponds to the freeze-out scenario with ${\langle \sigma v \rangle = 7.5 \times 10^{-25} {\,\mathrm{cm^3\,s^{-1}}}}$, whereas the bottom solid curve corresponds to the freeze-in scenario with ${\langle \sigma v \rangle = 6.7\times10^{-46} {\,\mathrm{cm^3\,s^{-1}}}}$. The dotted line shows the equilibrium dark matter density, ${\rho_{\chi,\mathrm{eq}} = \langle E_{\chi} \rangle n_{\chi,\mathrm{eq}}}$.[]{data-label="Fig:DensityEvolution"}](DensityEvolution.eps){width="3.4in"} $$\begin{aligned} n_{\chi,{\mathrm{eq}}} = \frac{g_{\chi}}{2\pi^2} \int_{m_{\chi}}^{\infty} \frac{\sqrt{E^2 - m_{\chi}^2}}{e^{E/T} +1} E \, \mathrm{d}E . \label{eq:Equilibrium}\end{aligned}$$ When evaluating the average energy of a dark matter particle, we make the approximation that ${\langle E_{\chi} \rangle \simeq \sqrt{m_{\chi}^2 + (3.151 \, T)^2}}$, which matches ${\rho_{\chi}/n_{\chi}}$ to within $10\%$. Figure \[Fig:DensityEvolution\] shows the evolution of the kinaton, radiation, and dark matter densities obtained by numerically solving Eq. (\[eq:Boltz\]). Initially, the kinaton’s energy density is dominant, but since it scales away more quickly than the radiation energy density, it eventually becomes subdominant. Figure \[Fig:DensityEvolution\] shows that the radiation energy density scales as $a^{-4}$ and is unaffected by dark matter annihilation or pair production. Since $\rho_r$ and $\rho_\phi$ evolve independently of $n_\chi$, we can solve for their evolution analytically. We then use these solutions to numerically solve Eq. (\[eq:Boltz\]b) and calculate the dark matter relic abundance. To accurately describe the evolution of $\rho_r$, we need to take into account the energy injection that occurs when Standard Model particles become nonrelativistic. When a particle species becomes nonrelativistic, its entropy is transferred to the remaining relativistic particles. Entropy is conserved during kination; therefore, the universal entropy $sa^3$ must remain constant, where $s$ is the entropy density: ${s \equiv (2\pi^2/45) \, T^3 g_{*s}(T)}$, and $g_{*s}(T)$ is the effective number of degrees of freedom that contribute to the entropy density. Due to the conservation of entropy, radiation cools during kination according to the same proportionality as during radiation domination: ${T \propto g_{*s}(T)^{-1/3} \, a^{-1}}$. To evaluate the temperature of the radiation bath, we set a maximum temperature of $T_{\mathrm{MAX}}$ at which ${\rho_{\chi} =0}$. We set ${T_{\mathrm{MAX}} = 8m_{\chi}}$ to ensure that if the dark matter is capable of reaching thermal equilibrium, it will have adequate time to do so. If the dark matter cannot reach thermal equilibrium, setting ${T_{\mathrm{MAX}} = 8m_{\chi}}$ ensures there will be enough time for maximal pair production. Therefore, the dark matter relic abundance will not be sensitive to $T_{\mathrm{MAX}}$. Using $T_{\mathrm{MAX}}$, we construct an expression for the temperature evolution during kination that accounts for changes in $g_{*s}(T)$: $$\begin{aligned} T = T_{\mathrm{MAX}} \left[\frac{g_{*s}(T_\mathrm{MAX})}{g_{*s}(T)} \right]^{1/3} \, \frac{a_I}{a}, \label{eq:Temperature}\end{aligned}$$ where $a_I$ is the scale factor value when ${T=T_{\mathrm{MAX}}}$. The final step in evaluating $\rho_r$ is to connect Eq. (\[eq:Temperature\]) and the definition of $\rho_r$. The radiation energy density is $\rho_r \equiv (\pi^2/30) \, g_*(T) \, T^4$, where $g_{*}(T)$ is the number of relativistic degrees of freedom at temperature $T$. Using this definition of $\rho_r$ and Eq. (\[eq:Temperature\]), we see that the evolution of $\rho_r$ during kination is $$\begin{aligned} \rho_r = \frac{\pi^2}{30} \, g_*(T) \, T_\mathrm{MAX}^4 \left[ \frac{g_{*s}(T_\mathrm{MAX})}{g_{*s}(T)} \right]^{4/3} \left(\frac{a_I}{a} \right)^4. \label{eq:RadiationEnergyDensity}\end{aligned}$$ Next, we analytically solve for $\rho_\phi$. Solving Eq. (\[eq:Boltz\]a) yields ${\rho_\phi = \rho_{\phi,I} \, (a_I/a)^6}$, where $\rho_{\phi,I}$ is $\rho_{\phi}$ when $a=a_I$. By defining $a_{{\mathrm{RH}}}$ as the scale factor value at the onset of radiation domination we see that $\rho_\phi$ evaluated at reheating equals ${\rho_{\phi,I} \, (a_I/a_{\mathrm{RH}})^6}$. Using Eq. (\[eq:RadiationEnergyDensity\]), we can evaluate $\rho_r$ at reheating. Considering that at reheating ${\rho_\phi = \rho_r}$, this implies that $$\begin{aligned} \rho_{\phi,I} = \frac{\pi^2}{30} \, g_*(T_{\mathrm{RH}}) \, T_\mathrm{MAX}^4 \left[ \frac{g_{*s}(T_\mathrm{MAX})}{g_{*s}(T_{\mathrm{RH}})} \right]^{4/3} \left(\frac{a_{\mathrm{RH}}}{a_I} \right)^2. \label{eq:InitialScalarEnergyDensity}\end{aligned}$$ Using Eq. (\[eq:Temperature\]) to relate $a_{\mathrm{RH}}$ to $T_{\mathrm{RH}}$, we obtain the evolution of $\rho_\phi$ during kination: $$\begin{aligned} \rho_{\phi} = \frac{\pi^2}{30} \left[ \frac{T_\mathrm{MAX}^3}{T_{\mathrm{RH}}} \right]^2 \left[\frac{g_{*s}(T_\mathrm{MAX})}{g_{*s}(T_{\mathrm{RH}})} \right]^2 g_*(T_{\mathrm{RH}}) \left(\frac{a_I}{a} \right)^6. \label{eq:ScalarEnergyDensity}\end{aligned}$$ Now that we have obtained expressions for $T(a)$, $\rho_r(a)$ and $\rho_\phi(a)$, we have the necessary components to numerically solve Eq. (\[eq:Boltz\]b) for $n_\chi(a)$, as shown in Figure \[Fig:FreezeoutDarkMatter\]. Figure \[Fig:OmegaDM\] shows the dark matter relic abundance as a function of $\langle \sigma v \rangle$ for several values of $T_{\mathrm{RH}}$ and $m_\chi$. For small $\langle \sigma v \rangle$ values, the dark matter cannot reach thermal equilibrium, and Figure \[Fig:OmegaDM\] shows that as $\langle \sigma v \rangle$ increases the dark matter relic abundance increases. Once $\langle \sigma v \rangle$ becomes large enough, pair production will bring $n_\chi$ up to its thermal equilibrium value. If dark matter reaches thermal equilibrium, we see from Figure \[Fig:OmegaDM\] that as $\langle \sigma v \rangle$ increases, the dark matter relic abundance decreases. In the following sections, we derive analytic expressions for the dark matter relic abundance generated during kination and analyze how the relic abundance is influenced by $T_{\mathrm{RH}}$. Freeze-Out {#sec:FreezeOut} ---------- If $\langle \sigma v \rangle$ is sufficiently large, then pair production brings dark matter into thermal equilibrium: ${n_{\chi} = n_{\chi,{\mathrm{eq}}}}$, as defined in Eq. (\[eq:Equilibrium\]). Once ${H \simeq \langle \sigma v \rangle n_{\chi,{\mathrm{eq}}}}$, the dark matter deviates from equilibrium and “freezes out". If dark matter freezes out during radiation domination, nearly all dark matter annihilations cease at freeze-out. However, if dark matter freezes out during kination, we see from Figure \[Fig:FreezeoutDarkMatter\] that we need to take dark matter annihilations between the time of freeze-out and reheating into account to get an accurate relic abundance. To analytically solve for the evolution of the dark matter number density between freeze-out and reheating, we define a dimensionless comoving number density ${Y \equiv n_{\chi} (a/a_I)^3 \, T_{{\mathrm{RH}}}^{-3}}$. Equation (\[eq:Boltz\]$\mathrm{b}$) is rewritten as ![The evolution of the comoving dark matter number density and equilibrium number density with ${T_{{\mathrm{RH}}} = 20 \, \mathrm{GeV}}$, ${m_{\chi} = 3000 \, \mathrm{GeV}}$, and ${\langle \sigma v \rangle = 10^{-32} {\,\mathrm{cm^3\,s^{-1}}}}$. The vertical line represents the point of reheating at ${a_{{\mathrm{RH}}} /a_I = 800}$. The solid horizontal line shows the comoving number density at the point of freeze-out solved from ${H{\left(T_\mathrm{F} \right)} = \langle \sigma v \rangle n_{\chi,{\mathrm{eq}}}}$. This figure demonstrates that dark matter annihilations after freeze-out significantly decrease the dark matter number density.[]{data-label="Fig:FreezeoutDarkMatter"}](DarkMatterDimensionless.eps){width="3.4in"} $$\begin{aligned} \frac{dY}{da} = \langle \sigma v \rangle \frac{T_{{\mathrm{RH}}}^3 \, a_I^3}{Ha^4} (Y_{{\mathrm{eq}}}^2 - Y^2). \label{eq:Dimensionless}\end{aligned}$$ After the dark matter freezes out, ${Y^2 \gg Y_{{\mathrm{eq}}}^2}$. Since during kination ${H = H(a_I) [a_I/a]^{3}}$, we simplify Eq. (\[eq:Dimensionless\]) to $$\begin{aligned} \frac{dY}{da} = \frac{-\lambda_{\mathrm{KD}}}{a} \, Y^2, \label{eq:ReducedBoltzEqn}\end{aligned}$$ where ${\lambda_{\mathrm{KD}} = T_{{\mathrm{RH}}}^3 \langle \sigma v \rangle / H(a_I)}$. Integrating Eq. (\[eq:ReducedBoltzEqn\]) from freeze-out to reheating yields $$\begin{aligned} \frac{1}{Y_\mathrm{F}} - \frac{1}{Y_{{\mathrm{RH}}}} = -\lambda_{\mathrm{KD}} \, \ln{\frac{a_{{\mathrm{RH}}}}{a_{\mathrm{F}}}} , \label{eq:SolvedBoltzEqn}\end{aligned}$$ where $Y_{\mathrm{F}}$ and $Y_{\mathrm{RH}}$ are the comoving dark matter number densities at freeze-out and reheating. Therefore, if freeze-out occurs during kination, the dark matter comoving number density experiences a logarithmic decrease between freeze-out and reheating. To evaluate the current dark matter density we need to reevaluate Eq. (\[eq:Dimensionless\]) during radiation domination and solve for $Y$ at some late time. During radiation domination ${H = H(a_{\mathrm{RH}}) [a_{\mathrm{RH}}/a]^{2}}$, and by defining ${\lambda_{\mathrm{RD}} = [T_{{\mathrm{RH}}}^3 \langle \sigma v \rangle / H(a_{\mathrm{RH}})]\times[a_I^3/a_{\mathrm{RH}}^2]}$, Eq. (\[eq:Dimensionless\]) simplifies to $$\begin{aligned} \frac{dY}{da} = \frac{-\lambda_{\mathrm{RD}}}{a^2} \, Y^2. \label{eq:ReducedBoltzEqnRD}\end{aligned}$$ Solving Eq. (\[eq:ReducedBoltzEqnRD\]) from reheating to a very late time yields $$\begin{aligned} \frac{1}{Y_\mathrm{RH}} - \frac{1}{Y_{\mathrm{LT}}} = -\lambda_{\mathrm{RD}} \left(\frac{1}{a_\mathrm{RH}}\right), \label{eq:SimplifySolRD}\end{aligned}$$ where $Y_{\mathrm{LT}}$ is the comoving dark matter number density at some late time $(a=a_\mathrm{LT})$. To obtain Eq. (\[eq:SimplifySolRD\]), we use the fact that ${a_{\mathrm{LT}} \gg a_{\mathrm{RH}}}$. Therefore, if dark matter freezes out during kination, $Y$ experiences a logarithmic decrease between freeze-out and reheating, after which $Y$ approaches a constant value. Utilizing $Y_{\mathrm{RH}}$ from Eq. (\[eq:SolvedBoltzEqn\]) and rewriting Eq. (\[eq:SimplifySolRD\]) in terms of the dark matter number density yields $$\begin{aligned} n_{\chi,\mathrm{LT}} =& \,\, \left[\frac{\langle \sigma v \rangle \, a_\mathrm{LT}^3}{H(a_I) \, a_I^3} \, \left(\ln \left[\frac{a_{{\mathrm{RH}}}}{a_\mathrm{F}} \right]+1\right)+\frac{a_\mathrm{LT}^3}{n_{\chi,\mathrm{F}} \, \, a_\mathrm{F}^3} \right]^{-1}. \label{eq:ReheatingNumberDensityFirst}\end{aligned}$$ We wish to express Eq. (\[eq:ReheatingNumberDensityFirst\]) in terms of our free parameters $m_{\chi}$, $T_{{\mathrm{RH}}}$, and $\langle \sigma v \rangle$. We can express $H(a_I)$ and $a_{{\mathrm{RH}}}$ in terms of $T_{{\mathrm{RH}}}$ and $T_{\mathrm{MAX}}$ using Eqs. (\[eq:Temperature\]) and (\[eq:InitialScalarEnergyDensity\]). In addition, since ${n_{\chi, \mathrm{F}} \simeq n_{\chi,\mathrm{eq}}}$, ${n_{\chi,\mathrm{F}} \simeq H{\left(T_\mathrm{F} \right)}/ \langle \sigma v \rangle}$. If freeze-out occurs during kination, ${H^2 \simeq \left(8 \pi G/3 \right) \rho_\phi}$; combining this with Eqs. (\[eq:Temperature\]) and (\[eq:ScalarEnergyDensity\]) allows us to solve for $H{\left(T_\mathrm{F} \right)}$: $$\begin{aligned} {H\!\left(T_\mathrm{F}\right)} = \,\, \left( \frac{4\pi^3}{45} \right)^{1/2} \left( \frac{T_\mathrm{F}^3}{m_\mathrm{pl} \, T_{\mathrm{RH}}} \right) \left( \frac{g_{*s\mathrm{F}}}{g_{*s\mathrm{RH}}} \right) g_{*\mathrm{RH}}^{1/2}, \label{eq:HubbleTemperature}\end{aligned}$$ where ${g_{*s{\mathrm{RH}}} = g_{*s}(T_{{\mathrm{RH}}})}$ and ${g_{*{\mathrm{RH}}} = g_{*}(T_{{\mathrm{RH}}})}$. These relations allow us to rewrite Eq. (\[eq:ReheatingNumberDensityFirst\]) as $$\begin{aligned} n_{\chi,\mathrm{LT}} =& \,\, \left(\frac{4\pi^3}{45} \right)^{1/2} \frac{T_{\mathrm{LT}}^3 \, g_{*s\mathrm{LT}} \, g_{*\mathrm{RH}}^{1/2}}{\langle \sigma v \rangle \, {m_{\mathrm{pl}}}\, T_{\mathrm{RH}}\, g_{*s{\mathrm{RH}}}} \\ &\times \left(\ln\left[\frac{T_\mathrm{F}}{T_{{\mathrm{RH}}}} \left(\frac{g_{*s\mathrm{F}}}{g_{*s{\mathrm{RH}}}}\right)^{1/3}\right] + 2\right)^{-1}, \nonumber \label{eq:ReheatingNumberDensity}\end{aligned}$$ where $T_\mathrm{F}$ is obtained by numerically solving ${H(T_\mathrm{F}) = \langle \sigma v \rangle n_{\chi,\mathrm{eq}}}$. For kination scenarios, ${m_\chi / T_{\mathrm{F}}}$ is roughly between 20 and 30. After $a_\mathrm{LT}$, ${n_\chi \propto a^{-3}}$, which allows us to relate the dark matter density at $a_\mathrm{LT}$ to today: $$\begin{aligned} \rho_{\chi,0} = \rho_{\chi,\mathrm{LT}} \left(\frac{a_{\mathrm{LT}}}{a_0}\right)^3 = \rho_{\chi,\mathrm{LT}} \left(\frac{T_0}{T_{\mathrm{LT}}}\right)^3 \left(\frac{g_{*s0}}{g_{*s\mathrm{LT}}}\right),\end{aligned}$$ where $T_0$ is the radiation temperature today and ${g_{*s0} = 3.91}$. Bringing all of the previous components together and scaling our analytic expression by a factor of $1.22$, thereby ensuring that it matches the numeric solution of Eq. (\[eq:Boltz\]b) within 20% for ${m_\chi / T_{\mathrm{RH}}> 100}$, we obtain an analytic expression for the freeze-out dark matter relic abundance: $$\begin{aligned} \Omega_{\chi} h^2 =& \,\, 6.06 \left(\frac{3\times10^{-26} {\,\mathrm{cm^3\,s^{-1}}}}{\langle \sigma v \rangle}\right) \frac{g_{*\mathrm{RH}}^{1/2}}{g_{*s\mathrm{RH}}} \left(\frac{m_{\chi}/T_{{\mathrm{RH}}}}{150}\right) \nonumber \\ &\times \left(\ln\left[\frac{T_{\mathrm{F}}/T_{\mathrm{RH}}}{10} \left(\frac{g_{*s{\mathrm{F}}}}{g_{*s{\mathrm{RH}}}}\right)^{1/3}\right] + 4.3\right)^{-1}. \label{eq:FreezeoutRelicAbundance}\end{aligned}$$ Equation (\[eq:FreezeoutRelicAbundance\]) indicates that decreasing $T_{{\mathrm{RH}}}$ increases the relic abundance. Decreasing $T_{{\mathrm{RH}}}$ requires increasing the kinaton energy density, which increases the Hubble parameter during kination. Since ${n_{\chi,\mathrm{F}} \simeq H(T_{\mathrm{F}})/\langle \sigma v \rangle}$, increasing the Hubble parameter increases the dark matter number density at freeze-out and thus increases the relic abundance. Furthermore, in our calculation of $Y_{{\mathrm{RH}}}$ we showed that dark matter annihilations do not cease during kination. As a result, Eq. (\[eq:FreezeoutRelicAbundance\]) includes an inverse logarithmic term that depends on the ratio ${T_{\mathrm{F}}/T_{\mathrm{RH}}}$. Figure \[Fig:OmegaDM\] shows the dark matter relic abundance as a function of $\langle \sigma v \rangle$ for several values of $T_{\mathrm{RH}}$ and $m_\chi$. In Figure \[Fig:OmegaDM\] we see that for sufficiently large $\langle \sigma v \rangle$ the freeze-out dark matter relic abundances from Eq. (\[eq:FreezeoutRelicAbundance\]), represented by the circular symbols, match the numeric solutions to Eq. (\[eq:Boltz\]$\mathrm{b}$), represented by the curves. We can solve for the minimum $\langle \sigma v \rangle$ that will result in the dark matter reaching thermal equilibrium. If dark matter freezes out, ${H{\left(T_\mathrm{F} \right)} = \langle \sigma v \rangle n_{\chi,{\mathrm{eq}}}}$, and we can rewrite this equation in terms of a new variable $x$, where $x \equiv m_{\chi}/T_\mathrm{F}$. Assuming that dark matter is nonrelativistic, ${H{\left(T_\mathrm{F} \right)} = \langle \sigma v \rangle n_{\chi,{\mathrm{eq}}}}$ can be rewritten as ${x^{-3/2} e^{x} g_{*s}\left(m_\chi /x\right) = constant \times \langle \sigma v \rangle}$. The left-hand side of this equation has a minimum value near ${x \sim 1.5}$ which implies that there is a minimum $\langle \sigma v \rangle$ for which a solution will exist. This lower bound on $\langle \sigma v \rangle$ is $$\begin{aligned} \langle \sigma v \rangle > & \,\, 9.37 \times 10^{-33} {\,\mathrm{cm^3\,s^{-1}}}\left(\frac{2}{g_\chi} \right) \nonumber \\ &\times \left( \frac{3 \, \mathrm{MeV}}{T_{{\mathrm{RH}}}} \right) \left( \frac{g_{*{\mathrm{RH}}}^{1/2}}{g_{*s{\mathrm{RH}}}}\right) g_{*s}(m_\chi/1.5) . \label{eq:CrossSectionBeginningFreezeOut}\end{aligned}$$ The horizontal line in Figure \[Fig:OmegaDM\] represents the Planck measurement of the observed dark matter abundance. To reproduce the observed dark matter abundance, freeze-out cases during kination require larger $\langle \sigma v \rangle$ than that required if freeze-out occurs during radiation domination. This comes from the fact that at a given temperature the Hubble parameter during kination is always higher than it is during radiation domination, which causes freeze-out to occur earlier. In order to compensate for the earlier freeze-out and reproduce the observed dark matter abundance, freeze-out scenarios during kination require ${\langle \sigma v \rangle > 3\times 10^{-26} {\,\mathrm{cm^3\,s^{-1}}}}$. Since this lower bound on $\langle \sigma v \rangle$ is more stringent than Eq. (\[eq:CrossSectionBeginningFreezeOut\]), Eq. (\[eq:FreezeoutRelicAbundance\]) is applicable to all freeze-out scenarios that generate the observed dark matter abundance. Freeze-In {#sec:FreezeIn} --------- For cross sections that violate Eq. (\[eq:CrossSectionBeginningFreezeOut\]), dark matter pair production is not sufficient to bring the dark matter into thermal equilibrium. Once pair production ceases, the dark matter “freezes in" and the comoving dark matter number density remains constant. The first step in determining the freeze-in dark matter relic abundance is to calculate the comoving dark matter number density when pair production ceases. In freeze-in scenarios, the dark matter number density does not reach thermal equilibrium, so ${n_{\chi} \ll n_{\chi,{\mathrm{eq}}}}$. The dimensionless comoving dark matter equilibrium number density is ${Y_{{\mathrm{eq}}} \equiv n_{\chi,{\mathrm{eq}}} T_{{\mathrm{RH}}}^{-3} (a/a_I)^3}$, and for freeze-in scenarios ${Y_{{\mathrm{eq}}} \gg Y}$. Therefore, Eq. (\[eq:Dimensionless\]) reduces to $$\begin{aligned} \frac{dY}{da} = \frac{\langle \sigma v \rangle \, T_{{\mathrm{RH}}}^3}{H\!\left(a_I\right) \, a } \, Y_{{\mathrm{eq}}}^2 \label{eq:FreezeInDimensionless}\end{aligned}$$ for freeze-in scenarios during kination. Equation (\[eq:FreezeInDimensionless\]) implies that $dY/da$ diverges as $a \rightarrow 0$ if $\langle \sigma v \rangle$ is independent of temperature. The same divergence occurs if the Universe is radiation dominated during dark matter production, and it would make the freeze-in abundance of dark matter dependent on $T_\mathrm{MAX}$. Previous analyses of the freeze-in process avoided this sensitivity to high-energy physics by assuming that ${\langle \sigma v \rangle \propto 1/ \, T^2}$ for relativistic particles [@Hall:2009; @Yaguna:2011]. We take the same approach and set ${\langle \sigma v \rangle = \langle \sigma v \rangle_s (m_\chi/T)^2}$ for ${T > m_\chi}$ and ${\langle \sigma v \rangle = \langle \sigma v \rangle_s}$ for ${T < m_\chi}$, where $\langle \sigma v \rangle_s$ is the *s*-wave dark matter annihilation cross section for massive particles. With this scaling, ${dY/da \rightarrow 0}$ as ${a \rightarrow 0}$, and the production of dark matter is finite during kination even if $T_\mathrm{MAX} \rightarrow \infty.$ Figure \[FreezeInDimensionless\] shows that $dY/da$ increases until ${a=a_*}$, which we define as the scale factor value at which pair production peaks. The temperature at which pair production peaks is $T_* = m_{\chi}$. Therefore, $dY/da$ reaches its maximum when ${a_*/a_{\mathrm{I}} = 8[g_{*s}(\mathrm{T_{MAX}})/g_{*s}(\mathrm{T_*})]^{1/3}}$. Figure \[FreezeInDimensionless\] also indicates that the integral of $dY/da$ converges; $Y$ will approach a constant value as pair production becomes less and less efficient. However, Eq. (\[eq:FreezeInDimensionless\]) is only valid during kination, so it will only provide an accurate dark matter density if nearly all the pair production occurs prior to reheating. Truncating the integration of $dY/da$ at $a_{\mathrm{PP}}$, where $T(a_\mathrm{PP}) = m_\chi/3.9$, reduces the value of $Y$ by less than 1% compared to integrating $dY/da$ out to $a=\infty$. Therefore, pair production has effectively halted when $T < m_{\chi}/3.9$, and we can use Eq. (\[eq:FreezeInDimensionless\]) to compute the relic abundance of dark matter provided that $T_\mathrm{RH} < m_{\chi}/3.9$. Integrating Eq. (\[eq:FreezeInDimensionless\]) from $0$ to $a_{\mathrm{PP}}$, while taking into account the fact that $\langle \sigma v \rangle$ changes from ${\langle \sigma v \rangle_s (m_\chi/T)^2}$ to ${\langle \sigma v \rangle_s}$ at ${T = m_\chi}$, gives $$\begin{aligned} Y_{\mathrm{PP}} =& \,\, 3.1\times 10^{-4} \left( \frac{T_{\mathrm{MAX}} /T_{\mathrm{RH}}}{150} \right)^3 \left(\frac{T_{\mathrm{RH}}}{5 \, \mathrm{GeV}}\right) \left(\frac{g_\chi}{2} \right)^2 \nonumber \\ &\times \left( \frac{ \langle \sigma v \rangle_s}{10^{-45} {\,\mathrm{cm^3\,s^{-1}}}}\right) \left( \frac{g_{*s{\mathrm{RH}}}}{g_{*{\mathrm{RH}}}^{1/2}} \right) \left(\frac{g_{*s\mathrm{MAX}}}{g_{*s}^2(m_\chi)}\right). \label{eq:EndPairProduction}\end{aligned}$$ After pair production ends, $Y$ remains nearly constant, and thus $Y_{\mathrm{PP}} = Y_{\mathrm{RH}}$. Therefore, the dark matter density at reheating can be written as $$\begin{aligned} \rho_{\chi,{{\mathrm{RH}}}} = m_{\chi} \, Y_{\mathrm{PP}} \, T_{{\mathrm{RH}}}^3 \left( a_{\mathrm{I}}/a_{{\mathrm{RH}}} \right)^3. \label{eq:FreezeInEnergyDensity}\end{aligned}$$ Equation (\[eq:Temperature\]) indicates that ${\left( a_{\mathrm{I}}/a_{{\mathrm{RH}}} \right)^3} \propto T_{\mathrm{MAX}}^{-3} \, g_{*s{\mathrm{MAX}}}^{-1}$. As a result, inserting Eq. (\[eq:EndPairProduction\]) into Eq. (\[eq:FreezeInEnergyDensity\]) seems to indicate that ${\rho_{\chi,{{\mathrm{RH}}}}}$ is independent of ${T_\mathrm{MAX}}$, but this is not generically true. When integrating Eq. (\[eq:FreezeInDimensionless\]) from $a=0$ to $a_{\mathrm{PP}}$ to obtain Eq. (\[eq:EndPairProduction\]), we effectively integrated from ${T = \infty}$ to ${T = m_\chi/3.9}$. However, integrating instead from ${T = 8 \, m_\chi}$ to ${T = m_\chi/3.9}$ does not significantly change the result. Therefore, if ${T_\mathrm{MAX} \geq 8 \, m_\chi}$, the freeze-in dark matter abundance does not depend on $T_\mathrm{MAX}$. Conversely, if ${T_\mathrm{MAX}< 8 \, m_\chi}$, ${\rho_{\chi,{{\mathrm{RH}}}}}$ will decrease as $T_\mathrm{MAX}$ decreases because maximal pair production is not reached. After reheating, $n_\chi \propto a^{-3}$, and we can use Eq. (15) to evolve $\rho_\chi$ from reheating to today. Combining the previous expressions and scaling our analytic expression by a factor of $0.95$ to match the numeric solution of Eq. (\[eq:Boltz\]b) provides an analytic expression for the freeze-in dark matter relic abundance: ![The evolution of $dY/da$ given ${T_{{\mathrm{RH}}}=1 \, \mathrm{GeV}}$, ${m_\chi = 5\times 10^4 \, \mathrm{GeV}}$, and ${\langle \sigma v \rangle_s = 10^{-47} {\,\mathrm{cm^3\,s^{-1}}}}$. The vertical line represents the scale factor at which pair production peaks, defined as $a_*$. For kination scenarios where dark matter freezes in and $T_\mathrm{MAX} = 8\, m_\chi$, ${a_*/a_{\mathrm{I}} \simeq 8}$.[]{data-label="FreezeInDimensionless"}](FreezeInDimensionless.eps){width="3.4in"} $$\begin{aligned} \Omega_{\chi} h^2 =& \, \, 1.08 \left( \frac{m_{\chi}}{1 \, \mathrm{GeV}} \right) \left(\frac{T_{\mathrm{RH}}}{100 \, \mathrm{GeV}}\right) \left(\frac{g_\chi}{2} \right)^2 \nonumber \\ &\times \left( \frac{ \langle \sigma v \rangle_s}{10^{-45} {\,\mathrm{cm^3\,s^{-1}}}}\right) \left(\frac{g_{*s{\mathrm{RH}}}}{g_{*{\mathrm{RH}}}^{1/2}} \right) g_{*s}^{-2}(m_\chi) . \label{eq:FreezeInRelicAbundance}\end{aligned}$$ Equation (\[eq:FreezeInRelicAbundance\]) indicates that increasing $T_{{\mathrm{RH}}}$ leads to a larger relic abundance. To understand how the freeze-in dark matter relic abundance relates to $T_{\mathrm{RH}}$ we need to investigate how $n_\chi$ relates to the Hubble parameter. The connection between the Hubble parameter and $n_\chi$ stems from the cooling rate $dT/dt$. During kination ${T\propto a^{-1}}$, and therefore ${dT/dt = -HT}$. Rewriting ${dn/dt}$ as a function of temperature yields ${dn/dt = \left(dT/dt\right) \left(dn/dT\right)} = f(T)$, where $f(T)$ is the right-hand side of Eq. (\[eq:Boltz\]b). This allows us to express $dn/dT$ as $$\begin{aligned} \frac{dn}{dT} = \frac{-f\left(T\right)}{HT}, \label{eq:CoolingRate}\end{aligned}$$ which implies that ${n \propto 1/H}$. Since increasing $T_{\mathrm{RH}}$ decreases $H$ during kination, it also decreases the cooling rate, leaving more time for pair production and thereby increasing the dark matter number density. In order to reach the observed dark matter abundance, scenarios in which dark matter freezes in during kination require larger $\langle \sigma v \rangle_s$ values than if freeze-in occurs during radiation domination. For example, given a dark matter mass of $100 \, \mathrm{GeV}$, a freeze-in scenario during radiation domination requires ${ \langle \sigma v \rangle_s = 10^{-47} {\,\mathrm{cm^3\,s^{-1}}}}$ in order for ${\Omega_{\chi} h^2 = 0.12}$ [@Dev:2013; @Planck:2015]. For the same $m_\chi$, a freeze-in scenario during kination with ${T_{\mathrm{RH}}= 0.033 \, \mathrm{GeV}}$ requires ${ \langle \sigma v \rangle_s = 10^{-41} {\,\mathrm{cm^3\,s^{-1}}}}$. Freeze-in scenarios during kination require larger $\langle \sigma v \rangle_s$ values to generate the observed dark matter abundance because the increased cooling rate during kination leaves less time for pair production. In Figure \[Fig:OmegaDM\], we see that for sufficiently small cross sections the dark matter relic abundances from Eq. (\[eq:FreezeInRelicAbundance\]), represented by the star symbols, match the numeric solutions to Eq. (\[eq:Boltz\]b), represented by the curves. We have already discussed how freeze-in requires the dark matter particles to never reach thermal equilibrium. For this to hold true, the dark matter number density at the end of pair production must be less than the dark matter equilibrium number density at the peak of pair production: ${Y(T_{\mathrm{PP}})<Y_{{\mathrm{eq}}}(T_*)}$. To solve for the largest $\langle \sigma v \rangle_s$ that will not result in dark matter reaching thermal equilibrium, we approximate the equilibrium number density as being nonrelativistic: $$\begin{aligned} Y_{{\mathrm{eq}}}(a) = g_\chi \left(\frac{a}{a_I} \right)^3 T_{{\mathrm{RH}}}^{-3} \left(\frac{m_{\chi} T}{2\pi}\right)^{3/2} e^{-m_{\chi}/T}. \label{eq:NonRelativisticEquilibrium}\end{aligned}$$ Evaluating Eq. (\[eq:NonRelativisticEquilibrium\]) at $a_*$ and equating that to Eq. (\[eq:EndPairProduction\]) gives us the cross sections that will result in dark matter freezing in during kination: $$\begin{aligned} \langle \sigma v \rangle_s \leq & \,\, 8.5\times 10^{-34} {\,\mathrm{cm^3\,s^{-1}}}\left( \frac{g_{*{\mathrm{RH}}}^{1/2}}{g_{*s{\mathrm{RH}}}} \right) \nonumber \\ &\times g_{*s}(m_\chi) \left( \frac{2}{g_{\chi}} \right) \left(\frac{3 \, \mathrm{MeV}}{T_{\mathrm{RH}}}\right). \label{eq:CrossSectionEndFreezeIn}\end{aligned}$$ Figure \[Fig:OmegaDM\] demonstrates that there is a range of $\langle \sigma v \rangle_s$ values for each reheat temperature where neither a freeze-out nor freeze-in scenario will result in the observed dark matter abundance. The left panel of Fig. \[Fig:OmegaDM\] shows that, at a fixed reheat temperature, decreasing $m_\chi$ increases the freeze-in cross section and decreases the freeze-out cross section that generates the observed dark matter abundance. However, once ${m_\chi \lesssim 3\,T_\mathrm{RH}}$, dark matter will no longer freeze in during kination, and as discussed in Section \[sec:FreezeOut\], $\langle \sigma v \rangle > 3\times 10^{-26} {\,\mathrm{cm^3\,s^{-1}}}$ is required to generate the observed dark matter abundance if dark matter freezes out during kination. The right panel of Fig. \[Fig:OmegaDM\] shows that decreasing $T_\mathrm{RH}$ increases the cross section that generates the observed dark matter abundance via the freeze-in mechanism. Therefore, setting ${T_{\mathrm{RH}}= 3 \, \mathrm{MeV}}$ gives an upper bound on the cross sections that can generate the observed dark matter abundance in freeze-in scenarios: $$\begin{aligned} \langle \sigma v \rangle_s < 2.66 \times 10^{-38} {\,\mathrm{cm^3\,s^{-1}}}\left( \frac{0.175 \, \mathrm{GeV}}{m_\chi} \right). \label{eq:MinForbiddenCrossSection}\end{aligned}$$ This maximal cross section is calculated using ${g_{*s}(m_\chi) = g_{*s}(0.175\, \mathrm{GeV})}$. When deriving Eq. (\[eq:FreezeInRelicAbundance\]) we assumed that $g_{*s}(T)$ was approximately constant during pair production, which implies that ${\Omega_\chi h^2 \propto g^{-2}_{*s}(m_\chi)}$. If dark matter freezes in during kination and $m_\chi$ is less than $0.17\, \mathrm{GeV}$, then the QCD phase transition occurs before the peak of pair production. At the QCD phase transition $g_{*s}(T)$ sharply decreases, resulting in an increase in the relic abundance as given by Eq. (\[eq:FreezeInRelicAbundance\]). To compensate for the increased relic abundance, freeze-in scenarios with ${m_\chi \leq 0.17\, \mathrm{GeV}}$ require cross sections smaller than the one calculated in Eq. (\[eq:MinForbiddenCrossSection\]). Therefore, the largest cross section that can generate the observed dark matter abundance in freeze-in scenarios is ${2.7\times 10^{-38} {\,\mathrm{cm^3\,s^{-1}}}}$. The relic abundances from Eq. (\[eq:FreezeInRelicAbundance\]) are within 20% of the solutions to Eq. (\[eq:Boltz\]b) for ${m_\chi > 0.17\, \mathrm{GeV}}$ and $T_{\mathrm{RH}}< m_\chi /3.9$. If ${m_\chi \leq 0.17\, \mathrm{GeV}}$, we need to take into consideration the evolution of $g_{*s}(T)$ to accurately calculate the relic abundance. Allowing for the evolution of $g_{*s}(T)$, $Y_{\mathrm{PP}}$ is rewritten as $$\begin{aligned} Y_{\mathrm{PP}} =& \,\, \left(\frac{45}{4\pi^3} \right)^{1/2} \frac{\langle \sigma v \rangle_s m_{pl}}{T_{\mathrm{RH}}^2 g^{1/2}_{*{\mathrm{RH}}}} \frac{T_\mathrm{MAX}^3}{m_\chi^6} \, g_{*s\mathrm{MAX}} \, g_{*s{\mathrm{RH}}} \nonumber \\ &\times \left(\int_{0}^{1} n^2_{\chi,\mathrm{eq}} \, x^7 \, g^{-2}_{*s}(m_\chi /x) \, dx \, \, + \right. \nonumber \\ &\left. \int_{1}^{x_\mathrm{PP}} n^2_{\chi,\mathrm{eq}} \, x^{5} \, g^{-2}_{*s}(m_\chi /x) \, dx \right). \label{eq:IntegralEndPairProduction}\end{aligned}$$ Using this expression for $Y_{\mathrm{PP}}$ and a scaling factor of $0.4$, we construct a modified relic abundance expression that takes into account the evolution of $g_{*s}(T)$. For freeze-in scenarios with ${m_\chi \leq 0.17\, \mathrm{GeV}}$ and $T_{\mathrm{RH}}< m_\chi /3.9$, this updated expression for $Y_{\mathrm{PP}}$ brings the analytic relic abundance solutions to within $25\%$ of the numeric solution to Eq. (\[eq:Boltz\]b). Constraints On Kination {#sec:Constraints} ======================= To constrain kination cosmologies, we first solve Eqs. (\[eq:FreezeoutRelicAbundance\]) and (\[eq:FreezeInRelicAbundance\]) for all combinations of the variables $m_{\chi}$, $T_{{\mathrm{RH}}}$, and $\langle \sigma v \rangle$ that produce the observed dark matter abundance of ${\Omega_{\chi} h^2 = 0.12}$ [@Planck:2015]. We set the minimum allowed reheat temperature to ${3 \, \mathrm{MeV}}$ to ensure that the period of kination does not alter the cosmic microwave background or the abundances of light elements [@Kawasaki:1999; @Kawasaki:2000; @Hannestad:2004; @Ichikawa:2005; @Ichikawa:2006].[^2] Next, we compare our allowed parameters to current constraints on $m_\chi$ and $\langle \sigma v \rangle$ from Fermi-LAT and H.E.S.S. Specifically, we use the Fermi-LAT PASS-8 constraints from observations of dwarf spheroidals [@Fermi:Constraints] and H.E.S.S. constraints from observations of the Galactic Center [@HESS:Constraints]. The Fermi-LAT data covers dark matter masses ranging from ${2 \, \mathrm{GeV} \leq m_{\chi} \leq 10^4 \, \mathrm{GeV}}$, while the H.E.S.S. data covers dark matter masses ranging from ${125 \, \mathrm{GeV} \leq m_{\chi} \leq 7 \times 10^4 \, \mathrm{GeV}}$. Figure \[NonEqConstraints\] shows the allowed parameter space for $m_\chi$ and $\langle \sigma v \rangle_s$ for scenarios in which dark matter freezes in during kination. To ensure that freeze-in occurs before the onset of radiation domination, we have restricted ourselves to ${m_{\chi}/3.9 > T_{{\mathrm{RH}}}}$. This restriction comes from the fact that the temperature at which pair production effectively ceases is $T_\mathrm{PP} = m_\chi /3.9$. Since the minimum reheat temperature is $3 \, \mathrm{MeV}$, we require that $m_\chi > 0.012 \, \mathrm{GeV}$ to ensure that freeze-in occurs before radiation domination. From Eq. (\[eq:MinForbiddenCrossSection\]) and Figure \[NonEqConstraints\], we see that scenarios in which dark matter freezes in during kination require ${\langle \sigma v \rangle < 2.7\times 10^{-38} {\,\mathrm{cm^3\,s^{-1}}}}$. Figure \[EqConstraints\] shows the allowed parameter space for $m_\chi$ and $\langle \sigma v \rangle$ for scenarios in which dark matter freezes out during kination. To obtain the observed dark matter abundance, freeze-out scenarios during kination must have an annihilation cross section greater than ${3\times10^{-26} {\,\mathrm{cm^3\,s^{-1}}}}$. As discussed in Section \[sec:FreezeOut\], freeze-out occurs earlier during kination than during radiation domination, and to compensate, freeze-out scenarios during kination require larger annihilation cross sections to generate the same dark matter density. In Figure \[EqConstraints\] we include the Fermi-LAT [@Fermi:Constraints] and H.E.S.S. [@HESS:Constraints] constraints for dark matter that annihilates via the $b\overline{b}$ channel. The Fermi-LAT bounds cover a range of dark matter masses from the mass of the bottom quark to a mass of ${10^4 \, \mathrm{GeV}}$. The H.E.S.S. bounds add additional constraints to dark matter masses ranging from ${200 \, \mathrm{GeV}}$ to ${7 \times 10^4 \, \mathrm{GeV}}$. Dark matter annihilation cross sections above the observational bounds are ruled out as these signals would have already been observed. Figure \[EqConstraints\] also includes the partial-wave unitarity bound, which requires ${\langle \sigma v \rangle \lesssim 1/m_\chi^2}$ [@Griest:1989; @Hui:2001; @Guo:2009]. We see from Figure \[EqConstraints\] that the unitarity bound rules out all kination scenarios with $\langle \sigma v \rangle$ values larger than $4.5\times10^{-23} {\,\mathrm{cm^3\,s^{-1}}}$. In addition, if dark matter annihilates via the $b\overline{b}$ channel, Fermi-LAT and H.E.S.S. observations constrain $\langle \sigma v \rangle$ to be less than $2 \times 10^{-25} {\,\mathrm{cm^3\,s^{-1}}}$ and $T_{\mathrm{RH}}$ to be greater than $1 \, \mathrm{GeV}$. ![Allowed freeze-in parameter space for $m_\chi$ and $\langle \sigma v \rangle_s$. Equation (\[eq:FreezeInRelicAbundance\]) is not applicable to scenarios with ${m_{\chi}/3.9 < T_{{\mathrm{RH}}}}$ because more than 1% of pair production occurs during radiation domination. Dark matter produced via freeze-in requires very small annihilation cross sections in order to reach the observed dark matter abundance. These small annihilation cross sections are not constrainable with current astrophysical observations.[]{data-label="NonEqConstraints"}](KinationConstraints_NonEquilibrium_Color.eps){width="3.4in" height="2.8in"} ![Allowed freeze-out parameter space for $m_\chi$ and $\langle \sigma v \rangle$. To obtain the observed dark matter abundance, scenarios with annihilation cross sections smaller than $3\times10^{-26} {\,\mathrm{cm^3\,s^{-1}}}$ require freeze-out to occur during radiation domination. The short dashed line represents the H.E.S.S. [@HESS:Constraints] constraints for annihilation in the $b\overline{b}$ channel, and the medium dashed line is the Fermi-LAT [@Fermi:Constraints] constraints, also for the $b\overline{b}$ annihilation channel. The long dashed line is the unitarity bound: ${\langle \sigma v \rangle \lesssim 1/m_\chi^2}$. All of the kination scenarios above the Fermi-LAT and H.E.S.S. constraint lines are ruled out as dark matter annihilations would have already been detected by the corresponding observations.[]{data-label="EqConstraints"}](KinationConstraints_Equilibrium_Color.eps){width="3.4in" height="2.8in"} Figure \[Fig:KinationConstraints\] shows the Fermi-LAT, H.E.S.S., and unitarity constraints on $m_\chi$ and $T_{\mathrm{RH}}$ for scenarios in which dark matter freezes out during kination for various annihilation channels. For every value of $m_\chi$ and $T_{{\mathrm{RH}}}$ we calculate the $\langle \sigma v \rangle$ that will produce the observed dark matter abundance via freeze-out using Eq. (\[eq:FreezeoutRelicAbundance\]). If the calculated $\langle \sigma v \rangle$ is above the Fermi-LAT or H.E.S.S. constraints, then that scenario is ruled out. The ruled-out area below ${T_{{\mathrm{RH}}} = 3 \, \mathrm{MeV}}$ represents the fact that, in order to produce the correct abundance of light elements, reheating must occur before a temperature of ${\sim 3 \, \mathrm{MeV}}$. The solid black line represents when ${m_\chi = 100 \, T_{{\mathrm{RH}}}}$. As discussed in Section \[sec:FreezeOut\], Eq. (\[eq:FreezeoutRelicAbundance\]) is accurate for ${T_{\mathrm{RH}}< m_\chi/100}$. As $T_{\mathrm{RH}}$ increases beyond $m_\chi/100$, numerical tests with ${m_\chi > 17 \, \mathrm{GeV}}$ indicate that the $\langle \sigma v \rangle$ value that yields the observed dark matter abundance rapidly decreases to $3\times 10^{-26} {\,\mathrm{cm^3\,s^{-1}}}$ as freeze-out occurs closer to radiation domination. Therefore, we make the conservative assumption that ${\langle \sigma v \rangle = 3\times 10^{-26} {\,\mathrm{cm^3\,s^{-1}}}}$ will give the observed dark matter abundance if ${m_\chi > 17 \, \mathrm{GeV}}$ and ${T_{\mathrm{RH}}> m_\chi/100}$. For ${m_\chi < 17 \, \mathrm{GeV}}$, numerical tests show that Eq. (\[eq:FreezeoutRelicAbundance\]) remains accurate for reheat temperatures slightly higher than ${m_\chi/100}$ if reheating occurs after the QCD phase transition. The QCD phase transition causes a sharp decrease in $g_*$ when ${T= 0.17 \, \mathrm{GeV}}$, and since ${T_\mathrm{RH} > 3 \, \mathrm{MeV}}$, ${g_{*s\mathrm{RH}} = g_{*\mathrm{RH}}}$ and ${H \propto g_{*s\mathrm{RH}}^{-1/2}}$ during kination. Consequently, the Hubble parameter at a given temperature during kination sharply increases as $T_\mathrm{RH}$ goes below ${0.17 \, \mathrm{GeV}}$, which causes freeze-out to occur earlier. Therefore, Eq. (\[eq:FreezeoutRelicAbundance\]) is applicable for scenarios with $T_\mathrm{RH}$ slightly higher than $m_\chi/100$ if ${m_\chi < 17 \, \mathrm{GeV}}$ because freeze-out still occurs during kination. For most annihilation channels, scenarios that generate the observed relic abundance with ${m_\chi <17 \, \mathrm{GeV}}$ are ruled out by Fermi-LAT constraints. The exception is dark matter annihilating via $\mu^+ \mu^-$. Figure \[Fig:KinationConstraints\] shows that, for dark matter annihilating via $\mu^+ \mu^-$, Fermi-LAT constraints rule out all scenarios with ${m_\chi \lesssim 8 \, \mathrm{GeV}}$. In addition, if $m_\chi$ is between ${8 \, \mathrm{GeV}}$ and ${17 \, \mathrm{GeV}}$, Fermi-LAT constraints rule out scenarios with ${T_{\mathrm{RH}}\lesssim 0.17 \, \mathrm{GeV}}$. In these scenarios, freeze-out occurs during kination even though $T_{\mathrm{RH}}$ may be higher than ${m_\chi /100}$. As $T_{\mathrm{RH}}$ increases beyond ${0.17 \, \mathrm{GeV}}$, numerical tests with ${8 \, \mathrm{GeV} < m_\chi < 17 \, \mathrm{GeV}}$ indicate that the ${\langle \sigma v \rangle}$ value required to obtain the observed dark matter abundance rapidly decreases to $3\times 10^{-26} {\,\mathrm{cm^3\,s^{-1}}}$ as freeze-out occurs closer to radiation domination. The resulting constraints on $m_\chi$ and $T_{\mathrm{RH}}$ are contingent on dark matter reaching thermal equilibrium during kination. Equation (\[eq:CrossSectionBeginningFreezeOut\]) indicates that decreasing $T_{\mathrm{RH}}$ and increasing $g_{*s}(m_\chi/1.5)$ increases the minimum value of $\langle \sigma v \rangle$ that results in dark matter reaching thermal equilibrium. Therefore, solving Eq. (\[eq:CrossSectionBeginningFreezeOut\]) with the minimum reheat temperature of $3 \, \mathrm{MeV}$ and ${g_{*s}(m_\chi/1.5) = 100}$ shows that, if ${\langle \sigma v \rangle > 3\times 10^{-31} {\,\mathrm{cm^3\,s^{-1}}}}$, dark matter will freeze out during kination regardless of $T_{\mathrm{RH}}$ or $m_\chi$. The Fermi-LAT and unitarity constraints establish an allowed mass range for each annihilation channel. The unitarity bound on $\langle \sigma v \rangle$ places an upper bound on the allowed dark matter mass of ${1.9 \times 10^4 \, \mathrm{GeV}}$ for all annihilation channels. The lower bound on the dark matter mass comes from the Fermi-LAT observations and is between ${8 \, \mathrm{GeV}}$ and ${160 \, \mathrm{GeV}}$, depending on the annihilation channel. As $T_{\mathrm{RH}}$ decreases, the range of viable masses decreases. The addition of the H.E.S.S. constraints restrict dark matter annihilating via $\tau^+ \tau^-$ to have a mass around either $250 \, \mathrm{GeV}$ or $9000 \, \mathrm{GeV}$. For dark matter masses between $470 \, \mathrm{GeV}$ and $2500 \, \mathrm{GeV}$, the H.E.S.S. observations constrain $\langle \sigma v \rangle$ to be less than $3\times 10^{-26} {\,\mathrm{cm^3\,s^{-1}}}$ for dark matter annihilating via $\tau^+ \tau^-$, thereby ruling out all scenarios where dark matter freezes out during kination or radiation domination. Figure \[Fig:KinationConstraints\] also shows that with the Fermi-LAT, H.E.S.S., and unitarity constraints we can place lower limits on $T_{\mathrm{RH}}$ if dark matter reaches thermal equilibrium during kination. For example, we can rule out kination scenarios with reheat temperatures below ${0.05 \, \mathrm{GeV}}$ for dark matter annihilating via the $e^+e^-$ or $\mu^+ \mu^-$ annihilation channel. We are also able to rule out kination scenarios with reheat temperatures below ${0.6 \, \mathrm{GeV}}$ for the $\tau^+ \tau^-$ and $u\overline{u}$ annihilation channels as well as reheat temperatures below ${1 \, \mathrm{GeV}}$ for the $b\overline{b}$ and $W^+W^-$ annihilation channels. In addition, kination scenarios where dark matter annihilates via the $b\overline{b}$, $\tau^+ \tau^-$, or $W^+W^-$ annihilation channel require $T_{\mathrm{RH}}$ be very close to $T_{\mathrm{F}}$, which implies that these kination scenarios are on the verge of being ruled out. Throughout this work, we assumed that dark matter consisted of one particle species. If dark matter consists of multiple particle species, then it is possible that only a fraction of the dark matter is thermally produced during kination. To determine what effect this has on the $T_\mathrm{RH}$ constraints shown in Figure \[Fig:KinationConstraints\], we neglect the ln${(T_\mathrm{F}/T_{\mathrm{RH}})}$ term in Eq. (\[eq:FreezeoutRelicAbundance\]) and make the rough estimate that for dark matter freezing out during kination, ${\Omega_\chi h^2 \propto m_\chi/\left(\langle \sigma v \rangle \, T_\mathrm{RH}\right)}$. If only a fraction of the dark matter consists of a thermal relic, such that ${\Omega_\chi h^2 = f \,\Omega_{dm} h^2}$, then to scale the relic abundance by a factor of $f$ for a fixed dark matter mass requires scaling the product of $\langle \sigma v \rangle$ and $T_\mathrm{RH}$ by a factor of $1/f$. In addition, since the Fermi-LAT and H.E.S.S. constraints are obtained using the dark matter annihilation rate $\Gamma = \langle \sigma v \rangle \rho_{\chi}^2 / m_\chi^2$, altering $\rho_{\chi}$ will subsequently reduce the annihilation rate by a factor of $f^2$ and raise the maximum allowed annihilation cross section $\langle \sigma v \rangle_\mathrm{max}$ by a factor of $1/f^2$. Therefore, the minimum allowed reheat temperature ${T_\mathrm{RH,min} \propto \langle \sigma v \rangle T_\mathrm{RH} / \langle \sigma v \rangle_{\mathrm{max}} \propto f^{-1}/f^{-2}\propto f}$. For example, for dark matter annihilating via $W^+W^-$, the minimum allowed reheat temperature is ${1 \, \mathrm{GeV}}$ if dark matter consists of a single particle species. If only a fraction of dark matter is thermally produced during kination and $\Omega_{\chi}h^2 = 0.05$, then $f = 0.42$ and the new minimum allowed reheat temperature is roughly $0.34 \, \mathrm{GeV}$. {width=".45\textwidth"} {width=".45\textwidth"} {width=".45\textwidth"} {width=".45\textwidth"} {width=".45\textwidth"} {width=".45\textwidth"} Conclusion {#sec:end} ========== Our uncertainty regarding the expansion history of the Universe between the end of inflation and the beginning of BBN allows for the possibility that within this period there was an intermittent era of kination. In this paper we have investigated the effects that a period of kination has on the thermal production of dark matter. Previous studies on this topic have required the use of specific kinaton potentials [@Pallis:2005; @Pallis:2nd2005; @Pallis:2006; @Lola:2009; @Pallis:2009]. Our analysis is independent of the kinaton potential, and therefore our constraints on $m_\chi$, $T_{\mathrm{RH}}$, and $\langle \sigma v \rangle$ are applicable to all kination models assuming that dark matter consists of one particle species that undergoes *s*-wave annihilation. In addition to numerically solving for the dark matter relic abundance, we have also derived analytic relic abundance equations for freeze-out and freeze-in kination scenarios. Our scenarios are determined by three parameters: the dark matter mass $m_\chi$, the reheat temperature $T_{\mathrm{RH}}$, and the dark matter annihilation cross section $\langle \sigma v \rangle$. In deriving relic abundance equations for freeze-out (Eq. \[eq:FreezeoutRelicAbundance\]) and freeze-in (Eq. \[eq:FreezeInRelicAbundance\]) scenarios, we have deduced physical relationships between our parameters and the dark matter relic abundance. For example, at a given temperature, the Hubble parameter during kination is higher than that during radiation domination. Therefore, freeze-out occurs earlier during kination, which increases the relic abundance. In order to compensate for this larger relic abundance, freeze-out scenarios require larger-than-canonical $\langle \sigma v \rangle$ values in order to increase the annihilation rate and subsequently decrease the dark matter abundance to the observed value. If on the other hand, dark matter freezes in during kination, an increase in $T_{{\mathrm{RH}}}$ will increase the relic abundance. Increasing $T_{{\mathrm{RH}}}$ decreases $\rho_{\phi}$, which decreases the Hubble parameter. Decreasing the Hubble parameter decreases the cooling rate, leaving more time for pair production and thereby increasing the relic abundance. To compensate for the increased cooling rate and bring the dark matter abundance into agreement with the observed value, freeze-in scenarios during kination require larger $\langle \sigma v \rangle$ values compared to during radiation domination. Overall, to reach the observed dark matter abundance, freeze-out scenarios during kination require $\langle \sigma v \rangle$ values that would underproduce dark matter during radiation domination, whereas freeze-in scenarios require $\langle \sigma v \rangle$ values that would overproduce dark matter during radiation domination. Therefore, the possibility that dark matter was thermally produced during kination significantly widens the field of potential dark matter candidates. In particular, thermally produced Winos and Higgsinos, which generally have $\langle \sigma v \rangle > 3\times10^{-26} {\,\mathrm{cm^3\,s^{-1}}}$, could constitute all the dark matter if they freeze out during a period of kination. Our analytic relic abundance equations allow us to efficiently determine the dark matter parameter space that would result in the observed dark matter abundance. To ensure that freeze-out occurs before reheating and that reheating occurs at a temperature above $3 \, \mathrm{MeV}$, $\langle \sigma v \rangle$ values between ${2.7 \times 10^{-38} {\,\mathrm{cm^3\,s^{-1}}}}$ and ${3\times 10^{-26} {\,\mathrm{cm^3\,s^{-1}}}}$ are forbidden for all dark matter masses. Using the PASS 8 Fermi-LAT observations of dwarf spheroidal galaxies, H.E.S.S. observations of the Galactic Center, and the unitarity bound on $\langle \sigma v \rangle$ we further constrain kination models. The observational bounds and subsequent constraints only apply to freeze-out scenarios because the required $\langle \sigma v \rangle$ values for freeze-in scenarios are below observational thresholds. From the unitarity constraint, we were able to rule out all kination scenarios with $\langle \sigma v \rangle$ greater than $4.5\times10^{-23} {\,\mathrm{cm^3\,s^{-1}}}$. These constraints also allowed us to rule out kination scenarios with reheat temperatures below ${0.05 \, \mathrm{GeV}}$ for dark matter annihilating via the $e^+e^-$ or $\mu^+ \mu^-$ annihilation channel. Similarly, we ruled out kination scenarios with reheat temperatures below ${0.6 \, \mathrm{GeV}}$ for the $\tau^+ \tau^-$ and $u\overline{u}$ annihilation channels as well as reheat temperatures below ${1 \, \mathrm{GeV}}$ for the $b\overline{b}$ and $W^+W^-$ annihilation channels. Since these new bounds on $T_{\mathrm{RH}}$ are below the electroweak phase transition, kination could facilitate baryogenesis [@Joyce:1996]. These bounds on $T_{\mathrm{RH}}$ are contingent on dark matter freezing out during kination. We have shown that, if ${\langle \sigma v \rangle > 3\times 10^{-31} {\,\mathrm{cm^3\,s^{-1}}}}$, dark matter will freeze out during kination regardless of $T_{\mathrm{RH}}$ or $m_\chi$. We also note that we only consider *s*-wave dark matter annihilation. If we consider a *p*-wave process, the annihilation rate in the galaxy would be suppressed relative to the annihilation rate at freeze-out and our bounds would no longer apply. We have shown that scenarios in which dark matter is thermally produced during kination are not ruled out by current observational limits on the dark matter annihilation cross section. In these scenarios, ${\langle \sigma v \rangle > 3\times 10^{-26} {\,\mathrm{cm^3\,s^{-1}}}}$ is required to generate the observed dark matter abundance via a freeze-out process. The observed dark matter abundance can be generated by a freeze-in process if ${\langle \sigma v \rangle < 2.7\times 10^{-38} {\,\mathrm{cm^3\,s^{-1}}}}$. Therefore, our uncertainty regarding the pre-BBN expansion history prevents us from knowing the dark matter annihilation cross section that yields the current dark matter abundance; there exists a degeneracy between the allowed values of $\langle \sigma v \rangle$ and $T_{\mathrm{RH}}$ that cannot be eliminated by relic abundance calculations alone. One possible approach to breaking this degeneracy involves studying the evolution of the dark matter perturbations during kination. The evolution of perturbations during kination will impact the small-scale matter power spectrum. Therefore, studying how the small-scale matter power spectrum and small-scale structure formation are affected by kination may provide a means to further constrain kination scenarios and reduce the range of viable annihilation cross sections. *Note added*: While we were finishing this paper, Ref. [@DEramo:2017] appeared on the arXiv. This paper also considers dark matter production during kination, and it similarly identifies the logarithmic decrease of the dark matter number density between freeze-out and reheating. Acknowledgments {#acknowledgments .unnumbered} =============== We thank Carisa Miller for her comments on our manuscript. This work was supported by NSF Grant No. PHY-1417446. [51]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , ****, (). , ****, (). , ****, (). , , , ****, (), . , , , ****, (), . , ****, (), . , , , ****, (), . , , , ****, (), . , , , ****, (), . , , , ****, (), . , , , , ****, (), . , , , ****, (), . , ****, (). , , , ****, (), . , ****, (). , , , ****, (), . , , , ****, (), . , ****, (), . , , , , ****, (), . , ****, (), . , ****, (), . , , , , ****, (), . , ****, (), . , ****, (), . , ****, (), . , ****, (), . , ****, (), . , ****, (), . , ****, (), . , , , , ****, (), . , , , ****, (), . , , , ****, (), . , ****, (), . , , , , (), . (), ****, (), . (), ****, (), . , ****, (), . , ****, (), . , ****, (), . , in ** (), pp. , . , , , , ****, (), . , , , ****, (), . , ****, (), . () (), . , , , , ****, (), . , ****, (), . , , , ****, (), . , ****, (). , ****, (), . , ****, (), . , , (), . [^1]: Throughout this work, we assume that the dark matter is composed of Majorana particles, and therefore ${\chi = \bar{\chi}}$. [^2]: These constraints on the reheat temperature were derived assuming that the radiation-dominated era was preceded by an early-matter-dominated era, but we expect that similar constraints would apply to kination.

ArXiv

--- abstract: 'We show the closed-form solution to the maximization of $\mathop{\mathrm{tr}}({\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{R}}})$, where $\ensuremath{\mathbf{A}}$ is given and $\ensuremath{\mathbf{R}}$ is unknown rotation matrix. This problem occurs in many computer vision tasks involving optimal rotation matrix estimation. The solution has been continuously reinvented in different fields as part of specific problems. We summarize the historical evolution of the problem and present the general proof of the solution. We contribute to the proof by considering the degenerate cases of $\ensuremath{\mathbf{A}}$ and discuss the uniqueness of $\ensuremath{\mathbf{R}}$.' author: - | Andriy Myronenko and Xubo Song\ [*Dept. of Science and Engineering, School of Medicine*]{}\ [*Oregon Health and Science University, Portland, OR*]{}\ [homepage:]{} [www.csee.ogi.edu/myron]{}\ [email:]{} [myron@csee.ogi.edu]{}\ title: 'On the closed-form solution of the rotation matrix arising in computer vision problems' --- Introduction {#sec:intro} ============ Many computer vision problems that require estimation of the optimal rotation matrix reduce to the maximization of ${\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{R}}})$[^1] for a given matrix ${\ensuremath{\mathbf{A}}}$: $$\label{eq:main} \text{max}\ {\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{R}}}), \ \ \text{s.t.}\ {\ensuremath{\mathbf{R}}}^T{\ensuremath{\mathbf{R}}}={\ensuremath{\mathbf{I}}},\ \det({\ensuremath{\mathbf{R}}})=1.$$ For instance, to estimate the closest rotation matrix ${\ensuremath{\mathbf{R}}}$ to the given matrix ${\ensuremath{\mathbf{A}}}$, we can minimize: $$\begin{gathered} \text{min}\ {\left\lVert{\ensuremath{\mathbf{R}}}-{\ensuremath{\mathbf{A}}}\right\rVert}^2_F={\mathop{\mathrm{tr}}}(({\ensuremath{\mathbf{R}}}-{\ensuremath{\mathbf{A}}})^T({\ensuremath{\mathbf{R}}}-{\ensuremath{\mathbf{A}}}))={\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{R}}}^T{\ensuremath{\mathbf{R}}}+{\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{A}}})-2{\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{R}}})=\\ {\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{I}}}+{\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{A}}})-2{\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{R}}})= -2{\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{R}}})+\text{const.} \end{gathered}$$ which is equivalent to the problem in Eq. \[eq:main\]. Historically, matrix ${\ensuremath{\mathbf{R}}}$ was first constrained to be only orthogonal ($\det({\ensuremath{\mathbf{R}}})=\pm 1$), which includes rotation and flip. A brief list of the optimization problems that simplify to the maximization of ${\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{R}}})$ include: - $\text{min}\ {\left\lVert{\ensuremath{\mathbf{R}}}-{\ensuremath{\mathbf{A}}}\right\rVert}^2_F$ : the closest orthogonal approximation problem [@Fan55; @Higham86],\ - $\text{min}\ \sum_i ({\ensuremath{\mathbf{x}}}_i-{\ensuremath{\mathbf{R}}}{\ensuremath{\mathbf{y}}}_i)^2 = {\left\lVert{\ensuremath{\mathbf{X}}}-{\ensuremath{\mathbf{R}}}{\ensuremath{\mathbf{Y}}}\right\rVert}^2_F$ : orthogonal Procrustes problem [@Schonemann66; @Hanson81], where ${\ensuremath{\mathbf{X}}}=({\ensuremath{\mathbf{x}}}_1,\dots,{\ensuremath{\mathbf{x}}}_N)^T,{\ensuremath{\mathbf{Y}}}=({\ensuremath{\mathbf{y}}}_1,\dots,{\ensuremath{\mathbf{y}}}_N)^T$ are matrices whose columns are formed from the point position vectors,\ - $\text{min}\ \sum_i ({\ensuremath{\mathbf{x}}}_i-(s{\ensuremath{\mathbf{R}}}{\ensuremath{\mathbf{y}}}_i+{\ensuremath{\mathbf{t}}}))^2 = {\left\lVert{\ensuremath{\mathbf{X}}}-(s{\ensuremath{\mathbf{R}}}{\ensuremath{\mathbf{Y}}}+{\ensuremath{\mathbf{T}}})\right\rVert}^2_F$ : Absolute orientation problem (generalized Procrustes problem) [@Arun87; @Umeyama91], where $s$ is a scaling constant and ${\ensuremath{\mathbf{t}}}$,${\ensuremath{\mathbf{T}}}$ are translation vector and matrix respectively,\ - $\text{max}\ {\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{R}}}^T{\ensuremath{\mathbf{A}}})$: Scott and Longuet-Higgins [@ScottHiggins91] correspondence estimation, where ${\ensuremath{\mathbf{A}}}$ is a proximity matrix. The Lemma ========= \[lem:trace\_AR\] Let ${\ensuremath{\mathbf{R}}}_{D\times D}$ be an unknown rotation matrix and ${\ensuremath{\mathbf{A}}}_{D\times D}$ be a known real square matrix. Let ${\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{S}}}{\ensuremath{\mathbf{V}}}^T$ be a Singular Value Decomposition (SVD) of ${\ensuremath{\mathbf{A}}}$, where ${\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{U}}}^T={\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{V}}}^T={\ensuremath{\mathbf{I}}}, {\ensuremath{\mathbf{S}}}={\mathop{\mathrm{d}}}(s_i), s_1 \geq s_2 \geq,\dots,\geq s_D,\geq 0.$ Then the optimal rotation matrix ${\ensuremath{\mathbf{R}}}$ that maximizes ${\mathop{\mathrm{tr}}}{({\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{R}}})}$ is $$\label{eq:lemtoprove} {\ensuremath{\mathbf{R}}}={\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{C}}}{\ensuremath{\mathbf{V}}}^T,\ \ \text{where} \ {\ensuremath{\mathbf{C}}}={\mathop{\mathrm{d}}}(1,1,\dots,1,\det({\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{V}}}^T)).$$ Matrix ${\ensuremath{\mathbf{R}}}$ is unique for any ${\ensuremath{\mathbf{A}}}$, except for two cases: 1. ${\mathop{\mathrm{rank}}}({\ensuremath{\mathbf{A}}}) < D-1$, 2. $\det({\ensuremath{\mathbf{A}}}) < 0$ and the smallest singular value, $s_D$, is not distinct. History of the problem ====================== The lemma has been reinvented repeatedly in various formulations in various fields. Historically, the problem was constrained to be only orthogonal. Here, we try to summarize the historical flow of the problems and its solutions that include the lemma. In 1952, Green [@Green52] showed the solution to orthogonal Procrustes problem in the special case of the full rank positive definite ${\ensuremath{\mathbf{A}}}$, where ${\ensuremath{\mathbf{R}}}$ is orthogonal. In 1966, Sch[ö]{}nemann [@Schonemann66] generalized the Green’s solution to the arbitrary ${\ensuremath{\mathbf{A}}}$ and discussed the uniqueness of ${\ensuremath{\mathbf{R}}}$. In 1981, Hanson and Norris [@Hanson81] presented the solution for strictly rotation matrix ${\ensuremath{\mathbf{R}}}$. Unfortunately, this work has not received the widespread attention. In the context of the closest orthogonal approximation problem, similar solution has been independently found in 1955 by Fan and Hoffman using *polar decomposition* [@Fan55; @Higham86]. In 1987, Arun et al. [@Arun87] presented the solution to the absolute orientation problem, and re-derived the lemma for orthogonal ${\ensuremath{\mathbf{R}}}$, presumably being not aware of the earlier works. In the same year, similar to Arun’s solution was independently obtained by Horn et al. [@Horn87]. In 1991, based on the Arun’s work, Umeyama [@Umeyama91] presented the proof for the optimal strictly rotational matrix, once again, being not aware of Hanson and Norris, and Sch[ö]{}nemann works. As we shall show, Umeyama did not consider all possible solutions, specifically for the degenerate cases of ${\ensuremath{\mathbf{A}}}$, which makes his proof slightly incomplete. Here, we prove the lemma in general case, mainly, following the Umeyama’s work [@Umeyama91]. In particular, we shall also consider the degenerate cases where ${\ensuremath{\mathbf{A}}}$ has not-distinct singular values, which was only briefly mentioned by Hanson and Norris [@Hanson81], but otherwise, to our best knowledge, never considered for the estimation of the optimal rotation matrix ${\ensuremath{\mathbf{R}}}$. Proof of the Lemma ================== We convert the constrained optimization problem into unconstrained using Lagrange multipliers. Define an objective function $f$ to be minimized as $$\text{min} \ f({\ensuremath{\mathbf{R}}})= -{\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{R}}})+{\mathop{\mathrm{tr}}}\left(({\ensuremath{\mathbf{R}}}^T{\ensuremath{\mathbf{R}}}-{\ensuremath{\mathbf{I}}})\Lambda\right)+\lambda(\det({\ensuremath{\mathbf{R}}})-1),$$ where $\Lambda$ is a symmetric matrix of unknown Lagrange multipliers and $\lambda$ is another unknown Lagrange multiplier. Equating to zero the partial derivatives of $f$ with respect to ${\ensuremath{\mathbf{R}}}$, we obtain the following system of equations: $$\label{eq:lebeq_deriv} \frac{\partial f}{\partial {\ensuremath{\mathbf{R}}}}=-{\ensuremath{\mathbf{A}}}+{\ensuremath{\mathbf{R}}}\Lambda+\lambda{\ensuremath{\mathbf{R}}}= {\ensuremath{\mathbf{R}}}{\ensuremath{\mathbf{B}}}-{\ensuremath{\mathbf{A}}}=0.$$ where ${\ensuremath{\mathbf{B}}}$ is symmetric by construction: ${\ensuremath{\mathbf{B}}}=\Lambda+\lambda{\ensuremath{\mathbf{I}}}$. Thus we need to solve a linear system of equations: $$\label{eq:polarproblem2} {\ensuremath{\mathbf{A}}}={\ensuremath{\mathbf{R}}}{\ensuremath{\mathbf{B}}}, \ \ \text{s.t.} \ {\ensuremath{\mathbf{R}}}^T{\ensuremath{\mathbf{R}}}={\ensuremath{\mathbf{I}}},\ \det({\ensuremath{\mathbf{R}}})=1.$$ Transposing Eq. \[eq:polarproblem2\] and multiplying from both sides we obtain: $$\label{eq:polarproblem3} {\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{A}}}={\ensuremath{\mathbf{B}}}^2.$$ The matrix ${\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{A}}}$ is guaranteed to be symmetric and positive definite (or semi-definite if ${\ensuremath{\mathbf{A}}}$ is singular), and we can decompose it using spectral decomposition: $$\label{eq:polarproblem4} {\ensuremath{\mathbf{B}}}^2={\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{A}}}={\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{S}}}^2{\ensuremath{\mathbf{V}}}^T,$$ where ${\ensuremath{\mathbf{S}}}^2$ is real non-negative diagonal matrix of eigenvalues of ${\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{A}}}$ as well as ${\ensuremath{\mathbf{B}}}^2$, so that $s_1^2 \geq s_2^2 \geq,\dots,\geq s_D^2,\geq 0$. Also, note that the matrix ${\ensuremath{\mathbf{S}}}$ is real non-negative diagonal matrix of the singular values of ${\ensuremath{\mathbf{A}}}$. Clearly, matrices ${\ensuremath{\mathbf{B}}}$ and ${\ensuremath{\mathbf{B}}}^2$ are both symmetric with commutative property: ${\ensuremath{\mathbf{B}}}{\ensuremath{\mathbf{B}}}^2={\ensuremath{\mathbf{B}}}^2{\ensuremath{\mathbf{B}}}$, hence both share the same eigenvectors, only when ${\ensuremath{\mathbf{B}}}^2$ is not degenerative[^2]. Thus matrix ${\ensuremath{\mathbf{B}}}$ is in the form: $${\ensuremath{\mathbf{B}}}={\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{M}}}{\ensuremath{\mathbf{V}}}^T$$ where ${\ensuremath{\mathbf{M}}}$ is real diagonal matrix with eigenvalues of ${\ensuremath{\mathbf{B}}}$, which must be in the form: ${\ensuremath{\mathbf{M}}}={\mathop{\mathrm{d}}}(\pm s_1,\pm s_2,\dots,\pm s_D)$. In the degenerate case of ${\ensuremath{\mathbf{A}}}$ (but still valid), ${\ensuremath{\mathbf{S}}}$, as well as ${\ensuremath{\mathbf{S}}}^2$, has repeated values, and matrix ${\ensuremath{\mathbf{M}}}$ does not have to be diagonal. ${\ensuremath{\mathbf{M}}}$ has symmetric block-diagonal structure with the number of blocks equal to the number of distinct values in ${\ensuremath{\mathbf{S}}}^2$. To see it happening, note that $$\begin{aligned} &{\ensuremath{\mathbf{S}}}^2{\ensuremath{\mathbf{M}}}={\ensuremath{\mathbf{V}}}^T{\ensuremath{\mathbf{B}}}^2{\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{V}}}^T{\ensuremath{\mathbf{B}}}{\ensuremath{\mathbf{V}}}={\ensuremath{\mathbf{V}}}^T{\ensuremath{\mathbf{B}}}^2{\ensuremath{\mathbf{B}}}{\ensuremath{\mathbf{V}}}={\ensuremath{\mathbf{V}}}^T{\ensuremath{\mathbf{B}}}{\ensuremath{\mathbf{B}}}^2{\ensuremath{\mathbf{V}}}={\ensuremath{\mathbf{M}}}{\ensuremath{\mathbf{S}}}^2, \Rightarrow \\ &{\ensuremath{\mathbf{S}}}^2{\ensuremath{\mathbf{M}}}-{\ensuremath{\mathbf{M}}}{\ensuremath{\mathbf{S}}}^2=0,\Rightarrow \\ &(s_i^2-s_j^2)m_{ij}=0,\ \forall i,j\end{aligned}$$ where $s_i^2$, $m_{ij}$ are the elements of ${\ensuremath{\mathbf{S}}}^2$ and ${\ensuremath{\mathbf{M}}}$ respectively. If all the $s_i^2$ are distinct, then we conclude that $m_{ij}=0,\forall i\neq j$ and ${\ensuremath{\mathbf{M}}}$ is diagonal. If not all $s_i^2$ are distinct, then $m_{ij}=0$ only if $s_i^2\neq s_j^2$, and thus ${\ensuremath{\mathbf{M}}}$ is block-diagonal formed from square symmetric blocks corresponding to repeated values $s_i$. Now, we consider the following cases separately: ${\ensuremath{\mathbf{A}}}$ is non-singular and non-degenerative (all singular values are distinct), ${\ensuremath{\mathbf{A}}}$ is non-singular and degenerative, ${\ensuremath{\mathbf{A}}}$ is singular. #### Non-degenerative case of ${\ensuremath{\mathbf{A}}}$: ${\ensuremath{\mathbf{M}}}$ is diagonal. Substituting ${\ensuremath{\mathbf{M}}}$ into equation Eq. \[eq:polarproblem2\] and then into the objective function, we obtain: $${\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{R}}})={\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{B}}}^T{\ensuremath{\mathbf{R}}}^T{\ensuremath{\mathbf{R}}})={\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{B}}})={\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{M}}}{\ensuremath{\mathbf{V}}}^T)={\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{M}}})$$ Taking into account that $\det({\ensuremath{\mathbf{R}}})=1$, from Eq. \[eq:polarproblem2\] we see that $$\det({\ensuremath{\mathbf{A}}})=\det({\ensuremath{\mathbf{R}}})\det({\ensuremath{\mathbf{B}}})=\det({\ensuremath{\mathbf{B}}})=\det({\ensuremath{\mathbf{V}}})\det({\ensuremath{\mathbf{M}}})\det({\ensuremath{\mathbf{V}}}^T)=\det({\ensuremath{\mathbf{M}}}),$$ hence $\det({\ensuremath{\mathbf{M}}})$ must have at least the same sign as $\det({\ensuremath{\mathbf{A}}})$. Clearly, matrix ${\ensuremath{\mathbf{M}}}$ that maximizes its trace is $$\begin{aligned} \label{eq:lemM1} {\ensuremath{\mathbf{M}}}&={\mathop{\mathrm{d}}}( s_1, s_2,\dots, s_D), &if \det({\ensuremath{\mathbf{A}}})>0,\\ \label{eq:lemM2} {\ensuremath{\mathbf{M}}}&={\mathop{\mathrm{d}}}( s_1, s_2,\dots, -s_D), &if \det({\ensuremath{\mathbf{A}}})<0.\end{aligned}$$ and the value of objective function at the optimum is $$\label{lem:eqvalue} {\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{R}}})={\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{M}}})=s_1+s_2+,\dots, +s_{D-1} \pm s_{D}$$ where the last sign depends on the determinant of ${\ensuremath{\mathbf{A}}}$. Now, we can find the optimal rotation matrix ${\ensuremath{\mathbf{R}}}$, from the Eq. \[eq:polarproblem2\]: $$\begin{aligned} {\ensuremath{\mathbf{A}}}&={\ensuremath{\mathbf{R}}}{\ensuremath{\mathbf{B}}},\\ {\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{S}}}{\ensuremath{\mathbf{V}}}^T&={\ensuremath{\mathbf{R}}}{\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{M}}}{\ensuremath{\mathbf{V}}}^T,\\ {\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{S}}}&={\ensuremath{\mathbf{R}}}{\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{M}}}.\end{aligned}$$ If ${\ensuremath{\mathbf{A}}}$ is non-singular (${\mathop{\mathrm{rank}}}({\ensuremath{\mathbf{A}}})=D$), then ${\ensuremath{\mathbf{M}}}$ is invertable, and the optimal ${\ensuremath{\mathbf{R}}}$ is $$\label{eq:nondegener} {\ensuremath{\mathbf{R}}}={\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{S}}}{\ensuremath{\mathbf{M}}}^{-1}{\ensuremath{\mathbf{V}}}^T={\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{C}}}{\ensuremath{\mathbf{V}}}^T, \ \ \text{where} \ {\ensuremath{\mathbf{C}}}={\mathop{\mathrm{d}}}(1,1,\dots,1,\det({\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{V}}}^T)).$$ where $\det({\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{V}}}^T)=\det({\ensuremath{\mathbf{U}}})\det({\ensuremath{\mathbf{V}}}^T)=sign(\det({\ensuremath{\mathbf{A}}}))=\pm 1$ depending on a sign of $\det({\ensuremath{\mathbf{A}}})$. #### Degenerative case of ${\ensuremath{\mathbf{A}}}$ : ${\ensuremath{\mathbf{M}}}$ is symmetric block diagonal. Since ${\ensuremath{\mathbf{M}}}$ is symmetric, it can be diagonalized using spectral decomposition: ${\ensuremath{\mathbf{M}}}={\ensuremath{\mathbf{Q}}}{\ensuremath{\mathbf{N}}}{\ensuremath{\mathbf{Q}}}^T$, where ${\ensuremath{\mathbf{Q}}}$ is orthogonal and also block-diagonal with the same block structure as ${\ensuremath{\mathbf{M}}}$. Matrix ${\ensuremath{\mathbf{N}}}$ is real and diagonal. $$\begin{aligned} {\ensuremath{\mathbf{B}}}^2&={\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{S}}}^2{\ensuremath{\mathbf{V}}}^T={\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{M}}}^2{\ensuremath{\mathbf{V}}}^T,\Rightarrow\\ {\ensuremath{\mathbf{S}}}^2&={\ensuremath{\mathbf{M}}}^2={\ensuremath{\mathbf{Q}}}{\ensuremath{\mathbf{N}}}^2{\ensuremath{\mathbf{Q}}}^T, \Rightarrow\\ {\ensuremath{\mathbf{N}}}^2&={\ensuremath{\mathbf{Q}}}^T{\ensuremath{\mathbf{S}}}^2{\ensuremath{\mathbf{Q}}}={\ensuremath{\mathbf{S}}}^2.\end{aligned}$$ The last equality holds, because ${\ensuremath{\mathbf{S}}}^2$ has multiples of identity along the diagonal, which correspond to the repeated values. Matrix ${\ensuremath{\mathbf{Q}}}$ is orthogonal block diagonal, where each block has a corresponding multiples of identity in ${\ensuremath{\mathbf{S}}}^2$. Using direct matrix multiplication you can see that ${\ensuremath{\mathbf{Q}}}^T{\ensuremath{\mathbf{S}}}^2{\ensuremath{\mathbf{Q}}}={\ensuremath{\mathbf{S}}}^2$. Thus ${\ensuremath{\mathbf{N}}}={\mathop{\mathrm{d}}}(\pm s_1,\pm s_2,\dots,\pm s_D)$, and the value of objective function at the optimum is ${\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{M}}})={\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{Q}}}{\ensuremath{\mathbf{N}}}{\ensuremath{\mathbf{Q}}}^T)={\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{N}}})$. Taking into account the sign of determinant of ${\ensuremath{\mathbf{A}}}$ we conclude that ${\ensuremath{\mathbf{N}}}$ that maximizes its trace is in the form: $$\begin{aligned} \label{eq:lemN1} {\ensuremath{\mathbf{N}}}&={\mathop{\mathrm{d}}}( s_1, s_2,\dots, s_D), &if \det({\ensuremath{\mathbf{A}}})>0,\\ \label{eq:lemN2} {\ensuremath{\mathbf{N}}}&={\mathop{\mathrm{d}}}( s_1, s_2,\dots, -s_D), &if \det({\ensuremath{\mathbf{A}}})<0.\end{aligned}$$ The objective function at the optimum is ${\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{N}}})=s_1+s_2+,\dots, +s_{D-1} \pm s_{D}$, which is exactly the same value as in Eq. \[lem:eqvalue\], when ${\ensuremath{\mathbf{M}}}$ is diagonal. Thus, in the degenerate case there is a set of block-diagonal matrices ${\ensuremath{\mathbf{M}}}$, which give the same objective function value as for the diagonal ${\ensuremath{\mathbf{M}}}$. Now, let us consider the form of ${\ensuremath{\mathbf{M}}}$ for the optimal choices of ${\ensuremath{\mathbf{N}}}$. When $\det({\ensuremath{\mathbf{A}}})>0$, from Eq. \[eq:lemN1\], we have: $${\ensuremath{\mathbf{M}}}={\ensuremath{\mathbf{Q}}}{\ensuremath{\mathbf{N}}}{\ensuremath{\mathbf{Q}}}^T={\ensuremath{\mathbf{Q}}}{\ensuremath{\mathbf{S}}}{\ensuremath{\mathbf{Q}}}^T={\ensuremath{\mathbf{S}}}={\ensuremath{\mathbf{N}}},$$ where the orthogonal matrix ${\ensuremath{\mathbf{Q}}}$ vanishes with corresponding multiples of identity in ${\ensuremath{\mathbf{S}}}$. Thus if $\det({\ensuremath{\mathbf{A}}})>0$ the optimal ${\ensuremath{\mathbf{M}}}$ is unique. In the case when $\det({\ensuremath{\mathbf{A}}})<0$ (Eq. \[eq:lemN2\]), equality ${\ensuremath{\mathbf{Q}}}{\ensuremath{\mathbf{N}}}{\ensuremath{\mathbf{Q}}}^T={\ensuremath{\mathbf{N}}}$ holds only if the smallest element $s_D$ is not repeated. If $s_D$ happen to be repeated, and $\det({\ensuremath{\mathbf{A}}})<0$, then ${\ensuremath{\mathbf{M}}}$ is not unique, and there is a set optimal solutions ${\ensuremath{\mathbf{M}}}={\ensuremath{\mathbf{Q}}}{\ensuremath{\mathbf{N}}}{\ensuremath{\mathbf{Q}}}^T$, which is unavoidable. However, even in this case, it is always possible to choose ${\ensuremath{\mathbf{M}}}$ to be diagonal (Eq. \[eq:lemM2\]). Similar to the non-degenerative case, if ${\ensuremath{\mathbf{A}}}$ is non-singular, the optimal ${\ensuremath{\mathbf{R}}}$ is found in the same way as in Eq. \[eq:nondegener\]: $$\label{eq:degener} {\ensuremath{\mathbf{R}}}={\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{S}}}{\ensuremath{\mathbf{M}}}^{-1}{\ensuremath{\mathbf{V}}}^T={\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{C}}}{\ensuremath{\mathbf{V}}}^T, \ \ \text{where} \ {\ensuremath{\mathbf{C}}}={\mathop{\mathrm{d}}}(1,1,\dots,1,\det({\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{V}}}^T)).$$ However, consider the uniqueness of SVD of ${\ensuremath{\mathbf{A}}}$ and computation of ${\ensuremath{\mathbf{R}}}$ in the degenerative case. We know that if the singular values of ${\ensuremath{\mathbf{A}}}$ are not distinct, then SVD of ${\ensuremath{\mathbf{A}}}$ is not unique. In particular, any normalized linear combination of singular vectors corresponding to the same singular value is also a singular vector. Consider SVD of degenerative ${\ensuremath{\mathbf{A}}}$: $$\begin{aligned} {\ensuremath{\mathbf{A}}}&={\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{S}}}{\ensuremath{\mathbf{V}}}^T,\\ {\ensuremath{\mathbf{U}}}&={\ensuremath{\mathbf{A}}}{\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{S}}}^{-1}\end{aligned}$$ Accordingly to Eq. \[eq:degener\]: $${\ensuremath{\mathbf{R}}}={\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{C}}}{\ensuremath{\mathbf{V}}}^T={\ensuremath{\mathbf{A}}}{\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{S}}}^{-1}{\ensuremath{\mathbf{C}}}{\ensuremath{\mathbf{V}}}^T$$ If $\det({\ensuremath{\mathbf{A}}})>0$, ${\ensuremath{\mathbf{R}}}={\ensuremath{\mathbf{A}}}{\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{S}}}^{-1}{\ensuremath{\mathbf{V}}}^T$, which means that ${\ensuremath{\mathbf{R}}}$ is unique, eventhough ${\ensuremath{\mathbf{V}}}$ is not. This is because ${\ensuremath{\mathbf{S}}}^{-1}$ has the same repeated elements as singular values in ${\ensuremath{\mathbf{S}}}$, then ${\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{S}}}^{-1}{\ensuremath{\mathbf{V}}}^T$ is a unique matrix, and thus ${\ensuremath{\mathbf{R}}}$ is uniquely determinted. If $\det({\ensuremath{\mathbf{A}}})<0$ and the smallest singular value is not distinct, then the rotation matrix ${\ensuremath{\mathbf{R}}}={\ensuremath{\mathbf{A}}}{\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{S}}}^{-1}{\ensuremath{\mathbf{C}}}{\ensuremath{\mathbf{V}}}^T$ is not unique, because different SVD of ${\ensuremath{\mathbf{A}}}$ produce different ${\ensuremath{\mathbf{V}}}$, and the matrix ${\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{S}}}^{-1}{\ensuremath{\mathbf{C}}}{\ensuremath{\mathbf{V}}}^T$ is not uniquely determined. Furthermore, even if the singular values of ${\ensuremath{\mathbf{A}}}$ are distinct but poor isolated (close to each other), a small perturbation to ${\ensuremath{\mathbf{A}}}$ can alter a singular vectors significantly [@Golub89], and thus ${\ensuremath{\mathbf{R}}}$ changes significantly as well. This means, that in case of $\det({\ensuremath{\mathbf{A}}})<0$ and degenerative ${\ensuremath{\mathbf{A}}}$ or close to degenerative, matrix ${\ensuremath{\mathbf{R}}}$ is extremely sensitive to any changes in ${\ensuremath{\mathbf{A}}}$. In particular, in this case, a round-off errors presented in computation of ${\ensuremath{\mathbf{A}}}$ and SVD of ${\ensuremath{\mathbf{A}}}$, can produce significantly different ${\ensuremath{\mathbf{R}}}$. We note, that Umeyama [@Umeyama91], in his derivation of the lemma, has not considered the case when ${\ensuremath{\mathbf{A}}}$ is degenerative. #### Singular case of ${\ensuremath{\mathbf{A}}}$: If ${\ensuremath{\mathbf{A}}}$ is singular and ${\mathop{\mathrm{rank}}}({\ensuremath{\mathbf{A}}})=D-1$ (only a single singular value is zero), then ${\ensuremath{\mathbf{M}}}={\ensuremath{\mathbf{S}}}={\mathop{\mathrm{d}}}(s_1,s_2,\dots,0)$ and $${\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{S}}}={\ensuremath{\mathbf{R}}}{\ensuremath{\mathbf{V}}}{\ensuremath{\mathbf{S}}}$$ If we define an orthogonal matrix ${\ensuremath{\mathbf{K}}}={\ensuremath{\mathbf{U}}}^T{\ensuremath{\mathbf{R}}}{\ensuremath{\mathbf{V}}}$, then $${\ensuremath{\mathbf{K}}}{\ensuremath{\mathbf{S}}}={\ensuremath{\mathbf{S}}}.$$ Since the column vectors of ${\ensuremath{\mathbf{K}}}$ are orthonormal, then they are in the form $$\begin{aligned} \label{eq:lemK1} &{\ensuremath{\mathbf{k}}}_i=(0,0,\dots,1_i,\dots,0)^T, & \text{for} \ 1\leq i\leq D-1\\ \label{eq:lemK2} &{\ensuremath{\mathbf{k}}}_D=(0,0,\dots,\pm 1)^T.\end{aligned}$$ Taking into account the constraint on determinant of ${\ensuremath{\mathbf{R}}}$, we have $$\det({\ensuremath{\mathbf{K}}})=\det({\ensuremath{\mathbf{U}}}^T)\det({\ensuremath{\mathbf{R}}})\det({\ensuremath{\mathbf{V}}})=\det({\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{V}}}^T)$$ Thus, we obtain: $$\label{eq:lemlast} {\ensuremath{\mathbf{R}}}={\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{K}}}{\ensuremath{\mathbf{V}}}^T={\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{C}}}{\ensuremath{\mathbf{V}}}^T, \ \ \text{where} \ {\ensuremath{\mathbf{C}}}={\mathop{\mathrm{d}}}(1,1,\dots,1,\det({\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{V}}}^T)).$$ Finally, if ${\ensuremath{\mathbf{A}}}$ is singular and ${\mathop{\mathrm{rank}}}({\ensuremath{\mathbf{A}}}) < D-1$ (${\ensuremath{\mathbf{A}}}$ has multiple zero singular values), then matrix ${\ensuremath{\mathbf{K}}}$ is not uniquely determined. Precisely, one can choose arbitrary last column-vectors of ${\ensuremath{\mathbf{K}}}$,(number of which is equivalent to the number of zero singular values) as far as they are orthonormal and $\det({\ensuremath{\mathbf{K}}})=\det({\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{V}}}^T)$. This gives a set of equivalent solutions for ${\ensuremath{\mathbf{R}}}$. Additional information or constraints require to make ${\ensuremath{\mathbf{R}}}$ unique. We note, that it is always possible to chose ${\ensuremath{\mathbf{K}}}$ according to Eq. \[eq:lemK1\],\[eq:lemK1\] and find ${\ensuremath{\mathbf{R}}}$ from Eq. \[eq:lemlast\]. Thus, we have considered all cases of ${\ensuremath{\mathbf{A}}}$, which concludes the lemma. Discussion and conclusion ========================= The lemma is of general interest and is usefull in many computer vision and machine learning problems that can be simplified to maximization of ${\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{R}}})$. The lemma shows the optimal solution for the rotation matrix ${\ensuremath{\mathbf{R}}}$. In most of the cases ${\ensuremath{\mathbf{R}}}$ is uniquely determined. In the case when ${\mathop{\mathrm{rank}}}({\ensuremath{\mathbf{A}}})< D-1$ and in the degenerate case, when the smallest singular value is not distinct and $\det({\ensuremath{\mathbf{A}}})<0$, the presented solution for ${\ensuremath{\mathbf{R}}}$ is still a global optimum of the function, but it is not unique. Also, we have shown, that in these degenerative cases, ${\ensuremath{\mathbf{R}}}$ is extremely sensitive to round-off errors in ${\ensuremath{\mathbf{A}}}$. In the cases when ${\ensuremath{\mathbf{R}}}$ is not unique, the solution given by Eq. \[eq:lemtoprove\] should be further justified by a particular problem. If we relax the constraint for ${\ensuremath{\mathbf{R}}}$ to be strictly rotational, and allow it to be any orthogonal (which allows for rotation and flip), then the derivation simplifies to the solution ${\ensuremath{\mathbf{R}}}={\ensuremath{\mathbf{U}}}{\ensuremath{\mathbf{V}}}^T$, which was established by Sch[ö]{}nemann [@Schonemann66], and it is unique for all non-singular ${\ensuremath{\mathbf{A}}}$. The lemma can be applied for the problems of arbitrary dimensions. [10]{} Ky Fan and Alan J. Hoffman. Some metric inequalities in the space of matrices. , 6(9):111–116, 1955. Nicholas J. Higham. Computing the polar decomposition–with applications. , 7(4):1160–1174, October 1986. Peter Sch[ö]{}nemann. A generalized solution of the orthogonal procrustes problem. , 31(1):1–10, March 1966. Richard J. Hanson and Michael J. Norris. Analysis of measurements based on the singular value decomposition. , 2(3):363–373, 1981. K.S. Arun, Thomas S. Huang, and Steven D. Blostein. Least-squares fitting of two 3-[D]{} point sets. , 9(5):698–700, 1987. Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. , 13(4):376–380, April 1991. Guy L. Scott and Christopher Longuet-Higgins. An algorithm for associating the features of two images. , 244(1309):21–26, April 1991. Bert F. Green. The orthogonal approximation of an oblique structure in factor analysis. , 17(4):429–440, December 1952. Berthold K. P. Horn, Hugh M. Hilden, and Shahrlar Negahdaripour. Closed form solutions of absolute orientation using orthonormal matrices. , 5(7):1127–1135, June 1997. Gene H. Golub and Chales F. Van Loan. . Johns Hopkins University Press, Baltimore, MD, second edition, 1989. [^1]: Matrix *trace*, ${\mathop{\mathrm{tr}}}()$, stands for a sum of diagonal elements of the matrix. ${\mathop{\mathrm{tr}}}({\ensuremath{\mathbf{A}}}^T{\ensuremath{\mathbf{R}}})$ also represents a Frobenius inner product, which is a sum of element-wise products of matrices ${\ensuremath{\mathbf{A}}}$ and ${\ensuremath{\mathbf{R}}}$. [^2]: Here by degenerative matrix we mean a matrix with not distinct (repeated) singular values. Note, that a matrix can be non-singular, but still degenerative.

ArXiv

--- abstract: 'We discuss selected topics on the topology of moduli spaces of curves and maps, emphasizing their relation with Gromov–Witten theory and integrable systems.' address: - 'Department of Mathematics, University of Utah, Salt Lake City, UT 84112-0090' - 'Department of Mathematics, Stanford University, Stanford, CA 94305-2125' author: - 'Y.-P. Lee' - 'R. Vakil' date: 'September 10, 2008.' title: Algebraic structures on the topology of moduli spaces of curves and maps --- Introduction ============ Forty years ago, Deligne and Mumford introduced their celebrated compactification of the moduli space of curves. In 1983, Mumford [@mumford] initiated a comprehensive study of intersection theory on the moduli space of curves, and in particular introduced the tautological ring. In many ways motivated by Witten’s conjecture [@witten], the last twenty years have seen a steadily growing understanding of rich algebraic structures on the cohomology of moduli spaces of curves, and related spaces, such as Gromov–Witten invariants and Hurwitz numbers. Even when the structures deal purely with the moduli space of curves, Gromov–Witten theory has provided a powerful tool for understanding them. The purpose of this article is to discuss some of these structures. [**1.**]{} [*Integrable systems.*]{} Witten’s conjecture determines all top intersections of $\psi$-classes on ${\overline{{{\mathcal{M}}}}}_{g,n}$, by showing that their generating function satisfies differential equations coming from integrable systems (the KdV hierarchy, or the Virasoro algebra). More precisely, let $$F_g = \sum_{n \geq 0} \frac 1 {n!} \sum_{k_1, \dots, k_n} \left( \int_{{\overline{{{\mathcal{M}}}}}_{g,n}} \psi_1^{k_1} \cdots \psi_n^{k_n} \right) t_{k_1} \cdots t_{k_n},$$ and let $F = \sum F_g \hbar^{2g-2}$ be the generating function for all genera. Then the KdV form of Witten’s conjecture is: $$\begin{aligned} & & (2n+1) \frac {\partial^3} {\partial t_n \partial t_0^2} F = \\ & & \quad \left( \frac {\partial^2} {\partial t_{n-1} \partial t_0} F \right) \left( \frac {\partial^3} {\partial t^3_0} F \right) + 2 \left( \frac {\partial^3} {\partial t_{n-1} \partial t_0^2} F \right) \left( \frac {\partial^2 } {\partial t_0^2} F \right) + \frac 1 4 \frac { \partial^5} {\partial t_{n-1} \partial t_0^4} F.\end{aligned}$$ Define differential operators ($n \geq -1$) $$\begin{aligned} L_{-1} &=& - \frac {\partial} {\partial t_0} + \frac {\hbar^{-2}} 2 t_0^2 + \sum_{i=0}^{\infty} t_{i+1} \frac \partial {\partial t_i} \\ L_0 &=& - \frac 3 2 \frac \partial {\partial t_1} + \sum_{i=0}^{\infty} \frac {2i+1} 2 t_i \frac {\partial} {\partial t_i} + \frac 1 {16} \\ L_n &=& \sum_{k=0}^{\infty} \frac { {\Gamma}(m+n+ \frac 3 2) }{ {\Gamma}(k+ \frac 1 2)} (t_k - {\delta}_{k,1}) \frac \partial {\partial t_{n+k}} \\ & & \quad \quad \quad + \frac {\hbar^2} 2 \sum_{k=1}^{n-1} (-1)^{k+1} \frac {{\Gamma}(n-k+ \frac 1 2)} {{\Gamma}( -k-\frac 1 2)} \frac \partial {\partial t_k} \frac \partial {\partial t_{n-k-1}} \quad \quad (n>0)\end{aligned}$$ These operators satisfy $[L_m, L_n] = (m-n) L_{m+n}$. The Virasoro form of Witten’s conjecture, due to Dijkgraaf, Verlinde, and Verlinde, is: $L_n e^F = 0$ for all $n$. These relations inductively determine the coefficients of $F$, and hence compute all intersection numbers. This unexpected relationship with integrable systems has since appeared repeatedly, and we now have a modest understanding as to why this relationship might exist. This subject is well-covered in the literature, so we content ourselves with a short discussion in §\[s:is\]. [**2.**]{} [*Faber-type Gorenstein conjectures.*]{} Based on numerical data found using Witten’s conjecture, Faber made a remarkable conjecture [@faber] on the tautological part of the cohomology of ${{\mathcal{M}}}_g$. Again, this structure was quite unexpected, and in some sense we still have little understanding “why” such a structure should exist. This conjectural structure seems to be shared by related moduli spaces. This topic is also well-covered (see for example Pandharipande’s ICM talk [@pandharipande]), so we describe just enough in §\[s:faber\] to motivate later discussion. [**3.**]{} [*Polynomiality.*]{} Proof of parts of the above conjectures often come through another “polynomiality” structure of certain invariants or intersection numbers, which is exemplified in their first appearance, the ELSV formula [@elsv1; @elsv2]. For example, the polynomiality of the ELSV formula is an easy-to-state fact about factoring permutations into transpositions in the symmetric group, but the only known proof requires the geometry of the moduli space of curves. There are other instances that look similar that are even less understood. In §\[s:pol\], we discuss the notion of polynomiality, and describe and explain a conjecture of Goulden, Jackson, and the second author that is sketched in [@gjv]: that there should be an ELSV-type formula relating certain double Hurwitz numbers to the intersection theory of some moduli space. We then describe work of Shadrin and Zvonkine [@sz] and Shadrin [@s] that these conjectural intersection numbers (arising from actual Hurwitz numbers) have much richer structure than suspected in [@gjv]. [**4.**]{} *Family topological field theory.* Madsen and Weiss’ celebrated proof [@MW] of Mumford’s conjecture on the stable (tautological) cohomology of the moduli space of curves has had an unexpected and powerful application to Gromov–Witten theory. Teleman has classified all cohomological 2D field theories based on a semi-simple complex Frobenius algebra $A$: they are controlled by a linear combination of ${\kappa}$-classes and by an extension datum to the Deligne-Mumford boundary. This leads to a proof of Givental’s conjecture that, roughly speaking, higher genus invariants are determined by genus $0$ invariants. This in turn implies the Virasoro conjecture for manifolds with semisimple quantum cohomology. We outline this important work in some detail in §\[s:teleman\]. [**5.**]{} *Witten’s conjecture on $r$-spin curves.* In the last section, we briefly outline Faber–Shadrin–Zvonkine’s proof [@FSZ] of Witten’s conjecture on $r$-spin curves, emphasizing on the parts which are most related to the tautological rings. Although this result is a consequence of Teleman’s theorem, it uses completely different ingredients and the proof itself is quite interesting in its own right. Two ingredients in their proof will be briefly explained: 1. The first author’s theorem that tautological relations hold for generating functions of axiomatic (semisimple) theories [@ypL2]. 2. Faber–Shadrin–Zvonkine’s reconstruction theorem which states that the Witten’s correlators in higher genus can be uniquely reconstructed from genus zero. Along the way, we also give some explanation of Givental’s formalism ([@aG1; @LP] and references therein) which has proved important in understanding the algebraic structure of Gromov–Witten theory. Note that the proof of the first result can be reduced to an elementary statement in the geometry of boundary divisors on moduli of curves via Givental’s theory. See Section \[s:5.3\]. [**Notation and background.**]{} We work over ${\mathbb{C}}$. General background may be found in [@hori Ch. 23-25], [@cime], and [@ypL4]. We assume the reader is familiar with the following notions: the moduli space of pointed curves ${{\mathcal{M}}}_{g,n}$ and its partial compactifications ${{\mathcal{M}}}_{g,n}^{rt}$ (curves with rational tails), ${{\mathcal{M}}}_{g,n}^{ct}$ (curves of compact type), and ${\overline{{{\mathcal{M}}}}}_{g,n}$ (Deligne-Mumford stable curves, the topic of this volume). the cohomology classes $\psi_i$ ($1 \leq i \leq n$), ${\kappa}_j$, and ${\lambda}_k$ ($0 \leq k \leq g$), and the tautological ring. We also note that [@s2] contains an excellent review of basic notions of cohomological field theories and Givental theory. Acknowledgements {#acknowledgements .unnumbered} ---------------- We would like to thank T. Jarvis, K. Liu, S. Shadrin, C. Teleman, and D. Zvonkine for their helpful comments on earlier versions of the paper. Both authors are partially supported by the NSF. Integrable systems {#s:is} ================== We content ourselves with a brief overview of Witten’s conjecture and related topics, sufficient to set the stage for later sections. It was only since the advent of Witten’s conjecture that we are able to compute all top intersections of $\psi$-classes in ${\overline{{{\mathcal{M}}}}}_{g,n}$, something we now take for granted. These allow one to compute all top intersections in the tautological ring, [@Falg]. It is also important when considering Witten’s conjecture to know not just [*that*]{} it is true, but also [*why*]{} it is true, i.e. why in retrospect integrable systems should be expected to control these top intersections. Of course, Witten’s original heuristic argument is one explanation. There are now a large number of proofs of Witten’s conjecture, and it is a sign of the richness of the conjecture that almost every proof has been fundamentally new, with fundamentally new insights. Kontsevich’s original proof ([@k], see also [@l]) remains unlike the others. Mirzakhani’s proof [@mirzakhani] (the third) gives an elegant interpretation of each summand in the Virasoro version of Witten’s conjecture. The remaining proofs pass through the ELSV formula (discussed in §\[s:pol\]), counting branched covers, and fundamentally about factoring permutations in the symmetric group. Okounkov and Pandharipande’s proof [@op] was part of their massive program in Gromov–Witten theory, including their proof of the Virasoro conjecture in dimension $1$. By a careful algebraic argument, Kim and K. Liu [@kimliu] give a direct proof of Witten’s conjecture through localization. Chen, Li, and Liu thereafter gave a different and very short proof, using ideas from [@lambdag] on the ${\lambda}_g$-conjecture (a sister conjecture to Witten’s conjecture, see the next section). Meanwhile, Kazarian and Lando also gave an algebro-geometric proof of Witten’s conjecture [@kazarianlando]. Shortly after that, also using ideas from [@lambdag], Kazarian gave a greatly streamlined proof [@kazarian]. At each stage, our understanding of the structure behind Witten’s conjectured deepened, and in some sense we now have quite a strong understanding of why it is true. One test of our understanding is how well it generalizes to other situations. We briefly mention two generalizations of Witten’s conjecture that will be relevant shortly. First, Witten gave a generalization of his conjecture to the moduli space of $r$-spin curves [@witten2]. Roughly speaking, an $r$-spin curve with marked points is an $r$th root of the canonical bundle twisted by given multiples of the points; see §\[wc\] for more detail. The Virasoro conjecture (mentioned earlier), due to Eguchi, Hori, Xiong, and also S. Katz gives Virasoro-related constraints on the Gromov–Witten invariants of a complex projective manifold [@ck §10.1.4]. The case of a point is the Virasoro form of Witten’s conjecture. More relatives of Witten’s conjecture will appear in the next section. Faber-type Gorenstein conjectures, and three sisters {#s:faber} ==================================================== Faber’s conjectures are well exposed in the literature; for example, the second author has discussed them in detail in [@cime §3.2]. So again we content ourselves with saying just enough to continue our story. Faber’s conjectures ------------------- Faber’s conjectures [@faber] describe an unexpected and in many ways still-unexplained structure on the tautological ring $R^*({{\mathcal{M}}}_g)$, which may be considered as a subring of the cohomology ring or the Chow ring (although this technically yields two different conjectures). Informally speaking, Faber’s conjectures on ${{\mathcal{M}}}_g$ state that $R^*({{\mathcal{M}}}_g)$ behaves like the cohomology ring of a $(g-2)$-dimensional complex projective manifold. Somewhat more precisely, this means: [**I.**]{} [*“Vanishing/socle” conjecture.*]{} $R^i({{\mathcal{M}}}_g) = 0$ for $i>g-2$, and $R^{g-2}({{\mathcal{M}}}_g) \cong {\mathbb{Q}}$. This was proved by Looijenga [@loo] and Faber [@faber Thm. 2]. [**II.**]{} [*Perfect pairing conjecture.*]{} The analogue of Poincaré duality holds: for $0\leq i \leq g-2$, the natural map $R^i({{\mathcal{M}}}_g) \times R^{g-2-i}({{\mathcal{M}}}_g) \rightarrow R^{g-2}({{\mathcal{M}}}_g) \cong {\mathbb{Q}}$ is a perfect pairing. This is currently open. [**III.**]{} [*Intersection number conjecture.*]{} Faber gives a formula for top (i.e. total degree $g-2$) intersections of generators of the tautological ring, as a multiple of a generator of $R^{g-2}({{\mathcal{M}}}_g)$. As we discuss in §\[twothree\], this is proved. The three conjectures above completely determine the tautological ring $R^*({{\mathcal{M}}}_g)$. There are three “sister” conjectures that parallel Faber’s, on three different spaces. See [@pandharipande] for more on the three sisters, and a more detailed history. The first sister: ${\overline{{{\mathcal{M}}}}}_{g,n}$ ------------------------------------------------------ On ${\overline{{{\mathcal{M}}}}}_{g,n}$, there is an analogous set of conjectures, with $g-2$ replaced by $3g-3+n$. This was first asked as a question by Hain and Looijenga [@hlo Question 5.5]; first stated as a speculation by Faber and Pandharipande [@FabP1 Speculation 3] (in the case $n=0$), and first stated as a conjecture by Pandharipande [@pandharipande Conj. 1]. In the cohomology ring, [ **I**]{} and [**II**]{} trivially hold, but in the Chow ring, this is far from clear. ([**I**]{} was first shown in [@gv].) Witten’s conjecture should be considered the intersection number part [ **III**]{}. The second sister: ${{\mathcal{M}}}_g$ {#twothree} -------------------------------------- The pointed version of Faber’s conjecture ${{\mathcal{M}}}_g$ isn’t for ${{\mathcal{M}}}_{g,n}$; it should be for ${{\mathcal{M}}}_{g,n}^{rt}$, the space of curves with “rational tails” (those stable curves with a component of geometric genus $g$). In this case, the role of $g-2$ is replaced by $g-2+n$. The intersection number portion of the conjecture (often called [*Faber’s intersection number conjecture*]{}) may be stated as follows: If all $a_i > 0$, then $$\label{Finc} \pi_* \psi_1^{a_1} \cdots \psi_n^{a_n} = \frac { (2g-3+n)! (2g-1)!!} { (2g-1)! \prod_{j=1}^n (2a_j-1)!!} {\kappa}_{g-2} \quad \quad \text{for $\sum a_i = g-2+n$.}$$ (This implies part [**III**]{} of Faber’s conjecture, see above, hence we use the same name.) This is now a theorem. Getzler and Pandharipande showed that the statement is a formal consequence of the Virasoro conjecture for ${\mathbb P}^2$ [@GeP], in fact the large volume limit. Givental thereafter described how to prove the Virasoro conjecture for projective space, and more generally Fano toric manifolds [@Gi1; @aG1]. This is a powerful approach to Faber’s conjecture, but perhaps somewhat roundabout. Recently, K. Liu and Xu have given an stunningly short and direct proof [@liuxu]. Their approach is quite different, and is part of their larger program for understanding the algebraic structure of these intersection numbers. Most notably, their approach yields a surprisingly simple explicit formula for Witten’s $n$-point function, and this has produced a number of interesting new results about intersection numbers, of which Faber’s intersection number conjecture is just one. As with Witten’s conjecture, the susceptibility of this problem to different approaches illustrates the richness of the question. One should hope that the story is not yet over, and more results should be obtainable from the successful earlier approaches. Furthermore, the beautiful form of equ.  clearly suggests that there should be a strong reason for it, and that further understanding should be sought. (The connection to double Hurwitz numbers, explored in [@gjv3], seems to be deeply related to this question.) The third sister: the ${\lambda}_g$-conjecture, for ${{\mathcal{M}}}_{g,n}^{ct}$ -------------------------------------------------------------------------------- The third sister of Faber’s conjecture is for the space ${{\mathcal{M}}}_{g,n}^{ct}$, the space of curves of compact type (stable curves whose dual graphs have no loop; those stable curves with compact Jacobian). The corresponding conjecture is due to Faber and Pandharipande ([@FabP1 Spec. 2], [@pandharipande Conj. 1]), with $g-2$ is replaced by $2g-3$. The socle portion was first proved as a consequence of Theorem $\star$ in [@thmstar]. The intersection number portion of the conjecture was first conjecture by Getzler and Pandharipande [@GeP]. It is called the ${\lambda}_g$-conjecture because of its incarnation as a statement about intersections on ${\overline{{{\mathcal{M}}}}}_{g,n}$ (rather than ${{\mathcal{M}}}_{g,n}^{ct}$): For $n,g \geq 1$, $$\int_{{\overline{{{\mathcal{M}}}}}_{g,n}} \psi_1^{b_1} \cdots \psi_n^{b_n} {\lambda}_g = \binom {2g-3+n} {b_1, \dots, b_n} c_g$$ where $\sum_{i=1}^n b_i = 2g-3+n$, $b_1, \dots, b_n \geq 0$, and $c_g$ is a constant depending only on $g$. The ${\lambda}_g$-conjecture has intriguingly proved more tractable than its sister intersection number conjectures, and a number of proofs now exist. Faber and Pandharipande gave the first proof [@FPlambdag]. A second proof is the same as the Getzler-Pandharipande-Givental proof of Faber’s intersection number conjecture: when making the conjecture, Getzler and Pandharipande showed that it is a formal consequence of the Virasoro conjecture for ${\mathbb P}^1$, which was later proved by Givental (and also by Okounkov and Pandharipande). K. Liu, C.-C. M. Liu, and Zhou gave a new proof [@llz] as a consequence of their proof of the Mariño-Vafa formula. Finally, Goulden, Jackson, and the second author gave a short direct (Gromov–Witten-free) proof in [@lambdag] by exploiting the “polynomiality” structure described in the next section, using the ELSV-formula (also described shortly). Kazarian and, independently at the same time, Kim and K. Liu showed that the algebraic structure introduced in [@gjv3], properly understood, also yield proofs (and explanations) of Witten’s conjecture. (This insight was certainly not known to the authors of [@gjv3].) Their two proofs (the most recent and shortest proofs, mentioned above) are quite distinct, and insightful. A complete understanding of the algebraic structures underlying Witten’s conjecture would presumably involve putting these two proofs into a common larger framework. Polynomiality {#s:pol} ============= We next describe the phenomenon of polynomiality of quasi-enumerative problems on moduli spaces of curves and maps. The central motivating example is the ELSV formula. The ELSV formula ---------------- Fix a genus $g$, a degree $d$, and a partition of $d$ into $n$ parts, ${\alpha}_1+ \cdots + {\alpha}_n = d$, and let $r = 2g+d+n-2$. Fix $r+1$ distinct points $p_1, \dots, p_n, \infty$ on ${\mathbb P}^1$. Define the [*Hurwitz number*]{} $H^g_{{\alpha}}$ as the number of branched covers of ${\mathbb P}^1$ by a (connected) Riemann surface, that are unbranched away from $p_1$, …, $p_n$, $\infty$, such that the branching over $\infty$ is given by ${\alpha}_1$, …, ${\alpha}_n$ (the monodromy lies in the conjugacy class corresponding to that partition), and the branching over each $p_i$ is $2+ 1+ \cdots + 1=d$ (the simplest nontrivial branching). We consider the $n$ preimages of $\infty$ to be labeled. Up to a straightforward combinatorial factor, this corresponds to the answer to the following combinatorial problem: given a permutation in conjugacy class ${\alpha}$, in how many ways can it be factored into $r$ transpositions that “connect” the numbers $1$ through $n$ ([ *transitive*]{} factorizations). (If the condition of transitivity seems unnatural, it is straightforward to connect to this to the problem without the transitive condition. This is equivalent to counting potentially disconnected covers. The algebraically simplest way to relate them: the exponential of the generating function counting connected covers is the generating function counting potentially disconnected covers.) Based on extensive evidence, the combinatorialists Goulden and Jackson [@gjconj Conj. 1.2] had conjectured that this combinatorial problem had a surprising polynomial behavior: fixing $g$ and $n$, $H^g_{{\alpha}}$ is a simple combinatorial term times a symmetric polynomial in ${\alpha}_1$, …, ${\alpha}_n$, with components in homogeneous degree between $2g-3+n$ and $3g-3+n$. This strongly suggests a connection between this combinatorial problem and the moduli space of curves! Ekedahl, Lando, M. Shapiro, and Vainshtein explained this polynomiality with their ground-breaking ELSV-formula: $$H^g_{{\alpha}} = r! \prod^n_{i=1}\left( \frac { {\alpha}_i^{{\alpha}_i}} {{\alpha}_i!} \right) \int_{{\overline{{{\mathcal{M}}}}}_{g,n}} \frac { 1 - {\lambda}_1 + \cdots + (-1)^g {\lambda}_g} { (1-{\alpha}_1 \psi_1) \cdots (1- {\alpha}_n \psi_n)}$$ Here the denominator should be considered formally inverted, i.e. $1/(1-{\alpha}_i \psi_i) = 1 + {\alpha}_i \psi_i + {\alpha}_i^2 \psi_i^2 + \cdots$, and the integral sign means to take the degree of the codimension $3g-3+n$ (dimension $0$) part of the integrand. Expanding the integral in the ELSV formula yields $$\sum_{a_1 + \cdots + a_n +k=3g-3+n} \left( (-1)^k \int_{{\overline{{{\mathcal{M}}}}}_{g,n}} \psi_1^{a_1} \cdots \psi_n^{a_n} \lambda_k \right) ( {\alpha}_1^{a_1} \cdots {\alpha}_n^{a_n} )$$ and thus the polynomiality is explained by interpreting these numbers as top intersections on the moduli space of curves! Better yet, the highest-degree terms are precisely the subject of Witten’s conjecture, and indeed the five proofs of Witten’s conjecture using the ELSV formula involve in different ways getting at these leading coefficients by asymptotic methods. Furthermore, the lowest-degree terms are precisely the subject of the ${\lambda}_g$-conjecture, which provides the entree for the proof of [@lambdag]. This polynomiality arises repeatedly, usually as a result of localization on spaces of stable maps, and was a key ingredient in, for example, work of Graber and the second author, e.g. [@gv; @thmstar]. By comparison, we give an example of polynomiality which has yet to be satisfactorily explained geometrically, which allows us to clarify a vague but suggestive conjecture of [@gjv]. A fourth sister? Conjectural geometry behind polynomiality of double Hurwitz numbers, [@gjv], and work of Shadrin and Zvonkine ------------------------------------------------------------------------------------------------------------------------------ We recall the history of the ELSV formula as motivation: we begin with an enumerative problem in geometry, of Hurwitz numbers, which can also be interpreted in terms of the combinatorics of the symmetric group. For fixed $g$ and $n$, these numbers are (up to a combinatorial factor) a polynomial in the parts of a partition (appearing in the definition of Hurwitz number). This polynomial is symmetric of degree $3g-3+n$, which is highly suggestive of the moduli space of curves, and indeed there is a genus $g$, $n$-pointed curve present in the enumerative problem. The generating function for these numbers is constrained by the KP (Kadomtsev-Petviashvili) hierarchy; to see this, a change of variables is necessary. Thanks to the ELSV formula, these numbers are related to the intersection theory of a fundamental moduli space, and as a result, surprising structure is known on the (tautological) cohomology ring of the moduli space. The identical story will apply in the situation we now describe, except that there is as of yet no moduli space, and no ELSV-type formula. There seems strong circumstantial evidence that there is a moduli space completing the story; this is the content of Conjecture 3.5 of [@gjv]. Instead of “single” Hurwitz numbers, we consider “one-part double” Hurwitz numbers $H^g_{(d),{\beta}}$, defined in the same way as single Hurwitz numbers, except that we require in addition complete branching over the point $0 \in {\mathbb P}^1$. Then by means of character theory, [@gjv Thm. 3.1] shows that for fixed $g$ and $n$, $H^g_{(d),{\beta}}$ is a symmetric polynomial in ${\beta}_1$, …, ${\beta}_n$, whose homogeneous pieces have even degree up to $4g-3+n$. The fundamental question this suggests is: [*is there a moduli space and ELSV-type formula explaining this polynomiality?*]{} Note that there is a moduli space of dimension $4g-3+n$ “present” in the problem: an $n$-pointed genus $g$ curve, along with a choice of line bundle. Motivated by this, one possible answer is the following. [**Conjecture [@gjv Conj. 3.5].**]{} For each $g \geq 0$, $n \geq 1$, $(g,n) \neq (0,1), (0,2)$, $$H^g_{(d),{\beta}} = r^g_{(d), {\beta}}! d \int_{\overline{\operatorname{Pic}}_{g,n}} \frac { \Lambda_0 - \Lambda_2 + \cdots \pm \Lambda_{2g}} { (1- {\beta}_1 \psi_1) \cdots (1 - {\beta}_n \psi_n)}$$ where $\overline{\operatorname{Pic}}_{g,n}$, $\psi_i$, and $\Lambda_{2k}$ satisfy a number of reasonable properties. For example, $\overline{\operatorname{Pic}}_{g,n}$ should be a compactification of the universal Picard variety over ${{\mathcal{M}}}_{g,n}$, which supports a (possibly virtual) fundamental class of the “expected” dimension $4g-3+n$. $\psi_i$ should be an extension of the pullback of $\psi_i$ from ${{\mathcal{M}}}_{g,n}$. See [@gjv] for a complete list of proposed properties. It should be emphasized that the fundamental question is that of finding the right space, and this conjecture should be seen as merely a proposal. Some of the properties suggested by geometry yield testable constraints on double Hurwitz numbers (such as the string and dilaton equation), and these indeed hold [@gjv Prop. 3.10]. Shadrin and Zvonkine show much more in [@sz] and [@s]. We emphasize that independent of the conjecture, their results are meaningful statements about double Hurwitz numbers. But their unpacking of the structure of these Hurwitz numbers is predicted precisely by the form of the conjecture. We state their results imprecisely in order to emphasize their form. Let $H$ be the generating function for the double Hurwitz numbers $H^g_{(d),{\beta}}$. Let $U$ be the generating function for just the highest-degree terms in the conjecture (i.e. those with no $\Lambda_i$ terms for $i>0$). Shadrin and Zvonkine define a linear differential operator $L$ which arises naturally in the cut-and-join formalism ($L = \sum b p_b \frac {\partial} {\partial p_b}$, where $p_b$ tracks parts of ${\alpha}$ of size $b$). The KP hierarchy is a system of partial differential equations on a power series $F$ in an infinite set of variables $p_1$, $p_2$, …. The first three equations are: $$\begin{aligned} F_{2,2} &=& - \frac 1 2 F^2_{1,1} + F_{3,1} - \frac 1 {12} F_{1,1,1,1} \\ F_{3,2} &=& - F_{1,1} F_{2,1} + F_{4,1} - \frac 1 6 F_{2,1,1,1} \\ F_{4,2} &=& - \frac 1 2 F^2_{2,1} - F_{1,1} F_{3,1} + F_{5,1} + \frac 1 8 F^2_{1,1,1} + \frac 1 {12} F_{1,1} F_{1,1,1,1} \\ & & \quad \quad - \frac 1 4 F_{3,1,1,1} + \frac 1 {120} F_{1,1,1,1,1,1}\end{aligned}$$ Subscript $i$ refers to differentiation by $p_i$. The exponent $\tau = e^F$ of any solution is called a $\tau$-function of the hierarchy. For readable expositions on the KP hierarchy, see [@kazarian §6] and (influenced by this) [@gjkp]. Shadrin and Zvonkine prove that under a scaling and renaming of variables, [**(i)**]{} $L^2 H$ is a $\tau$-function of the KP hierarchy, i.e. it satisfies the bilinear Hirota equations. Furthermore, $L^2 H$ satisfies the linearized KP equations. [**(ii)**]{} $U$ is a $\tau$-function for the KP hierarchy (in unusual variables), and thus satisfies the bilinear Hirota equations. Furthermore, it satisfies the linearized KP equations in the same variables. This is a complete analogue of Witten’s Conjecture. Statement [**(i)**]{} follows in a standard way from the general theory of integrable systems, but the proof of [**(ii)**]{} is quite subtle. Shadrin has recently taken this further. In [@s], he applies Kazarian’s techniques to this series, relating the conjectured “intersection numbers” to the Hirota equations. The computations turn out to be simpler, and he gives explicit and rather transparent formulas for the generating series. One might take Shadrin and Zvonkine’s results to suggest that the moduli space behind the integrable system is precisely that suggested by [@gjv Conj. 3.5]. However, this may not be the case: their methods are quite robust, and similar spaces (for example, with $n+1$ marked points, with the last point the preimage of $0$) should yield similar results. Teleman’s work on family topological field theory {#s:teleman} ================================================= Terminology ----------- For this section only, dimensions here mean dimensions over $\mathbb{R}$ and surfaces mean topological oriented 2-$\dim_{\mathbb{R}}$ “manifolds with nodal singularities”. Curves, however, mean algebraic (or complex) curves! Topological field theory and moduli of curves --------------------------------------------- C. Teleman introduced the notion of *family topological field theory* (FTFT) into Gromov–Witten theory. In the semisimple case, he completely classified all FTFTs. As an important application, he proved *Givental’s conjecture* [@aG1] on the quantization formula, which in particular gives an explicit reconstruction of higher genus semisimple Gromov–Witten theory from the genus zero theory. Teleman’s result is the first instance of applying two powerful results from *topology* of moduli spaces of curves to Gromov–Witten theory: *Harer stability* and the *Madsen–Weiss theorem* (Mumford’s conjecture). Harer stability is purely topological in its formulation, and Mumford’s conjecture so far has no algebro-geometric proof. In this section, we will explain how these two results can be applied to Gromov–Witten theory. Two dimensional topological field theory and semisimplicity ----------------------------------------------------------- A 2-$\dim_{\mathbb{R}}$ topological field theory (TFT) is a symmetric, (strong, monoidal) functor of topological tensor categories $Z: \mathbf{C} \to \mathbf{Vec}$. $\mathbf{Vec}$ is the usual tensor category of vector spaces. $\mathbf{C}$ is the category whose objects are 1-$\dim_{\mathbb{R}}$ oriented closed manifolds, i.e. a disjoint union of oriented $S^1$’s. The morphisms are oriented cobordisms of the objects, with obvious compositions. The tensor structure is defined by disjoint union of objects. The notion of TFT is equivalent to the notion of Frobenius algebra. A (commutative) *Frobenius algebra* is a $k$-algebra with an identity $1$, a pairing $(\cdot, \cdot) : A^{\otimes 2} \to A$ which is symmetric and nondegenerate, and satisfies the *Frobenius property* $$(a * b, c) = (a, b*c),$$ where $*$ stands for the multiplicative structure of $A$. The equivalence of TFT and Frobenius algebra can be seen as follows. $A = Z (S^1)$, and the identity element $1 \in A$ is defined by $Z(\text{cap})$, (0 input, 1 output). The nondegenerate pairing $(\cdot, \cdot)$ is defined by $Z(\text{bent cylinder})$ (2 inputs, 0 output), and the ring structure is defined by $Z(\text{pair of pants})$ (2 inputs, 1 output). \#1\#2\#3\#4\#5[ @font ]{} (7366,4092)(0,-10) (120,3428) (3683,3690) (3683,2753) (7283,3878) (7283,2752) (5408,3353) (5708,465)(1058,465) (1178.000,495.000)(1058.000,465.000)(1178.000,435.000) (158,3915)(160,3914)(164,3912) (171,3908)(182,3902)(197,3894) (217,3883)(240,3870)(267,3855) (296,3838)(328,3819)(361,3798) (395,3777)(429,3754)(462,3730) (495,3706)(526,3680)(555,3653) (583,3626)(608,3597)(630,3566) (650,3534)(665,3500)(676,3464) (683,3428)(683,3390)(677,3353) (665,3317)(648,3284)(628,3252) (604,3223)(577,3196)(548,3170) (516,3146)(483,3124)(448,3103) (412,3083)(376,3064)(338,3045) (302,3028)(266,3013)(231,2998) (200,2985)(171,2973)(146,2964) (125,2956)(109,2949)(97,2945) (89,2942)(85,2941)(83,2940) (3683,3915)(3682,3915)(3678,3914) (3672,3912)(3662,3910)(3648,3906) (3630,3901)(3607,3895)(3579,3888) (3546,3879)(3508,3869)(3466,3857) (3420,3845)(3371,3831)(3319,3816) (3264,3799)(3208,3783)(3150,3765) (3092,3746)(3034,3727)(2976,3707) (2919,3687)(2862,3666)(2807,3644) (2754,3621)(2702,3598)(2653,3574) (2605,3550)(2561,3524)(2519,3497) (2480,3469)(2445,3440)(2413,3409) (2386,3377)(2364,3344)(2347,3310) (2337,3275)(2333,3240)(2337,3205) (2347,3170)(2364,3136)(2386,3103) (2413,3071)(2445,3040)(2480,3011) (2519,2983)(2561,2956)(2605,2930) (2653,2906)(2702,2882)(2754,2859) (2807,2836)(2862,2814)(2919,2793) (2976,2773)(3034,2753)(3092,2734) (3150,2715)(3208,2697)(3264,2681) (3319,2664)(3371,2649)(3420,2635) (3466,2623)(3508,2611)(3546,2601) (3579,2592)(3607,2585)(3630,2579) (3648,2574)(3662,2570)(3672,2568) (3678,2566)(3682,2565)(3683,2565) (3683,3465)(3681,3465)(3676,3464) (3667,3463)(3654,3462)(3635,3460) (3611,3458)(3580,3454)(3545,3451) (3505,3446)(3460,3441)(3412,3436) (3362,3429)(3310,3423)(3256,3416) (3203,3408)(3150,3400)(3098,3392) (3048,3383)(2999,3374)(2953,3364) (2909,3354)(2868,3343)(2831,3331) (2797,3318)(2767,3304)(2743,3290) (2724,3274)(2712,3257)(2708,3240) (2712,3222)(2724,3205)(2743,3188) (2767,3171)(2797,3155)(2831,3140) (2868,3125)(2909,3111)(2953,3097) (2999,3083)(3048,3070)(3098,3058) (3150,3046)(3203,3034)(3256,3022) (3310,3011)(3362,3000)(3412,2990) (3460,2981)(3505,2973)(3545,2965) (3580,2958)(3611,2953)(3635,2949) (3654,2945)(3667,2943)(3676,2941) (3681,2940)(3683,2940) (5408,3540)(5409,3540)(5411,3540) (5416,3539)(5423,3539)(5433,3538) (5446,3537)(5462,3536)(5481,3534) (5504,3532)(5529,3531)(5557,3529) (5588,3527)(5621,3526)(5655,3525) (5692,3524)(5729,3523)(5768,3523) (5808,3523)(5849,3524)(5891,3525) (5934,3527)(5979,3530)(6025,3534) (6073,3539)(6123,3546)(6174,3553) (6228,3562)(6283,3573)(6341,3585) (6399,3599)(6458,3615)(6520,3633) (6579,3653)(6636,3673)(6689,3693) (6739,3714)(6785,3734)(6829,3755) (6869,3775)(6908,3796)(6944,3816) (6978,3837)(7010,3857)(7041,3878) (7070,3898)(7098,3917)(7124,3937) (7149,3955)(7172,3973)(7193,3990) (7212,4005)(7229,4019)(7243,4031) (7255,4041)(7265,4049)(7272,4055) (7277,4060)(7280,4063)(7282,4064)(7283,4065) (5408,3165)(5409,3165)(5411,3165) (5416,3165)(5423,3165)(5433,3165) (5446,3165)(5462,3164)(5481,3164) (5504,3164)(5529,3163)(5557,3162) (5588,3161)(5621,3159)(5655,3158) (5692,3156)(5729,3153)(5768,3150) (5808,3147)(5849,3143)(5891,3138) (5934,3133)(5979,3126)(6025,3119) (6073,3111)(6123,3101)(6174,3091) (6228,3079)(6283,3065)(6341,3050) (6399,3033)(6458,3015)(6520,2994) (6579,2973)(6636,2951)(6689,2930) (6739,2908)(6785,2887)(6829,2866) (6869,2845)(6908,2824)(6944,2804) (6978,2784)(7010,2764)(7041,2744) (7070,2725)(7098,2706)(7124,2687) (7149,2669)(7172,2652)(7193,2636) (7212,2622)(7229,2609)(7243,2597) (7255,2588)(7265,2580)(7272,2574) (7277,2570)(7280,2567)(7282,2566)(7283,2565) (7283,3690)(7281,3689)(7276,3688) (7267,3685)(7253,3681)(7234,3676) (7210,3669)(7181,3660)(7148,3650) (7111,3638)(7071,3625)(7030,3611) (6988,3596)(6945,3581)(6903,3564) (6862,3547)(6822,3529)(6785,3510) (6750,3490)(6717,3469)(6687,3447) (6661,3423)(6639,3398)(6623,3372) (6612,3344)(6608,3315)(6612,3286) (6623,3258)(6639,3232)(6661,3207) (6687,3183)(6717,3161)(6750,3140) (6785,3120)(6822,3101)(6862,3083) (6903,3066)(6945,3049)(6988,3034) (7030,3019)(7071,3005)(7111,2992) (7148,2980)(7181,2970)(7210,2961) (7234,2954)(7253,2949)(7267,2945) (7276,2942)(7281,2941)(7283,2940) (83,1815)[(0,0)\[lb\]]{} (2558,1815)[(0,0)\[lb\]]{} (5483,1815)[(0,0)\[lb\]]{} (2483,15)[(0,0)\[lb\]]{} We will mainly be concerned with the *semisimple* case, where the product structure is diagonalizable. That is, $A \cong \oplus_i k \epsilon_i$ as an $k$-algebra, where $\epsilon_i * \epsilon_j = \delta_{ij} \epsilon_i$ is the *canonical basis*. Therefore, up to isomorphisms, $A$ is classified by $\Delta_i := (\epsilon_i, \epsilon_i)$. Let $$\tilde{\epsilon}_i := \frac{1}{\sqrt{\Delta_i}} \epsilon_i$$ be the *normalized canonical basis*. \[p:1\] Let ${}^qC_g^p$ be an oriented surface of genus $g$, with $p$ inputs and $q$ outputs. Then $Z({}^qC_g^p)$ defines a map $${}^q\mu_g^p: A^{\otimes p} \to A^{\otimes q},$$ such that ${}^q\mu_g^p$ is “diagonal” in the (tensor power of the) normalized canonical basis. Furthermore, the entry of ${}^q\mu_g^p$ in $\tilde{\epsilon}_i^{\otimes p} \mapsto \tilde{\epsilon}_i^{\otimes q}$ is $\Delta_i^{\chi({}^qC_g^p)/2}$, where $\chi({}^qC_g^p)$ is the (topological) Euler characteristic of ${}^qC_g^p$. This proposition can be easily proved by decomposing ${}^qC_g^p$ into pairs of pants and applying the facts that $Z({}^1C_0^1)$ is the identity and that $Z({}^1C_0^2)$ defines the algebra multiplication and is diagonal with entries $\Delta_i^{1/2}$ in the normalized canonical basis. Let ${}^1C_1^1$ be the torus with one outgoing and one incoming boundary. Since the ring structure is determined by $\tilde{Z}$ on a pair of pants, one can piece together 2 pairs of pants and obtain ${}^1C_1^1$. For future references, denote $\alpha$ (the diagonal matrix) ${}^1\mu_1^1=\tilde{Z}({}^1C_1^1)$. Family topological field theories --------------------------------- A FTFT assigns a family of topological $(m,n)$-pointed 2-$\dim_{\mathbb{R}}$ surfaces $\mathcal{C} \to B$ a cohomology class $\bar{Z}(B) \in H^*(B, \operatorname{Hom}(A^m, A^n))$, where $A^m$ are, in general, local systems on $B$. $\bar{Z}$ must satisfy two additional properties, in addition to being a “fiberwise TFT:” 1. functoriality with respect to the (topological) base change; 2. a strong gluing axiom. The first property ensures that all FTFTs are pullbacks of the universal setting over the classifying spaces. The second property requires a little more explanation. The gluing requires parameterizations of boundary circles, instead of the punctures. The strong gluing axiom asserts that the gluing axiom has to hold for *any* lifting to families with parameterized boundaries. The $(m,n)$-pointed family can be lifted to a family of $m$ incoming and $n$ outgoing circles in the following way. Take the torus bundle $\tilde{B} \to B$ with fiber $(S^1)^m \times (S^1)^n$ the product of unit tangent spaces at the marked points. (Since the space of Riemannian metrics is contractible, the choice of the metric does not matter.) Because $\operatorname{Diff}_+(S^1)$, the orientation-preserving diffeomorphisms, is homotopy equivalent to its subgroup of rigid rotations, $\tilde{B}$ is homotopy equivalent to the base space of the corresponding family with parameterized circle boundaries. Note that $\mathcal{C}$ allows nodal singularities. Teleman classified all FTFTs for which the fiberwise TFT is generically semisimple. Teleman established this classification in 3 steps. **Step 1.** Classification of FTFTs $\tilde{Z}$ associated to smooth families of surfaces with parameterized boundaries. **Step 2.** Going from parameterized boundaries to punctures. This requires a new piece of information, which is denoted $E(\psi)$. Roughly, at each outgoing boundary circle, $E(\psi)$ (or $E(\psi)^{-1}$ for an incoming circle) “transforms” the parameterized boundary to a puncture. **Step 3.** Allowing nodal (stable) degenerations of curves. The corresponding FTFT is denoted $\bar{Z}$. $\bar{Z}$ requires again one new piece of data: $L:=\bar{Z}(\text{pinched cylinder})$. In fact, this classification can be generalized to, what Teleman calls “Lefschetz theory”, where the family of curves does not have to be stable, and with more general gluing axioms. Interested readers are referred to Teleman’s original article [@cT]. \#1\#2\#3\#4\#5[ @font ]{} (2716,1971)(0,-10) (158,1382) (2558,1382) (158,1944)(160,1944)(163,1943) (169,1941)(180,1938)(194,1934) (213,1929)(236,1922)(265,1914) (298,1905)(336,1894)(378,1882) (423,1869)(471,1854)(522,1839) (575,1822)(628,1805)(683,1787) (738,1768)(792,1749)(846,1729) (898,1709)(950,1688)(999,1667) (1047,1645)(1093,1622)(1136,1598) (1177,1574)(1214,1548)(1249,1522) (1279,1495)(1306,1466)(1328,1437) (1344,1406)(1354,1375)(1358,1344) (1354,1313)(1344,1283)(1328,1253) (1306,1225)(1279,1198)(1249,1173) (1214,1149)(1177,1126)(1136,1104) (1093,1083)(1047,1063)(999,1044) (950,1025)(898,1008)(846,991) (792,974)(738,958)(683,943) (628,929)(575,915)(522,902) (471,889)(423,878)(378,867) (336,858)(298,849)(265,842) (236,836)(213,831)(194,826) (180,823)(169,821)(163,820) (160,819)(158,819) (2558,1944)(2556,1944)(2553,1943) (2547,1941)(2536,1938)(2522,1934) (2503,1929)(2480,1922)(2451,1914) (2418,1905)(2380,1894)(2338,1882) (2293,1869)(2245,1854)(2194,1839) (2141,1822)(2088,1805)(2033,1787) (1978,1768)(1924,1749)(1870,1729) (1818,1709)(1766,1688)(1717,1667) (1669,1645)(1623,1622)(1580,1598) (1539,1574)(1502,1548)(1467,1522) (1437,1495)(1410,1466)(1388,1437) (1372,1406)(1362,1375)(1358,1344) (1362,1313)(1372,1283)(1388,1253) (1410,1225)(1437,1198)(1467,1173) (1502,1149)(1539,1126)(1580,1104) (1623,1083)(1669,1063)(1717,1044) (1766,1025)(1818,1008)(1870,991) (1924,974)(1978,958)(2033,943) (2088,929)(2141,915)(2194,902) (2245,889)(2293,878)(2338,867) (2380,858)(2418,849)(2451,842) (2480,836)(2503,831)(2522,826) (2536,823)(2547,821)(2553,820) (2556,819)(2558,819) (383,69)[(0,0)\[lb\]]{} Teleman proved that $\tilde{Z}$, $E(\psi)$, and $L$ are sufficient to “piece together” $\bar{Z}$. Roughly, one can replace punctures by parameterized boundaries via $E(\psi)$, and smooth the nodes via $L$. The remaining theory is then $\tilde{Z}$. Among the 3 steps, Step 1 involves the aforementioned Harer stability and Madsen–Weiss Theorem and is most relevant to this article’s audience. In the remaining of this section, we will give some explanation of Step 1. But we first give a short digression on how Teleman’s classification proves Givental’s conjecture. Proof of Givental’s conjecture ------------------------------ We will use the following terminology proposed in [@ypL4]. A *geometric* Gromov–Witten theory (GWT) is the “usual” GWT constructed from moduli spaces of stable maps of *all genera*. Each geometric GWT is encoded in a single generating function $$\label{e:tauGW} \tau_{GW} := \exp \left( \sum_{g=0}^{\infty} \hbar^{g-1} F_g \right)$$ where $F_g$ is the genus $g$ generating function of GW invariants. An *axiomatic* GWT is the one constructed from Givental’s quantization formalism, when the Frobenius structure is semisimple. Roughly, the generating function of an axiomatic GWT is *defined* to be $$\label{e:tauG} \tau_G := \hat{\operatorname{O}} \tau^{N pt},$$ where $\tau^{N pt}$ is the $N$ copies of Witten–Kontsevich generating function (or equivalently GWT of $N$ points), and $\hat{\operatorname{O}}$ is an operator defined by quantizing a quadratic functions defined via *genus zero* GW invariants (or equivalently the Frobenius structure). When the Frobenius structure is semisimple, one can ask whether $\tau_G$ is equal to $\tau_{GW}$. The equality is termed *Givental’s conjecture*. The conjecture in particular implies that all higher genus GW invariants can be reconstructed from genus zero invariants, as $\hat{O}$ involves only genus zero data. One can prove that $\tau_G$ can be written as $$\tau_G = \exp \left( \sum_{g=0}^{\infty} \hbar^{g-1} G_g \right).$$ Furthermore, it can be shown without much difficulty that $G_0=F_0$, which implies that the genus zero truncation of Givental’s conjecture holds. See [@LP Part II] for details. Teleman’s classification implies Givental’s conjecture. Unfortunately, the brief account given here is mostly redundant for those who are familiar with the basic definitions. A meaningful explanation of this involves the detailed construction of Givental’s quantization formalism and is beyond the scope of this article. Interested readers are referred to Teleman’s original paper [@cT]. First of all, it is easy to see that axiomatic GWTs are FTFT. For those who are familiar with the definition of *cohomological field theory*, GWTs are CohFTs and CohFTs satisfy the axioms of FTFTs. In fact, Teleman’s classification implies a *stronger* versions of Givental’s conjecture: The corresponding axiomatic and geometric CohFTs are equal. Indeed, CohFTs yield cohomology classes on ${\overline{\mathcal{M}}}_{g,n}$, and the integration of these classes (with monomials of $\psi$-classes) gives GW invariants. To prove Givental’s conjecture, Teleman traces through Givental’s quantization formalism and shows that all semisimple GWTs can be reconstructed from the genus zero data, and from known facts about the tautological classes of the moduli space of curves. (Very naively speaking, one can see all discussion in this section involves only genus zero data and moduli of curves.) Since axiomatic and geometric theories are, by construction, identical in genus zero, the proof is then complete. Wheeled PROP structure ---------------------- Let ${}^q{M}_g^p$ be the moduli space of the Riemann surfaces of genus $g$ with $p$ incoming and $q$ outgoing marked points. Let ${}^q\tilde{M}_g^p$ be the torus bundle over ${}^q{M}_g^p$ as discussed earlier. Homotopically, ${}^q{M}_g^p \sim B \Gamma_g^{p+q}$, where $\Gamma_g^{p+q}$ is the corresponding mapping class group, and one can think of ${}^q{M}_g^p$ as a classifying space of bordered Riemann surfaces of genus $g$ with $p$ incoming and $q$ outgoing parameterized boundaries. Gluing the outgoing boundary circles of a surface with incoming ones of another gives the structure of a *PROP* on ${}^q\tilde{M}_g^p$. Roughly, a PROP is something like an operad, but with multiple outputs.[^1] In addition to the PROP structure, it also allows *contraction*, via gluing boundary circles of the same surface. The additional contraction operation makes it a *wheeled PROP* [@MMS]. Harer stability --------------- We recall Harer stability. For the notational convenience, we will *not* distinguish the outgoing and incoming boundary, but use $r := p+q$. Let $\tilde{M}_{g,r}$ be the (homotopy type of the) moduli space of oriented surfaces of genus $g$, and $r$ boundary circles. (The number of punctures will be fixed during the discussion, and will be omitted from the notations.) Let $C_{g,r}$ be a bordered oriented surface of genus $g$ with $r$ boundary circles. Define three operations $$\begin{aligned} \phi_1 &: C_{g,r} \to C_{g+1, r-1}, &r \ge 2, \\ \phi_2 &: C_{g,r} \to C_{g, r+1}, &r \ge 1, \\ \phi_3 &: C_{g,r} \to C_{g+1, r-2}, &r \ge 2 \end{aligned}$$ as follows: $\phi_1$ is defined by gluing a pair of pants along 2 boundary circles; $\phi_2$ is defined by gluing a pair of pants along 1 boundary circle; $\phi_3$ is defined by gluing 2 boundary circles together. \[t:harer\] $(\phi_1)_* : H_k (\tilde{M}_g^r) \to H_k (\tilde{M}_{g+1}^{r-1})$, and similarly $(\phi_2)_*$ and $(\phi_3)_*$, are isomorphisms for $k \le [g/3]$ *(the stable range)*. The stable range can be improved, but we only need this estimate. Thanks to Harer stability, one can talk about $H_k({M}_{\infty})$. Here $r$ is redundant in the stable range by a combination of $\phi_i$’s, and hence $\tilde{M}_{\infty} = M_{\infty}$. Classification of ${}^1\tilde{Z}$ --------------------------------- We will first study ${}^1\tilde{Z}$, i.e. FTFT with one outgoing and no incoming boundary. One can increase genus by applying the composition $\phi_2 \circ \phi_1$. By semisimplicity, the effect of this operation on ${}^1\tilde{Z}_g$ is the multiplication by (the diagonal matrix) $\alpha$. Therefore, ${}^1\tilde{Z}_{g+G} = (\alpha^G) {}^1\tilde{Z}_g$ in the stable range and $(\alpha^{-G}) {}^1\tilde{Z}_G$ stabilizes to $\tilde{Z}^+ \in H^*(M_{\infty})$. \[p:2\] (i) The degree zero component of $\tilde{Z}^+ \in H^*(M_{\infty})$ is the identity. \(ii) $H^*(M_{\infty})$ carries a Hopf algebra structure. \(iii) $\log \tilde{Z}^+ =\sum_{l \ge 1} a_l \kappa_l$, where $a_l \in A$. \(i) can be seen by restricting our family of oriented smooth surfaces to a point, and apply base change property: For an inclusion $\iota: pt \to M$, the pullback in cohomology in degree zero $\iota^* :H^0(M) \to H^0(pt)$ induces isomorphism on degree zero component of $\tilde{Z}^+$. \(ii) is a consequence of the wheeled PROP structure. \(iii) is a consequence of the wheeled PROP structure, gluing axiom, and Mumford’s conjecture. We will explain the idea now. An element in a Hopf algebra is called [*primitive*]{} if $\Delta (x) = 1 \otimes x + x \otimes 1$. It is [*group-like*]{} if $\Delta (g) = g \otimes g$. It is easy to check that group-like elements are exponentials of primitive elements. $\tilde{Z}^+$ is a group-like element. This follows from the monoidal structure (in PROP) on $M_{\infty}$ defined by gluing two surfaces into a pair of pants: $$\tilde{M}^1_{g_1} \times \tilde{M}^1_{g_2} \to \tilde{M}^1_{g_1+ g_2}.$$ Note that the multiplicative factor $\alpha^{-(g_1+g_2)}$ is consistent on both sides. Therefore, $\log \tilde{Z}^+$ is a primitive element. Now, Madsen–Weiss’s theorem: $$H^*(M_{\infty}, \mathbb{Q}) = \mathbb{Q} [\kappa_0, \kappa_1, \kappa_2, ...].$$ Thanks to the polynomial structure of the Hopf algebra, the only primitive elements are $\sum_{l \ge 0} a_l \kappa_l$. Applying Proposition \[p:2\] (i), $a_0=0$ and $\log \tilde{Z}^+ =\sum_{l \ge 1} a_l \kappa_l$. Note that the converse of the statement also holds: All elements of the form $\sum_{l \ge 1} a_l \kappa_l$ serve as FTFT on smooth surfaces with parameterized boundaries. This completely classifies $\log \tilde{Z}^+$ and hence ${}^1\tilde{Z}$. The final steps: Classifying ${}^q\tilde{Z}^p$ ---------------------------------------------- One can increase genus by gluing ${}^1C^1$, and can increase $p$ and $q$ by gluing a genus zero surface with $p+1$ inputs and $q$ outputs to the existing one output. The latter gives a map from ${}^1\tilde{M}_g$ to ${}^q\tilde{M}^p_g$. Thanks to Harer stability, both operations are homology equivalences in the stable range. By functoriality of base change, ${}^q\tilde{Z}_g^p$ is determined by its pullback to ${}^1\tilde{M}_g$. On ${}^1\tilde{M}_g$, this can be seen as feeding the output of ${}^1\tilde{M}_g$ into ${}^q\mu^{p+1}: A^{\otimes p+1} \to A^{\otimes q}$. Therefore: ${}^q \tilde{Z}_g^p$ is diagonal in the (tensor power of the) normalized basis, and the entry for $\tilde{\epsilon}_i^{\otimes p} \mapsto \tilde{\epsilon}_i^{\otimes q}$ is $\Delta_i^{\chi/2} \exp \left( \sum_{l \ge 1} a_{il} \kappa_l \right)$. This completes the classification of $\tilde{Z}$. Witten’s conjecture on $r$-spin curves {#s:invariance} ====================================== In this section, we briefly describe Faber–Shadrin–Zvonkine’s proof [@FSZ] of Witten’s conjecture on $r$-spin curves. This conjecture also follows from combining Teleman’s more general result in the previous section, and Givental’s result in [@aG2]. [@FSZ] takes a different approach. Semisimplicity of the Frobenius structure is tacitly assumed throughout this section. Overview -------- Witten’s conjecture states that a certain kind of Gromov–Witten theory can be constructed from $r$-spin curves, whose moduli spaces are certain branched covers of ${\overline{\mathcal{M}}}_{g,n}$. Furthermore, the generating function $\tau_W$ constructed from Witten’s correlators satisfies the Gelfand–Dickii hierarchy. The proof has several ingredients. Here we explain *one* aspect of the proof, which is closer to our own. **Step 0.** Witten [@witten2] and Jarvis–Kimura–Vaintrob [@JKV] show, building on Jarvis’ earlier work [@tJ1; @tJ2], that the genus zero truncation of the conjecture holds. Several rigorous constructions of Witten’s correlators for all genera were defined (e.g. [@tM; @PV]). **Step 1.** The first author proposed to prove Witten’s conjecture via a version of Givental’s conjecture: First, show that the axiomatic generating function $\tau_G$ satisfies Gelfand–Dickii hierarchy. Then, show $$\label{e:WG} \tau_W = \tau_G,$$ where $\tau_W$ is Witten’s generating function. **Step 2.** Givental showed that $\tau_G$ satisfies Gelfand–Dickii hierarchy [@aG2]. **Step 3.** The first author proved a $g \le 2$ truncation of by introducing an algorithm, termed *invariance constraints*, to compute universal relations in axiomatic GWT. The *universal relations* are, be definition, the relations which hold for all axiomatic GWTs. Furthermore, it is shown that the tautological relations are universal relations [@ypL1; @ypL2; @ypL3]. **Step 4.** Faber–Shadrin–Zvonkine proved in full generality by showing a reconstruction theorem which allows one to reduce the comparison of the geometric GWT and the axiomatic GWT to a genus zero version. It can be considered as a special case of Teleman’s theorem in the previous section. This proof uses the idea of invariance constraints and Ionel’s theorem [@eI], and is entirely different from Teleman’s approach. It is worth noting that Witten’s conjecture is related to the Frobenius structure of type $A_r$ singularities. A generalization to type $D$ and $E$ has been carried out by H. Fan, T. Jarvis, and Y. Ruan (in preparation). We will briefly explain Steps 0, 3 and 4. Witten’s correlators {#wc} -------------------- An $r$-spin structure on a smooth curve $C$ of genus $g$ with $n$ marked points $x_i$ is a line bundle $L$ together with the identification $$L^{\otimes r} = K_C \left( \sum_i m_i x_i \right),$$ where $0 \le m_i \le r-1$ are integers such that $r | (2g-2-\sum_i m_i)$. The moduli space of $r$-spin structures has a natural compactification ${\overline{\mathcal{M}}}_{g,m}^{1/r}$ with a natural morphism $\pi$ to ${\overline{\mathcal{M}}}_{g,n}$, which is a degree $r^{2g-1}$ “branched cover”. Note that there is a global automorphism (multiplication by $r$th roots of unity) of any point in ${\overline{\mathcal{M}}}^{1/r}_{g,n}$. The degree is $r^{2g}$ for the corresponding coarse moduli spaces. Witten’s correlators are defined by a virtual fundamental class construction [@tM; @PV; @aC] on ${\overline{\mathcal{M}}}_{g,(m_1,...,m_n)}^{1/r}$ $$\int_{[{\overline{\mathcal{M}}}_{g,(m_1,...,m_n)}^{1/r}]^{{\operatorname{vir}}}} \pi^* \Psi,$$ where $\Psi$ is a monomial of $\psi$-classes on ${\overline{\mathcal{M}}}_{g,n}$. By the projection formula, it is equal to $\int_{[{\overline{\mathcal{M}}}_{g,n}]} c_W \Psi$, where $c_W (m) \cap {\overline{\mathcal{M}}}_{g,n}:= \pi_*([{\overline{\mathcal{M}}}_{g,a}^{1/r}]^{{\operatorname{vir}}})$. (Caution: This is slightly different from the definition of $c_W$ in [@FSZ].) Invariance constraints and universal relations {#s:5.3} ---------------------------------------------- The discussion here in intended to get across general ideas, and will be imprecise. Interested readers may consult [@ypL4; @LP] or Givental’s original papers ([@aG2] and references therein) for details. The starting point of this is Givental’s axiomatic (semisimple) GWTs. A major breakthrough in Givental’s discovery is that the “moduli space” of Frobenius manifolds (or equivalently genus zero axiomatic theories) of *a fixed dimension* has an action by a “twisted loop group”, or *Givental’s group*. Furthermore, all semisimple theories form a single orbit under this group action. One of the simplest semisimple theories of dimension $N$ is the GWT of $N$ points, which is given by $N$ copies of GWT of one point. This implies that all semisimple axiomatic theories can be obtained by a group action on GWT of $N$ points. That is, for each semisimple axiomatic theory $T$, one can find an element $O_T$ in Givental’s group such that $T$ is equal to $O_T$ acting on the moduli point defined by GWT of $N$ points. $O_T$ is uniquely determined if a “homogeneity” condition holds. We shall not go further into the details, but only point out that Witten’s generating function $\tau_W$ satisfies the homogeneity condition. A geometric GWT involves all genera and is encoded in a single generating function $\tau_{GW}$ as in . For axiomatic theories, the generating functions $\tau_G$ are obtained via “quantizing” $O_T$. $\hat{O}_T$ is an operator which then acts on the generating function of GWT of $N$ points $\tau^{N pt}$ as in . One may define a *universal relation* to be an equation of tautological classes which holds for all axiomatic GWTs. In particular it must contain tautological relations for moduli of curves as it has to hold for $\tau^{N pt}$. To show that a tautological relation for moduli curves holds for all axiomatic GWTs, we can check whether it is “invariant” under the action of $\hat{O}_T$ for all axiomatic GWTs $T$. In fact, since $O_T$’s form Givental group, we only have to check the invariance at the level of Lie algebra. This gives very strong constraints on the form of the possible universal relations. It was checked that all tautological equations are universal, [@ypL2]. A simple geometric proof was later discovered by Faber–Shadrin–Zvonkine and independently R. Pandharipande and the first author (see [@FSZ §3]). The Faber–Shadrin–Zvonkine Uniqueness Theorem --------------------------------------------- Let $\mu^i$ be partitions of $d$. Let ${H}_{g}(\mu^1, \mu^2)$ be the moduli space of Hurwitz covers of degree $d$ from genus $g$ curves to $\mathbb{P^1}$, with fixed profiles $(\mu^1, \mu^2)$ at $(0, \infty)$ in $\mathbb{P^1}$, and otherwise simple ramification. Let $\overline{H}_g(\mu^1, \mu^2)$ be the compactification by admissible covers. There is a natural morphism $$\rho: \overline{H}_g(\mu^1, \mu^2) \to {\overline{\mathcal{M}}}_{g, \sum_i l(\mu^i)}$$ by forgetting the covering maps and stabilizing the domain curves. One can consider variations of the above constructions. First, one can allow partitions $\mu^i =(\mu^i_1, \mu^i_2, \ldots)$ to contain zeros, which correspond to marked unramified points. Second, $\rho$ can be composed with forgetful maps (by forgetting marked points.) For the purpose of this section, we are only interested in forgetting points which are *not* marked unramified points. By abusing notation, we will denote them by $\rho: \overline{H}_g \to {\overline{\mathcal{M}}}_{g,n}$. Let $p := \sum_i l (\mu^i) -n$. (Note that a marked unramified point increases the length by 1.) We will call the images of $\rho$ *double Hurwitz cycles*. Witten’s correlators, by definition, can be expressed as integrals of the following type $$\int_{{\overline{\mathcal{M}}}_{g,n}} c_W \Psi,$$ where $\Psi$ is a polynomial of $\psi$-classes. Also, by construction, (complex) degree of $\Psi$ is at least $g$. Now Ionel’s theorem [@eI] states that any monomials in $\psi$ and $\kappa$ of Chow degrees greater than $g-1$ can be represented as a linear combination of classes of the form $$q_*(DHC_1 \times DHC_2 \times \ldots)$$ where $q: \prod_i {\overline{\mathcal{M}}}_{g_i, n_i} \to {\overline{\mathcal{M}}}_{g,n}$ is the gluing morphism and $DHC_i$ is a certain double Hurwitz cycle in ${\overline{\mathcal{M}}}_{g_i, n_i}$. Since it is known [@FP] that double Hurwitz cycles are tautological, and tautological classes are linearly generated by monomials of $\psi$ and $\kappa$ classes on (the vertices of) the dual graphs, the above arguments imply that *any monomial in $\psi$ and $\kappa$ on ${\overline{\mathcal{M}}}_{g,n}$ of Chow degrees at least $g$ can be represented as a linear combination of dual graphs with at least one edge*. This is called *$g$-reduction* in [@FSZ]. By $g$-reduction on $\Psi$ and the splitting principle of $c_W$, Witten’s correlators can then be reduced, with simple dimension counts and manipulations of tautological classes, to genus zero correlators. Note that the above arguments apply equally to axiomatic and geometric theories. For axiomatic theories, one has to show that all tautological relations holds, so that the same form of $g$-reduction applies. As explained in the previous subsection, this is accomplished by showing they satisfy the invariance constraints. Now, the fact that both theories have 1. the same reduction to genus zero correlators, and 2. the same genus zero correlators implies the geometric theory equals the axiomatic one. This, combined with Givental’s theorem [@aG2] then proves Witten’s conjecture. [ELSV2]{} L. Chen, Y. Li, and K. Liu, [*Localization, Hurwitz numbers and the Witten conjecture*]{}, preprint 2006, math.AG/0609263. A. Chiodo, *The Witten top Chern class via $K$-theory*, J. Algebraic Geom. **15** (2006), no. 4, 681–707. D. Cox and S. Katz, [*Mirror Symmetry and Algebraic Geometry*]{}, Mathematical surveys and Monographs [**68**]{}, Amer.Math. Soc., Providence, RI, 1999. T. Ekedahl, S. Lando, M. Shapiro, and A. Vainshtein, [*On Hurwitz numbers and Hodge integrals*]{}, C. R. Acad. Sci.  Paris Sér. I Math. [**328**]{} (1999), 1175–1180. T. Ekedahl, S. Lando, M. Shapiro, and A. Vainshtein, [*Hurwitz numbers and intersections on moduli spaces of curves*]{}, Invent. Math. [**146**]{} (2001), 297–327. C. Faber, [*A conjectural description of the tautological ring of the moduli space of curves*]{}, in [*Moduli of Curves and Abelian Varieties*]{}, 109–129, Aspects Math., [**E33**]{}, Vieweg, Braunschweig, 1999. C. Faber, [*Algorithms for computing intersection numbers on moduli spaces of curves, with an application to the class of the locus of Jacobians*]{}, in [*New trends in algebraic geometry (Warwick, 1996)*]{}, 93–109, London Math. Soc. Lecture Note Ser., 264, Cambridge Univ. Press, Cambridge, 1999. C. Faber and R. Pandharipande, [ *Logarithmic series and Hodge integrals in the tautological ring*]{}, Michigan Math. J. (Fulton volume) [**48**]{} (2000), 215–252. C. Faber and R. Pandharipande, [*Hodge integrals, partition matrices, and the ${\lambda}_g$ conjecture*]{}, Ann. Math. (2) [**157**]{} (2003), no. 1, 97–124. C. Faber and R. Pandharipande, *Relative maps and tautological classes*, J. Eur. Math. Soc.  7 (2005), no. 1, 13–49. C. Faber, S. Shadrin, and D. Zvonkine, *Tautological relations and the $r$-spin Witten conjecture*, arXiv:math/0612510. E. Getzler and R. Pandharipande, [*Virasoro constraints and the Chern classes of the Hodge bundle*]{}, Nuclear Phys. B [**530**]{} (1998), 701–714. A. Givental, [*Semisimple Frobenius structures at higher genus*]{}, IMRN [**2001**]{}, no. 23, 1265–1286. A. Givental, *Gromov–Witten invariants and quantization of quadratic Hamiltonians*, Dedicated to the memory of I. G. Petrovskii on the occasion of his 100th anniversary, Mosc. Math. J.  1 (2001), no. 4, 551–568, 645. A. Givental, *$A\sb {n-1}$ singularities and $n$KdV hierarchies*, Dedicated to Vladimir I. Arnold on the occasion of his 65th birthday, Mosc. Math. J.  **3** (2003), no. 2, 475–505, 743. I. P. Goulden and D. M. Jackson, [*The number of ramified coverings of the sphere by the double torus, and a general form for higher genera*]{}, J. Combin. Theory A [**88**]{} (1999), 259–275. I. P. Goulden and D. M. Jackson, [*The KP hierarchy, branched covers, and triangulations*]{}, preprint 2008, arXiv:0803.3980v1, Adv. in Math., to appear. I. P. Goulden, D. M. Jackson, and R. Vakil, [*Towards the geometry of double Hurwitz numbers*]{}, Adv. Math. (Artin issue) [**198**]{} (2005), 43–92. I. P. Goulden, D. M. Jackson, and R. Vakil, [*A short proof of the ${\lambda}_g$-conjecture without Gromov-Witten theory: Hurwitz theory and the moduli of curves*]{}, submitted for publication, preprint 2007, earlier version math.AG/0604297. I. P. Goulden, D. M. Jackson, and R. Vakil, [*The moduli space of curves, double Hurwitz numbers, and Faber’s intersection number conjecture*]{}, submitted for publication, preprint 2006, math.AG/0611659. T.  Graber and R. Vakil, [*On the tautological ring of ${\overline{{{\mathcal{M}}}}}_{g,n}$*]{}, in [*Proceedings of the Seventh Gökova Geometry-Topology Conference 2000*]{}, International Press, 2000. T. Graber and R. Vakil, [*Relative virtual localization and vanishing of tautological classes on moduli spaces of curves*]{}, Duke Math. J. [**130**]{} (2005), no. 1, 1–37. R. Hain and E. Looijenga, [*Mapping class groups and moduli spaces of curves*]{}, Proc. Sympos. Pure Math. [ **62**]{} Part 2, pp. 97–142, Amer. Math. Soc., Providence, RI, 1997. K. Hori, S. Katz, A. Klemm, R. Pandharipande, R. Thomas, C. Vafa, R. Vakil, and E. Zaslow, *Mirror Symmetry*, Clay Math. Inst., Amer. Math. Soc., 2002. E. Ionel, *Topological recursive relations in $H\sp {2g}(\mathcal{M}\sb {g,n})$*, Invent. Math.  **148** (2002), no. 3, 627–658. T. Jarvis, *Geometry of the moduli of higher spin curves*, Internat. J. Math.  **11** (2000), no. 5, 637–663. T. Jarvis, *Picard Group of the Moduli of Higher Spin Curves*, New York Journal of Mathematics, **7** (2001), 23–47. T. Jarvis, T. Kimura, and A. Vaintrob, *Moduli spaces of higher spin curves and integrable hierarchies*, Compositio Math.  **126** (2001), no. 2, 157–212. M. E. Kazarian, [*KP hierarchy for Hodge integrals*]{}, preprint 2007, `http://www.mi.ras.ru/~kazarian/papers/newwit0703.pdf`. M. E. Kazarian and S. K. Lando, [*An algebro-geometric proof of Witten’s conjecture*]{}, J. Amer. Math. Soc. [**20**]{} (2007), 1079-1089. Y.-S. Kim and K. Liu, [*A simple proof of Witten conjecture through localization*]{}, preprint 2005, math.AG/0508384. M. Kontsevich, [*Intersection theory on the moduli space of curves and the matrix Airy function*]{}, Comm. Math. Phys.  [**147**]{} (1992), 1–23. Y.-P. Lee, *Invariance of tautological equations I: conjectures and applications*, J. Eur. Math. Soc. [**10**]{} (2008), no. 2, 399–413. Y.-P. Lee, *Invariance of tautological equations II: Gromov–Witten theory*, (appendix with Y. Iwao,) math.AG/0605708, to appear in J. Amer. Math. Soc. Y.-P. Lee, *Witten’s conjecture and Virasoro conjecture up to genus two*, in [*Gromov–Witten theory of spin curves and orbifolds*]{}, 31–42, Contemp. Math., 403, Amer. Math. Soc., Providence, RI, 2006. Y.-P. Lee, *Notes on axiomatic Gromov–Witten theory and applications*, arXiv:0710.4349, to appear in the Proceedings of the 2005 AMS Summer Research Institute on Algebraic Geometry. Y.-P. Lee and R. Pandharipande, *Frobenius manifolds, Gromov–Witten theory, and Virasoro constraints*, in preparation.  Material needed for this paper are available from `http://www.math.princeton.edu/~rahulp/`. C.-C. M.  Liu, K. Liu, and J. Zhou, [*A proof of a conjecture of Mariño-Vafa on Hodge integrals*]{}, J. Diff. Geom. [**65**]{} (2003), no. 2, 289–340. K. Liu and H. Xu, [*A proof of the Faber intersection number conjecture*]{}, preprint 2008, arXiv:0803.2204. E. Looijenga, [*Intersection theory on Deligne-Mumford compactifications after Witten and Kontsevich*]{}, Séminaire Bourbaki, Vol. 1992/93, Astérisque No. 216 (1993), Exp. No.  768, 4, 187–212. E. Looijenga, [*On the tautological ring of ${{\mathcal{M}}}_g$*]{}, Invent. Math. [**121**]{} (1995), no. 2, 411–419. M. Markl, *Operads and Props*, preprint 2006, arXiv:math/0601129. M. Markl, S. Merkulov, and S. Shadrin, *Wheeled PROPs, graph complexes and the master equations*, preprint, arXiv:math/0610683. I. Madsen and M. Weiss, *The stable moduli space of Riemann surfaces: Mumford’s conjecture*, Ann. of Math. (2) **165** (2007), no. 3, 843–941. M. Mirzakhani, [*Weil-Petersson volumes and intersection theory on the moduli space of curves*]{}, J. Amer. Math. Soc. [**20**]{} (2007), no. 1, 1–23. T. Mochizuki, *The virtual class of the moduli stack of stable $r$-spin curves*, Comm. Math. Phys. **264** (2006), no. 1, 1–40. D. Mumford, [*Toward an enumerative geometry of the moduli space of curves*]{}, in [*Arithmetic and Geometry*]{}, Vol. II, M. Artin and J. Tate ed., 271–328, Prog.Math. [**36**]{}, Birk. Boston, Boston, MA, 1983.  A. Okounkov and R. Pandharipande, [*Gromov–Witten theory, Hurwitz numbers, and matrix models, I*]{}, math.AG/0101147. R. Pandharipande, [*Three questions in Gromov–Witten theory*]{}, in [*Proceedings of the International Congress of Mathematicians, Vol. II (Beijing, 2002)*]{}, 503–512, Higher Ed. Press, Beijing, 2002. A. Polishchuk and A. Vaintrob, *Algebraic construction of Witten’s top Chern class*, in [*Advances in algebraic geometry motivated by physics (Lowell, MA, 2000)*]{}, 229–249, Contemp. Math., 276, Amer. Math. Soc., Providence, RI, 2001. S. Shadrin, [*On the structure of Goulden-Jackson-Vakil formula*]{}, preprint 2008, `http://staff.science.uva.nl/~sshadrin/papers/gjvform.pdf`. S. Shadrin, [*BCOV theory via Givental group action on CohFT*]{}, preprint 2008, `http://staff.science.uva.nl/~sshadrin/papers/bcovgiv.pdf`. S.  Shadrin and D. Zvonkine, [*Changes of variables in ELSV-type formulas*]{}, Michigan Math. J. [**55**]{} (2007), no. 1, 209–228. C. Teleman, *The structure of 2D semi-simple field theories*, preprint 2007, arXiv:0712.0160. R. Vakil, [*The moduli space of curves and Gromov–Witten theory*]{}, in [*Enumerative Invariants in Algebraic Geometry and String Theory*]{} (Behrend and Manetti eds.), Lecture Notes in Math., Springer-Verlag, to appear, arXiv:0602347. E. Witten, [*Two dimensional gravity and intersection theory on moduli space*]{}, Surveys in Diff. Geom. [**1**]{} (1991), 243–310. E. Witten, [*Algebraic geometry associated with matrix models of two-dimensional gravity*]{}, in [*Topological methods in modern mathematics (Stony Brook, NY, 1991)*]{}, 235–269, Publish or Perish, Houston, TX (1993). [^1]: The formal definition of PROP, or product and permutation category, can be found in, e.g. [@mM]. Briefly, a PROP $(P,*,S, 1))$ is a symmetric strict monoidal category such that 1. the objects $Ob(P)$ are identified with the set $\mathbb{Z}_{\ge 0}$; 2. the product satisfies $m*n = m+n$, for any $m,n \in Ob(P)$ (hence the unit $1=0$); 3. $S$ is the permutation symmetry; 4. each hom-set $Mor_P(m,n)$ is a $k$-module and the operations of the monoidal category $P$ are compatible with this $k$-linear structure.

ArXiv

--- abstract: | Linear predictors form a rich class of hypotheses used in a variety of learning algorithms. We present a tight analysis of the empirical Rademacher complexity of the family of linear hypothesis classes with weight vectors bounded in $\ell_p$-norm for any $p \geq 1$. This provides a tight analysis of generalization using these hypothesis sets and helps derive sharp data-dependent learning guarantees. We give both upper and lower bounds on the Rademacher complexity of these families and show that our bounds improve upon or match existing bounds, which are known only for $1 \leq p \leq 2$. author: - | Pranjal Awasthi pranjalawasthi@google.com\ Google Research & Rutgers University\ New York, NY 10011, USA Natalie Frank nf1066@nyu.edu\ Department of Mathematics\ Courant Institute of Mathematical Sciences\ New York, NY 10012, USA Mehryar Mohri mohri@google.com\ Google Research &\ Courant Institute of Mathematical Sciences\ New York, NY 10011, USA bibliography: - 'lrad.bib' title: On the Rademacher Complexity of Linear Hypothesis Sets --- Introduction ============ Linear predictors form a rich class of hypotheses used in a variety of learning algorithms, including SVM [@CortesVapnik1995], logistic regression or conditional maximum entropy models [@BergerDellaPietraDellaPietra1996], ridge regression [@HoerlKennard1970], and Lasso [@Tibshirani1996]. Different regularizations or $\ell_p$-norm conditions are used to constrain the family of linear predictors. This short note gives a sharp analysis of the generalization properties of linear predictors for arbitrary $\ell_p$-norm upper bound constraints. To do so, we give tight upper bounds on the empirical Rademacher complexity of these hypothesis sets which we show are matched by lower bounds, modulo some constants. The notion of Rademacher complexity is a general complexity measure used to derive sharp data-dependent learning guarantees for different hypothesis sets, including margin bounds, which are key in the analysis of generalization for classification [@KoltchinskiiPanchenko2002; @BartlettMendelson2002; @MohriRostamizadehTalwalkar2018]. There are known upper bounds on the Rademacher complexity of linear hypothesis sets for some values of $p$, including $p = 1$ or $p = 2$ [@BartlettMendelson2002; @MohriRostamizadehTalwalkar2018], as well as $1 < p < 2$ [@KakadeSridharantTewari2008]. Our upper bounds on the empirical Rademacher complexity are tighter than those known for $1 \leq p < 2$ and match the existing one for $p = 2$. We further give upper bounds on the Rademacher complexity for other values of $p$ ($p > 2$). Our upper bounds are expressed in terms of $\| {{\mathbf X}}^\top \|_{2, p^*}$, where ${{\mathbf X}}$ is the matrix whose columns are the sample points and where $p^*$ conjugate number associated to $p$. We give matching lower bounds in terms of the same quantity for all values of $p$, which suggest the key role played by this quantity in the analysis of complexity. Much of the results presented here already appeared in [@AwasthiFrankMohri2020], in the context of the analysis of adversarial Rademacher complexity. Here, we present a more self-contained and detailed analysis, including the statement and proof of lower bounds. In Section \[sec:preliminaries\], we introduce some preliminary definitions and notation. We present our new upper and lower bounds on the Rademacher complexity of linear hypothesis sets in Section \[sec:linear\_function\_classes\] (Theorem \[th:main\] and Theorem \[th:lowerbound\]). The proof of the upper bounds is given in Appendix \[app:main\] and that of the lower bounds in Appendix \[app:lowerbound\]. Lastly, in Appendix \[app:compare\] we give a detailed analysis of how our bounds improve upon existing ones. Preliminaries {#sec:preliminaries} ============= We will denote vectors as lowercase bold letters (e.g., ${{\mathbf x}}$) and matrices as uppercase bold (e.g., ${{\mathbf X}}$). The all-ones vector is denote by ${\mathbf{1}}$. The Hölder conjugate of $p \geq 1$ is denoted by $p^*$. For a matrix ${{\mathbf M}}$, the $(p, q)$-*group norm* is defined as the $q$-norm of the $p$-norm of the columns of ${{\mathbf M}}$, that is $\| {{\mathbf M}}\|_{p, q} = \| (\|{{\mathbf M}}_1\|_1, \ldots, \|{{\mathbf M}}_d\|_p) \|_q$, where ${{\mathbf M}}_i$s are the columns of ${{\mathbf M}}$. Let ${\mathcal{F}}$ be a family of functions mapping from $\Rset^d$ to $\Rset$. Then, the *empirical Rademacher complexity* of ${\mathcal{F}}$ for a sample ${{\mathscr S}}= ({{\mathbf x}}_1, \ldots {{\mathbf x}}_m)$, is defined by $$\begin{aligned} \label{eq:erc_def} {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}) = \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\left[ \sup_{f\in {\mathcal{F}}}\frac 1 m \sum_{i=1}^m \sigma_i f({{\mathbf x}}_i) \right],\end{aligned}$$ where ${{\boldsymbol \sigma}}= (\sigma_1, \ldots, \sigma_m)$ is a vector of i.i.d.Rademacher variables, that is independent uniform random variables taking values in ${ \{ -1, +1 \} }$. The *Rademacher complexity* of ${\mathcal{F}}$, ${\mathfrak R}_m({\mathcal{F}})$, is defined as the expectation of this quantity: ${\mathfrak R}_m({\mathcal{F}}) = \operatorname*{\mathbb{E}}_{{{\mathscr S}}\sim {{\mathscr D}}^m}[{\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}})]$, where ${{\mathscr D}}$ is a distribution over the input space $\Rset^d$. The empirical Rademacher complexity is a key data-dependent complexity measure. For a family of functions ${\mathcal{F}}$ taking values in $[0, 1]$, the following learning guarantee holds: for any $\delta > 0$, with probability at least $1 - \delta$ over the draw of a sample $S \sim {{\mathscr D}}^m$, the following inequality holds for all $f \in {\mathcal{F}}$ [@MohriRostamizadehTalwalkar2018]: $$\operatorname*{\mathbb{E}}_{x \sim {{\mathscr D}}}[f(x)] \leq \operatorname*{\mathbb{E}}_{x \sim {{\mathscr S}}}[f(x)] + 2 {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}) + 3 \sqrt{\frac{\log \frac{2}{\delta}}{2m}},$$ where we denote by $\operatorname*{\mathbb{E}}_{x \sim {{\mathscr S}}}[f(x)]$ the empirical average of $f$, that is $\operatorname*{\mathbb{E}}_{x \sim {{\mathscr S}}}[f(x)] = \frac{1}{m} \sum_{i = 1}^m f(x_i)$. A similar inequality holds for the average Rademacher complexity ${\mathfrak R}_m({\mathcal{F}}_p)=\operatorname*{\mathbb{E}}_{{{\mathscr S}}\sim {{\mathscr D}}^m} [{\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}})]$: $$\operatorname*{\mathbb{E}}_{x \sim {{\mathscr D}}}[f(x)] \leq \operatorname*{\mathbb{E}}_{x \sim {{\mathscr S}}}[f(x)] + 2 {\mathfrak R}_m({\mathcal{F}}) + \sqrt{\frac{\log \frac{1}{\delta}}{2m}}.$$ An important application of these bounds is the derivation of margin bounds which are crucial in the analysis of classification. Fix $\rho > 0$. Then, for any $\delta > 0$, with probability at least $1 - \delta$ over the draw of a sample $S \sim {{\mathscr D}}^m$, the following inequality holds for all $f \in {\mathcal{F}}$ [@KoltchinskiiPanchenko2002; @MohriRostamizadehTalwalkar2018]: $$\begin{aligned} \operatorname*{\mathbb{E}}_{(x, y) \sim {{\mathscr D}}}[1_{y f(x) \leq 0}] & \leq \operatorname*{\mathbb{E}}_{(x, y) \sim {{\mathscr S}}}\left[\min \left(1, \Big( 1 - \tfrac{y f(x)}{\rho} \Big)_+ \right) \right] + \frac {2} \rho {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}) + 3\sqrt{\frac{\log \frac 2\delta}{2m}}\\ & \leq \frac{1}{m} \sum_{i = 1}^m 1_{y_i f(x_i) \leq \rho} + \frac {2} \rho {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}) + 3\sqrt{\frac{\log \frac 2\delta}{2m}}.\end{aligned}$$ Finer margin guarantees were recently presented by @CortesMohriSuresh2020 in terms of Rademacher complexity and other complexity measures. Furthermore, the Rademacher complexity of a hypothesis set also appears as a lower bound in generalization. As an example, for a symmetric family of functions ${{\mathscr G}}$ taking values in $[-1, +1]$, the following holds [@VaartWellner1996]: $$\frac{1}{2} \left[ {\mathfrak R}_m({{\mathscr G}}) - \frac{1}{\sqrt{m}} \right] \leq \operatorname*{\mathbb{E}}_{{{\mathscr S}}\sim {{\mathscr D}}^m} \left[ \sup_{f \in {{\mathscr G}}} \left| \operatorname*{\mathbb{E}}_{x \sim {{\mathscr D}}}[f(x)] - \operatorname*{\mathbb{E}}_{x \sim {{\mathscr S}}}[f(x)] \right| \right] \leq 2 {\mathfrak R}_m({{\mathscr G}}).$$ The hypothesis set we will analyze in this paper is that of linear predictors whose weight vector is bounded in $\ell_p$-norm: $$\begin{aligned} \label{eq:linear_function_class} {\mathcal{F}}_p = {\big \{ {{\mathbf x}}\mapsto {{\mathbf w}}\cdot {{\mathbf x}}\colon \| {{\mathbf w}}\|_p \leq W \big \} }.\end{aligned}$$ Empirical Rademacher Complexity of Linear Hypothesis Sets {#sec:linear_function_classes} ========================================================= The main results of this note are the following upper and lower bounds on the empirical Rademacher complexity of linear hypothesis sets. [theorem]{}[maintheorem]{} \[th:main\] Let ${\mathcal{F}}_p = { \{ {{\mathbf x}}\mapsto {{\mathbf w}}\cdot {{\mathbf x}}\, \colon \| {{\mathbf w}}\|_p \leq W \} }$ be a family of linear functions defined over $\Rset^d$ with bounded weight in $\ell_p$-norm. Then, the empirical Rademacher complexity of ${\mathcal{F}}_p$ for a sample ${{\mathscr S}}= ({{\mathbf x}}_1, \ldots, {{\mathbf x}}_m)$ admits the following upper bounds: $$\begin{aligned} {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_p) \leq \begin{cases} \frac W m\sqrt{{2\log(2d)}} \, \| {{{\mathbf X}}^\top}\|_{2, p^*} & \text{if $p = 1$} \nonumber \\ \frac{\sqrt{2}W}{m} \Bigg[\frac{\Gamma \left( \tfrac{p^* + 1}{2} \right)}{\sqrt{\pi}} \Bigg]^{\frac{1}{p^*}} \!\! \| {{\mathbf X}}^\top \|_{2, p^*} & \text{if $1 <p \le 2$} \\ \frac{W}{m}\| {{\mathbf X}}^\top \|_{2, p^*}, & \text{if $p \ge 2$} \nonumber \end{cases}\end{aligned}$$ where ${{\mathbf X}}$ is the $d \times m$-matrix with ${{\mathbf x}}_i$s as columns: ${{\mathbf X}}= [{{\mathbf x}}_1 \,\ldots\, {{\mathbf x}}_m]$. Furthermore, the constant factor in the inequality for the case $1 < p \leq 2$ can be bounded as follows: $$e^{-\frac 12}\sqrt{p^*} \leq \sqrt 2\bigg[\frac{\Gamma( \tfrac{p^* + 1}{2} )}{\sqrt{\pi}} \bigg]^{\frac{1}{p^*}} \leq e^{-\frac 12} \sqrt{p^* + 1}.$$ The proof is given in Appendix \[app:main\]. Both the statement of the theorem and its proof first appeared in [@AwasthiFrankMohri2020] in the context of the analysis of adversarial Rademacher complexity. We present a self-contained analysis in this note to make the results more easily accessible, as we believe these results are of a wider interest. The next theorem is new and provides a lower bound for ${\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_p)$ which, modulo a constant factor, matches the upper bounds stated above. [theorem]{}[lowerboundtheorem]{} \[th:lowerbound\] Let ${\mathcal{F}}_p = { \{ {{\mathbf x}}\mapsto {{\mathbf w}}\cdot {{\mathbf x}}\, \colon \| {{\mathbf w}}\|_p \leq W \} }$ be a family of linear functions defined over $\Rset^d$ with bounded weight in $\ell_p$-norm. Then, the empirical Rademacher complexity of ${\mathcal{F}}_p$ for a sample ${{\mathscr S}}= ({{\mathbf x}}_1, \ldots, {{\mathbf x}}_m)$ admits the following lower bound, where ${{\mathbf X}}= [{{\mathbf x}}_1 \,\ldots\, {{\mathbf x}}_m]$: $$\begin{aligned} {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_p) \geq \frac { W}{\sqrt 2 m} \| {{\mathbf X}}^\top \|_{2, p^*}.\end{aligned}$$ This lower bound is in tight in terms of dependence on sample size $m$ and dimension $d$. The proof is given in Appendix \[app:lowerbound\]. The following corollary presents somewhat looser upper bounds that may be more convenient in various contexts, such as that of kernel-based hypothesis sets. The corollary can be derived directly by combining Theorem \[th:main\] and Proposition \[prop:norm\_ratio\] (see Section \[sec:comparison\]). Let ${\mathcal{F}}_p = { \{ {{\mathbf x}}\mapsto {{\mathbf w}}\cdot {{\mathbf x}}\, \colon \| {{\mathbf w}}\|_p \leq W \} }$ be a family of linear functions defined over $\Rset^d$ with bounded weight in $\ell_p$-norm. Then, the empirical Rademacher complexity of ${\mathcal{F}}_p$ for a sample ${{\mathscr S}}= ({{\mathbf x}}_1, \ldots, {{\mathbf x}}_m)$ admits the following upper bounds, where ${{\mathbf X}}= [{{\mathbf x}}_1 \,\ldots\, {{\mathbf x}}_m]$:\ $$\begin{aligned} \text{for } p = 1, \qquad {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_p) & \leq \frac Wm \sqrt {{2\log(2d)}}\|{{\mathbf X}}\|_{p^*, 2};\\ \text{for } 1 < p \leq 2, \qquad {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_p) & \leq e^{-\frac 12}\sqrt{p^*+1}\frac Wm \|{{\mathbf X}}^\top\|_{2, p^*}\\ {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_p) & \leq e^{-\frac 12}\sqrt{p^*+1}\frac Wm \|{{\mathbf X}}\|_{p^*,2};\\ \text{for } p \geq 2, \qquad {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_p) & \leq \frac W m \|{{\mathbf X}}^\top\|_{2, p^*}\\ {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_p) & \leq \frac {W \min(m,d)^{\frac 1 {p^*}-\frac 12}}m \|{{\mathbf X}}\|_{p^*,2}.\end{aligned}$$ Discussion {#sec:discussion} ---------- We now make a few remarks about Theorem \[th:main\] and present the proof in Appendix \[app:main\]. The theorem states that for any data set, ${\widehat}{\mathcal{R}}_{{\mathscr S}}({\mathcal{F}}_p)$ is a constant times $\frac 1m \|{{\mathbf X}}^\top\|_{2, p^*}$. This is in contrast to the quantity $\|{{\mathbf X}}^\top\|_{p^*, 2}$ that appears in the existing analysis available in the literature for linear hypothesis sets [@KakadeSridharantTewari2008]. However, as we will soon see in Theorem \[th:rc\_linear\_comparison\] using $\|{{\mathbf X}}^\top\|_{2, p^*}$ always leads to a better upper bound. The quantity $\|{{\mathbf X}}^\top\|_{2, p^*}$ is thus the key term defining the behavior of the Rademacher complexity. It’s somewhat counterintuitive that $ \|{{\mathbf X}}^\top\|_{2, p^*}$ rather than $\|{{\mathbf X}}\|_{p^*,2}$ defines the size of the Rademacher complexity. To illustrate why this should be the case we start with an example. Consider a distribution ${\mathcal{D}}$ that puts probability 1 on a single $x$-value. A data set drawn from ${\mathcal{D}}$ will have the ${{\mathbf x}}_i$s constant, so we take ${{\mathbf x}}_i={{\mathbf x}}$. In this scenario we are trying to classify a single point, so we expect the Rademacher complexity to be independent of dimension. Further, because ${\mathcal{D}}$ has support $\{({{\mathbf x}},1), ({{\mathbf x}}, -1)$, one can show that ${\widehat}R_{\mathcal{S}}(f)\to R(f)$ at the rate $\frac 1 {sqrt m}$ for any function $f$. Now since this empirical Rademacher complexity is a quantity that defines rate of convergence of an estimator for a single point, one would expect that it should asymptotically behave like $\frac 1 {\sqrt m}$. However, $$\frac 1m \|{{\mathbf X}}^\top\|_{2, p^*}=\frac 1 {\sqrt m} \|{{\mathbf x}}\|_{p^*}$$ while $$\frac 1m \|{{\mathbf X}}^\top\|_{p^*,2}=\frac {\sqrt d} m \|{{\mathbf x}}\|_{p^*}$$ so we see that the $\|{{\mathbf X}}^\top\|_{2, p^*}$ is the norm that exhibits the desired behavior. The reason for this phenomenon is that trying to estimate the linear function ${{\mathbf w}}\in \Rset^d$ is qualitatively similar to trying to estimate $d$ different means, each which should have a rate of convergence roughly like $\frac 1 {\sqrt m}$. Just like the example above, taking the 2-norm of the *rows* of ${{\mathbf X}}$ captures this behavior. The $p^*$ component of the group norm then embodies how the magnitudes of the components of ${{\mathbf w}}$ can relate in the class ${\mathcal{F}}_p$. Another interesting aspect of the upper bound is the dimension dependence of the constant in front of $\|{{\mathbf X}}^\top\|_{2, p^*}$. This constant is independent of dimension only for $p > 1$. For $p = 1$, the $\sqrt{\log (d)}$ dependence on dimension is tight, which can be seen from the correspondence tightness of the maximal inequality and thus that of Massart’s inequality [@BoucheronLugosiMassart2013]. We also provide a simple example further illustrating this dependence in Appendix \[app:sqrt\_log\_d\]. This observation also explains why the constant for $p > 1$ approaches infinity as $p\to 1$: if we had that $${\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_p)\leq c(p) \|{{\mathbf X}}^\top\|_{2, p^*}$$ for $p>1$, then by continuity $${\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_1)\leq \lim_{p\to 1} c(p) \|{{\mathbf X}}^\top\|_{2,\infty}$$ If $c(p)$ were dimension independent and $\lim_{p\to \infty} c(p)$ were finite, then the constant for $p=1$ would be finite and dimension independent as well. Since we just showed that the constant for $p=1$ must have dimension dependence, we must have that $\lim_{p\to 1} c(p)=\infty$. This observation suggests that finding dimension-dependent constant for $1<p<2$ could greatly improve the upper bound of Theorem \[th:main\]. However, our example where the dimension dependence was tight for $p=1$ had $d=2^m$, which is unrealistic for most applications. It’s possible that with some reasonable assumption on the relationship between $m$ and $d$, one could find a far better constant for $1<p<2$. Comparison with Previous Work {#sec:comparison} ----------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------  A plot comparing two norms of the $4 \times 4$ identity matrix, $\|{{\mathbf I}}^\top\|_{2, p^*}$ and $\|{{\mathbf I}}\|_{p^*,2}$; the lower bound on the ratio of the two norms (\[eq:q\_leq\_p\]) in Proposition \[prop:norm\_ratio\] holds for this matrix. (b) Same as (a), but for Gaussian matrices.](matrix_norms_of_identity_vs_p_star.pdf "fig:")  A plot comparing two norms of the $4 \times 4$ identity matrix, $\|{{\mathbf I}}^\top\|_{2, p^*}$ and $\|{{\mathbf I}}\|_{p^*,2}$; the lower bound on the ratio of the two norms (\[eq:q\_leq\_p\]) in Proposition \[prop:norm\_ratio\] holds for this matrix. (b) Same as (a), but for Gaussian matrices.](normal_matrices_norm_plot.pdf "fig:") (a) (b) ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -.1in -.15in We are not aware of any existing bound for the empirical Rademacher complexity of linear hypothesis sets for $p > 2$ before this work. For other values of $p$, the best existing upper bounds were given by [@KakadeSridharantTewari2008] for $1 <p \leq 2$ and by [@BartlettMendelson2001] (see also [@MohriRostamizadehTalwalkar2018]) for $p = 1$: $$\begin{aligned} \label{eq:previous_linear_rc} {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_p)\leq \begin{cases} W \sqrt{\frac {2\log(2d)}m}\| {{\mathbf X}}^\top\|_{+\infty, +\infty} & \text{ if $p=1$} \\ \frac W m \sqrt{p^* - 1} \|{{{\mathbf X}}}\|_{p^*, 2} & \text{ if $1 < p \leq 2$} \end{cases}\end{aligned}$$ Our new upper bound coincides with (\[eq:previous\_linear\_rc\]) when $p=2$ and is strictly tighter otherwise. Readers familiar with Rademacher complexity bounds for linear hypothesis sets will notice that our bound in this case depends on the norm $\| {{\mathbf X}}^\top \|_{2, p^*}$. In contrast, the previously known bounds depend on $\| {{\mathbf X}}\|_{p^*, 2}$. In fact, one can show that the $\|{{\mathbf X}}^\top\|_{2,p^*}$ is always smaller than $\|{{\mathbf X}}\|_{p^*, 2}$ for $p \in (1,2]$, that is $p^*\geq 2$, as shown by the last inequality of in the following proposition. [proposition]{}[propnormratioprop]{} \[prop:norm\_ratio\] Let ${{\mathbf M}}$ be a $d\times m$ matrix. If $q\leq p$, then $$\label{eq:q_leq_p} \min(m,d)^{\frac 1p-\frac 1q} \|{{\mathbf M}}^\top\|_{p,q} \leq \|{{\mathbf M}}\|_{q,p} \leq \|{{\mathbf M}}^\top\|_{p,q}$$ If $q\geq p$, then $$\label{eq:p_leq_q} \min(m,d)^{\frac 1p-\frac 1q} \|{{\mathbf M}}^\top\|_{p,q}\geq \|{{\mathbf M}}\|_{q,p}\geq \|{{\mathbf M}}^\top\|_{p,q}$$ These bounds are tight. The proof is presented in Appendix \[sec:norm\_comparison\]. To visualize the ratio between these two norms, we plot the two norms for various values of $p^*$ in figure \[fig:norms\]. For convenience, in the discussion below, we set $c_1(p)=\sqrt{p^*-1}$ and $c_2(p)=\sqrt 2\big[\frac{\Gamma( \tfrac{p^* + 1}{2} )}{\sqrt{\pi}} \big]^{\frac{1}{p^*}}$. Regarding the growth of the constant in our bound, Theorem \[th:main\] implies that as $p^* \to \infty$, $c_2(p)$ grows asymptotically like $e^{-\frac 12}\sqrt {p^*} $. Furthermore, $c_2(p) \leq c_1(p)$ in the relevant region (See Appendix \[sec:const\_bound\]). In Figure \[fig:constants\] we plot $c_1(p),c_2(p)$ and the bounds on $c_2(p)$ to illustrate the growth rate of these constants with $p^*$. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -- ![A plot of $c_1(p)$, $c_2(p)$, and the bounds from Lemma \[lemma:constant-comparison\]. Note that $c_1(2) = c_2(2)$ and that the upper and lower bounds on $c_2$ are tight.](constants_figure.pdf "fig:") ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -- -.15in -.15in \[fig:constants\] Proposition \[prop:norm\_ratio\] and the inequality $c_2(p) \leq c_1(p)$ imply the following result. [theorem]{}[normratioprop]{} \[th:rc\_linear\_comparison\] For $p \leq 2$, the following inequality holds: $$\frac{\sqrt{2}W}{m} \bigg[\frac{\Gamma( \tfrac{p^* + 1}{2} )}{\sqrt{\pi}} \bigg]^{\frac{1}{p^*}} \|{{{\mathbf X}}^\top}\|_{2, p^*}\leq \frac W m \sqrt{p^*-1} \|{{{\mathbf X}}}\|_{p^*,2}$$ Thus, for $1 < p \leq 2$, the bound of Theorem \[th:main\] is tighter than . Conclusion ========== We presented tight bounds on the empirical Rademacher complexity of linear hypothesis sets constrained by an $\ell_p$-norm bound on the weight vector. These bounds can be used to derive sharp generalization guarantees for these hypothesis sets in a variety of different contexts, by plugging them in existing Rademacher complexity learning bounds. Our proofs and guarantees suggest an extension beyond $\ell_p$-norm constrained hypothesis sets that we will discuss elsewhere. Proof of Theorem \[th:main\] {#app:main} ============================ In this section, we present the proof of Theorem \[th:main\]. The proof proceeds in several steps. First, in Appendix \[sec:linear\_p\_1\] we upper bound the Rademacher complexity of ${\mathcal{F}}_1$. Next, in Appendix \[sec:new\_linear\_rc\_proof\], we establish the upper bound for $p > 1$. Lastly, in Appendix \[sec:const\_bound\], we prove the inequalities for the constant terms in the case $1 < p \leq 2$. Proof of the upper bound, case $p=1$ {#sec:linear_p_1} ------------------------------------ The bound on the Rademacher complexity for $p=1$ was previously known but we reproduce the proof of this theorem for completeness. We closely follow the proof given in [@MohriRostamizadehTalwalkar2018]. For any $i \in [m]$, $x_{ij}$ denotes the $j$th component of ${{\mathbf x}}_i$. $$\begin{aligned} {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_1) & = \frac{1}{m} \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\left[ \sup_{\| {{\mathbf w}}\|_1 \leq W} {{\mathbf w}}\cdot \sum_{i = 1}^m \sigma_i {{\mathbf x}}_i \right]\\ & = \frac{W}{m} \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\left[ \Big\| \sum_{i = 1}^m \sigma_i {{\mathbf x}}_i \Big\|_\infty \right] & \text{(by definition of the dual norm)}\\ & = \frac{W}{m} \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\left[ \max_{j \in [d]} \left| \sum_{i = 1}^m \sigma_i x_{ij} \right| \right] & \text{(by definition of $\| \cdot \|_\infty$)}\\ & = \frac{W}{m} \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\left[ \max_{j \in [d]} \max_{s \in { \{ -1, +1 \} }} s \sum_{i = 1}^m \sigma_i x_{ij} \right] & \text{(by definition of $| \cdot |$)}\\ & = \frac{W}{m} \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\left[ \sup_{{{\mathbf z}}\in {\mathcal{A}}} \sum_{i = 1}^m \sigma_i z_i \right] ,\end{aligned}$$ where ${\mathcal{A}}$ denotes the set of $d$ vectors ${ \{ s (x_{1j}, \ldots, x_{mj})^\top \colon j \in [d], s \in { \{ -1, +1 \} } \} }$. For any ${{\mathbf z}}\in A$, we have $\| {{\mathbf z}}\|_2 \leq \sup_{{{\mathbf z}}\in A} \|{{\mathbf z}}\|_2=\|{{\mathbf X}}^\top\|_{2,\infty}$. Further, ${\mathcal{A}}$ contains at most $2d$ elements. Thus, by Massart’s Lemma [@Massart2000; @MohriRostamizadehTalwalkar2018], $$\begin{aligned} {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_1) & \leq W \|{{\mathbf X}}^\top\|_{2, \infty} \frac{\sqrt{2 \log (2d)}}{m},\end{aligned}$$ which concludes the proof. Proof of upper bound, case $p > 1$ {#sec:new_linear_rc_proof} ---------------------------------- Here again, we use the shorthand ${{\mathbf u}}_{{\boldsymbol \sigma}}=\sum_{i = 1}^m \sigma_i{{\mathbf x}}_i$. By definition of the dual norm, we can write: $$\begin{aligned} {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_p) & = \frac{1}{m} \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\Bigg[ \sup_{\| {{\mathbf w}}\|_p \leq W } {{\mathbf w}}\cdot \sum_{i = 1}^m \sigma_i {{\mathbf x}}_i \Bigg]\\ & = \frac{W}{m} \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\big[ \| {{\mathbf u}}_{{\boldsymbol \sigma}}\|_{p^*} \big] & (\text{dual norm property})\\ & \leq \frac{W}{m} \Big[ \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\big[ \| {{\mathbf u}}_{{\boldsymbol \sigma}}\|_{p^*}^{p^*} \big]\Big]^{\frac{1}{p^*}}. & (\text{Jensen's inequality, $p^* \in [1, +\infty)$})\\ & = \frac{W}{m} \Big[ \sum_{j = 1}^d \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\big[ |{{\mathbf u}}_{{{\boldsymbol \sigma}}, j}|^{p^*} \big] \Big]^{\frac{1}{p^*}}.\end{aligned}$$ Next, by Khintchine’s inequality [@Haagerup1981], the following holds: $$\begin{aligned} \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\big[ |{{\mathbf u}}_{{{\boldsymbol \sigma}}, j}|^{p^*} \big] &\leq B_{p^*} \Big[ \sum_{i = 1}^m x_{i, j}^2 \Big]^{\frac{p^*}{2}},\end{aligned}$$ where $B_{p^*} = 1$ for $p^* \in [1, 2]$ and $$\begin{aligned} B_{p^*} &= 2^{\frac{p^*}{2}}\frac{\Gamma \big( \frac{p^* + 1}{2} \big)}{\sqrt{\pi}},\end{aligned}$$ for $p \in [2, +\infty)$. This yields the following bound on the Rademacher complexity: $${\widehat}{\mathfrak R}_S({\mathcal{F}}_p) \leq \begin{cases} \frac{W}{m}\| {{\mathbf X}}^\top \|_{2, p^*} & \text{if }p^* \in [1, 2], \\[.25cm] \frac{\sqrt{2}W}{m} \bigg[\frac{\Gamma \big( \tfrac{p^* + 1}{2} \big)}{\sqrt{\pi}} \bigg]^{\frac{1}{p^*}} \| {{\mathbf X}}^\top \|_{2, p^*} & \text{if } p^* \in [2, +\infty). \end{cases}$$ Bounding the Constant {#sec:const_bound} --------------------- For convenience, set $c_2(p)\colon = \sqrt 2 \big(\frac{\Gamma(\frac{p^*+1} 2 )}{\sqrt \pi}\big)^\frac 1 {p^*}$. We establish upper and lower bound on $c_2(p)$. \[lemma:const\_bound+\] \[lemma:f2\_bound\]Let $c_2(p)=\sqrt 2 \big(\frac{\Gamma(\frac{p^*+1} 2 )}{\sqrt \pi}\big)^\frac 1 {p^*}$. Then the following inequalities hold: $$e^{-\frac 12}\sqrt{ p^*}\leq c_2(p)\leq e^{-\frac 12} \sqrt {p^*+1}.$$ For convenience, we set $q = p^*$, $f_1(q)=c_1(p)$, $f_2(q)=c_2(p)$. Next, we recall a useful inequality [@OlverLozierBoisvertClark2010] bounding the gamma function: $$\begin{aligned} 1<(2\pi)^{-\frac 12}x^{\frac 12-x}e^x\Gamma(x)<e^{\frac 1 {12x}}.\label{eq:gamma_bound}\end{aligned}$$ We start with the upper bound. If we apply the right-hand side inequality of (\[eq:gamma\_bound\]) to $\Gamma(\frac {q+1}2)$ we get the following bound on $f_2(q)$: $$\label{eq:f2_bound} f_2(q)\leq 2^\frac 1 {2q} e^{-\frac 12} \sqrt{q+1} e^{-\frac 1{2q}+\frac 1 {6(q+1)q}}$$ It is easy to verify that, $$\label{eq:upper_bound_factor} 2^\frac 1{2q}e^{-\frac 1 {2q} +\frac 1 {6q(q+1)}}=e^{\frac 1q(\frac {\ln 2-1}2+\frac 1 {6q(q+1)})}.$$ Furthermore, the expression $(\frac {\ln 2-1}2+\frac 1 {6q(q+1)})$ decreases with increasing $q$. At $q=2$, it is negative, which implies that (\[eq:upper\_bound\_factor\]) is less than 1 for $q\geq 2$. Hence $$f_2(q) \leq e^{-\frac 12} \sqrt{q+1}$$ Next, we prove the lower bound. Applying the lower bound of (\[eq:gamma\_bound\]) to $\Gamma(\frac{q+1}2)$ results in $$f_2(q)\geq e^{-\frac 12}\sqrt q \left(e^{-\frac 1{2q}(\log 2 -1)}\sqrt{1+\frac 1 q}\right).$$ We will establish that $\Big(e^{-\frac 1{2q}(\log 2 -1)}\sqrt{1+\frac 1 q}\Big)\geq 1$, which will complete the proof of the lower bound. We prove this statement by showing that $$\left(e^{-\frac 1{2q}(\log 2 -1)}\sqrt{1+\frac 1 q}\right)^2 = e^{-\frac 1 q(\log 2 -1)} \left( 1+\frac 1q \right)\geq 1.$$ By applying some elementary inequalities $$\begin{aligned} e^{-\frac 1 q(\log 2 -1)}\left(1+\frac 1q\right)&\geq \left( \frac 1q(\log 2-1)+1\right) \left( 1 + \frac 1q \right) & (\text{using }e^x\geq 1+x)\\ &=1+\frac 1q \left( \log (2) -\frac{1-\log (2)}q\right)\\ &\geq 1\end{aligned}$$ The last inequality follows since $\Big( \log (2) -\frac{1-\log (2)}q\Big)$ increases with $q$, and is positive at $q = 2$. Proof of Theorem \[th:lowerbound\] {#app:lowerbound} ================================== In this section, we prove the lower bound of Theorem \[th:lowerbound\]. For any vector ${{\mathbf u}}$, let $|{{\mathbf u}}|$ denote the vector derived from ${{\mathbf u}}$ by taking the absolute value of each of its components. Starting as in the proof of Theorem \[th:main\], using the dual norm property, we can write: $$\begin{aligned} {\widehat}{\mathcal{R}}_{{\mathscr S}}({\mathcal{F}}_p) & = \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\left[\sup_{\|{{\mathbf w}}\|_p\leq W} {{\mathbf w}}\cdot \sum_{i = 1}^m \sigma_i{{\mathbf x}}_i \right] \\ & = \frac{W}{m} \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\left[\left\| \left| \sum_{i = 1}^m \sigma_i{{\mathbf x}}_i \right| \right\|_{p^*} \right]& (\text{dual norm property})\\ & \geq \frac{W}{m} \left\| \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\left[ \left| \sum_{i = 1}^m \sigma_i{{\mathbf x}}_i \right| \right] \right\|_{p^*} & (\text{norm sub-additivity})\\ & = \frac{W}{m} \left[\sum_{j = 1}^d \left(\operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\left[ \left| \sum_{i = 1}^m \sigma_i{{\mathbf x}}_{ij} \right| \right]\right)^{p^*} \right]^{\frac{1}{p^*}} \\ & \geq \frac{W}{m} \left[\sum_{j = 1}^d \left(\frac{1}{\sqrt{2}} \left| \sum_{i = 1}^m {{\mathbf x}}_{ij}^2 \right|^{\frac{1}{2}} \right)^{p^*} \right]^{\frac{1}{p^*}} & (\text{Khintchine's ineq. \citep{Haagerup1981}})\\ & = \frac{W}{\sqrt{2} \, m} \left[\sum_{j = 1}^d \left[ \left| \sum_{i = 1}^m {{\mathbf x}}_{ij}^2 \right| \right]^{\frac{p^*}{2}} \right]^{\frac{1}{p^*}} \\ & = \frac{W}{\sqrt{2} \, m} \| {{\mathbf X}}^\top \|_{2, p^*}.\end{aligned}$$ Proof of Proposition \[prop:norm\_ratio\] {#sec:norm_comparison} ========================================= In this section, we prove Proposition \[prop:norm\_ratio\]. This result implies that for $p \in (1, 2)$, the group norm $\|{{\mathbf X}}^\top\|_{2, p^*}$, is always a lower bound on the term $\|{{\mathbf X}}\|_{p^*,2}$ that appears in existing upper bounds. We first present a simple lemma helpful for the proof. \[lemma:norm\_ratio\] Let $1\leq p,r\leq \infty$ and $d$ be dimension. Then $$\sup_{\| {{\mathbf w}}\|_p\leq 1}\| {{\mathbf w}}\|_{{{r}^*}}=\max(1, d^{1-\frac 1r -\frac 1p})$$ We prove that, if $p\geq {{r}^*}$, then the following equality holds: $$\sup_{\| {{\mathbf w}}\|_p\leq 1}\| {{\mathbf w}}\|_{{{r}^*}}=d^{1-\frac 1r -\frac 1p},$$ and otherwise that the following holds: $$\sup_{\| {{\mathbf w}}\|_p\leq 1}\| {{\mathbf w}}\|_{{{r}^*}}= 1.$$ If $p\geq {{r}^*}$, by H[ö]{}lder’s generalized inequality with $\frac 1 {r^*} = \frac 1p + \frac 1s$, $$\sup_{\| {{\mathbf w}}\|_p\leq 1} \| {{\mathbf w}}\|_{{r}^*} \leq \sup_{\| {{\mathbf w}}\| _p \leq 1} \| {\mathbf{1}}\|_s \| {{\mathbf w}}\|_p = \| {\mathbf{1}}\|_s = d^\frac 1s = d^{\frac 1 {r^*} - \frac 1 p} = d^{1 - \frac 1r - \frac 1p}.$$ Note that equality holds at the vector $\frac 1 {d^\frac 1p}{\mathbf{1}}$, and this implies that the inequality in the line above is an equality. Now for $p\leq {{r}^*}$, $\| {{\mathbf w}}\|_p\geq \| {{\mathbf w}}\|_{{r}^*}$, implying that $\sup_{\| {{\mathbf w}}\|_p\leq 1}\| {{\mathbf w}}\|_{{{r}^*}}\leq 1$. Here, equality is achieved at a unit vector ${{\mathbf e}}_1$. We now present the proof of Proposition \[prop:norm\_ratio\]. First, follows from by substituting ${{\mathbf M}}={{\mathbf A}}^\top$ for a matrix ${{\mathbf A}}$: For $q\leq p$, $$\min(m,d)^{\frac 1p -\frac 1q} \|{{\mathbf A}}\|_{p,q}\leq \|{{\mathbf A}}^\top\|_{q,p}\leq \|{{\mathbf A}}\|_{p,q}$$ which implies that $$\|{{\mathbf A}}^\top\|_{q,p}\leq \|{{\mathbf A}}\|_{p, q}\leq \min(m,d)^{\frac 1q-\frac 1p}\|{{\mathbf A}}^\top\|_{q, p}$$ However, now $p$ and $q$ are swapped in comparison to . Now after swapping them again, for $p\leq q$, $$\|{{\mathbf A}}^\top\|_{p,q}\leq \|{{\mathbf A}}\|_{q,p}\leq \min(m,d)^{\frac 1p-\frac 1q}\|{{\mathbf A}}^\top\|_{p,q}$$ The rest of this proof will be devoted to showing . Next, if $p=q$, then $\| {{\mathbf M}}\|_{q,p}=\| {{{\mathbf M}}^\top}\|_{p,q}$. For the rest of the proof, we will assume that $q < p$. Specifically, $q < +\infty$ which allows us to consider fractions like $\frac pq$. We will show that for $q < p$, the following inequality holds: $\|{{\mathbf M}}\|_{q, p} \leq \|{{\mathbf M}}^\top\|_{p, q}$, or equivalently, $\|{{\mathbf M}}\|^q_{q, p} \leq \|{{\mathbf M}}^\top\|^q_{p, q}$. We will use the shorthand $r = \tfrac{p}{q} > 1$. By definition of the group norm and using the notation ${{\mathbf U}}_{ij} = |{{\mathbf M}}_{ij}|^p$, we can write $$\begin{aligned} \| {{\mathbf M}}\|^q_{q, p} = \bigg[ \sum_{i = 1}^m \Big[\sum_{j = 1}^d |{{\mathbf M}}_{ij}|^q \Big]^{\frac{p}{q}} \bigg]^{\frac{q}{p}} = \bigg[ \sum_{i = 1}^m \Big[ \sum_{j = 1}^d{{\mathbf U}}_{ij} \Big]^{r} \bigg]^{\frac{1}{r}} \mspace{-10mu} & = \left\| \left[ \begin{smallmatrix} \sum_{j = 1}^d {{\mathbf U}}_{1j}\\ \vdots\\ \sum_{j = 1}^d {{\mathbf U}}_{mj} \end{smallmatrix} \right] \right\|_{r}\\ & \leq \sum_{j = 1}^d \left\| \left[ \begin{smallmatrix} {{\mathbf U}}_{1j}\\ \vdots\\ {{\mathbf U}}_{mj} \end{smallmatrix} \right] \right\|_r = \sum_{j = 1}^d \Big[ \sum_{i = 1}^m|{{\mathbf M}}_{ij}|^p \Big]^{\frac{q}{p}} = \|{{\mathbf M}}^\top\|^q_{p, q}.\end{aligned}$$ To show that this inequality is tight, note that equality holds for an all-ones matrix. Next, we prove the inequality $$\min(m, d)^{\frac 1q - \frac 1p} \|{{\mathbf M}}^\top\|_{p, q} \leq \|{{\mathbf M}}\|_{q, p},$$ for $q \leq p$. Applying Lemma \[lemma:norm\_ratio\] twice gives $$\label{eq:lin_app_d} \|{{\mathbf M}}^\top \|_{p,q}\leq \| {{\mathbf M}}^\top\|_{q,q}=\|{{\mathbf M}}\|_{q,q}\leq d^{\frac 1q-\frac 1p}\|{{\mathbf M}}\|_{p,q}.$$ Again applying Lemma \[lemma:norm\_ratio\] twice gives $$\label{eq:lin_app_m} \|{{\mathbf M}}^\top\|_{p,q}\leq m^{\frac 1q-\frac 1p}\|{{\mathbf M}}^\top\|_{p,p}=m^{\frac 1q-\frac 1p}\|{{\mathbf M}}\|_{p,p}\leq m^{\frac 1q-\frac 1p}\|{{\mathbf M}}\|_{p,q}.$$ Next, we show that is tight if $d\leq m$ and that is tight if $d \geq m$. If $d\leq m$, the bound is tight for the block matrix ${{\mathbf M}}= \left[ \begin{smallmatrix} {{\mathbf I}}_{d \times d} \ | \ \mathbf 0 \end{smallmatrix} \right]$, and, if $d\geq m$, then the bound is tight for the block matrix ${{\mathbf M}}= \left[ \begin{smallmatrix} {{\mathbf I}}_{d\times d} \\[.075cm] \hline\\ \mathbf 0 \end{smallmatrix} \right].$ Proof of Theorem \[th:rc\_linear\_comparison\] {#app:compare} ============================================== Both Theorem \[th:main\] and equation present upper bounds on ${\widehat}{\mathfrak R}_{{{\mathscr S}}}({\mathcal{F}}_p)$ for $1 < p \leq 2$. Both of these bounds are of the form a constant times a matrix norm of ${{\mathbf X}}$. In Appendix \[sec:norm\_comparison\], we compared the two matrix norms and proved the inequality $\|{{\mathbf X}}^\top\|_{2, p^*}\leq \|{{\mathbf X}}\|_{p^*,2}$ in the relevant region (Lemma \[lemma:norm\_ratio\]). Here, we compare the two constants and show that the constant associated with Theorem \[th:main\] is smaller than the one appearing in (Lemma \[lemma:constant-comparison\]). These lemmas combined directly prove Theorem \[th:rc\_linear\_comparison\]. In this section, we study the constants in the two known bounds on the Rademacher complexity of linear classes for $1 < p \leq 2$. Specifically, [[R]{}\_[[S]{}]{}(\_p)]{} W m \_[p\^\*,2]{} & \[eq:previous\_linear\_rc\_2\]\ \^ \_[2, p\^\*]{} & \[eq:new\_linear\_rc\] We will compare the constants in equations (\[eq:previous\_linear\_rc\_2\]) and (\[eq:new\_linear\_rc\]), namely $\frac{\sqrt 2 W}m \big(\frac{\Gamma(\frac{p^*+1} 2 )}{\sqrt \pi}\big)^\frac 1 {p^*} $ and $\frac W m\sqrt{p^*-1}$. Since $\frac Wm$ divides both of these constants, we drop this factor and work with the expressions $c_1(p) \colon = \sqrt{p^* - 1}$ and $c_2(p)\colon = \sqrt 2 \big(\frac{\Gamma(\frac{p^*+1} 2 )}{\sqrt \pi}\big)^\frac 1 {p^*}$. Here we establish our main claim that $c_2(p) \leq c_1(p)$. \[lemma:constant-comparison\] Let $c_1(p)=\sqrt{p^*-1}$ and $c_2(p)=\sqrt 2 \big(\frac{\Gamma(\frac{p^*+1} 2 )}{\sqrt \pi}\big)^\frac 1 {p^*}$. Then $$c_2(p) \leq c_1(p),$$ for all $1 \leq p \leq 2$. First note that $c_1(2)=c_2(2)$. For convenience, set $q=p^*$, $f_1(q)=c_1(p)$, and $f_2(q)=c_2(p)$. We claim $\frac d {dq} f_1(q)\geq \frac d {dq} f_2(q)$ for $q\geq 2$, and this implies that $c_2(p)\leq c_1(p)$ for $1\leq p\leq 2$. The rest of this proof is devoted to showing that $\frac d {dq} f_1(q)\geq \frac d {dq} f_2(q)$. Upon differentiating we get that $f_1'(q)=\frac 1 {2\sqrt{q-1}}$. Next, we will differentiate $f_2$. To start, we state a useful inequality (see Equation $2.2$ in [@Alzer1997]) bounding the digamma function, $\psi(x)$. $$\begin{aligned} \psi(x)\leq \log(x)-\frac 1{2x} \label{eq:digamma_bound}\end{aligned}$$ Recall that the digamma function is the logarithmic derivative of the gamma function, $\psi(x) = \frac d {dx} (\log\Gamma(x))=\frac{\Gamma'(x)}{\Gamma(x)}$. Now we differentiate $\ln f_2$: $$\begin{aligned} \frac d {dq} (\ln f_2(q)) & =\frac{ \frac q2 \psi(\frac {q+1} 2)-(\ln(\Gamma(\frac {q+1} 2))-\ln(\sqrt \pi))}{q^2}\\ & \leq\frac{ \frac q2(\log(\frac {q+1}2-\frac 1 {q+1})-(\ln(\Gamma(\frac{q+1}2))-\ln\sqrt \pi)}{q^2} &\,\text{(by (\ref{eq:digamma_bound}))}\\ & \leq \frac {\frac q2(\log \frac {q+1}2 -\frac 1 {q+1})-(\frac 12 \ln2 +\frac q2\log \frac {q+1}2-\frac {q+1}2)}{q^2}&\text{(by the left-hand equality in (\ref{eq:gamma_bound}))}\\ & = \frac 1 {2q}+\frac 1 {q^2}\Big(\frac 1 {2(q+1)}-\frac 12\log2\Big)\\ & \leq \frac 1 {2q}.\end{aligned}$$ The last line follows since we only consider $q\geq 2$ and $\frac 1 {2(q+1)}-\frac 12\ln 2\leq 0$ in this range. Finally, the fact that $\frac d {dq}(\ln f_2(q))=f_2'(q)/f_2(q)$ implies $$\begin{aligned} f_2'(q) & = f_2(q) \frac d {dq} (\ln f_2(q))\\ & \leq \frac 1 {2q} f_2(q)&\,\text{(by }\frac d {dq}(\ln f_2(q))\leq \frac 1 {2q} \text{)}\\ & \leq \frac{e^{-\frac 12} \sqrt{q+1}}{2q}&\,\text{(by applying the upper bound in Lemma~\ref{lemma:f2_bound})}\\ & =\frac 1 {2\sqrt{q-1}}\frac{e^{-\frac 12}\sqrt{(q+1)(q-1)}}q\\ & \leq e^{-\frac 12} \frac 1{2\sqrt{q-1}}&\,\text{(using }q^2-1\leq q^2)\\ & \leq \frac 1 {2\sqrt{q-1}}=f_1'(q)&\,(\text{using }e^{-\frac 12}<1).\end{aligned}$$ The Tightness of the $\sqrt{\log(d)}$ factor for $p=1$ {#app:sqrt_log_d} ====================================================== Here, we provide an example showing that the dimension dependence of $\sqrt{\log(d)}$ in our upper bound on the Rademacher complexity of linear functions bounded in $\ell_1$ norm is tight. Consider a data set with $d=2^m$. Then the data matrix ${{\mathbf X}}$ has $2^m$ rows. We pick the data $\{{{\mathbf x}}_i\}$ so that the rows of ${{\mathbf X}}$ are the set $\{-1,+1\}^m$. This means that $\|{{\mathbf X}}^\top\|_{2, p^*}=\sqrt m$ and we can compute the Rademacher complexity as $$\begin{aligned} {\widehat}{\mathfrak R}_{{\mathscr S}}({\mathcal{F}}_1)&= \frac 1m \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\left[ \sup_{\|{{\mathbf w}}\|_1 \leq W}{{\mathbf w}}\cdot \sum_{i=1}^m \sigma_i {{\mathbf x}}_i\right]= \frac 1m \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\left[ \left\|\sum_{i=1}^m \sigma_i {{\mathbf x}}_i\right\|_\infty \right]&\text{(definition of dual norm)}\\ &=\frac 1m \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\left[ \max_{1\leq j\leq d} \sum_{i=1}^m \sigma_i ({{\mathbf x}}_i)_j\right]= \frac 1m \operatorname*{\mathbb{E}}_{{\boldsymbol \sigma}}\left[ m\right]&\text{(tightness of Cauchy-Schwartz)}\\ &=\frac m m=\frac 1m \sqrt m \sqrt m= \frac 1m \sqrt{\log (d)} \|{{\mathbf X}}^\top\|_{2,\infty} &(d=2^m,\|{{\mathbf X}}^\top\|_{2,\infty}=\sqrt m) \end{aligned}$$ Therefore, the $\sqrt{\log(d)}$ dependence in the constant for $p=1$ is tight.

ArXiv

--- author: - | $^1$, Andrea Lapi$^{1,2}$, Zhen-Yi Cai$^{3}$, Mattia Negrello$^{4}$, Gianfranco De Zotti$^{1,4}$, Francesca Perrotta$^{1}$, Luigi Danese$^{1}$\ $^1$Astrophysics Sector, SISSA, Via Bonomea 265, I-34136 Trieste ; $^2$Dipartimento di Fisica, Università ‘Tor Vergata’, Via della Ricerca Scientifica 1, I-00133 Roma, Italy ; $^3$Center for Astrophysics, University of Science and Technology of China, Hefei, 230026, China; $^4$INAF - Osservatorio Astronomico di Padova, Vicolo dell’Osservatorio 5, I-35122 Padova, Italy\ E-mail: title: Radio Observations of Star Forming Galaxies in the SKA era --- Introduction {#sect:intro} ============ The huge amount of infrared (IR) to millimeter-wave data that has been accumulating in the last several years thanks to *Spitzer*, *Herschel*, SCUBA and South Pole Telescope (SPT) surveys has made possible spectacular advances in our understanding of galaxy evolution. In fact the interstellar dust absorbs and re-emits at IR wavelengths about half of the starlight of the Universe [@Casey2014]. Hence, the evolution of the IR luminosity function directly maps that of the dust-enshrouded star formation rate (SFR). *Herschel* data are particularly powerful in this respect as they probe the dust emission peak, thus providing the best estimates of the total IR luminosity. However, the IR emission misses the starlight not absorbed by dust and therefore underestimates the SFR whenever the absorption optical depth is not very high. This is the case for the earliest phases of galaxy evolution when the metal enrichment of the interstellar medium is just beginning, as well as for dust-poor dwarf galaxies and metal-poor regions of more-massive galaxies [@KennicuttEvans2012]. A complete inventory of the SFR requires a combination of IR and UV photometry and thus suffers from limitations in both wavebands. An important alternative tool for measuring the cosmic star formation history of the Universe is provided by deep radio continuum surveys. A tight relationship between low-frequency radio (synchrotron) and IR luminosity (hence SFR) has long been established [@Helou1985; @Condon1992; @Yun2001; @Ivison2010; @Jarvis2010; @Bourne2011; @Mao2011]. However the physical basis of this relationship is not yet totally clear. In fact many physical processes (propagation of relativistic electrons, strength and structure of the magnetic field, size and composition of dust grains) must conspire together to produce this relation [@Bell2003; @Helou1993; @Niklas1997; @Murphy2009; @Lacki2010; @Hippelein2003]. Hence it is not granted that the relation also applies to redshift/luminosity ranges where the available data are insufficient to test it accurately. Moreover we cannot be sure that the observed synchrotron emission is not contaminated by faint nuclear activity. There is a second process contributing to the radio emission of star-forming galaxies: the free-free emission from hot electrons, which is directly proportional to the production rate of ionising photon by young, massive stars. It shows up at rest frame frequencies of tens of GHz, where it is generally optically thin, and thus offers a clean way to quantify the current star formation activity in galaxies. This picture could be complicated by the presence of anomalous dust emission [@PlanckCollaborationXX2011 and references therein] which occurs at similar frequencies and is thought to arise from spinning dust grains [e.g., @DraineLazarian1998]. However there is currently no evidence that this component contributes significantly to globally integrated measurements [@Murphy2012]. Thus high frequency radio observations may be particularly powerful for precisely measuring the star formation history of the Universe. Very deep radio surveys have shown that, at GHz frequencies, the counts below 100–$200\,\mu$Jy are dominated by star-forming galaxies [@Padovani2011 and references therein]. Current surveys only extend to a few tens of $\mu$Jy, i.e. cover a flux density range where, at low radio frequencies, the detected radio emission is of synchrotron origin. Only with the advent of the Square Kilometer Array (SKA) we expect that the high-$z$ star-forming galaxies can be seen via their free-free emission (see also Murphy, E. J., et al. 2015, “The Astrophysics of Star Formation Across Cosmic Time at z $\geq$ 10 GHz with the Square Kilometre Array”, in proceedings of “Advancing Astrophysics with the Square Kilometre Array”, ). In this paper we carry out a thorough investigation of the radio counts of star forming galaxies. In Sect. \[sect:model\] we present a short outline of the model, that builds on the work by @Cai2013 and @Cai2014. In Sect. \[sect:calibration\] we discuss the calibration of the relation between radio emissions and SFR. The South Pole Telescope (SPT) surveys at mm wavelengths are especially useful to test the SFR/free-free relation; this motivated a re-analysis of the SPT 95 GHz sample of dusty galaxies. These relations allow us to extend to radio frequencies the @Cai2013 model for the cosmological evolution of star forming galaxies and exploit it, in Sect. \[sect:SKAcounts\], to work out predictions for the counts of such galaxies in the range 1.4 – 30 GHz. We also compare the coverage of the SFR–$z$ plane by *Herschel*, UV surveys and SKA. Finally, Sect. \[sect:conclusions\] summarizes our main conclusions. Throughout this paper we adopt a flat $\Lambda \rm CDM$ cosmology with matter density $\Omega_{\rm m} = 0.32$, $\Omega_{\rm b} = 0.049$, $\Omega_{\Lambda} = 0.68$, Hubble constant $h=H_0/100\, \rm km\,s^{-1}\,Mpc^{-1} = 0.67$, spectrum of primordial perturbations with index $n = 0.96$ and normalization $\sigma_8 = 0.83$ [@PlanckCollaborationXVI2013]. Outline of the model {#sect:model} ==================== A direct tracer of recent star formation is the UV emission of galaxies, coming from the photospheric emission of massive young stars. In recent years a great effort has been made to measure the UV luminosity functions up to high redshifts [@Bouwens2008; @Bouwens2011; @Smit2012; @Oesch2012; @Oesch2013a; @Oesch2013b; @Schenker2013; @McLure2013], with the aim of reconstructing the history of cosmic re-ionization. However, as the chemical enrichment of the ISM proceeds and, correspondingly, the dust abundance increases, a larger and larger fraction of starlight is absorbed and re-emitted at far-IR wavelengths. The most active star-formation phases of high-$z$ galaxies indeed suffer by strong dust obscuration and are most effectively studied in the far-IR/sub-mm region. A comprehensive investigation of the evolution of the IR luminosity functions has been recently carried out by @Cai2013 based on a “hybrid” approach that reflects the observed dichotomy in the ages of stellar populations of early-type galaxies on one side and late-type galaxies on the other [cf. @Bernardi2010 their Fig. 10]. Early-type galaxies and massive bulges of Sa galaxies are composed of relatively old stellar populations with mass-weighted ages $\gtrsim 8$–9Gyr (corresponding to formation redshifts $z\gtrsim 1$–1.5), while the disk components of spirals and the irregular galaxies are characterized by significantly younger stellar populations. Thus the progenitors of early-type galaxies, referred to as proto-spheroidal galaxies or protospheroids, are the dominant star-forming population at $z\gtrsim 1.5$, while IR galaxies at $z\lesssim 1.5$ are mostly late-type “cold” (normal) and “warm” (starburst) galaxies. The @Cai2013 model accurately fits a broad variety of data[^1]: multi-frequency and multi-epoch luminosity functions of galaxies and AGNs, redshift distributions, number counts (total and per redshift bins). Moreover, it accurately accounts for the recently determined counts and redshift distribution of strongly lensed galaxies detected by the South Pole Telescope [SPT; @Mocanu2013; @Weiss2013], published after the paper was completed [@Bonato2014]. In general the total (8–$1000\,\mu$m) IR luminosity, $L_{\rm IR}$, is a good proxy of the obscured SFR when the dust heating is dominated by young stars. In disks of normal galaxies, however, the IR luminosity is the sum of a “warm” component heated by young stars and of a “cold” (or “cirrus”) component, heated by the general radiation field that may be dominated by older stars. This issue was investigated by @Clemens2013 using a complete sample of local star-forming galaxies detected by *Planck*. We adopt the relation between SFR and $L_{\rm IR}$ derived by these authors, for a Chabrier initial mass function (IMF). As mentioned in Sect. \[sect:intro\], to get a complete census of the cosmic SFR we need to complement IR measurements with UV SFR tracers to measure the unattenuated starlight. For the high-$z$ proto-spheroidal galaxies we adopted the accurately tested physical model for the evolution of the UV luminosity function worked out by @Cai2014 in the framework of the scenario proposed by @Granato2004 and further elaborated by @Lapi2006 [@Lapi2011], @Mao2007, and @Cai2013. For the low-$z$ galaxies we have supplemented the @Cai2013 model for evolution of late-type galaxies in the IR with a parametric model for evolution in the UV. Briefly, the UV ($\lambda = 1500\,$Å) luminosity function is described by: $$\label{eq:LF} \Phi (\log L_{1500}, z) \hbox{d}\,\log L_{1500} = \Phi^* \Big(\frac{L_{1500}}{L^*}\Big)^{1-\alpha} \times \exp \Big[- \frac{\log ^2 (1+L_{1500}/L^*)}{2\sigma^2} \Big] \hbox{d}\,\log L_{1500}.$$ A simple pure luminosity evolution model ($L^*(z) = L^*_0 (1+z)^{\alpha_L}$ up to $z = 1$) turned out to provide a sufficiently good description of the data (Mancuso et al., in preparation). The best-fit values of the parameters are $\log(\Phi^*_0/[\rm dex^{-1}\,Mpc^{-3}]) = -2.150 \pm 0.095$, $\log(L^*_0/L_\odot) = 9.436 \pm 0.119$, $\alpha = 1.477 \pm 0.050$, $\sigma = 0.326 \pm 0.035$, and $\alpha_L = 2.025 \pm 0.063$. The @KennicuttEvans2012 calibrations were adopted to convert the UV luminosity functions into SFR functions. The total redshift-dependent SFR functions were then computed summing the IR-based and the UV-based ones. [ ![*left panel*: Euclidean normalized differential counts at 1.4 GHz compared with model predictions. The dashed brown line shows the best fit model for radio AGNs by @Massardi2010. The other dashed lines show the contributions of the star-forming populations considered in this paper. The two triple-dot dashed lines show the total synchrotron and free-free emissions from these populations. The solid vertical lines correspond, from left to right, to three $5\,\sigma$ detection limits of preliminarily planned SKA1-MID surveys. The vertical blue dash-dotted line and the pink dashed line are respectively the flux limits for Ultra-Deep SKA2 (MID) and for the 50$\%$ Wide SKA1-MID surveys, while the green dash-dot-dotted line is the limit for both 50$\%$ Deep SKA1-MID and Wide SKA2 (MID) surveys. *Right panel*: our re-assessment of the integral counts of dusty galaxies at $\nu$=95 GHz (asterisks with error bars) compared with the expectations from the @Cai2013 model. []{data-label="fig:1.4_95GHz_counts"}](conteggi_1_4_paper_ska.eps "fig:"){width="50.00000%"} ![*left panel*: Euclidean normalized differential counts at 1.4 GHz compared with model predictions. The dashed brown line shows the best fit model for radio AGNs by @Massardi2010. The other dashed lines show the contributions of the star-forming populations considered in this paper. The two triple-dot dashed lines show the total synchrotron and free-free emissions from these populations. The solid vertical lines correspond, from left to right, to three $5\,\sigma$ detection limits of preliminarily planned SKA1-MID surveys. The vertical blue dash-dotted line and the pink dashed line are respectively the flux limits for Ultra-Deep SKA2 (MID) and for the 50$\%$ Wide SKA1-MID surveys, while the green dash-dot-dotted line is the limit for both 50$\%$ Deep SKA1-MID and Wide SKA2 (MID) surveys. *Right panel*: our re-assessment of the integral counts of dusty galaxies at $\nu$=95 GHz (asterisks with error bars) compared with the expectations from the @Cai2013 model. []{data-label="fig:1.4_95GHz_counts"}](conteggi_ska_95_mu2.eps "fig:"){width="50.00000%"} ]{} Calibration of the relation between radio emission and SFR {#sect:calibration} ========================================================== A tight linear correlation between the 1.4 GHz luminosity, dominated by synchrotron, and the IR luminosity has been established since many years [@Condon1992]. A calibration of the relation between the SFR and the synchrotron emission was calculated by @Murphy2011. In order to take into account electron ageing effects [@BandayWolfendale1991] we have adopted a steepening by $\Delta \alpha=0.5$ above a break frequency of 20 GHz. The relationship between synchrotron luminosity and SFR then writes: $$\label{eq:Lsync} L_{\rm sync}\simeq 1.9\times 10^{28} \left(\frac{\hbox{SFR}}{\hbox{M}_{\odot}\hbox{yr}^{-1}}\right) \left(\frac{\nu}{\hbox{GHz}}\right)^{-0.85}\left[1+\left(\frac{\nu}{ 20\rm GHz}\right)^{0.5}\right]^{-1}\, \hbox{erg}\,\hbox{s}^{-1}\,\hbox{Hz}^{-1}.$$ Coupling this relation with the redshift dependent SFR functions yielded by the model outlined in Sect. \[sect:model\] we get a good fit to the sub-mJy 1.4 GHz counts (Fig. \[fig:1.4\_95GHz\_counts\], left panel) without any adjustment of the parameters. The high-frequency synchrotron emission is increasingly suppressed with increasing $z$, as the timescale for energy losses of relativistic electrons by inverse Compton scattering off the Cosmic Microwave Background photons decreases as $(1+z)^{-4}$ [@Norris2013; @Carilli2008; @Murphy2009]. This may lower the counts at the faintest flux densities, but not by a large factor since, as illustrated by Fig. \[fig:1.4\_95GHz\_counts\] (left panel), below the limits of current surveys the free-free contribution is comparable to the synchrotron one. A relationship between SFR and free-free emission was derived by @Murphy2012. We have reformulated it as: $$\label{eq:Lff} L_{\rm ff}=3.75\times 10^{26} \left(\frac{\hbox{SFR}}{M_\odot/\hbox{yr}}\right) \, \left(\frac{T}{10^4\,\hbox{K}}\right)^{-0.5}\, \hbox{g}(\nu,\hbox{T})\,\exp{\left(-\frac{h\nu}{ k\hbox{T}}\right)}\, \hspace{3pt} \hbox{erg}\,\hbox{s}^{-1}\,\hbox{Hz}^{-1}$$ where T is the temperature of the emitting plasma and $\hbox{g}(\nu,\hbox{T})$ is the Gaunt factor for which we adopt the approximation proposed by @Draine2011 which is more accurate than the one used by @Murphy2012. The coefficient of eq. (\[eq:Lff\]) was computed requiring that this equation equals that by @Murphy2012 for $\nu=33\,$GHz (the frequency at which the relation was calibrated), $T= 10^4\,$K and a pure hydrogen plasma. We note that the calibrations of the above relationships are based on the Kroupa IMF while that between the IR emission and the SFR (Sect. \[sect:model\]) relies on the Chabrier IMF. However, as shown by @ChomiukPovich2011 the two IMFs give almost identical calibrations. To test the $L_{\rm ff}$-SFR relation we used the South Pole Telescope (SPT) observations of dusty galaxies at 95GHz [@Mocanu2013], since we expect that, at this frequency, the free-free emission shows up clearly in local galaxies not hosting a radio loud AGN. To estimate the 95GHz counts of dusty galaxies @Mocanu2013 adopted a statistical approach. They choose the local minimum in the distribution of the $\alpha^{150}_{220}$ spectral indices, $\alpha^{150}_{220}=1.5$, as the threshold for source classification and computed, for each source, the probability that their posterior $\alpha^{150}_{220}$ is greater than the threshold value, $P(\alpha^{150}_{220} > 1.5)$. This quantity was interpreted as the probability that a source is dust-dominated. The $95$ GHz differential counts were computed as the sum of probabilities $P(\alpha^{150}_{220} > 1.5)$. Since the fraction of dusty galaxies is much lower than that of synchrotron dominated sources, this statistical approach is endowed with large uncertainties and may strongly overestimate the counts of dusty galaxies. We have then re-estimated the counts using a lengthier but safer approach, i.e. we have checked the SED of each 95GHz source brighter than the 95% completeness limit of 12.6 mJy, collecting all the photometric data available in the literature. To model the SEDs we considered both synchrotron and free-free emission, as well as thermal dust emission. We found that only 4 sources, all with $P(\alpha^{150}_{220} > 1.5)\simeq 1$, are indeed dusty galaxies. This is a factor $\simeq 3$ below the number found by @Mocanu2013. The corresponding integral count is shown in the right panel of Fig. \[fig:1.4\_95GHz\_counts\]. Further tests and/or predictions of the model are provided by counts at 4.8, 8.4 ,15 and 30 GHz (Fig. \[fig:RadioCounts\]). The relative importance of free-free compared to synchrotron obviously increases with increasing frequency. While at 1.4 GHz the free-free contribution is always below the synchrotron one, at 8.4 GHz it takes over at tens of $\mu$Jy levels. 1 [ {width="76.40000%"} ]{} ![*left panel:* integral number counts of proto-spheroids at 1.4 GHz. Solid black line: unlensed proto-spheroids; solid red line: strongly lensed protospheroids as a function of the total flux density (sum of all images of a single source); dot-dashed red line: counts of strongly lensed proto-spheroids as a function of the flux density of the second brightest image. The vertical dashed lines mark the flux density limits of the possible surveys for SKA1-MID. *Right panel:* minimum SFR detectable by SKA1-MID (dashed magenta line), 50$\%$ SKA1-MID (dashed light blue line) and SKA2 (MID) (dashed blue line) Ultra-Deep surveys, as a function of $z$, compared with the minimum SFR detected by *Herschel* surveys (dashed-dot-dot-dot lines, green for 250 $\mu$m, orange for 350 $\mu$m and red for 500 $\mu$m) and by UV/H$\alpha$ surveys (horizontal yellow line; see text).[]{data-label="fig:lensing_negrello"}](NgtF_1d4GHz_17Oct2014.eps "fig:"){width="44.00000%"} ![*left panel:* integral number counts of proto-spheroids at 1.4 GHz. Solid black line: unlensed proto-spheroids; solid red line: strongly lensed protospheroids as a function of the total flux density (sum of all images of a single source); dot-dashed red line: counts of strongly lensed proto-spheroids as a function of the flux density of the second brightest image. The vertical dashed lines mark the flux density limits of the possible surveys for SKA1-MID. *Right panel:* minimum SFR detectable by SKA1-MID (dashed magenta line), 50$\%$ SKA1-MID (dashed light blue line) and SKA2 (MID) (dashed blue line) Ultra-Deep surveys, as a function of $z$, compared with the minimum SFR detected by *Herschel* surveys (dashed-dot-dot-dot lines, green for 250 $\mu$m, orange for 350 $\mu$m and red for 500 $\mu$m) and by UV/H$\alpha$ surveys (horizontal yellow line; see text).[]{data-label="fig:lensing_negrello"}](compare_herschel_ska_aaska.eps "fig:"){width="53.00000%"} Predictions for surveys with the Square Kilometer Array {#sect:SKAcounts} ======================================================= Preliminary plans for the phase1 SKA-MID include a set of surveys at $\sim$ 1-1.4 GHz aimed at investigating the galaxy evolution: an Ultra-Deep survey over 1 deg$^2$ with rms $\sim$ 50 nJy/beam, a Deep survey over 10-30 deg$^2$ with rms $\sim$ 0.2 $\mu$Jy/beam, and a Wide survey over 1000-5000 deg$^2$ with rms $\sim$ 1 $\mu$Jy/beam (see @Dewdney2013, @Braun2013). The corresponding $5\,\sigma$ limits are indicated by vertical solid lines in the left panel of Fig. \[fig:1.4\_95GHz\_counts\]; the other vertical lines indicate the detection limits achievable with 50$\%$ of SKA1-MID Ultra-Deep and Wide surveys and (full) SKA2 (MID) Ultra-Deep and Wide surveys sensitivity. We also show the contributions to the 1.4 GHz Euclidean normalized differential number counts of the three populations of dusty galaxies considered by @Cai2013. The main contributors to the “bump” at tens of $\mu$Jy levels are late-type galaxies at $z\simeq 1$–1.5. Higher $z$ proto-spheroidal galaxies become increasingly important at lower flux densities, down to a few hundred nJy’s. The predicted redshift distributions for surveys at the SKA1-MID flux density limits are shown in Fig. \[fig:RedshiftDistr\]. The fraction of galaxies at very high redshifts ($z\ge 6$) increases rapidly with decreasing flux density. At a few hundred nJy levels we expect detections of galaxies at $z$ of up to 10, making possible to investigate the cosmic SFR across the re-ionization epoch. Note that, although the deepest surveys with the HST are getting close to that, they inevitably miss the dust-obscured star formation, while dust obscuration does not affect SKA measurements. The fraction of strongly lensed galaxies in flux-limited surveys increases with redshift, to the point that these sources are dominant at the highest redshifts (see Fig. \[fig:RedshiftDistr\]). The predicted counts for magnifications $\mu \ge 2$ are illustrated in the left panel of Fig. \[fig:lensing\_negrello\], where the solid black line represents the unlensed proto-spheroidal galaxies, while the red lines represent the lensed galaxies as a function of the total flux density (solid) or of the flux density of the second brightest image (dot-dashed). The model predicts 655, 204 and 40 strongly lensed galaxies per deg$^2$ brighter than $0.25$, 1 and $5\,\mu$Jy, respectively; for 250, 100 and 20 of them, respectively, the SKA1-MID will directly detect at least two images. A more detailed discussion on the gravitational lens statistics with SKA is presented in McKean, J., et al. 2015, “Strong Gravitational Lensing with the SKA”, in proceedings of “Advancing Astrophysics with the Square Kilometre Array”, . The right panel of Fig. \[fig:lensing\_negrello\] compares the SKA potential in measuring the evolution of the cosmic SFR to the outcome of *Herschel* and of the deepest UV and H$\alpha$ surveys. The minimum luminosities, hence the minimum SFRs, reached by the latter surveys vary little with redshift. The yellow horizontal line corresponds to their average. Planned SKA surveys can detect galaxies with SFRs from tens to hundred $M_\odot/$yr, up to the highest redshifts, extending the SFR functions by up to 3 orders of magnitude compared with *Herschel*, thus encompassing SFRs typical of $L_\star$ galaxies. Conclusions {#sect:conclusions} =========== We have worked out detailed predictions of the counts and redshift distributions for planned SKA surveys, distinguishing the contributions of the different populations of star-forming galaxies: normal late-type, starburst and proto-spheroidal galaxies. The predictions are based on models by @Cai2013 and @Cai2014, that fit a broad variety of UV and far-IR/sub-mm data relevant to determine the epoch-dependent SFR function. These models, however, do not include the contribution to the SFR functions of moderate to low redshift late-type and starburst galaxies. We have upgraded them adding these populations. The upgraded models were combined with the relationships between SFR and radio (synchrotron and free-free) emission derived by @Murphy2011 [@Murphy2012]. Such relationships has been checked exploiting the deepest 1.4GHz counts and with data from the 95GHz SPT survey, that we have re-analyzed finding that the published SPT counts of dusty galaxies are overestimated by a factor $\simeq 3$. We have shown that the SKA will allow us to get information, not affected by dust extinction, on galaxy SFRs down to tens of $M_\odot/\hbox{yr}$ up to the highest redshifts, thus extending by up to 3 orders of magnitude the high-$z$ SFR functions derived from *Herschel* surveys. Acknowledgements {#acknowledgements .unnumbered} ================ We gratefully acknowledge many constructive comments by an anonymous referee, that helped us improving this paper. Work supported in part by ASI/INAF Agreement 2014-024-R.0 for the [*Planck*]{} LFI activity of Phase E2 and by PRIN INAF 2012, project “Looking into the dust-obscured phase of galaxy formation through cosmic zoom lenses in the Herschel Astrophysical Large Area Survey”. Banday, A. J., Wolfendale, A. W.  1991, MNRAS, 248, 705 Bell, E. F. 2003, ApJ, 586, 794 Bernardi, M., Shankar, F., Hyde, J. B., et al. 2010, [MNRAS]{}, 404, 2087 Bonato, M., Negrello, M., Cai, Z.-Y., et al. 2014, [MNRAS]{}, 438, 2547 Bourne, N., Dunne, L., Ivison, R. J., et al.,  2011, MNRAS, 410, 1155 Bouwens, R. J., Illingworth, G. D., Franx, M., Ford, H. 2008, [ApJ]{}, 686, 230 Bouwens, R. J., Illingworth, G. D., Oesch, P. A., et al. 2011, [ApJ]{}, 737, 90 Braun, R.  2013, “SKA1 Imaging Science Performance”, Document number SKA-TEL-SK0-DD-XXX Revision A Draft 2 Cai, Z.-Y., Lapi, A., Bressan, A., et al. 2014, [ApJ]{}, 785, 65 Cai, Z.-Y., Lapi, A., Xia, J.-Q., et al. 2013, [ApJ]{}, 768, 21 Carilli, C. L., Lee, N., Capack, P., et al. 2008, [ApJ]{}, 689, 883 Casey, C. M., Narayanan, D., Cooray, A. 2014, [PhR]{}, 541, 45 Chomiuk, L., Povich, M. S. 2011, [AJ]{}, 142, 197 Clemens, M. S., Negrello, M., De Zotti, G., et al. 2013, [MNRAS]{}, 433, 695 Condon, J. J. 1992, ARA&A, 30, 575 Dewdney, P., Turner, W., Millenaar, R., et al.  2013, “SKA1 System Baseline Design”, Document number SKA-TEL-SK0-DD-001 Revision1 De Zotti, G., Ricci, R., Mesa, D., et al. 2005, A&A, 431, 893 Draine, B. T. 2011, Physics of the Interstellar and Intergalactic Medium by Bruce T. Draine. Princeton University Press, 2011. ISBN: 978-0-691-12214-4, Draine, B. T. & Lazarian, A. 1998, ApJ, 494, L19 Granato, G. L., De Zotti, G., Silva, L., Bressan, A., Danese, L. 2004, [ApJ]{}, 600, 580 Helou, G., Soifer, B. T., Rowan-Robinson, M. 1985, ApJ, 298, L7 Helou, G., Bicay, M. D. 1993, ApJ, 415, 93 Hippelein, H., Haas, M., Tuffs, R. J., et al. 2003, [A$\&$A]{}, 407, 137 Ivison, R. J., Swinbank, A. M.;,Swinyard, B., et al. 2010, A&A, 518, L31 Jarvis, M. J., Smith, D. J. B., Bonfield, D. G., et al. 2010, MNRAS, 409, 92 Kennicutt, R. C., & Evans, N. J. 2012, [ARA$\&$A]{}, 50, 531 Lacki, B. C, Thompson, T. A., 2010, [ApJ]{}, 717, 196 Lapi, A., Gonz[á]{}lez-Nuevo, J., Fan, L., et al. 2011, [ApJ]{}, 742, 24 Lapi, A., Shankar, F., Mao, J., et al. 2006, [ApJ]{}, 650, 42 Mao, M. Y., Huynh, M. T., Norris, R. P., et al. 2011, ApJ, 731, 79 Mao, J., Lapi, A., Granato, G. L., De Zotti, G., Danese, L. 2007, [ApJ]{}, 667, 655 Massardi, M., Bonaldi, A., Negrello, M., et al. 2010, MNRAS, 404, 532 McLure, R. J., Dunlop, J. S., Bowler, R. A. A., et al. 2013, [MNRAS]{}, 432, 2696 Mocanu, L. M., Crawford, T. M., Vieira, J. D., et al. 2013, [ApJ]{}, 779, 61 Murphy, E. J., Bremseth, J., Mason, B. S., et al. 2012, [ApJ]{}, 761, 97 Murphy, E. J., Condon, J. J., Schinnerer, E., et al. 2011, [ApJ]{}, 737, 67 Murphy, E. J., 2009, [ApJ]{}, 706, 482 Niklas, S., Beck, R. 1997, A&A, 320, 54 Norris, R. P., Afonso, J., Bacon, D., et al. 2013, [PASA]{}, 30, 20 Oesch, P. A., Bouwens, R. J., Illingworth, G. D., et al. 2012, [ApJ]{}, 759, 135 Oesch, P. A., Bouwens, R. J., Illingworth, G. D., et al. 2013a, [ApJ]{}, 773, 75 Oesch, P. A., Bouwens, R. J., Illingworth, G. D., et al. 2013b, arXiv:1309.2280 Padovani, P., Miller, N., Kellermann, K. I., et al. 2011, ApJ, 740, 20 Planck Collaboration XX 2011, A&A, 536, A20 Planck Collaboration XVI 2013, arXiv:1303.5076 Schenker, M. A., Robertson, B. E., Ellis, R. S., et al. 2013, [ApJ]{}, 768, 196 Smit, R., Bouwens, R. J., Franx, M., et al. 2012, [ApJ]{}, 756, 14 Wei[ß]{}, A., De Breuck, C., Marrone, D. P., et al. 2013, [ApJ]{}, 767, 88 Yun, M. S., Reddy, N. A., Condon, J. J. 2001, ApJ, 554, 803 [^1]: See figures in <http://people.sissa.it/~zcai/galaxy_agn/index.html>.

ArXiv

--- abstract: | The Fisher-matrix formalism is used routinely in the literature on gravitational-wave detection to characterize the parameter-estimation performance of gravitational-wave measurements, given parametrized models of the waveforms, and assuming detector noise of known colored Gaussian distribution. Unfortunately, the Fisher matrix can be a poor predictor of the amount of information obtained from typical observations, especially for waveforms with several parameters and relatively low expected signal-to-noise ratios (SNR), or for waveforms depending weakly on one or more parameters, when their priors are not taken into proper consideration. In this paper I discuss these pitfalls; show how they occur, even for relatively strong signals, with a commonly used template family for binary-inspiral waveforms; and describe practical recipes to recognize them and cope with them. Specifically, I answer the following questions: (i) What is the significance of (quasi-)singular Fisher matrices, and how must we deal with them? (ii) When is it necessary to take into account prior probability distributions for the source parameters? (iii) When is the signal-to-noise ratio high enough to believe the Fisher-matrix result? In addition, I provide general expressions for the higher-order, beyond–Fisher-matrix terms in the 1/SNR expansions for the expected parameter accuracies. author: - Michele Vallisneri date: Jun 20 2007 title: 'Use and Abuse of the Fisher Information Matrix in the Assessment of Gravitational-Wave Parameter-Estimation Prospects' --- Introduction ============ Over the last two decades, the prevailing attitude in the gravitational-wave (GW) source-modeling community has been one of *pre-data positioning*: in the absence of confirmed detections, the emphasis has been on exploring which astrophysical systems, and which of their properties, would become accessible to GW observations with the sensitivities afforded by planned (or desired) future experiments, with the purpose of committing theoretical effort to the most promising sources, and of directing public advocacy to the most promising detectors. In this positioning and in this exploration, the *expected accuracy* of GW source parameters, as determined from the signals yet to be observed, is often employed as a proxy for the amount of physical information that could be gained from detection campaigns. However, predicting the parameter-estimation performance of future observations is a complex matter, even with the benefit of accurate theoretical descriptions of the expected waveforms and of faithful characterizations of the noise and response of detectors; in practice, the typical source modeler has had much less to go with. The main problem is that there are few analytical tools that can be applied generally to the problem, before resorting to relatively cumbersome numerical simulations that involve multiple explicit realizations of signal-plus-noise datasets. In the source-modeling community, the analytical tool of choice has been the Fisher information matrix $F_{ij}[h] = (h_i,h_j)$: here $h_i(t)$ is the partial derivative of the gravitational waveform $h(t)$ of interest with respect to the $i$-th source parameter $\theta_i$, and “$(\cdot,\cdot)$” is a signal product weighted by the expected power spectral density of detector noise, as described in Sec. \[sec:gaussianlikelihood\]. Now, it is usually claimed that the inverse Fisher matrix $F^{-1}_{ij}[h_0]$ represents the covariance matrix of parameter errors in the parameter-estimation problem for the true signal $h_0(t)$. This statement can be interpreted in three slightly different ways (all correct), which we examine in detail in Sec. \[sec:threeways\], and preview here: 1. The inverse Fisher matrix $F^{-1}_{ij}[h_0]$ is a lower bound (generally known as the *Cramér–Rao bound*) for the error covariance of any *unbiased estimator* of the true source parameters. Thus, it is a *frequentist* error (see Sec. \[sec:freqnbayes\]): for any *experiment* characterized by the true signal $h_0(t)$ and a certain realization $n(t)$ of detector noise, the parameter estimator $\hat{\theta}$ is a vector function of the total detector output $s = n + h_0$, and $F^{-1}_{ij}[h_0]$ is a lower bound on the covariance (i.e., the fluctuations) of $\hat{\theta}$ in an imaginary infinite sequence of experiments with different realizations of noise. The Cramér–Rao bound is discussed in Sec. \[sec:cramerrao\]. 2. The inverse Fisher matrix $F^{-1}_{ij}[h_0]$ is the frequentist error covariance for the *maximum-likelihood* (ML) parameter estimator $\hat{\theta}^\mathrm{ML}$, assuming Gaussian noise, in the limit of strong signals (i.e., high signal-to-noise ratio SNR) or, equivalently, in the limit in which the waveforms can be considered as linear functions of source parameters. We shall refer to this limit as the linearized-signal approximation, or LSA. This well-known result is rederived in Sec. \[sec:frequentisthighsn\]. 3. The inverse Fisher matrix $F^{-1}_{ij}[h_0]$ represents the covariance (i.e., the multidimensional spread around the mode) of the *posterior probability distribution* $p(\theta_0|s)$ for the true source parameters $\theta_0$, as inferred (in *Bayesian* fashion) from a *single* experiment with true signal $h_0$, assuming Gaussian noise, in the high-SNR limit (or in the LSA), and in the case where any *prior probabilities* for the parameters are constant over the parameter range of interest. Properly speaking, the inverse Fisher matrix is a measure of uncertainty rather than error, since in any experiment the mode will be displaced from the true parameters by an unknown amount due to noise.[^1] See Sec. \[sec:bayeshighsn\] for a rederivation of this result. As pointed out by Jaynes [@jaynes2003], while the numerical identity of these three different error-like quantities has given rise to much confusion, it arises almost trivially from the fact that in a neighborhood of its maximum, the signal likelihood $p(s|\theta_0)$ is approximated by a normal probability distribution with covariance $F^{-1}_{ij}$. In this paper, I argue that the Cramér–Rao bound is seldom useful in the work of GW analysts (Sec. \[sec:cramerrao\]), and while the high-SNR/LSA frequentist and Bayesian results are legitimate, they raise the question of whether the signals of interest are strong (or linear) enough to warrant the limit, and of what happens if they are not. In addition, if we possess significant information about the prior distributions (or even the allowed ranges) of source parameters, it is really only in the Bayesian framework that we can fold this information reliably into the Fisher result (Sec. \[sec:frequentisthighsn\]). Thus, I recommend the Bayesian viewpoint as the most fruitful way of thinking about the Fisher-matrix result (although I will also derive parallel results from the frequentist viewpoint). Of course, the study of Bayesian inference for GW parameter-estimation problems need not stop at the leading-order (Fisher-matrix) expression for the posterior likelihood: Markov Chain Monte Carlo (MCMC) algorithms [@mcmc] can provide very reliable results, immune from any considerations about signal strength, but they require a significant investment of time to implement them, and of computational resources to run them, since they necessarily involve explicit realizations of the noise. More rigorous Bayesian bounds (such as the Weiss–Weinstein and Ziv–Zakai bounds examined by Nicholson and Vecchio [@nv1998]) can also be derived, but they require a careful appraisal of the nonlocal structure of the likelihood function. By contrast, the Fisher-matrix formalism is singularly economical, and it seems clear that it will always be the first recourse of the GW data analyst. To use it reliably, however, we must understand the limits of its applicability. The purpose of this paper is to explore these limits. I do so by providing practical solutions to three issues that were already raised in the seminal treatments of GW detection by Finn [@finn1992] and by Cutler and Flanagan [@cf1994], but that seem to have been almost ignored after that: 1. What is the significance of the singular or ill-conditioned Fisher matrices that often appear in estimation problems with several source parameters, and how do we deal with them? Can we still believe the Fisher result in those cases? (See Sec. \[sec:disappearing\].) 2. When is it necessary to take into account the prior probability distributions for the parameters, even if specified trivially by their allowed ranges? (See Sec. \[sec:priors\].) 3. When is the high-SNR/LSA approximation warranted? (As anticipated above, the high-SNR limit is *equivalent* to the LSA, as we shall show in Secs. \[sec:frequentisthighsn\] and \[sec:bayeshighsn\].) That is, how strong a signal will we need to measure if we are to believe the Fisher-matrix result for its uncertainty? (See Sec. \[sec:howhigh\].) Last, I discuss the extension of the LSA beyond the leading order, in both the frequentist and Bayesian parameter-estimation frameworks (Sec. \[sec:higherorder\]), in a form that the adventurous GW analyst can use to test the reliability of the Fisher result (but higher-order derivatives and many-indexed expressions start to mount rapidly, even at the next-to-leading order). By contrast, I do not address the reduction in parameter-estimation accuracy due to the presence of secondary maxima in the likelihood function, as noticed [@bsd1996] and carefully modeled [@bd1998] by Balasubramanian and colleagues in their extensive Monte Carlo simulations of ML estimation for inspiraling binaries using Newtonian and first post-Newtonian waveforms. My treatment follows Refs. [@finn1992; @cf1994], as well as the classic texts on the statistical analysis of noisy data (e.g., Refs. [@wainstein62; @oppenheim83; @kay1993]). I am indebted to Jaynes and Bretthorst [@jaynes2003; @brett1988] for their enlightening, if occasionally blunt, perspective on frequentist and Bayesian parameter estimation. The reader already familiar with the standing of the Fisher-matrix formalism in the frequentist and Bayesian frameworks can skip Secs. \[sec:freqnbayes\] (a refresher on the difference between the frequentist and Bayesian viewpoints) and \[sec:cramerrao\]–\[sec:bayeshighsn\] (a pedagogical derivation of the three approaches to the inverse–Fisher-matrix result that were introduced at the beginning of this section), and move directly to discussion of the three *issues* in Secs. \[sec:disappearing\]–\[sec:howhigh\], and to the higher-order formalism in Sec. \[sec:higherorder\], referring back to Sec. \[sec:threeways\] as needed to establish notation. Whenever my discussion requires a practical example, I consider signals from inspiraling binaries of two black holes, both of mass $10 M_\odot$, as described by the restricted post-Newtonian approximation for adiabatic, circular inspirals (see Sec. \[sec:standardsignal\]); in my examples, I assume detection and parameter estimation are performed on Initial-LIGO [@ligo] data, and I adopt the LIGO noise curve of Table IV in Ref. [@DIS3]. Throughout, I use geometric units; I assume the Einstein summation convention for repeated indices; and I do not distinguish between covariant and contravariant indices, except in Sec. \[sec:higherorder\]. Three roads to the Fisher matrix {#sec:threeways} ================================ In this section I discuss the “three roads” to the inverse Fisher matrix as a measure of uncertainty for GW observations: the Cramér–Rao bound (Sec. \[sec:cramerrao\]), the high-SNR/LSA limit for the frequentist covariance of the ML estimator (Sec. \[sec:frequentisthighsn\]), and the high-SNR/LSA limit for the single-experiment covariance of the Bayesian posterior distribution (Sec. \[sec:bayeshighsn\]). Sections \[sec:freqnbayes\] and \[sec:gaussianlikelihood\] are refreshers about frequentist and Bayesian parameter estimation, and about the analytical expression for the likelihood of GW signals in Gaussian noise. A refresher on the frequentist and Bayesian frameworks {#sec:freqnbayes} ------------------------------------------------------ The *frequentist* (or orthodox) approach to parameter estimation for GW signals can be summed up as follows: 1. We are given the detector data $s$ and we take it to consist of the true signal $h_0 = h(\theta_0)$ (where $\theta_0$ is the vector of the *true system parameters*) plus additive noise $n$. 2. We select a *point estimator* $\hat{\theta}(s)$: that is, a vector function of detector data that (it is hoped) approximates the true values of source parameters, except for the statistical error due to the presence of noise. One important example of point estimator is the ML estimator $\hat{\theta}^\mathrm{ML}$, which maximizes the *likelihood* $p(s|\theta)$ of observing the measured data $s$ given a value $\theta$ of the true parameters. For additive noise, this likelihood coincides with the probability of a noise realization $n = s - h(\theta)$, and for Gaussian noise it is given below in Sec. \[sec:gaussianlikelihood\]. 3. We characterize statistical error as the fluctuations of $\hat{\theta}(s)$, computed over a very long series of independent *experiments* where the source parameters are kept fixed, while detector noise $n$ is sampled from its assumed probability distribution (often called the *sampling* distribution). The estimator $\hat{\theta}$ is usually chosen according to one or more criteria of optimality: for instance, *unbiasedness* requires that $\langle \hat{\theta}(s) \rangle_n$ (the average of the estimator over the noise probability distribution) be equal to $\theta_0$. A rather different approach is that of *Bayesian* inference: 1. We do not assume a true value of the system parameters, but we posit their *prior probability distribution* $p(\theta)$. 2. Given the data $s$, we do not compute estimators, but rather the full *posterior probability distribution* $p(\theta|s)$, using *Bayes’ theorem* $p(\theta|s) = p(s|\theta) \times p(\theta) / p(s)$, where $p(s) = \int p(s|\theta) \, p(\theta) \, d\theta$. 3. We characterize statistical error *in a single experiment* by the spread of the posterior distribution $p(\theta|s)$. The differences between the frequentist and Bayesian approaches are not only mathematical, but also epistemic: as their name indicates, “frequentists” view probabilities essentially as the relative frequencies of outcomes in repeated experiments, while “Bayesians” view them as subjective[^2] indices of certainty for alternative propositions. For an introduction to the contrasting views, I refer the reader to the excellent treatise (very partial to the Bayesian worldview) by Jaynes [@jaynes2003], and to Ref. [@cf1994] for a more GW-detection–oriented discussion. Once actual detections are made, the Bayesian approach of computing posterior probability distributions for the signal parameters *given the observed data* seems more powerful than the frequentist usage of somewhat arbitrary point estimators; the latter will always result in throwing away useful information, unless the chosen estimators are *sufficient statistics* (i.e., unless the likelihood depends on the data only through the estimators). As for statistical error, it seems preferable to characterize it from the data *we have* (actually, from the posterior distributions that we infer from that data), rather than from the data *we could have obtained* (i.e., from the sampling distribution of estimators in a hypothetical *ensemble* of experiments). As Cutler and Flanagan [@cf1994] point out, however, it is in the current *pre-data* regime that we seek to compute expected parameter accuracies; in the absence of actual confirmed-detection datasets, it seems acceptable to consider *ensembles* of possible parameter-estimation experiments, and to use frequentist statistical error as an inverse measure of potential physical insight. The best solution, bridging the two approaches, would undoubtedly be to examine the frequentist distribution of some definite measure of Bayesian statistical error; unfortunately, such a hybrid study is generally unfeasible, given the considerable computational requirements of even single-dataset Bayesian analyses. Likelihood for GW signals in Gaussian noise {#sec:gaussianlikelihood} ------------------------------------------- Under the assumption of stationary and Gaussian detector noise, the likelihood $\log p(s|\theta)$ can be obtained very simply from a noise-weighted inner product of the detector output and of the signal $h(\theta)$ (see for instance Eq. (2.3) in Ref. [@cf1994]): $$p(s|\theta) \propto e^{-(s-h(\theta),s-h(\theta))/2}; \label{eq:pstheta}$$ the weighting is performed with respect to the expected power spectral density of detector noise by defining the noise-weighted inner product of two real-valued signals as $$\label{eq:noiseproduct} (h,g) = 4 \, \mathrm{Re} \int_0^{+\infty} \frac{\tilde{h}(f)^* \tilde{g}(f)}{S_n(f)} \, df,$$ where $\tilde{h}(f)$ and $\tilde{g}(f)$ are the Fourier transforms of $h(t)$ and $g(t)$, “$*$” denotes complex conjugation, and $S_n(f)$ is the one-sided power spectral density of the noise. From the definition of $S_n(f)$ as $\langle \tilde{n}^*(f) \tilde{n}(f') \rangle_n = \frac{1}{2} S_n(|f|) \, \delta(f - f')$, we get the useful property $$\label{eq:noiseprod} \bigl\langle (h,n) (n,g) \bigr\rangle_n = (h,g),$$ where again “$\langle \cdot \rangle_n$” denotes averaging over the probability distribution of the noise. First road: Derivation and critique of the Cramér–Rao bound {#sec:cramerrao} ----------------------------------------------------------- The derivation in this section is inspired by the treatment of Ref. [@jaynes2003 p. 518], and it is given for simplicity in the case of one source parameter. We wish to pose a bound on the frequentist estimator variance $$\mathrm{var} \, \hat{\theta} = \Bigl\langle \Bigl( \hat{\theta}(s) - \bigl\langle \hat{\theta}(s) \bigr\rangle \Bigr)^2 \Bigr\rangle_n:$$ to do this, we consider the ensemble product $$\bigl\langle u(s), v(s) \bigr\rangle_n = \int u(s) \, v(s) \, p(s|\theta_0) \, ds,$$ where $p(s|\theta_0)$ is the likelihood of observing the detector output $s$ given the true source parameter $\theta_0$, or equivalently the likelihood of observing the noise realization $n = s - h_0$. Setting $v(s) = \hat{\theta}(s) - \bigl\langle \hat{\theta}(s) \bigr\rangle_n$, we obtain a bound on $\langle v,v \rangle_n \equiv \mathrm{var}\, \hat{\theta}$ from the Schwarz inequality: $$\mathrm{var} \, \hat{\theta} \equiv \langle v, v \rangle_n \geq \frac{\langle u, v \rangle_n^2}{\langle u, u \rangle_n}. \label{eq:schwarz}$$ This inequality is true for any function $u(s)$ of the data, and it becomes an equality when $u(s) \propto v(s)$. Since we wish to derive a bound that applies generally to all estimators, we should not have (or try) to provide too much detail about $\hat{\theta}$ (and therefore $v(s)$). A simple assumption to make on $\hat{\theta}$ is that it is an unbiased estimator: $$\bigl\langle \hat{\theta}(s) \bigr\rangle_n = \theta_0 \quad \Rightarrow \quad \partial_{\theta_0} \bigl\langle \hat{\theta}(s) \bigr\rangle_n = 1.$$ How does this help us? It turns out that we can write a function $d(s)$ whose ensemble product with any other function $w(s)$ yields the derivative $\partial_{\theta_0} \langle w(s) \rangle_n$; this function is just $d(s) = \partial_{\theta_0} \log p(s|\theta_0)$, because $$\int w(s) [\partial_{\theta_0} \log p(s|\theta_0)] \, p(s|\theta_0) \, ds = \int w(s) \, \partial_{\theta_0} p(s|\theta_0) \, ds = \partial_{\theta_0} \int w(s) p(s|\theta_0) \, ds = \partial_{\theta_0} \langle w(s) \rangle_n, \label{eq:dercramer}$$ assuming of course[^3] that we can exchange integration and differentiation with respect to $\theta_0$. For any $s$, $d(s)$ encodes the *local relative change* in the likelihood function as $\theta_0$ is changed. It follows that $\langle d(s), v(s) \rangle_n = \partial_{\theta_0} \bigl\langle \hat{\theta}(s) \bigr\rangle_n = 1$, so from Eq.  we get[^4] $$\mathrm{var} \, \hat{\theta} \geq \frac{1}{\bigl\langle d(s), d(s) \bigr\rangle} \equiv \frac{1}{\bigl\langle \partial_{\theta_0} \log p(s|\theta_0) , \partial_{\theta_0} \log p(s|\theta_0) \bigr\rangle_{n}}, \label{eq:cramersimple}$$ which is the unbiased-estimator version of the Cramér–Rao bound. If the estimator is biased, we can still use the Schwarz inequality by providing the derivative of the *bias* $b(\theta_0)$ with respect to $\theta_0$: $$\langle \hat{\theta}(s) \rangle_n = \theta_0 + b(\theta_0) \quad \Rightarrow \quad \partial_{\theta_0} \langle \hat{\theta}(s) \rangle_n = 1 + \partial_{\theta_0} b(\theta_0),$$ and therefore $$\label{eq:cramersingle} \mathrm{var} \, \hat{\theta} \geq \frac{(1 + \partial_{\theta_0} b)^2}{\bigl\langle \partial_{\theta_0} \log p(s|\theta_0), \partial_{\theta_0} \log p(s|\theta_0) \bigr\rangle_{n}}.$$ Generalizing to a multidimensional expression is straightforward, if verbose (see, e.g., Ref. [@kay1993]): $$\mathrm{covar}_n(\hat{\theta}_i,\hat{\theta}_l) \geq \bigl(\delta_{im} + \partial_m b_i(\theta_0)\bigr) F^{-1}_{mj} \bigl(\delta_{jl} + \partial_j b_l(\theta_0)\bigr), \label{eq:cramerfull}$$ where the *Fisher information matrix* is defined by $$F_{il} = \Bigl\langle \bigl(\partial_i \log p(s|\theta_0)\bigr), \bigl(\partial_l \log p(s|\theta_0)\bigr) \Bigr\rangle_{n} = - \Bigl\langle \partial_i \partial_l \log p(s|\theta_0) \Bigr\rangle_{n}\,. \label{eq:fisherdef}$$ The second equality is established by taking the gradient of $\int (\partial_i \log p(s|\theta_0)) \, p(s|\theta_0) \, ds$, and remembering that $\partial_i \int p(s|\theta_0) \, ds = \partial_i 1 = 0$. With the help of Eqs.  and , we can compute the Fisher matrix for GW signals in additive Gaussian noise, which is the familiar expression $F_{ij} = (\partial_i h,\partial_j h)$. The full expression for the Cramér–Rao bound, which includes the effects of bias, has interesting consequences, for it implies that biased estimators can actually *outperform*[^5] unbiased estimators, since the $\partial_m b_i(\theta_0)$ can be negative. Unfortunately, we have no handle on these derivatives without explicitly choosing a particular estimator (which goes against the idea of having a generic *bound*), so the Cramér–Rao bound can only give us a definite result for the subclass of unbiased estimators. As pointed out by Cutler and Flanagan [@cf1994 App. A 5], it follows that the bound cannot be used to place absolute limits on the accuracy of estimators (i.e., lower bounds on frequentist error)—limits that would exclude or severely limit the possibility of inferring the physical properties of sources from their emitted GWs. Even if the lower bound for unbiased estimators is very discouraging, there is always a chance that a biased estimator could do much better, so we cannot use the bound to prove “no go” theorems. Going back to Eq. , we note that the bound is satisfied as an equality when $$\label{eq:exponentialequation} u(s) \propto v(s) \quad \Rightarrow \quad d(s) \equiv \partial_{\theta_0} \log p(s|\theta_0) = q(\theta_0) [\hat{\theta}(s) - \langle \hat{\theta}(s) \rangle_n].$$ By integrating, we obtain a relation between the likelihood and the estimator: $$\label{eq:exponentialfamily} p(s|\theta_0) = \frac{m(s)}{Z(\theta_0)} e^{-l(\theta_0) \hat{\theta}(s)};$$ the estimation problems (i.e., the pairings of given likelihoods and chosen estimators) for which this relation holds true are said to belong to the *exponential family*, and these problems are the only ones for which the Cramér–Rao bound is satisfied exactly as an equality. Equation generalizes trivially to multidimensional problems by replacing the exponential with $\exp \, \{-l_k(\theta_0) \hat{\theta}_k(s)\}$. Unfortunately, for a given $p(s|\theta_0)$ there is no guarantee that any unbiased estimator exists that satisfies Eq.  and that therefore can actually *achieve* the bound; all we can say in general about the performance of *unbiased* estimators is that they will underperform the Cramér–Rao bias, but we do not know how badly. As discussed above, the bound tells us nothing in general about *biased* estimators. It follows that the bound cannot be used to establish guaranteed levels of accuracy (i.e., upper bounds on frequentist error), which would prove the possibility of inferring the physical properties of sources from their GWs. We can only do so if we can identify a specific estimator that achieves the bound. In the next section we shall see that the ML estimator[^6] does so in the high-SNR limit, where waveforms can be approximated accurately as *linear* functions of their parameters within the region of parameter space where $p(s|\theta)$ is not negligible (so the high-SNR limit coincides with the limit in which the LSA is accurate). We conclude that the Cramér–Rao bound is seldom useful to the GW analyst as a proper bound, whether to make positive or negative expected-accuracy statements; where it is useful, it reduces to the high-SNR/LSA result for the ML estimator. Second road: Derivation and critique of the frequentist high-SNR/LSA result {#sec:frequentisthighsn} --------------------------------------------------------------------------- We denote the true signal as $h_0$ (so $s = h_0 + n$), and expand the generic waveform $h(\theta)$ around $h_0$, normalizing signals by the *optimal signal-to-noise ratio* of the true signal, $A = \sqrt{(h_0,h_0)}$ (also known in this context as *signal strength*): $$h(\theta) = h_0 + \theta_k h_k + \theta_j \theta_k h_{jk} / 2 + \cdots = A (\bar{h}_0 + \theta_k \bar{h}_k + \theta_j \theta_k \bar{h}_{jk} / 2 + \cdots);$$ here we are translating source parameters as needed to have $h(0) = h_0$, defining $h_i = \partial_i h|_{\theta = 0}$, $h_{ij} = \partial_{ij} h|_{\theta = 0}$ (and so on), and $\bar{h}_0 = h_0 / A$, $\bar{h}_k = h_k / A$ (and so on).[^7] The likelihood is then given by[^8] $$\begin{aligned} \label{eq:likexpanded} p(s|\theta) \propto e^{-(s - h(\theta),s - h(\theta))/2} = \exp \bigr\{ & -(n,n)/2 - A^2 \bigl[ \theta_j \theta_k (\bar{h}_j,\bar{h}_k) + \theta_j \theta_k \theta_l (\bar{h}_j,\bar{h}_{kl}) + \cdots \bigr] / 2 \\ & + A \bigl[\theta_j (n,\bar{h}_j) + \theta_j \theta_k (n,\bar{h}_{jk}) / 2 + \theta_j \theta_k \theta_l (n,\bar{h}_{jkl}) / 3! + \cdots \bigr] \bigr\}. \end{aligned}$$ The ML equations $\partial_j p(s|\theta^\mathrm{ML}) = 0$ are given by $$\label{eq:mlexpanded} 0 = \frac{1}{A} \left[ (n,\bar{h}_j) + \hat{\theta}_k (n,\bar{h}_{jk}) + \hat{\theta}_k \hat{\theta}_l (n,\bar{h}_{jkl}) / 2 + \cdots \right] - \left[ \hat{\theta}_k (\bar{h}_j,\bar{h}_k) + \hat{\theta}_k \hat{\theta}_l \left( (\bar{h}_j,\bar{h}_{kl})/2 + (\bar{h}_k,\bar{h}_{jl}) \right) + \cdots \right],$$ where we have divided everything by $A^2$, and we omit the “${}^\mathrm{ML}$” superscript for conciseness. A careful study of Eq.  shows that it can be solved in perturbative fashion by writing $\hat{\theta}_j^\mathrm{ML}$ as a series in $1/A$, $$\label{eq:mlestimatorexpand} \hat{\theta}_j^\mathrm{ML} = \hat{\theta}_j^{(1)} / A + \hat{\theta}_j^{(2)} / A^2 + \hat{\theta}_j^{(3)} / A^3 + \cdots,$$ and by collecting the terms of the same order in Eq. , $$\label{eq:expandsystem} \begin{aligned} O(1/A): \quad & (n,\bar{h}_j) - \hat{\theta}_k^{(1)} (\bar{h}_j,\bar{h}_k) = 0, \\ O(1/A^2): \quad & \hat{\theta}_k^{(1)} (n,\bar{h}_{jk}) - \hat{\theta}_k^{(1)} \hat{\theta}_l^{(1)} \left( (\bar{h}_j,\bar{h}_{kl})/2 + (\bar{h}_k,\bar{h}_{jl}) \right) - \hat{\theta}_k^{(2)} (\bar{h}_j,\bar{h}_k) = 0, \\ O(1/A^3): \quad & \ldots \end{aligned}$$ thus the ML solution $\hat{\theta}_j^\mathrm{ML}$ is given by $$\label{eq:solbyorders} \begin{aligned} \hat{\theta}_j^\mathrm{ML} = \, & \frac{1}{A} \, (\bar{h}_j,\bar{h}_k)^{-1} (\bar{h}_k,n) \, + \\ & \frac{1}{A^2} \, (\bar{h}_j,\bar{h}_i)^{-1} \bigl\{ (n,\bar{h}_{ik}) (\bar{h}_k,\bar{h}_l)^{-1} (\bar{h}_l,n) - \left( (\bar{h}_i,\bar{h}_{kl})/2 + (\bar{h}_k,\bar{h}_{il}) \right) (\bar{h}_k,\bar{h}_m)^{-1} (\bar{h}_m,n) (\bar{h}_l,\bar{h}_n)^{-1} (\bar{h}_n,n) \bigr\} \, + \\ & \frac{1}{A^3} \, \bigl\{ \cdots \bigr\} + \cdots \end{aligned}$$ Thus we see that the limit of large $A$ (i.e., high SNR) coincides with the *linearized-signal approximation* (LSA) where only the first derivatives of the signals are included. In the LSA, the likelihood is just $$\begin{aligned} \label{eq:standardlikelihood} p(s|\theta) &\propto \exp \bigl\{ -(n,n)/2 - \theta_j \theta_k (h_j,h_k)/2 + \theta_j (h_j,n) \bigr\} \\ &= \exp \bigl\{ -(n,n)/2 - A^2 \theta_j \theta_k (\bar{h}_j,\bar{h}_k)/2 + A \theta_j \,(\bar{h}_j,n) \bigr\} \quad \text{(LSA)}, \end{aligned}$$ and the ML estimator is given by $$\label{eq:lsaest} \hat{\theta}_j^\mathrm{ML} = (1/A) (\bar{h}_j,\bar{h}_k)^{-1} (\bar{h}_k,n) \quad \text{(LSA)}.$$ Since $\langle(\bar{h}_k,n)\rangle_n = 0$, we see also that the ML estimator is unbiased. The variance of $\hat{\theta}^\mathrm{ML}$ is then obtained by averaging $\hat{\theta}_j^\mathrm{ML} \hat{\theta}_k^\mathrm{ML}$ over noise realizations, $$\label{eq:lsavar} \begin{aligned} \bigl< \hat{\theta}_j^\mathrm{ML} \hat{\theta}_k^\mathrm{ML} \bigr>_n & = \frac{1}{A^2} \, (\bar{h}_j,\bar{h}_l)^{-1} \bigl< (\bar{h}_l,n) (n, \bar{h}_m,n) \bigr>_n (\bar{h}_m,\bar{h}_k)^{-1} = \\ &= \frac{1}{A^2} \, (\bar{h}_j,\bar{h}_l)^{-1} (\bar{h}_l,\bar{h}_m) (\bar{h}_m,\bar{h}_k)^{-1} = \frac{1}{A^2} \, (\bar{h}_j,\bar{h}_k)^{-1} \quad \text{(LSA)}, \end{aligned}$$ and it coincides with the mean quadratic error in the frequentist sense. In Eq. , the second equality follows from Eq. . The interpretation of the limit is that, for strong signals, the typical $\hat{\theta}^\mathrm{ML}_j - \theta_{0j}$ becomes small enough that the log-likelihood is accurately described by the product of detector data and a linearized signal. Equation is the standard Fisher-information–matrix result, and it implies that in the high-SNR/LSA limit the ML estimator achieves the Cramér–Rao bound. As we shall see in Sec. \[sec:higherorder\], the next-order correction to the variance scales as $1/A^4$, not $1/A^3$. This is because all $O(1/A^3)$ terms contain odd numbers of $n$, whose products vanish under the ensemble average. The fact itself that there is a next-order correction shows that for generic $A$ the ML estimator does not achieve the bound. The fact that the Cramér–Rao bound is achieved in the high-SNR/LSA limit, but not beyond it, can also be seen in the light of Eq. , which encodes a standard form for the estimation problems in the exponential family. To express the LSA likelihood in this form, we can set $m(s) = e^{-(s - h_0,s - h_0)/2}$ and $Z(\theta) = e^{\theta_j \theta_k (h_j,h_k)/2}$; it remains to establish that $$-l_j(\theta) \hat{\theta}^\mathrm{ML}_j(s) = \theta_j (h_j, s - h_0),$$ which is satisfied by $l_j(\theta) = -(h_j,h_k) \theta_k$ \[see Eq. \]. Now, if additional terms are added to Eq. , beginning with terms cubic in the $\theta_i$, $\hat{\theta}_j^\mathrm{ML}(s)$ comes to be a nonlinear function of the signal, such that no $-l_j(\theta)$ can multiply it in the right way to reconstruct the likelihood. It then follows that the estimation problem moves outside the exponential family, and the Cramér–Rao bound cannot be achieved. It is possible (but perhaps not desirable, as we shall see shortly) to modify the ML estimator to take into account prior knowledge about the expected distribution of sources. The resulting maximum-posterior estimator $\hat{\theta}^\mathrm{MP}$ is defined as the mode of the posterior probability $p(\theta|s) = p(s|\theta) p(\theta) / p(s)$, $$\hat{\theta}^\mathrm{MP} = \mathrm{maxloc}_\theta \, p(\theta|s) = \mathrm{maxloc}_\theta \, p(s|\theta) p(\theta).$$ This is a biased estimator: in the high-SNR/LSA limit, and with a Gaussian prior $p(\theta) \propto \exp \{-P_{ij} (\theta_i - \theta^P_i) (\theta_j - \theta^P_j) / 2\}$ centered at $\theta^P$ (the only prior that can be easily handled analytically), we find $$\label{eq:biaswithprior} b^\mathrm{MP}_i = \left\langle \theta_i^\mathrm{MP} \right\rangle_n = \left[(\bar{h}_i,\bar{h}_j) + P_{ij}/A^2\right]^{-1} (P_{jk}/A^2) \, \theta^P_k \quad \text{(LSA/Gaussian prior)};$$ thus the $\hat{\theta}^\mathrm{MP}$ becomes unbiased for $A \rightarrow \infty$ (indeed, in that limit $\hat{\theta}^\mathrm{MP}$ tends to $\hat{\theta}^\mathrm{ML}$). For the frequentist variance we find $$\label{eq:varwithprior} \begin{aligned} \bigl\langle \hat{\theta}_i^\mathrm{MP} \hat{\theta}_j^\mathrm{MP} \bigr\rangle_n - \bigl\langle \hat{\theta}_i^\mathrm{MP} \bigr\rangle_n \bigl\langle \hat{\theta}_j^\mathrm{MP} \bigr\rangle_n & = \bigl\langle \hat{\theta}_i^\mathrm{MP} \hat{\theta}_j^\mathrm{MP} \bigr\rangle_n - b^\mathrm{MP}_i b^\mathrm{MP}_j \\ & = \frac{1}{A^2} \left[(\bar{h}_i,\bar{h}_k) + P_{ik} / A^2 \right]^{-1} (\bar{h}_k,\bar{h}_l) \left[(\bar{h}_l,\bar{h}_j) + P_{lj} / A^2 \right]^{-1} \quad \text{(LSA w/prior)}, \end{aligned}$$ which coincides[^9] with the generalized Cram ér–Rao bound of Eq. , proving that the estimation problem defined by the LSA likelihood and $\hat{\theta}^\mathrm{MP}$ belongs to the exponential family. The reason why $\hat{\theta}^\mathrm{MP}$ is not too useful to characterize future parameter-estimation performance is that we expect a reasonable measure of error to converge to the effective width of the prior in the limit of vanishing signal strength. Instead, in the absence of any information from the experiment, $\hat{\theta}^\mathrm{MP}$ becomes stuck at the mode of the prior, and its variance \[in Eq. \] tends to zero. This behavior occurs for any nonuniform prior.[^10] Third-road Derivation of the Bayesian high-SNR/LSA result {#sec:bayeshighsn} --------------------------------------------------------- We now wish to show that in any single experiment, if the high-SNR/LSA limit is warranted (and if the parameter priors are uniform over the parameter region of interest), the inverse Fisher-information matrix yields the variance of the Bayesian posterior probability distribution. To do so, we rewrite Eq.  in terms of *normalized parameters* $\bar{\theta}_i = A \, \theta_i$: $$\begin{gathered} \label{eq:anotherp} p(s|\theta) \propto \exp \Bigl\{ -(n,n)/2 + \Bigl[ (n,\bar{h}_j) \bar{\theta}_j + \frac{1}{A} (n,\bar{h}_{jk}) \bar{\theta}_{j} \bar{\theta}_{k} / 2 + \frac{1}{A^2} (n,\bar{h}_{jkl}) \bar{\theta}_{j} \bar{\theta}_{k} \bar{\theta}_{l} / 3! + O(1/A^3) \Bigr] \\ - \Bigl[ (\bar{h}_j,\bar{h}_k) \bar{\theta}_j \bar{\theta}_k + \frac{1}{A} (\bar{h}_j,\bar{h}_{kl}) \bar{\theta}_j \bar{\theta}_{k} \bar{\theta}_{l} + \frac{1}{A^2} (\bar{h}_{jk},\bar{h}_{lm}) \bar{\theta}_{j} \bar{\theta}_{k} \bar{\theta}_{l} \bar{\theta}_{m} / 4 + \frac{2}{A^2} (\bar{h}_{j},\bar{h}_{klm}) \bar{\theta}_{j} \bar{\theta}_{k} \bar{\theta}_{l} \bar{\theta}_{m} / 3! + O(1/A^3) \Bigr]/2 \Bigr\}.\end{gathered}$$ We can build the variance from the posterior mean $$\label{eq:bayesianmean} \bigl< \bar{\theta}_i \bigr>_p \equiv \int \bar{\theta}_i \, p(s|\theta) \, d\theta \bigg/\!\! \int p(s|\theta) \, d\theta$$ and the quadratic moment $$\bigl< \bar{\theta}_i \bar{\theta}_j \bigr>_p = \int \bar{\theta}_i \bar{\theta}_j \, p(s|\theta) \, d\theta \bigg/\!\! \int p(s|\theta) \, d\theta$$ where “$\langle \cdot \rangle_p$” denotes integration over $p(s|\theta)$. The idea is to proceed in perturbative fashion, writing the moments as series in $\epsilon = 1/A$: taking $\langle \bar{\theta}_{i} \rangle_{p}$ as an example, $$\big\langle \bar{\theta}_{i} \big\rangle_{p} = \sum_{n=0}^\infty \frac{\epsilon^n}{n!} \big\langle \bar{\theta}_{i} \big\rangle_{p}^{\!(n)} \quad \Rightarrow \quad \langle \bar{\theta}_i \rangle_p^{(n)} = \frac{\partial^n \langle \bar{\theta}_i \rangle_p}{\partial \epsilon^n} \bigg|_{\epsilon = 0}.$$ Since $\epsilon$ appears at both the numerator and denominator of Eq. , we write $$\big\langle \bar{\theta}_{i} \big\rangle_{p} = \frac{ \int \bar{\theta}_{i} \, p(0) \, d \theta + \epsilon \int \bar{\theta}_{i} \, \frac{\partial p(0)}{\partial \epsilon} \, d \theta + \frac{\epsilon^2}{2} \int \bar{\theta}_{i} \, \frac{\partial^2 p(0)}{\partial \epsilon^2} d \theta + \cdots }{ \int p(0) \, d \theta + \epsilon \int \frac{\partial p(0)}{\partial \epsilon} \, d \theta + \frac{\epsilon^2}{2} \int \frac{\partial^2 p(0)}{\partial \epsilon^2} d \theta + \cdots }$$ (where the argument of $p$ implies that the $(n)$-th derivative is evaluated at $\epsilon = 0$), and therefore $$\begin{aligned} \big\langle \bar{\theta}_{i} \big\rangle_{p}^{(0)} &= \int \bar{\theta}_{i} \, p(0) \, d \theta \bigg/\!\! \int p(0) \, d \theta, \\ \big\langle \bar{\theta}_{i} \big\rangle_{p}^{(1)} &= \biggl[ \int \bar{\theta}_{i} \, \frac{\partial p(0)}{\partial \epsilon} \, d \theta - \big\langle \bar{\theta}_{i} \big\rangle_{p}^{(0)} \int \frac{\partial p(0)}{\partial \epsilon} \, d \theta \biggr] \bigg/\!\! \int p(0) \, d \theta; \\ & \ldots \end{aligned} \label{eq:bayesianseries}$$ similar expressions hold for $\langle \bar{\theta}_i \bar{\theta}_j \rangle_p$, and a general expression for the $(n)$-th–order contribution is given in Sec. \[sec:higherbayesian\]. The $\epsilon \rightarrow 0$ limit coincides with the limit of large signal strengths, or of vanishing derivatives higher than the first, since in that case Eq.  truncates to Eq. . In this limit, $$\label{eq:bayesavg} \bigl< \bar{\theta}_i \bigr>_p = \bigl< \bar{\theta}_i \bigr>_p^{(0)} = (\bar{h}_i,\bar{h}_j)^{-1} (n,\bar{h}_j) \quad \text{(LSA)}$$ and $$\label{eq:bayesvar} \bigl\langle \Delta \bar{\theta}_i \Delta \bar{\theta}_j \bigr\rangle_p = \bigl\langle (\bar{\theta}_i - \bigl\langle \bar{\theta}_i \bigr\rangle_p^{(0)}) (\bar{\theta}_j - \bigl\langle \bar{\theta}_j \bigr\rangle_p^{(0)}) \bigr\rangle_p^{(0)} = (\bar{h}_i,\bar{h}_j)^{-1} \quad \text{(LSA)},$$ and therefore $$\bigl\langle \theta_i \theta_j \bigr\rangle_p = \frac{1}{A^2} (\bar{h}_i,\bar{h}_j)^{-1} = (h_i,h_j)^{-1} \quad \text{(LSA)},$$ as can be seen by rewriting the exponential of Eq.  as $$p(s|\bar{\theta}) \propto \exp \bigl\{ - (\bar{h}_i,\bar{h}_j) (\bar{\theta}_i - \langle \bar{\theta}_i \rangle_p) (\bar{\theta}_j - \langle \bar{\theta}_j \rangle_p) / 2 \bigr\},$$ where we have omitted factors independent from $\bar{\theta}$ that cancel out in the normalization of $p(s|\bar{\theta})$. Reinstating $A$ in Eq.  we see that in the high-SNR/LSA limit the mean of the posterior distribution coincides with the ML estimator, as is reasonable, since the average of a normal distribution coincides with its mode. The two however differ when higher-order terms are included, as we shall see in Sec. \[sec:higherorder\]. From Eq.  we see also that, to leading order, the variance of the posterior distribution is experiment-independent, and it coincides with the variance of the ML estimator (remember however that the two have very different interpretations[^11]). With the addition of a Gaussian prior $p(\theta) \propto e^{-P_{ij} \theta_i \theta_j / 2}$ centered at $\theta = 0$, Eqs.  and change only slightly:[^12] $$\label{eq:resummedprior} \begin{aligned} \bigl\langle \bar{\theta}_i \bigr\rangle_p &= [(\bar{h}_i,\bar{h}_j) + P_{ij}/A^2]^{-1} (n,\bar{h}_j) \\ \bigl\langle \Delta \bar{\theta}_i \Delta \bar{\theta}_j \bigr\rangle_p &= [(\bar{h}_i,\bar{h}_j) + P_{ij}/A^2]^{-1} \end{aligned} \quad \text{(LSA/Gaussian prior)}.$$ Note that $p(\theta) \propto e^{-(1/A^2) P_{ij} \bar{\theta}_i \bar{\theta}_j/2}$ is formally an $O(1/A^2)$ contribution to the likelihood exponential that would enter the $1/A$ expansion beginning at that order. However, if $P_{ij}$ is large enough to matter at the signal strengths of interest, it probably makes sense to bundle it with the zeroth-order terms as we did here. In contrast with Eq.  for the frequentist variance of $\hat{\theta}^\mathrm{MP}$, we see that in the limit of vanishing signal strength the variance of the posterior goes to the variance $P_{ij}$ of the prior. Standard compact-binary signal model {#sec:standardsignal} ==================================== Throughout the rest of this paper, our fiducial model for compact-binary signals will be simple stationary-phase–approximated (SPA) waveforms including phasing terms from the spin–orbit and spin–spin interactions of parallel or antiparallel component spins. Parameter estimation with these waveforms was studied by Poisson and Will [@pw1995]. In this paper we adopt second-order post-Newtonian[^13] (2PN) Fourier-domain waveforms as written by Arun and colleagues [@arun2005]: $$\label{eq:pnwaveform} \tilde{h}(M_c,\eta,\beta,\sigma,\phi_0,t_0;f) \propto f^{-7/6} \exp \, i \{ \psi(M_c,\eta,\beta,\sigma;f) + \phi_0 + 2 \pi f t_0 \},$$ with $$\label{eq:pnphasing} \begin{aligned} \psi(M_c,\eta,\beta,\sigma;f) = \frac{3}{128 \, \eta \, v^5} \biggl\{ 1 & + \frac{20}{9} \biggl(\frac{743}{336} + \frac{11}{4} \eta\biggr) v^2 + (4 \beta - 16 \pi) v^3 + 10 \biggl(\frac{3 058 673}{1 016 064} + \frac{5 429}{1 008}\eta + \frac{617}{144}\eta^2 - \sigma \biggr) v^4 \biggr\}, \end{aligned}$$ where $v = (\pi M f)^{1/3}$, $M = m_1 + m_2$ is the total mass of the binary, $\eta = m_1 m_2 / M^2$ is the symmetric mass ratio, $M_c = M \eta^{3/5}$ is the chirp mass. The spin–orbit parameter $\beta$ and the spin–spin parameter $\sigma$ [@pw1995; @spinpapers] are given by $$\beta = \sum_{i=1}^{2} \frac{\hat{\mathbf{L}} \cdot \mathbf{S}_i}{12 \, m_i^2} \biggl[ 113 \Bigl(\frac{m_i}{M}\Bigr)^2 + 75 \, \eta \biggr] = \frac{113 \, \hat{\mathbf{L}} \cdot [\mathbf{S}_1 + \mathbf{S}_2] + 75 \, \hat{\mathbf{L}} \cdot [(m_2/m_1) \mathbf{S}_1 + (m_1/m_2) \mathbf{S}_2]}{12 \, M^2},$$ and $$\sigma = \frac{721 (\hat{\mathbf{L}}\cdot\mathbf{S}_1)(\hat{\mathbf{L}}\cdot\mathbf{S}_2) - 247 \,\mathbf{S}_1 \cdot \mathbf{S}_2}{48 \, m_1 \, m_2 \, M^2},$$ with $\mathbf{S}_1$ and $\mathbf{S}_2$ the spins of the binary components. We truncate waveforms at the (Keplerian) last stable circular orbit ($v = 1/\sqrt{6}$). For simplicity, in this article we do not discuss the estimation of the amplitude parameter $\mathcal{A}$ that would multiply the right-hand side of Eq. . \[From Eqs.  and it follows that $(\partial_\mathcal{A} h,\partial_i h) = 0$ for $i \neq \mathcal{A}$, so the amplitude $\mathcal{A}$ effectively decouples from all other parameters in the Fisher matrix.\] However, all discussions to follow can accommodate the addition of $\mathcal{A}$ with trivial modifications. The singular case of the disappearing parameter {#sec:disappearing} =============================================== In Sec. \[sec:threeways\] we have examined the interpretation of the inverse Fisher matrix as a frequentist or Bayesian measure of error or uncertainty. In this section, we discuss what happens when the Fisher is matrix singular, or almost so, so that the attempts to invert it numerically yield warnings that it is badly conditioned. It is pedagogical to begin this discussion by considering the case where the matrix is exactly singular (Sec. \[sec:singularfisher\]), and then to widen our scope to approximate singularity (Sec. \[sec:approxfisher\]). The conclusion is that a singular Fisher matrix is almost always a symptom that the high-SNR/LSA limit is not to be trusted, that prior probabilities play an important role, or both. Singular Fisher matrix {#sec:singularfisher} ---------------------- A singular Fisher matrix implies that the corresponding LSA likelihood is a *singular normal* distribution [@gupta], which is constant along the directions of the Fisher-matrix eigenvectors with null eigenvalues[^14] (henceforth, somewhat improperly, we shall call these *null eigenvectors*), so the ML equation has no solutions, and the even moments of the distribution are infinite, even for parameters that do not appear in the null eigenvectors. Thus, the frequentist variance of the ML estimator and the Bayesian variance of the posterior distribution are (formally) infinite for all parameters. How to deal with this? If the signal is *really* linear, so that the LSA expressions are exact, it is possible to discard the combinations of parameters that correspond to the null eigenvectors, and characterize the variance of the remaining parameters. Let us see how, in the frequentist and Bayesian frameworks. In what follows, we denote the total number of source parameters by $N$, and the number of non-null eigenvectors by $R$. In the frequentist ML framework, we write $(\bar{h}_i,\bar{h}_j)$ in the singular-value (SV) decomposition[^15] [@golub] as $\sum_{\lambda^{(k)} \neq 0} \theta^{(k)}_i \lambda^{(k)} \theta^{(k)}_j$ (with $(k) = 1,\ldots,R$), or $\Theta \Sigma \Theta^T$ in matrix notation (with $\Theta$ an $N \times R$ matrix with orthonormal columns, and $\Sigma$ a diagonal matrix formed from the $R$ non-zero eigenvalues). We can then refactor the ML equation as $$\label{eq:svmoment} \bigl(\Theta \Sigma \Theta^T\bigr) \boldsymbol{\hat{\theta}} = A^{-1} \mathbf{\hat{n}} \quad \Rightarrow \quad \bigl(\Theta^T \boldsymbol{\hat{\theta}}\bigr) = A^{-1} \bigl(\Sigma^{-1} \Theta^T \mathbf{\hat{n}}\bigr) \quad \Rightarrow \quad \hat{c}^{(k)} = A^{-1} (\lambda^{(k)})^{-1} n^{(k)},$$ where $\hat{c}^{(k)}$ and $n^{(k)}$ denote the coefficients of the decompositions of $\hat{\theta}$ and $(\bar{h}_i,n)$ with respect to the normalized non-null eigenvectors of $(\bar{h}_i,\bar{h}_j)$. Since the ensemble average $\langle n^{(k)} n^{(l)} \rangle_n$ is just $\lambda^{(k)} \delta^{(k)(l)}$ (where $\delta$ is Kronecker’s delta), the frequentist covariance of the ML estimators $\hat{c}^{(k)}$ is the diagonal matrix $A^{-2} (\lambda^{(k)})^{-1} \delta^{(k)(l)}$. In the Bayesian framework, the quantities of interest are the moments of the $c^{(k)}$ over infinite ranges of the $c^{(k)}$ *and* of the coefficients $C^{(K)}$ (with $(K) = 1,\ldots,N-R$) corresponding to the null eigenvectors, which are not included in the SV decomposition. Formally, these moments are ratios of two infinities, because the LSA likelihood is not a function of the $C$ \[not even through the $n^{(K)} \equiv \theta_i^{(K)} (\bar{h}_i,n)$ terms, which are zero since $(\theta_i^{(K)} \bar{h}_i,\theta_j^{(K)} \bar{h}_j) = 0$\], but they may be evaluated as improper integrals, *in the limit* of the ranges for the $c^{(K)}$ extending to infinity: $$\label{eq:limitmoment} \langle \Delta c^{(k)} \Delta c^{(l)} \rangle_p = \frac{\int c^{(k)} c^{(l)} p(s|c) \, dc \, dC}{\int p(s|c) \, dc \, dC} = \lim_{\Delta C^{(K)} \rightarrow \infty} \frac{\left(\int_{-\Delta C^{(K)}}^{+\Delta C^{(K)}} dC \right) \int c^{(k)} c^{(l)} p(s|c) \, dc}{\left( \int_{-\Delta C^{(K)}}^{+\Delta C^{(K)}} dC \right) \int p(s|c) \, dc} = A^{-2} (\lambda^{(k)})^{-1} \delta^{(k)(l)}.$$ We can then work back to the frequentist components of the covariance matrix (or the Bayesian posterior moments) that involve any $\hat{\theta}_i$ that *do not appear* in the null eigenvectors. All other $\hat{\theta}_i$, however, are *completely* indeterminate.[^16] In the frequentist framework, it may be possible to work back to *interval* estimates of their values by combining a ML estimate of the $\hat{c}^{(k)}$ with finite allowed ranges for some of the $\theta_i$; however, this would constitute a form of prior distribution for the $\theta_i$, which is not entirely compatible with the ML estimator (what happens if the solution of the ML equation falls outside the allowed range?). In the Bayesian framework, salvation may come from the prior probability distributions that make the posterior integrable.[^17] Unless the priors are also normal, though, the resulting moments cannot be expressed simply as analytical expressions of the Fisher matrix. The most benign outcome occurs when the null eigenvectors correspond individually to one or more of the original parameters, or when the subspace spanned by null eigenvectors corresponds to a subset of the original parameters. The null-eigenvector combinations of parameters may also have clear physical interpretations: for instance, for a monochromatic, continuous sinusoid of frequency $f$, the absolute time offset $t_0$ and the initial phase $\phi_0$ are essentially the same parameter, so the Fisher matrix has a null eigenvector along the parameter combination $f t_0 - \phi_0$, which can be discarded, while $f t_0 + \phi_0$ remains well determined. A similar case is the degeneracy between luminosity distance and a certain function of the sky-position angles in the analysis of short GW chirps with a single ground-based detector.[^18] Other combinations of parameters can be more ambiguous and troubling—what is the meaning of estimating a parameter equal to a mass plus a spin? In those cases, our best hope is again that the degeneracy will be cured by prior probabilities, or by higher-order corrections in the $1/A$ expansion, in which cases the Fisher-matrix formalism is certainly insufficient. Ill-conditioned Fisher matrix {#sec:approxfisher} ----------------------------- All nonsingular matrices have well-defined inverses, although these *may* be difficult to compute. The notion of *ill conditioning* from the theory of linear systems of equations [@golub] can be invoked here to provide a bound (valid under reasonable conditions) on the perturbation of the inverse of a perturbed matrix, $$\label{eq:conditionm} \frac{||(M + \delta M)^{-1} - M^{-1}||}{||M^{-1}||} \leq \kappa(M) \frac{||\delta M||}{||M||} + O(||\delta M||^2);$$ here “$||\cdot||$” is a matrix norm (e.g., the 2-norm $||M||_2 = \sup_\mathbf{x} ||M \mathbf{x}||_2 / ||\mathbf{x}||_2$ derived from the vector 2-norm $||\mathbf{x}||_2 = (\sum_i x_i^2)^{1/2}$), and $\kappa(M) = ||M||\,||M^{-1}||$ is the *condition number*. Since $||M||_2$ is equal to $M$’s largest eigenvalue, $\kappa_2(M)$ is given by the ratio of its largest- to smallest-modulus eigenvalues. From a numerical-analysis perspective, as Finn [@finn1992] points out, the gist of Eq.  is that, roughly speaking, matrix inversion can amplify roundoff error by a factor $\kappa$, leading to the loss of up to $\log_{10} \kappa$ digits of precision. The same amplification will apply to any inaccuracies in our knowledge of $M$. Taken at face value, this means that the Fisher-matrix results of Eqs.  and may be inaccurate at a 100% level if the components of the Fisher matrix are not known to a fractional accuracy better than $\kappa^{-1}(F)$. Of course, Eq.  is only an upper bound, and this doomsday scenario needs not be realized in practice. One way to check whether the matrix-inversion sensitivity is a concern is to add small random perturbations, Monte Carlo-style, to the Fisher matrix elements, and then verify the change in the covariance matrix. Such an experiment for our standard SPA model (with $m_1 = m_2 = 10 M_\odot$ and no spins) shows that perturbing the *12th* significant digits of the $F_{ij}$ components is already enough to engender 100% changes in the diagonal elements of $(F^{-1})_{ij}$ (i.e., the predicted parameter variances). This behavior is $\sim 100$ times less severe than what is predicted by Eq. , but it still tells a rather cautionary tale about numerical sensitivity in the inversion of that particular Fisher matrix.[^19] These problems can be cured, somewhat trivially, by adopting higher-precision arithmetics, and by computing the Fisher matrix to better accuracy. It may also be possible to improve the condition number by changing the units of source parameters, which may reduce the magnitude gap between the largest- and smallest-modulus eigenvalues. More to the point, it is the consequences of the Fisher-matrix condition number on the substance (rather than the numerical accuracy) of Eqs.  and that should attract our attention. A large condition number implies one or more small Fisher-matrix eigenvalues, and consequently large statistical fluctuations for the combinations of source parameters corresponding to the small-eigenvalue eigenvectors, at least according to the LSA. The interpretation is that large parameter changes in the direction of the small-eigenvalue eigenvectors are needed to produce changes in the waveform comparable to typical noise fluctuations. Under this condition, we have to worry whether the LSA can really describe the likelihood over the entire parameter ranges of interest: of course, these depend on the SNR available at detection (at leading order, their extent is inversely proportional to signal strength). In Sec. \[sec:howhigh\] we describe a numerical criterion to decide when the SNR is high enough to believe the LSA. We also have to worry whether prior probability distributions for the parameters (perhaps in the simple form of allowed ranges) already restrict the estimated (for frequentists) or probable (for Bayesians) values of parameters beyond what is predicted by the Fisher-matrix variance. In the next section we discuss a simple test to decide whether priors should be included. The burden of prior commitments {#sec:priors} =============================== As Cutler and Flanagan point out [@cf1994 p. 2691], “it is not necessary for *a priori* information to be very detailed or restrictive in order that it have a significant effect on parameter-extraction accuracy. All that is necessary is that it be more restrictive than the information contained in the waveform, for some of the parameters \[…\] what is more surprising is that due to the effects of correlations, the rms errors obtained for the other parameters may also be overestimated by large factors.” Roughly speaking, this happens because as we move in parameter space, the change in the signal can be partially absorbed by changing correlated parameters together; thus, limiting the range available to one parameter also limits the range over which a correlated parameter can run while not significantly modifying the signal. In this section we seek a practical recipe to determine, in the context of a parameter-estimation problem specified by a family of waveforms and a fiducial SNR, whether it is necessary to take priors into consideration when evaluating projected parameter accuracies. Since prior probabilities can only be discussed consistently in the framework of Bayesian parameter estimation, in this section we will restrict ourselves to that context. The Gaussian priors examined at the end of Sec. \[sec:bayeshighsn\] are rarely appropriate in actual practice, but they do provide a quick test to see if the prior-less Fisher result can be taken as it stands, or whether a more careful analysis is needed that includes the effects of priors. In Sec. \[sec:priorexample\] we try out this quick test on the SPA model of Sec. \[sec:standardsignal\]. For simplicity, we shall consider the effects of priors as logically independent from the sufficiency of the LSA, although the two problems clearly come into play together in real situations. Testing for the influence of priors (normal true-parameter–centered priors) {#sec:priorexample} --------------------------------------------------------------------------- We shall discuss our quick test by way of an example. The standard SPA model of Sec. \[sec:standardsignal\] has six parameters: $M_c$, $\eta$, $\beta$, $\sigma$, $\phi_0$, and $t_0$ (plus $A$, which we disregard). We work at 2PN with SNR = $A$ = 10, with true parameters $m_1 = m_2 = 10 M_\odot$ (corresponding to $M_c = 8.71 M_\odot$, $\eta = 0.25$), and $\beta = \sigma = \phi_0 = t_0 = 0$. We wish to examine the effect of priors for three related parameter-estimation problems involving different subsets of parameters: a 4-parameter problem (4pp) where we disregard spin parameters (i.e., where we assume we know *a priori* that the true binary has no spin); a 5pp where spin–orbit coupling \[as represented by $\beta$ in Eq. \] is important, but spin–spin interactions can be neglected; and a 6pp where we include also spin–spin interactions \[as represented by $\sigma$ in Eq. \]. As we shall see, priors become increasingly important as the number of parameters increases. In each problem, we compute the expected covariance matrix of the posterior distribution as the inverse of (a submatrix of) the Fisher matrix, neglecting any non-LSA effects. We represent priors as normal distributions centered around null parameter displacements (i.e., the true parameter value), with standard deviations of 0.25 for $\eta$ and, following Poisson and Will [@pw1995] 8.5 for $\beta$ and 5 for $\sigma$ (in Ref. [@baw2005], Berti and colleagues derive and adopt approximate priors for $\beta$ and $\sigma$ with standard deviations $\Delta \beta = 9.4$ and $\Delta \sigma = 2.5$). This representation is very crude, but it is the only one that leads to a simple analytical result \[Eq. \] for the posterior covariance, and it should give at least a qualitative idea of the effect of imposing rectangular priors covering the allowed parameter ranges. Results are shown in the upper section of Table \[tab:priors1010\], and are as follows. $\Delta M_c / M_c$ $\Delta \sigma$ $\Delta \phi_0$ ------------------------------------------------ --------------------- --------------------- ----- ----------------- ------------------- -- ----- 4pp, no priors $2.9\times 10^{-2}$ $8.3\times 10^{-2}$ 7.3 3.0 4pp, NTC prior on $\eta$ $2.7\times 10^{-2}$ $7.9\times 10^{-2}$ 7.0 2.8 5pp, no priors $3.3 \times 10^2$ 6.9 5pp, NTC prior on $\beta$ $3.0\times 10^{-2}$ 8.5 7.7 3.0 5pp, NTC prior on $\eta$, $\beta$ $3.0\times 10^{-2}$ $2.1\times 10^{-1}$ 4.9 7.7 3.0 6pp, NTC priors on $\eta$, $\beta$ $4.3\times 10^{-2}$ $2.5\times 10^{-1}$ 8.4 $4.3 \times 10^1$ 5.3 6pp, NTC priors on $\eta$, $\beta$, $\sigma$ $3.0\times 10^{-2}$ $2.1\times 10^{-1}$ 5.1 4.9 $1.1 \times 10^1$ 3.1 4pp, exact priors on $\eta$ $1.8\times 10^{-2}$ $5.0\times 10^{-2}$ 4.4 1.9 5pp, exact priors on $\eta$, $\beta$ $2.9\times 10^{-2}$ $7.1\times 10^{-2}$ 2.4 7.5 2.9 6pp, exact priors on $\eta$, $\beta$, $\sigma$ $2.9\times 10^{-2}$ $7.1\times 10^{-2}$ 2.6 2.9 9.0 3.0 : Fisher-matrix rms errors in the 4-, 5-, and 6-parameter–estimation problems for a $(10+10)M_\odot$ binary with $\beta = \sigma = 0$ and $\mathrm{SNR} = 10$, evaluated under different combinations of normal true-parameter–centered (NTC) priors (upper section of table) and of the exact priors of Sec. \[sec:exactprior\] (lower section). The underlined errors are larger than the physical range for the parameter. \[tab:priors1010\] The first line of Table \[tab:priors1010\] shows the 4pp no-prior $1\sigma$ values for the single-parameter rms errors (i.e., the square roots of the diagonal elements in the covariance matrix). Among these, $\Delta M_c$ and $\Delta \eta$ seem reasonable, but we get hung up on the value of $\Delta \phi_0$. Can the $1\sigma$ error region be larger than the physically meaningful range for this angle? On general grounds, we should worry that the LSA cannot know that the waveforms are exactly periodic (and therefore nonlinear) in the angular parameters, so it blithely extrapolates small-angle effects to infinite ranges. However, as pointed out by Cutler [@cpc2007], this extrapolation is roughly correct for a simple complex phase such as $\phi_0$ \[see Eq. \], for which the main correlated-parameter effect is to absorb the global phase shifts due to changes in the other parameters.[^20] A large $\Delta \phi_0$ indicates that this absorption can happen through several cycles of phasing. We conclude that $\phi_0$ is essentially undetermined, but we have no reason to distrust the errors for $\Delta M_c$, $\Delta \eta$, and $\Delta t_0$. Applying a prior to $\eta$ does not change the picture significantly, but priors do matter once we add the spin parameters, which are very poorly determined at this SNR. In the 5pp, we find unphysically large errors for both $\beta$ *and* $\eta$, which are cured only by imposing priors on *both* parameters. In the 6pp, we find that a prior is needed also for $\sigma$; adding it engenders measurable changes in $\Delta M_c$ and $\Delta \eta$. As a rule of thumb, we should expect such effects whenever the signal derivatives show significant correlations, and when the magnitudes of the priors, measured crudely as the squared inverses $(\theta_i^\mathrm{max} - \theta_i^\mathrm{min})^{-2}$ of the effective parameter ranges induced by the priors, are comparable to the corresponding diagonal Fisher-matrix elements $F_{ii}$. Testing for the influence of priors (exact priors) {#sec:exactprior} -------------------------------------------------- We can perform an even better test by evaluating the effects of *exact* priors while still working in the LSA. Doing this requires some numerics, which are however very manageable on a workstation-class system. The idea is to integrate $\langle \Delta \theta_i \Delta \theta_j \rangle_p$ as a Monte Carlo sum, which can be accomplished as follows. First, we need to fix a reference experiment by drawing the random variable $n_j \equiv (n,h_j)$ from its ensemble distribution, which in Gaussian noise is a normal distribution with mean zero and covariance matrix $F_{ij} \equiv (h_i,h_j)$ \[see Eq. \]. To do so, we generate a zero-mean, unit-variance, normal $N$-tuple, and multiply it by $\sqrt{F_{ij}}$ (where the square root is taken in the linear-operator sense and exists for nonsingular Fisher matrices). Note that we cannot work with SNR-invariant expressions (e.g., normalized parameter errors $\bar{\theta}_i$), since the priors set a scale for the strength of the signal. We can now draw samples distributed with the LSA likelihood $p(s|\theta) \propto \exp \{ -(h_j,h_k) \theta_j \theta_l / 2 + n_j \theta_j \}$. To do so, we generate zero-mean, unit-variance, normal $N$-tuples, multiply them by $(F_{ij}^{-1})^{1/2}$, and offset them by $F_{ij}^{-1} n^j$. We include the effects of priors (therefore obtaining a population $\{\theta^{(i)}\}$ distributed according to the LSA *posterior* probability) by going through the samples, and discarding each of them with a probability $1 - p(\theta) / \max_\theta p$ (for rectangular priors the probability of discarding is always 0 or 1). The covariance matrix of the posterior distribution can then be computed from the surviving samples. We repeat this procedure for many different experiments (i.e., $n_j$’s), and take a frequentist average of the covariance-matrix components (or study their frequentist distribution). Again, the Bayesian interpretation of this entire procedure is as follows: we place the true signal at $\theta = 0$; we draw from the possible noise realizations according to their ensemble probability; and we compute the variance of the posterior distribution for each noise realization. If the priors are very restrictive compared to Fisher-matrix–only errors, we may find that we are discarding a very large percentage of the samples. To avoid this, we can incorporate a normal approximation to the priors in the probability distribution used to generate the samples (i.e., by multiplying normal variates by $\sqrt{[(h_i,h_j) + P_{ij}]^{-1}}$, and offsetting them by $[(h_i,h_j) + P_{ij}]^{-1} n^j$), and then sieve the resulting samples with respect to $\propto p(\theta) e^{P_{ij} \theta_i \theta_j / 2}$ instead of $p(\theta)$. It is also possible to use rejection sampling [@nrc], as we did for the results reported in this section, or the Metropolis algorithm [@metropolis] with the likelihood or likelihood-plus-NTC-prior as proposal distribution, and the full posterior as the target distribution. Applying the procedure outlined above to our $(10+10)M_\odot$ system yields the results listed in the lower section of Table \[tab:priors1010\]. We adopt exact priors given by rectangular probability distributions covering the intervals $[0,\infty]$ for $M_c$, $[0,0.25]$ for $\eta$, $[-8.5,8.5]$ for $\beta$, and $[-5,5]$ for $\sigma$. Each quoted error is a frequentist average of 200 independent Monte Carlo estimates, each computed for a different realization of noise from an initial sampling of $10^6$ parameter sets, reduced to $5\times 10^4$–$2 \times 10^5$ samples after rejection sampling, depending on the estimation problem. The expected errors are significantly reduced compared to the NTC-prior estimates. These reductions stem mainly from the greater tightness of the rectangular priors, and are especially significant for for $\eta$, for which the symmetric NTC prior is indeed very crude. The lesson is that we can use $[(h_i,h_j) + P_{ij}]^{-1}$ (i.e., the quick test) to decide whether priors are important, but we need something more sophisticated, such as the procedure described in this section, to gauge their effects accurately. Of course, this gain in accuracy may be only virtual if the LSA is not warranted for our problem. Deciding that question is the object of the next section. The unbearable lightness of signal to noise {#sec:howhigh} =========================================== As we have seen in Sec. \[sec:threeways\], the high-SNR and LSA limits coincide because larger signal strengths correspond to smaller statistical errors, which in turn imply that the linearized-signal expression for the likelihood is more accurate. The equivalence of the two limits is manifest in the $1/A$ expansions such as Eqs.  and . Indeed, Finn [@finn1992] cautions that “it is important that the probability contours of interest (e.g. 90%) do not involve \[errors\] so large that the linearization of \[the likelihood\] is a poor approximation.” In practice, given a family of waveforms and the true parameter values, we need to ask how high an SNR is needed for the limits to yield accurate expected errors. One approach involves comparing the Fisher-matrix results with errors computed at the next order in the $1/A$ expansions: in App. A 5 of Ref. [@cf1994], Cutler and Flanagan provide next-order formulas for the frequentist variance (although they do not apply them to the Fisher-matrix estimates in the same article). In Sec. \[sec:higherorder\] we provide the mathematicals tools to do so in our notation; we must however warn the reader that the calculation is rather cumbersome, except for simple waveforms, and the verdict is still not definitive: the smallness of a term in a series does not guarantee that the series is converging. A simpler approach involves working with the ratio $r(\theta,A)$ of the LSA likelihood to the exact likelihood to build a consistency criterion for the Fisher-matrix formalism. In this section we shall see that under reasonable conditions, the ratio $r$ is given in logarithm by $$\label{eq:sntest} |\log r(\theta,A)| = \bigl(\theta_j h_j - \Delta h(\theta) , \theta_k h_k - \Delta h(\theta) \bigr) / 2,$$ where $\Delta h(\theta) = h(\theta) - h(0)$, $A$ is the signal strength, and $\theta$ is the error (in a sense to be made precise shortly). Since $h(\theta) = h(0) + \theta_i h_i + \cdots$, the product in Eq.  represents the noise-weighted norm of the higher-than-linear contributions to $h(\theta)$, expanded around the true source parameters. The idea of the criterion is to choose an isoprobability surface (say, the $1\sigma$ surface), as predicted by the Fisher matrix, and then explore it to verify that the mismatch between the LSA and exact likelihoods is smaller than a fiducial value (say, $|\log r| < 0.1$), so that we can actually believe the LSA in predicting the $1\sigma$ surface to begin with. We stress that this is just a criterion of *consistency*. Even if the Fisher-matrix result is internally consistent, it may still be inaccurate; conversely, the structure of the ambiguity function across parameter space could conspire in such a way as to make the LSA results correct, although we have no reason to expect that in general. In the rest of this section, we explain how the criterion comes about in the frequentist (Sec. \[sec:criterionfreq\]) and Bayesian (Sec. \[sec:criterionbayes\]) frameworks, and we show a concrete example of the criterion in use (Sec. \[sec:criterionpractical\]). Frequentist justification of the maximum-mismatch criterion {#sec:criterionfreq} ----------------------------------------------------------- As in Secs. \[sec:threeways\] and \[sec:priors\], we assume that the detector output is $s = A \bar{h}_0 + n$. In the LSA, the ML estimator $\hat{\theta}^\mathrm{ML}$ is a normal variable with mean zero and covariance matrix $(1/A^2) (\bar{h}_j,\bar{h}_k)^{-1}$ \[Eq. \]. For a given signal amplitude $A$, let $\hat{\theta}^\mathrm{ML}$ take on the specific value ${\theta^{1\sigma}}$ on its $1\sigma$ surface. From Eqs. , , and , the mismatch ratio $r$ is given by $$r({\theta^{1\sigma}}) = \exp \left\{-\bigl(s - A (\bar{h}_0 + {\theta^{1\sigma}}_j \bar{h}_j),s - A (\bar{h}_0 + {\theta^{1\sigma}}_k \bar{h}_k)\bigr)/2\right\} \Big/ \exp\left\{-\bigl(s - A \bar{h}({\theta^{1\sigma}}), s - A \bar{h}({\theta^{1\sigma}}) \bigr)/2\right\};$$ writing $s$ out, we eliminate all instances of $\bar{h}_0$: $$\log r({\theta^{1\sigma}}) = - A^2 {\theta^{1\sigma}}_j {\theta^{1\sigma}}_k \bigl( \bar{h}_j, \bar{h}_k \bigr) / 2 + A^2 \bigl( \Delta \bar{h}({\theta^{1\sigma}}), \Delta \bar{h}({\theta^{1\sigma}}) \bigr) / 2 + A \, {\theta^{1\sigma}}_j \bigl( \bar{h}_j, n \bigr) - A \bigl( \Delta \bar{h}({\theta^{1\sigma}}), n \bigr).$$ The first two terms in the exponent can be computed given ${\theta^{1\sigma}}$; not so the two products involving $n$. To obtain the first, we note that if $\hat{\theta}^\mathrm{ML} = {\theta^{1\sigma}}$, then the noise must be such that ${\theta^{1\sigma}}= (1/A) (\bar{h}_j,\bar{h}_k)^{-1} (\bar{h}_k,n)$ \[Eq. \], so $(\bar{h}_j,n) = A (\bar{h}_j,\bar{h}_k) {\theta^{1\sigma}}_k$. To obtain the second, we change our perspective slightly, and average $\log r({\theta^{1\sigma}})$ over all noise realizations $n$ compatible with ${\theta^{1\sigma}}$. This is how. Let $x_j \equiv (\bar{h}_j,n)$, $y \equiv ( \Delta \bar{h}({\theta^{1\sigma}}), n )$: separately, $x_j$ and $y$ are normal random variables with mean zero and covariances equal to $(\bar{h}_j,\bar{h}_k)$ and $( \Delta \bar{h}({\theta^{1\sigma}}), \Delta \bar{h}({\theta^{1\sigma}}) )$, respectively; taken together, they are jointly normal variables with covariance $(\bar{h}_j,\Delta \bar{h}({\theta^{1\sigma}}))$ \[Eq.  again\]. We now know enough to build $p(x,y)$, from which we can derive the conditional probability $p(y|x)$ and compute the conditional mean of $y$, which (after the algebra of App. \[sec:conditional\]) turns out to be $A \, {\theta^{1\sigma}}_j (\bar{h}_j,\Delta \bar{h}({\theta^{1\sigma}}))$. Altogether, we find $$\begin{aligned} \langle \log r({\theta^{1\sigma}}) \rangle_{n({\theta^{1\sigma}})} &= A^2 {\theta^{1\sigma}}_j {\theta^{1\sigma}}_k \bigl( \bar{h}_j, \bar{h}_k \bigr) / 2 + A^2 \bigl( \Delta \bar{h}({\theta^{1\sigma}}), \Delta \bar{h}({\theta^{1\sigma}}) \bigr) / 2 - A^2 {\theta^{1\sigma}}_j \bigl( \bar{h}_j,\Delta \bar{h}({\theta^{1\sigma}}) \bigr) \\ &= \bigl( {\theta^{1\sigma}}_j h_j - \Delta h({\theta^{1\sigma}}), {\theta^{1\sigma}}_k h_k - \Delta h({\theta^{1\sigma}}) \bigr) / 2, \label{eq:frequentistresult} \end{aligned}$$ just as anticipated in Eq. . The signal strength $A$ enters Eq.  explicitly, but also implicitly through the parameter width of the Fisher-matrix $1\sigma$ surface. Thus Eq.  can be solved for the $A$ that corresponds to ${\theta^{1\sigma}}$ small enough to yield $r$ as close to unity as desired. Since to leading order ${\theta^{1\sigma}}_j - \Delta h({\theta^{1\sigma}}) = A \bar{h}_{jk} {\theta^{1\sigma}}_j {\theta^{1\sigma}}_k$, and since to leading order ${\theta^{1\sigma}}$ scales as $1/A$, we expect $\log r$ to scale as $1/A^2$ for large enough $A$. In summary, the maximum-mismatch criterion is justified from a frequentist viewpoint as a constraint on the ratio $r$ at points on a constant–LSA-probability surface, averaged over all realizations of noise compatible with finding the ML estimator at those points. Bayesian justification of the maximum-mismatch criterion {#sec:criterionbayes} -------------------------------------------------------- The justification of the maximum-mismatch criterion from a Bayesian viewpoint requires another slight change of perspective. Again we assume $s = A \bar{h}_0 + n$; this time, however, we expand the waveform not with respect to the true parameters ($\theta = 0$), but to the observed location $\hat{\theta}^\mathrm{ML}(n) \equiv \theta^n_0$ of maximum LSA likelihood for a given experiment. In the absence of priors, it is with respect to this location that the uncertainty of the posterior would be judged in a single experiment. Thus we write $h(\theta) \equiv h(\theta^n_0 + \theta^n) \simeq h^n_0 + \theta^n_j h^n_j$, where $h^n_0 \equiv h(\theta^n_0)$: the “${}^n$” superscripts serve to remind us that the parameter displacements $\theta^n$ (and the waveform derivatives $h^n_j$) are evaluated from (at) $\theta^n_0$. We also write $\Delta h^n_0$ for $h_0 - h^n_0$. The $1\sigma$ surface over which we are going to evaluate the ratio $r$ will be a surface of equiprobable true-signal locations, given the observed location $\theta^n_0$ of maximum LSA likelihood. In the LSA, the distribution over experiments of the true-signal location with respect to $\theta^n_0$ is again normal with covariance matrix $(h^n_i,h^n_j)^{-1}$. Thus we have $\theta^n = -\theta^n_0$, and the mismatch ratio $r$ is given by $$r(\theta_0) = \exp \left\{ -\bigl( s - A(\bar{h}_0^n + \theta^n_j \bar{h}^n_j), s - A(\bar{h}_0^n + \theta^n_k \bar{h}^n_k) \bigr) / 2 \right\} \Big/ \exp \left\{ -\bigl( s - A \bar{h}_0, s - A \bar{h}_0 \bigr) \right\};$$ writing $s$ out, the denominator reduces to $\exp -(n,n)/2$, and $h_0$ enters the numerator only through $\Delta h^n_0$: $$\label{eq:tempbayesianr} \log r(\theta_0) = - A^2 \theta^n_j \theta^n_k \bigl( \bar{h}^n_j, \bar{h}^n_k \bigr) / 2 - A^2 \bigl( \Delta \bar{h}^n_0, \Delta \bar{h}^n_0 \bigr) / 2 + A^2 \theta^n_i \bigl( \bar{h}^n_i, \Delta \bar{h}^n_0 \bigr) + A \, \theta^n_i \bigl( \bar{h}^n_i, n \bigr) - A \bigl(\Delta \bar{h}^n_0, n \bigr).$$ Now, since the LSA likelihood can be written as $p(s|\theta^n) \propto \exp -(n + \Delta h_0^n - \theta^n_i h_i^n,n + \Delta h_0^n - \theta^n_j h_j^n) / 2$, the ML equation $\partial p/\partial \theta^n_i = 0$ at $\theta^n_i = 0$ implies $(h^n_i,n) = -(h^n_i,\Delta h^n_0)$. We handle the last term of the equation by evaluating the conditional mean of $y^n \equiv -A (\Delta \bar{h}^n_0, n)$ given $x^n = (\bar{h}^n_j,n) = -A (\bar{h}^n_j,\Delta \bar{h}^n_0)$, producing $-A (\bar{h}_j^n,n) (\bar{h}^n_j,\bar{h}^n_k)^{-1} (\bar{h}^n_k, \Delta \bar{h}^n_0) = A (\bar{h}_j^n,n) (h^n_j,h^n_k)^{-1} (h^n_k, n)$ (again, see App. \[sec:conditional\]). We can then use Eq.  to replace $(h^n_j,h^n_k)^{-1} (h^n_k, n)$ with $\theta^n_{0j} = -\theta^n_j$ (working to leading order), so that the last two terms of Eq.  end up canceling out: $$\begin{aligned} \log r(\theta_0) &= - A^2 \theta^n_j \theta^n_k \bigl( \bar{h}^n_j, \bar{h}^n_k \bigr) / 2 + A^2 \theta^n_j (\bar{h}^n_j, \Delta \bar{h}^n_0) - A^2 \bigl( \Delta \bar{h}^n_0, \Delta \bar{h}^n_0 \bigr) / 2 \\ &= -\bigl( \theta^n_j h^n_j - \Delta h(\theta^n), \theta^n_k h^n_k - \Delta h(\theta^n) \bigr) / 2. \label{eq:bayesianresult} \end{aligned}$$ Again, this equation can be solved for the $A$ that corresponds to $1\sigma$ true-signal locations $\theta^n$ small enough to yield $r$ close to unity. Interestingly, the signs of the frequentist and Bayesian expressions and are opposite, indicating (at least *prima facie*) that the likelihood is overestimated in the frequentist case, underestimated in the Bayesian case. Given the conditions under which we have obtained Eqs.  and , it is perhaps best to consider only their absolute value as rough indicators of the appropriateness of the high-SNR/LSA limit. In summary, the maximum-mismatch criterion is justified from a Bayesian viewpoint by fixing the location of maximum LSA likelihood, and then exploring a surface of equiprobable true-signal locations, evaluating for each the average of $\log r$ over all experiments (i.e., realizations of noise) compatible with having the true signal there. Practical usage of the maximum-mismatch criterion {#sec:criterionpractical} ------------------------------------------------- In both the frequentist and the Bayesian pictures, Eq.  yields the noise-averaged logarithm of the likelihood mismatch, $|\log r|$, as a function of the signal strength $A$ and of a direction in parameter space that identifies a point on the $1\sigma$ surface, given by the solutions of the LSA equation $A^2 (\bar{h}_j,\bar{h}_k) \theta_j \theta_k = 1$, and interpreted as equiprobable locations for the ML estimator given the true signal $\theta = 0$ (in the frequentist picture), or for the true signal given the mode of the likelihood at $\theta = 0$ (in the Bayesian picture). We use Eq.  by fixing the signal strength to what is reasonably expected in observations, perhaps close to the minimum detection SNR, although the astronomical distribution and intrinsic strengths of sources may prompt other choices (e.g., the the supermassive–black-hole binaries to be observed by LISA have typical SNRs in the hundreds); and then by evaluating $|\log r|$ as a function of direction in parameter space. Figure \[fig:simplelogr\] shows an example of this procedure for a very simple and benign one-dimensional estimation problem (a sinusoid of known amplitude and frequency in Gaussian stationary noise), where the only parameter left to estimate is the initial phase $\phi_0$ ($=0$ for the true signal). For each value of $\mathrm{SNR} \equiv A$, the expected $1\sigma$ surface consists of just the two points $\phi_0^{1\sigma} = 1/A$. Figure \[fig:simplelogr\] shows $|\log r|$ as a function of $\phi_0^{1\sigma}$, and therefore of $A$. If we set a threshold of $|\log r| = 0.1$ (the dashed line) to claim the high-SNR/LSA limit as consistent, we see that the consistency criterion is not satisfied for $A=1$, where $|\log r| \simeq 0.12$, but it begins to be satisfied for $A \gtrsim 1.09$. Once again, for a given SNR, $|\log r|$ at $\phi_0^{1\sigma}$ is an index of the closeness of the LSA and exact likelihoods at a typical values of the errors, and averaged among compatible noise realizations. ![Consistency criterion for the simple waveform model $h(\phi_0) = A \cos(2 \pi f t + \phi_0)$ in Gaussian stationary noise, with fixed (known) $A$ and $f$. The curve plots $|\log r|$ as a function of the $1\sigma$ error $\phi_0^{1\sigma} = \pm 1/A$, with specific values of $A$ called out by the circles. For the threshold $|\log r| = 0.1$, consistency is achieved for $A \gtrsim 1.09$. (To generate this graph, the integration time was set to $1000/f$, and the variance of noise adjusted so that $(\bar{h}(\phi_0),\bar{h}(\phi_0)) = 1$.)\[fig:simplelogr\]](simplelogr-r.eps) The principle is the same for multiparameter estimation problems, where we have the additional task of sampling the entire $1\sigma$ surface in a manner consistent with the LSA distribution at $1\sigma$. One way to do so is to obtain the eigenvalues $\lambda^{(i)}$ and eigenvectors $\theta^{(i)}_j$ of $(\bar{h}_j,\bar{h}_k)$, and then sample the parameter values $\theta = \sum_{(i)=1}^N \tilde{c}^{(i)} \theta^{(i)}_j / (A \sqrt{\lambda^{(i)}})$, with $\tilde{c}^{(i)}$ distributed uniformly on the $N$-dimensional unit sphere.[^21] We then obtain the cumulative distribution function for the values of $|\log r|$, which we plot in Fig. \[fig:logrhist\] for our reference model. If we consider the high-SNR/LSA limit to be sufficiently realized when $|\log r| < 0.1$ over 90% of the $1\sigma$ surface, we conclude that the Fisher-matrix formalism (with no priors) is self-consistent for SNRs between 10 and 20 in the 4-parameter problem, between 100 and 200 in the 5pp, and between 4000 and 10000 in the 6pp. ![Cumulative distribution function for $|\log r|$ on the $1 \sigma$ surface at various SNRs for our reference SPA model with $m_1 = m_2 = 10 M_\odot$. The SNR required to have 90% of the $1\sigma$ points at $|\log r| = 0.1$ (dashed lines) increases considerably (in fact, to unrealistic values) as we move from the 4pp to 5pp and 6pp. This figure was produced without imposing any priors on the source parameters.\[fig:logrhist\]](histlogr-r2.eps){width="5in"} The eigenvector directions that push the required SNR toward higher values are usually those associated with the smallest-magnitude eigenvalues. To confirm this, and to get some clues about the beyond-LSA structure of the likelihood, we can fix the maximum acceptable value of $|\log r|$ (say, again to 0.1) and then solve Eq.  for $A$ as a function of direction in parameter space. We do so for the 4pp in Fig. \[fig:maxsnsecs\], where we show all two-dimensional parameter subspaces along pairs of eigenvectors (strictly speaking, were are not sampling a single $1\sigma$ surface, but considering the set of such surfaces for all SNRs, and determining on which of them $|\log r| = 0.1$, as a function of parameter angle). ![SNR values at which $|\log r| = 0.1$ in the 4pp for our reference SPA model with $m_1 = m_2 = 10 M_\odot$. The six subplots display sections of parameter space corresponding to all distinct pairs of eigenvectors of the Fisher matrix; the polar radius of the curves shows the required SNR (plotted logarithmically from $\log \mathrm{SNR} = 0$), while the polar angle is computed between pairs of renormalized eigenvector coefficients $\tilde{c}^{(i)}$. The graph at the bottom right shows the composition of the four Fisher-matrix eigenvectors in terms of the source parameters, as well as their respective eigenvalues. This figure was produced without imposing any priors on the source parameters.\[fig:maxsnsecs\]](polarsn4p-2-r.eps){width="6in"} In the Bayesian framework, it is also possible to combine the maximum-mismatch criterion with the normal prior test of Sec. \[sec:priorexample\], by investigating the values of $|\log r|$ on the $1\sigma$ surface given by the solutions of the LSA-*cum*-*prior* equation $[(h_j,h_k) + P_{jk}]\theta_j \theta_k = 1$. This test can help decide whether the LSA is warranted once priors are factored in: it can be shown that Eq.  continues to hold with NTC priors, although its interpretation is not as clean, because their mode moves around the $1\sigma$ surface as we explore it. The maximum-mismatch criterion may indicate that, at the signal strengths of interest, the LSA becomes consistent only with the priors; in that case, the reliable predictor of source parameter accuracy would not be $[(h_j,h_k) + P_{jk}]^{-1}$ (given the crudeness of NTC priors), but rather the result of an LSA Monte Carlo procedure such as that described in Sec. \[sec:exactprior\]. It is very hard to make a general statement about the errors in the expected accuracies when the LSA Fisher-matrix result is not self-consistent. Such errors are strongly SNR-dependent, and it is usually necessary to include parameter priors into consideration. As anecdotal evidence, I offer that for our reference model at $\mathrm{SNR} = 10$, a full-blown Monte Carlo sampling of the posterior distribution, involving an explicit time-domain realization of noise and adopting the priors of Sec. \[sec:exactprior\], reports posterior variances that differ from the last three rows of Tab. \[tab:priors1010\] by few tens percent for the 4pp, and by factors of a few (for $M_c$ and $\eta$ only, since $\Delta \beta$ and $\Delta \sigma$ are dominated by the priors) for the 5pp and 6pp. In conclusion, I submit that graphs like those of Fig. \[fig:logrhist\] can be useful to assess the consistency of the “straight” Fisher-matrix formalism, are easy to produce with little additional machinery, and should be included in all articles that use the formalism to predict the future parameter-estimation performance of GW observations. If a single number must be quoted, it could be the SNR at which 90% of the $1\sigma$ surface yields $|\log r| < 0.1$. Beyond the linearized-signal approximation {#sec:higherorder} ========================================== In this section we develop mathematical tools to derive higher-than-LSA expressions for the frequentist mean and variance of the ML estimator over an ensemble of noise realizations (Sec. \[sec:higherfrequentist\]), and for the Bayesian mean and variance of the posterior distribution (without priors) in a single experiment (Sec. \[sec:higherbayesian\]). These expressions provide corrections to the Fisher-matrix result, and can therefore be used to check its accuracy, as suggested by Cutler and Flanagan [@cf1994], who derive a general expression for the $1/A^4$ correction to the frequentist variance. A formal treatment of the $1/A$ expansion for the frequentist moments can be found in Barndorff-Nielsen and Cox [@bnc1994] and in Zanolin, Naftali, and Makris [@znm2007], who use the expansion to determine conditions for the ML estimate to become unbiased and attain the Cramér–Rao bound [@nm2001]. However, computing higher-order corrections involves a considerable amount of tensorial algebra that calls for the use of specialized software, such as <span style="font-variant:small-caps;">MathTensor</span> [@mathtensor]; they also involve higher-than-first derivatives of the waveforms and products of several inverse Fisher matrices, which may raise concerns about the numerical accuracy of the computations. Throughout this section, we distinguish between covariant and contravariant indices (as in $n_i = (\bar{h}_i,n)$ and $\theta^i$, respectively); in fact, we find it convenient to use the inverse normalized Fisher matrix $(\bar{h}^i,\bar{h}^j)^{-1}$ to raise indices, therefore hiding its repeated appearance in tensor expressions. In the frequentist framework {#sec:higherfrequentist} ---------------------------- Using the $1/A$ expansion of Eqs.  and , the perturbative ML equations can be written in general as $$\begin{aligned} H_{ij} \theta^j_{(1)} & = N_i, && \\ H_{ij} \theta^j_{(2)} & = N_{ij} \theta^j_{(1)} & -\,& H_{ijk} \theta^{jk}_{(2)}, \\ H_{ij} \theta^j_{(3)} & = N_{ij} \theta^j_{(2)} + N_{ijk} \theta^{jk}_{(2)} & -\,& H_{ijk} \theta^{jk}_{(3)} - H_{ijkl} \theta^{jkl}_{(3)}, \\ H_{ij} \theta^j_{(4)} & = N_{ij} \theta^j_{(3)} + N_{ijk} \theta^{jk}_{(3)} + N_{ijkl} \theta^{jkl}_{(3)} \!\!\!\!\! & -\,& H_{ijk} \theta^{jk}_{(4)} - H_{ijkl} \theta^{jkl}_{(4)} - H_{ijklm} \theta^{jklm}_{(3)}, \\ & \cdots && \end{aligned} \label{eq:mleqsexpand}$$ with $N_i = (n,\bar{h}_i) / 0!$, $N_{ij} = (\bar{h}_{ij},n) / 1!$, $N_{ijk} = (\bar{h}_{ijk},n) / 2!$ (and so on), and $$\begin{aligned} H_{ij} & = \frac{\bar{h}_{i,j}}{0! \, 1!}, \\ H_{ijk} & = \frac{\bar{h}_{i,jk}}{0! \, 2!} + \frac{\bar{h}_{ij,k}}{1! \, 1!}, \\ H_{ijkl} & = \frac{\bar{h}_{i,jkl}}{0! \, 3!} + \frac{\bar{h}_{ij,kl}}{1! \, 2!} + \frac{\bar{h}_{ijk,l}}{2! \, 1!}, \\ & \cdots \end{aligned} \label{eq:defbigh}$$ where $\bar{h}_{i,j} = (\bar{h}_i,\bar{h}_j)$, $\bar{h}_{i,jk} = (\bar{h}_i,\bar{h}_{jk})$ (and so on), and where the two factorials at each denominator are those (respectively) of the number of indices before the comma minus one, and of the number of indices after the comma. Also, the $\theta^j_{(n)}$ of Eq.  are the unknown $1/A^n$ contributions to $\hat{\theta}^\mathrm{ML}$ (as in Eq. , dropping hats for simplicity), while the multi-index parameter objects such as $\theta^{jk}_{(2)}$ are given by $$\begin{array}{lll} \theta^{jk}_{(2)} = \theta^j_{(1)} \theta^k_{(1)}, \quad & \theta^{jk\phantom{l}}_{(3)} = \theta^j_{(1)} \theta^k_{(2)} + \theta^j_{(2)} \theta^k_{(1)}, \quad & \theta^{jk\phantom{lm}}_{(4)} = \theta^j_{(1)} \theta^k_{(3)} + \theta^j_{(2)} \theta^k_{(2)} + \theta^j_{(3)} \theta^k_{(1)}, \\[2mm] & \theta^{jkl}_{(3)} = \theta^j_{(1)} \theta^k_{(1)} \theta^l_{(1)}, & \theta^{jkl\phantom{m}}_{(4)} = \theta^j_{(1)} \theta^k_{(1)} \theta^l_{(2)} + \theta^j_{(1)} \theta^k_{(2)} \theta^l_{(1)} + \theta^j_{(2)} \theta^k_{(1)} \theta^l_{(1)}, \\[2mm] & & \theta^{jklm}_{(4)} = \theta^j_{(1)} \theta^k_{(1)} \theta^l_{(1)} \theta^m_{(1)}, \end{array}$$ and so on. In general, the object $\theta^{j_1{\cdots}j_m}_{(n)}$ will consist of as many addends as there are partitions of $n$ into $m$ integers, including all permutations of each partition. For instance, the $n=5$, $m=3$ object $\theta_{(5)}^{jkl}$ would have terms for each of the partitions $1+1+3$, $1+3+1$, $3+1+1$, $1+2+2$, $2+1+2$, $2+2+1$. The solution of each equation in Eq.  is trivial given the solutions of all equations of lower order. Since the inverse matrix $(H^{-1})^{ij} \equiv (\bar{h}_i,\bar{h}_j)^{-1} = A^2 F_{ij}^{-1}$ appears multiple times in the solutions (because $H_{ij}$ multiplies the unknown $\theta^j_{(n)}$ in each equation), it is convenient to adopt a compact notation that hides the $(H^{-1})^{ij}$ by raising every index into which they are contracted. We then find $$\begin{aligned} \theta_{(1)}^i & = N^i, \\ \theta_{(2)}^i & = {N^i}_j \theta_{(1)}^j - {H^i}_{jk} \theta_{(2)}^{jk} = {N^i}_j N^j - {H^i}_{jk} N^j N^k, \\ \theta_{(3)}^i & = {N^i}_j \theta_{(2)}^j + {N^i}_{jk} \theta_{(1)}^j \theta_{(1)}^k - {H^i}_{jk} (\theta_{(1)}^j \theta_{(2)}^k + \theta_{(2)}^j \theta_{(1)}^k) - {H^i}_{jkl} \theta_{(1)}^j \theta_{(1)}^k \theta_{(1)}^l = \cdots \\ & \ldots \end{aligned}$$ The frequentist mean and covariance of the $\theta^i$ can be built from these solutions, remembering the Wick-product rule [@zinnjustin] for the ensemble average of products of Gaussian variables: $$\begin{aligned} \langle(a,n)\rangle_n & = 0, \\ \langle(a,n)(b,n)\rangle_n & = (a,b), \\ \langle(a,n)(b,n)(c,n)\rangle_n & = 0, \\ \langle(a,n)(b,n)(c,n)(d,n)\rangle_n & = (a,b)(c,d) + (a,c)(b,d) + (a,d)(b,c) \\ & \ldots \end{aligned} \label{eq:wick}$$ (for any signals $a$, $b$, $c$, and $d$), where all the products with an odd number of factors vanish, while the products with an even number of factors are given by the sum of terms corresponding to all distinct pairings of signals into inner products. Thus we find that all the $\langle \theta^i_{\mathrm{odd}\,(k)} \rangle_n$ vanish, while the first non-zero correction to $\langle \theta^i \rangle_n$ is $$\langle \theta_{(2)}^i \rangle_n = \langle {N^i}_j N^j \rangle_n - {H^i}_{jk} \langle N^j N^k \rangle_n = \bar{h}^{ij}_{\phantom{ij},j} - {H^i}_{jk} \bar{h}^{j,k}.$$ As for the covariance, $$\begin{gathered} \langle \theta^i \theta^j \rangle_n - \langle \theta^i \rangle_n \langle \theta^j \rangle_n = \biggl[ \frac{1}{A^2} \bigl\langle \theta^i_{(1)} \theta^j_{(1)} \bigr\rangle_n + \frac{1}{A^3} \bigl\langle \cancel{\theta^i_{(1)} \theta^j_{(2)}} + \cancel{\theta^i_{(2)} \theta^j_{(1)}} \bigr\rangle_n + \frac{1}{A^4} \bigl\langle \theta^i_{(1)} \theta^j_{(3)} + \theta^i_{(2)} \theta^j_{(2)} + \theta^i_{(3)} \theta^j_{(1)} \bigr\rangle_n + \cdots \biggr] \\ - \biggl[ \frac{1}{A^2} \bigl\langle \cancel{\theta^i_{(1)}} \rangle_n \langle \cancel{\theta^j_{(1)}} \bigr\rangle_n + \frac{1}{A^3} \Bigl( \bigl\langle \cancel{\theta^i_{(1)}} \bigr\rangle_n \langle \theta^j_{(2)} \bigr\rangle_n + \bigl\langle \theta^i_{(2)} \bigr\rangle_n \langle \cancel{\theta^j_{(1)}} \bigr\rangle_n \Bigr) \\ + \frac{1}{A^4} \Bigl( \bigl\langle \cancel{\theta^i_{(1)}} \bigr\rangle_n \langle \cancel{\theta^j_{(3)}} \bigr\rangle_n + \bigl\langle \theta^i_{(2)} \bigr\rangle_n \langle \theta^j_{(2)} \bigr\rangle_n + \bigl\langle \cancel{\theta^i_{(3)}} \bigr\rangle_n \langle \cancel{\theta^j_{(1)}} \bigr\rangle_n \Bigr) + \cdots \biggr], \label{eq:fourthcovariance}\end{gathered}$$ where all the stricken-through terms vanish because they are proportional to ensemble products of an odd number of $n$ terms. The surviving contributions are given by $$\begin{aligned} \bigl\langle \theta^i_{(1)} \theta^j_{(1)} \bigr\rangle_n =\, & \bigl\langle N^i N^j \bigr\rangle_n = \bar{h}^{i,j}, \\ \bigl\langle \theta^i_{(1)} \theta^j_{(3)} \bigr\rangle_n =\, & \bigl\langle N^i {N^j}_k {N^k}_l N^l \bigr\rangle_n - {H^k}_{lm} \bigl\langle N^i {N^j}_k N^l N^m \bigr\rangle_n + \bigl\langle N^i {N^j}_{kl} N^k N^l \bigr\rangle_n - {H^j}_{klm} \bigl\langle N^i N^k N^l N^m \bigr\rangle_n \\ & - {H^j}_{kl} \Bigl( \bigl\langle N^i N^k {N^l}_m N^m \bigr\rangle_n - {H^l}_{mq} \bigl\langle N^i N^k N^m N^q \bigr\rangle_n + \bigl\langle N^i {N^k}_m N^m N^l \bigr\rangle_n - {H^k}_{mq} \bigl\langle N^i N^m N^q N^l \bigr\rangle_n \Bigr), \\ \bigl\langle \theta^i_{(2)} \theta^j_{(2)} \bigr\rangle_n =\, & \bigl\langle {N^i}_k N^k {N^j}_m N^m \bigr\rangle_n - {H^i}_{kl} \bigl\langle N^k N^l {N^j}_m N^m \bigr\rangle_n - {H^j}_{mq} \bigl\langle N^m N^q {N^i}_k N^k \bigr\rangle_n + {H^i}_{kl} {H^j}_{mq} \bigl\langle N^k N^l N^m N^q \bigr\rangle_n, \end{aligned} \label{eq:lotsofn}$$ and of course $\langle \theta^i_{(3)} \theta^j_{(1)} \rangle_n = \langle \theta^j_{(1)} \theta^i_{(3)} \rangle_n$. The first of these equations reproduces the standard Fisher-matrix result. The four-$N$ products in Eq.  follow from Eq. . For instance, the last two products are given by $$\begin{aligned} \bigl\langle N^k N^l N^m N^q \bigr\rangle_n &= \bigl\langle N^k N^l \bigr\rangle_n \bigl\langle N^m N^q \bigr\rangle_n + \bigl\langle N^k N^m \bigr\rangle_n \bigl\langle N^l N^q \bigr\rangle_n + \bigl\langle N^k N^q \bigr\rangle_n \bigl\langle N^l N^m \bigr\rangle_n \\ & = \bar{h}^{k,l}\bar{h}^{m,q} + \bar{h}^{k,m}\bar{h}^{l,q} + \bar{h}^{k,q}\bar{h}^{l,m}, \\ \bigl\langle N^m N^q {N^i}_k N^k \bigr\rangle_n &= \bar{h}^{m,q} \, \bar{h}^{ik,}_{\phantom{ik,}k} + \bar{h}^{m,i}_{\phantom{m,i}k} \, \bar{h}^{q,k} + \bar{h}^{m,k} \, \bar{h}^{q,i}_{\phantom{q,i}k} . \end{aligned} \label{eq:nproducts}$$ These expressions can be substituted into those of Eq. , and those into Eq. , yielding the frequentist variance to order $1/A^4$. Unfortunately, this requires computing second- and third-order waveform derivatives (the latter for $H_{jklm}$). In the Bayesian framework {#sec:higherbayesian} ------------------------- To generalize Eq. , we write $$\mathcal{I}^{(n)} = \int \bar{\theta}^i \frac{\partial^n p(0)}{\partial \epsilon^n} d\theta, \quad \mathcal{N}^{(n)} = \int \frac{\partial^n p(0)}{\partial \epsilon^n} d\theta,$$ and find the recurrence relation $$\langle \bar{\theta}^i \rangle_p^{(n)} = \biggl(\mathcal{I}^{(n)} - \sum_{j=1}^n \biggl[ \biggl(\begin{array}{c} n \\ j \end{array} \biggr) \langle \bar{\theta}_i \rangle_p^{(n-j)} \times \mathcal{N}^{(j)} \biggr] \biggr) \Big/ \mathcal{N}^{(0)}, \label{eq:recurrence}$$ which we may prove by expanding the identity $\mathcal{I}^{(0)} + \epsilon \, \mathcal{I}^{(1)} + \frac{\epsilon^2}{2} \mathcal{I}^{(2)} + \cdots = [\langle \bar{\theta}^i \rangle_p^{(0)} + \epsilon \langle \bar{\theta}^i \rangle_p^{(1)} + \cdots] \times [\mathcal{N}^{(0)} + \epsilon \, \mathcal{N}^{(1)} + \cdots]$ on both sides as a series of $\epsilon$, leading to $$\mathcal{I}^{(n)} = \sum_{j=0}^n \biggl(\begin{array}{c} n \\ j \end{array} \biggr) \mathcal{N}^{(j)} \langle \bar{\theta}_i \rangle_p^{(n-j)},$$ whence Eq. . To obtain all needed derivatives with respect to $\epsilon$, we rewrite Eq.  as $$\begin{aligned} p(s|\theta) \propto \exp \Bigl\{ & -(n,n)/2 + \Bigl[ N_i \bar{\theta}^i + \epsilon N_{ij} \bar{\theta}^i \bar{\theta}^j + \epsilon^2 N_{ijk} \bar{\theta}^i \bar{\theta}^j \bar{\theta}^k + \cdots \Bigr] \\ & - \Bigl[ H_{jk} \bar{\theta}^j \bar{\theta}^k + \epsilon H'_{jkl} \bar{\theta}^j \bar{\theta}^k \bar{\theta}^l + \epsilon^2 H'_{jklm} \bar{\theta}^j \bar{\theta}^k \bar{\theta}^l \bar{\theta}^m + \cdots \Bigr] / 2 \Bigr\}, \end{aligned}$$ where $N_i = (n,\bar{h}_i) / 1! = n_i$, $N_{ij} = (\bar{h}_{ij},n) / 2! = n_{ij}/2$, $N_{ijk} = (\bar{h}_{ijk},n) / 3!$ (and so on), and also the $H'_{j_1{\cdots}j_n}$ have slightly different denominators than the $H_{j_1{\cdots}j_n}$ of Eq. : $$\begin{aligned} H'_{ijk} & = \frac{\bar{h}_{i,jk}}{1! \, 2!} + \frac{\bar{h}_{ij,k}}{2! \, 1!}, \\ H'_{ijkl} & = \frac{\bar{h}_{i,jkl}}{1! \, 3!} + \frac{\bar{h}_{ij,kl}}{2! \, 2!} + \frac{\bar{h}_{ijk,l}}{3! \, 1!}; \\ & \cdots \end{aligned}$$ namely, the denominator is $m! \, l!$ for the product $h_{j_1{\cdots}j_m,j_1{\cdots}j_l} \equiv (\bar{h}_{j_1{\cdots}j_m},\bar{h}_{j_1{\cdots}j_l})$. Expanding as a series of $\epsilon$ yields $$\begin{aligned} p(s|\theta) \propto e^{-H_{jk} \bar{\theta}^j \bar{\theta}^k / 2 + n_j \bar{\theta}^j} \times \biggl\{ 1 & + \epsilon \Bigl( N_{jk} \bar{\theta}^j \bar{\theta}^k - \frac{1}{2} H'_{jkl} \bar{\theta}^j \bar{\theta}^k \bar{\theta}^l \Bigr) \\ & + \frac{\epsilon^2}{2} \biggl( \Bigl( N_{jk} \bar{\theta}^j \bar{\theta}^k - \frac{1}{2} H'_{jkl} \bar{\theta}^j \bar{\theta}^k \bar{\theta}^l \Bigr)^2 + 2 \Bigl( N_{jkl} \bar{\theta}^j \bar{\theta}^k \bar{\theta}^l - \frac{1}{2} H'_{jklm} \bar{\theta}^j \bar{\theta}^k \bar{\theta}^l \bar{\theta}^m \Bigr) \biggr) + \cdots \biggr\}, \end{aligned}$$ so that the $\mathcal{I}^{(n)}$ and $\mathcal{N}^{(n)}$ are given by expressions akin to $$\mathcal{I}^{(1)} / \mathcal{N}^{(0)} = \int \bar{\theta}^i \Bigl( N_{jk} \bar{\theta}^j \bar{\theta}^k - \frac{1}{2} H'_{jkl} \bar{\theta}^j \bar{\theta}^k \bar{\theta}^l \Bigr) e^{-H_{jk} \bar{\theta}^j \bar{\theta}^k / 2 + n_j \bar{\theta}^j} d \theta \bigg/\!\! \int e^{-H_{jk} \bar{\theta}^j \bar{\theta}^k / 2 + n_j \bar{\theta}^j} d \theta.$$ Now, the integrals of the general form $$\langle \bar{\theta}^{i_1} \cdots \bar{\theta}^{i_m} \rangle_p^{(0)} = \int \bar{\theta}^{i_1} \cdots \bar{\theta}^{i_m} e^{-H_{jk} \bar{\theta}^j \bar{\theta}^k / 2 + n_j \bar{\theta}^j} d \theta \bigg/\!\! \int e^{-H_{jk} \bar{\theta}^j \bar{\theta}^k / 2 + n_j \bar{\theta}^j} d \theta, \label{eq:intgenform}$$ can be computed with the Wick identity[^22] [@zinnjustin] $$\bigl\langle F(\bar{\theta}) \bigr\rangle_p^{(0)} = F\biggl(\frac{\partial}{\partial n}\biggr) \exp \bigl\{n_i (H^{ij})^{-1} n_j / 2\bigr\};$$ in particular (again using $(H^{ij})^{-1}$ to raise indices), $$\begin{aligned} \langle \bar{\theta}^i \rangle^{(0)}_p & = n^i, \\ \langle \bar{\theta}^i \bar{\theta}^j \rangle^{(0)}_p & = (H^{ij})^{-1} + n^i n^j, \\ \langle \bar{\theta}^i \bar{\theta}^j \bar{\theta}^k \rangle^{(0)}_p & = (H^{ij})^{-1} n^k + (H^{ik})^{-1} n^j + n^i (H^{jk})^{-1} + n^i n^j n^k. \\ & \ldots \end{aligned}$$ Unfortunately, the $1/A^4$ (i.e., $\epsilon^2$) corrections to the variance turn out to be rather unwieldy, and belong in a symbolic-manipulation software package rather than on these pages. We content ourselves with the $1/A^2$ correction to the posterior mean (remember that the normalized parameters $\bar{\theta}$ carry an $A$), $$\langle \bar{\theta}^i \rangle_p = n^i + \epsilon \Bigl[ {n^i}_k n^k - \bigl( \tfrac{1}{2} \bar{h}^{i,}_{\phantom{i,}kl} + \bar{h}^{i}_{\phantom{i}k,l} \bigr) \bigl( n^k n^l + \bar{h}^{kl} \bigr) \Bigr] + O(\epsilon^2),$$ and the $1/A^3$ correction to the variance, $$\begin{gathered} \langle \bar{\theta}^i \bar{\theta}^j \rangle_p - \langle \bar{\theta}^i \rangle_p \langle \bar{\theta}^j \rangle_p = \bar{h}^{ij} + \epsilon \Bigl[ n^{ij} + \tfrac{1}{2} n^i n^j n_{kl} \bigl( n^k n^l - \bar{h}^{kl} \bigr) - n^k \bigl( \bar{h}^{i,j}_{\phantom{i,j}k} + \bar{h}^{j,i}_{\phantom{j,i}k} + \bar{h}_{k,}^{\phantom{k,}ij} \bigr) \\ - n^i n^j n^k \bigl( \tfrac{1}{2} \bar{h}_{k,l}^{\phantom{k,l}l} + \bar{h}^{l,}_{\phantom{l,}kl} + \tfrac{1}{2} \bar{h}_{k,lm} n^l n^m \bigr) \Bigr] + O(\epsilon^2).\end{gathered}$$ Thus we see that the $1/A^3$ contribution to the variance does not vanish in any single experiment (unless $n^i = 0$). It does vanish, however, under frequentist average, since it involves products of odd numbers of noises. Conclusion ========== In this article I tried to provide, as it were, a user’s manual for the Fisher information matrix. It seems clear that the Fisher-matrix formalism will continue to be featured prominently in research dealing with the parameter-estimation prospects of future GW observations, because of its compactness and accessibility, and because of the difficulty of computing higher-order corrections and running full-blown simulations. Yet the three questions posed in the introduction loom over the credibility of Fisher-matrix results, which is all the more worrisome when these results are used to justify choices in science policy or experiment design. The recipes provided in this paper to answer the initial questions can help assert (or falsify) the accuracy of the formalism for specific signal models. In particular: 1. As discussed in Sec. \[sec:disappearing\], ill-conditioned or singular Fisher matrices point to the need for increased numerical accuracy, and occasionally to a case for discarding a parameter or combination of parameters, but more often to suspicions about the appropriateness of the high-SNR/LSA limit. Section \[sec:singularfisher\] describes how to use the singular value decomposition of the Fisher matrix to discard truly degenerate linear combinations of parameters; Sec. \[sec:approxfisher\] describes how to roughly assess the sensitivity of the Fisher-matrix inverse to numerical error by means of the Fisher-matrix condition number, and more carefully by a simple Monte Carlo test. 2. The necessity of including prior distributions for the source parameters, perhaps in as simple a form as uniform distributions over the physically allowed ranges, can be roughly assessed by verifying whether Fisher-matrix results change with the addition of simple Gaussian priors, as shown in Sec. \[sec:priorexample\]; more accurate estimates of the effect of priors can be obtained by integrating the variance of an exact-prior–LSA-likelihood posterior with the simple Monte Carlo algorithm of Sec. \[sec:exactprior\]. 3. The detected-signal strength (i.e., the SNR) necessary for Fisher-matrix results to be internally consistent can be evaluated with the likelihood-mismatch criterion that follows from Eqs.  and of Sec. \[sec:howhigh\], or (at the price of some algebra) by computing the higher-order corrections presented in Sec. \[sec:higherorder\]. If the Fisher-matrix formalism remains inconsistent at the SNRs of interest, even with the help of priors, there is little recourse but to embark in explicit Monte Carlo simulations of frequentist [@bsd1996] or Bayesian [@mcmc] parameter estimation. Such simulations can consistently include sophisticated priors, and explore the secondary maxima of the posterior (or likelihood, in the frequentist case). They are the gold standard of this trade, but as such they are expensive in human effort and CPU resources. The recipes given in this paper can help establish when they are truly needed. I would like to thank Yanbei Chen, Curt Cutler, Yi Pan, and Michele Zanolin for useful discussions; for reviewing this manuscript, I am grateful to John Armstrong, Emanuele Berti, Steve Drasco, Frank Estabrook, Sam Finn, Éanna Flanagan, and especially Alessandra Buonanno. My work was supported by the LISA Mission Science Office and by the Human Resources Development Fund at the Jet Propulsion Laboratory, California Institute of Technology, where it was performed under contract with the National Aeronautics and Space Administration. Lemma for the conditional average of jointly normal random variables {#sec:conditional} ==================================================================== Assume the vector $x_j$ and the scalar $y$ are jointly normal random variables with mean zero and covariance matrix $$\mathcal{C} = \left( \begin{array}{cc} F_{ij} & H_{i} \\ H_{j} & G \end{array} \right). \label{eq:jointnormal}$$ From the standard Frobenius–Schur formula for the inverse of a block matrix [@bodewig], $$\label{eq:schur} \left( \begin{array}{cc} A & B \\ C & D \end{array} \right)^{-1} = \left( \begin{array}{cc} A^{-1} + A^{-1} B S^{-1}_A C A^{-1} & -A^{-1} B S^{-1}_A \\ -S^{-1}_A C A^{-1} & S^{-1}_A \end{array} \right)$$ (with $S_A = D - C A^{-1} B$ the *Schur complement* of $A$), we find $$\mathcal{C}^{-1} = \left( \begin{array}{cc} F^{-1}_{ij} + S^{-1}_A (F^{-1}_{ik} H_k) (F^{-1}_{jl} H_l) & -S^{-1}_A (F^{-1}_{jk} H_k) \\ -S^{-1}_A (F^{-1}_{ik} H_k) & S^{-1}_A \end{array} \right),$$ since in our case $F^{-1}_{ij}$ is symmetric and $S_A$ is the scalar $G - (F^{-1}_{ij} H_i H_j)$. Now, the joint distribution of $x_j$ and $y$ is given by $$p(x,y) \propto \exp - \left\{ \left( x_i \; y \right) \cdot \mathcal{C}^{-1} \cdot \left( \begin{array}{c} x_i \\ y \end{array} \right) \right\} / \, 2,$$ while the conditional distribution of $y$ given $x_j$ is $p(y|x) = p(x,y)/p(x) = p(x,y) / \left[\int p(x,y) \, dy\right]$. Since however $p(x)$ can be a function only of $x$, by the properties of Gaussian integrals it must be that $p(x) \propto \exp \, (\cdots)_{ij} x_i x_j$. It follows that $p(y|x)$ must be of the form $$p(y|x) \propto \exp - \left\{ S_A^{-1} y^2 - 2 S_A^{-1} (x_i F_{ij}^{-1} H_j) y + (\cdots)_{ij} x_i x_j \right\} / \, 2,$$ from which, by inspection, we conclude that $$\langle y \rangle_{x_i} = \int y \, p(y|x)\, dy = x_i F_{ij}^{-1} H_j$$ and that $$\mathrm{var}_{x_i} y = \int (y - \langle y \rangle_{x_i})^2 \, p(y|x) \, dx = S_A = G - F_{ij}^{-1} H_i H_j.$$ [99]{} E. T. Jaynes and G. L. Bretthorst (ed.), *Probability theory: the logic of science* (Cambridge University Press, Cambridge, 2003). See, e.g., N. Christensen and R. Meyer, Phys. Rev. D **58**, 082001 (1998); N. Christensen, R. J. Dupuis, G. Woan, and R. Meyer, Phys. Rev. D **70**, 022001 (2004). D. Nicholson and A. Vecchio, Phys. Rev. D **57**, 4588 (1998). L. S. Finn, Phys. Rev. D **46**, 5236 (1992). C. Cutler and É. E. Flanagan, Phys. Rev. D **49**, 2658 (1994). R. Balasubramanian, B. S. Sathyaprakash, and S. V. Dhurandhar, Phys. Rev. D **53**, 3033 (1996). R. Balasubramanian and S. V. Dhurandhar, Phys. Rev. D **57**, 3408 (1998). L. A. Wainstein and L. D. Zubakov, *Extraction of signals from noise* (Prentice-Hall, Englewood Cliffs, NJ, 1962). A. V. Oppenheim, A. S. Willsky, and I. T. Young, *Signals and systems* (Prentice-Hall, Englewood Cliffs, NJ, 1983). S. M. Kay, *Fundamentals of statistical signal processing: estimation theory* (Prentice-Hall, Englewood Cliffs, NJ, 1993). G. L. Bretthorst, *Bayesian spectrum analysis and parameter estimation* (Springer-Verlag, New York, 1988). A. Abramovici et al., [*Science*]{} [**256**]{}, 325 (1992). T. Damour, B. R. Iyer and B. S. Sathyaprakash, Phys. Rev. D **63**, 044023 (2001); [**66**]{}, 027502 (2002). E. Poisson and C. M. Will, Phys. Rev. D **52**, 848 (1995). K. G. Arun, B. R. Iyer, B. S. Sathyaprakash, and P. A. Sundararajan, Phys. Rev. D **71**, 084008 (2005); erratum, **72**, 069903 (2005). L. Blanchet, T. Damour, G. Esposito-Farèse, and B. R. Iyer, Phys. Rev. Lett. **93**, 091101 (2004). L. E. Kidder, C. M. Will, and A. G. Wiseman, Phys. Rev. D **47**, R4183 (1993). See, e.g., A. K. Gupta and D. K. Nagar, *Matrix Variate Distributions* (Chapman and Hall/CRC, Boca Raton, FL, 2000). A. Tarantola, *Inverse problem theory and methods for model parameter estimation* (SIAM, Philadelphia, PA, 2005). G. Golub and C. van Loan, *Matrix computations*, 3rd ed. (Johns Hopkins Univ. Press, London, 1996). E. Berti, A. Buonanno, and C. M. Will, Class. Quant. Grav. **22**, S943 (2005). C. Cutler (private communication). W. H. Press, B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling, *Numerical Recipes in C* (Cambridge University Press, Cambridge, 1988). N. Metropolis et al., J. Chem. Phys. **21**, 1087 (1953); W. K. Hastings, Biometrika **57**, 97 (1970). O. E. Barndorff-Nielsen and D. R. Cox, *Inference and Asymptotics* (Chapman and Hall, London, 1994). M. Zanolin, E. Naftali, and N. C. Makris, in preparation. E. Naftali and N. C. Makris, J. Acoust. Soc. Am. **110**, 1917 (2001); A. Thode, M. Zanolin, E. Naftali, I. Ingram, P. Ratilal, and N. C. Makris, *ibid.* **112**, 1890 (2002). L. Parker and S. M. Christensen, *MathTensor: A System for Doing Tensor Analysis by Computer* (Addison-Wesley, Reading, MA, 1994). J. Zinn-Justin, *Path Integrals in Quantum Mechanics* (Oxford University Press, Oxford, 2005). E. Bodewig, *Matrix Calculus* (North-Holland, Amsterdam, 1959). [^1]: In the high-SNR/LSA limit with negligible priors, the posterior probability mode, seen as a frequentist statistic, coincides with the ML estimator; thus its fluctuations are again described by the inverse Fisher matrix. [^2]: Only in the sense that subjects with different prior assumptions could come to different conclusions after seeing the same data; indeed, Bayesian statistics describes how prior assumptions become *deterministically* modified by the observation of data. [^3]: This assumption fails for some (mildly) pathological likelihood functions, which can provide counterexamples to the Cramér–Rao bound. [^4]: To obtain Eq. , we need to notice also that for any $w(s)$, $\langle d(s), \langle w(s) \rangle_n \rangle_n = 0$, since $\langle w(s) \rangle_n$ does not depend on $s$ (but only on $\theta_0$), and the integral of Eq.  reduces to $\langle w(s) \rangle_n \int \partial_{\theta_0} p(s|\theta_0) ds = \langle w(s) \rangle_n \partial_{\theta_0} 1 = 0$. [^5]: This is true even if we evaluate the performance of estimators on the basis of their quadratic error $$\langle (\hat{\theta}_i - \theta_{0i}) (\hat{\theta}_l - \theta_{0l}) \rangle_n \geq b_i(\theta_0) b_l(\theta_0) + \bigl(\delta_{im} + \partial_m b_i(\theta_0)\bigr) F^{-1}_{mj} \bigl(\delta_{jl} + \partial_j b_l(\theta_0)\bigr)$$ rather than on the basis of their variance. [^6]: Indeed, Eq.  implies that if both an *efficient* (i.e., bound-achieving) unbiased estimator and the ML estimator exist, they must coincide. To show this, we notice that if the ML estimator exists, the log-derivative $\partial_i \log p(s|\theta) = -\partial_i l_k(\theta) (\hat{\theta}_k - \theta_k)$ must be zero at $\theta = \hat{\theta}^\mathrm{ML}$, from which it follows that $\hat{\theta}_k = \hat{\theta}^\mathrm{ML}_k$. [^7]: The statistical uncertainty in the estimated signal strength can still be handled in this notation by taking one of the $\bar{h}_k$ to lie along $\bar{h}_0$; the corresponding $\theta_k$ represents a fractional correction to the true $A$. [^8]: Formally, it is troubling to truncate the series expression for the exponent at any order beyond quadratic, since the integral of the truncated likelihood may become infinite; the important thing to keep in mind, however, is that the series need converge only within a limited parameter range determined self-consistently by the truncated-likelihood estimator, by compact parameter ranges, or (in the Bayesian case) by parameter priors. Similar considerations apply to the derivation of the higher-order corrections given in Sec. \[sec:higherorder\]. [^9]: Note that the Fisher matrix that must be substituted into Eq.  is still $-\langle \partial_j \partial_k p(s|\theta) \rangle_n = (h_j,h_k)$, and not $-\langle \partial_j \partial_k [p(s|\theta) p(\theta)] \rangle_n = (h_j|h_k) + P_{jk}$. The prior distribution does not concern the Cramér–Rao bound, which is computed from the likelihood alone for a *fixed known* value of the true source parameters. Instead, we happen to be using an estimator that takes into in account prior information, which enters into the Cramér–Rao bound via the derivative of the bias. [^10]: For uniform priors (e.g., rectangular distributions corresponding to the allowed parameter ranges), $\hat{\theta}^\mathrm{MP}$ actually becomes undefined in the $A \rightarrow 0$ limit. [^11]: If we define the quadratic error of the posterior distribution as $\bigl< \bar{\theta}_i \bar{\theta}_j \bigr>_p$ (which is appropriate given that the true signal is at $\theta = 0$), we must increment $(\bar{h}_i,\bar{h}_j)^{-1}$ by the experiment-dependent quantity $\bigl< \bar{\theta}_i \bigr> \bigl< \bar{\theta}_j \bigr> = (\bar{h}_i,\bar{h}_l)^{-1} (n,\bar{h}_l) (\bar{h}_m,n) (\bar{h}_m,\bar{h}_j)^{-1}$. Interestingly, the frequentist average of the Bayesian error $\bigl< \bigl< \bar{\theta}_i \bar{\theta}_j \bigr>_p \bigr>_n$ is $2 (\bar{h}_i,\bar{h}_j)^{-1}$, twice the frequentist variance of $\hat{\theta}^\mathrm{ML}$. [^12]: With the Gaussian prior, the quadratic error $\bigl< \bigl< \bar{\theta}_i \bar{\theta}_j \bigr>_p \bigr>_n$ becomes $[(\bar{h}_i,\bar{h}_l) + P_{il}/A^2]^{-1} (\bar{h}_l,\bar{h}_m) [(\bar{h}_m,\bar{h}_j) + P_{mj}/A^2]^{-1}$. [^13]: Waveform phasing expressions accurate to 3.5PN order are also provided in Ref. [@arun2005]. We do not use these in this article for the sake of simplicity, since they would not change the qualitative picture of parameter estimation presented here. For the reader’s reference, however, the higher-than-2PN corrections to Eq. , including the errata to Ref. [@arun2005], are $$\begin{gathered} \pi \biggl(\frac{38645}{252} - \frac{65}{9}\eta^3\biggr) v^5 \log v + \biggl[ \biggl(\frac{11 583 231 236 531}{4 694 215 680} - \frac{640}{3} \pi^2 - \frac{6 848}{21} \gamma - \frac{6 848}{21}\log(4) \biggr) + \biggl(-\frac{15 335 597 827}{3 048 192} + \frac{2 255}{12}\pi^2 \\ - \frac{1 760}{3} \theta + \frac{12 320}{9} \lambda\biggr) \eta + \frac{76 055}{1 728}\eta^2 - \frac{127 825}{1 296}\eta^3 \biggr] v^6 -\frac{6 848}{21} v^6 \log v + \pi \biggl(\frac{77 096 675}{254 016} + \frac{378 515}{1 512}\eta - \frac{74 045}{756}\eta^2\biggr) v^7, \nonumber\end{gathered}$$ where $\gamma = 0.57721\cdots$ is Euler’s constant, and $\lambda = -1987/3080$ and $\theta = -11831/9240$ are recently determined constants in the PN expansion [@blanchet2004]. [^14]: \[note:coordinates\]A reasonable objection to computing the eigensystem of the Fisher matrix is that it leads to taking linear combinations of parameters that may have different units. It is possible to avoid this problem by looking at the Fisher matrix more abstractly as a linear operator, and talking of its range and null space [@tarantola]; or more pragmatically, by dividing all parameters by their typical range; or perhaps by taking their logarithm (since we are working with errors, units can be forgotten as additive constants), which in the linearized theory is equivalent to dividing by the true parameters. We are going to largely ignore this issue, treating the parameters as pure numbers resulting from adopting a God-given system of units. [^15]: For square matrices, the SV decomposition is essentially equivalent to an eigenvector–diagonal-matrix decomposition where we drop the rows and columns corresponding to the null eigenvalues and eigenvectors. [^16]: In a truly linear system, this is true no matter how small the eigenvector component in that parameter direction; clearly, this raises a problem of accuracy in the numerical computation of eigenvectors. [^17]: Even a single prior in the form of a rectangle function will regularize the integration over all the null-eigenvector coordinates that include that parameter. For normal priors, whether the posterior becomes integrable depends on the eigenstructure of $A^2 (\bar{h}_i,\bar{h}_j) + P_{ij}$. [^18]: Although neither of these examples is a linear model described exactly by the LSA, the degeneracy persists in the exact likelihood, so its Fisher-matrix diagnosis is correct. For such “perfect” degeneracies to occur, the two parameters must appear in all waveform expressions only as a sum or product; this would imply that their units can be sensibly summed, or that their combination has direct physical meaning. [^19]: Augmenting the Fisher matrix with normal priors for $\eta$, $\beta$, and $\sigma$, as described in Sec. \[sec:priorexample\], can somewhat cure this instability to inversion, although the result is SNR-dependent: for $\mathrm{SNR}=10$, the errors in $(F_{ij} + P_{ij})^{-1}$ become intolerable for fractional perturbations in $F_{ij}$ of order $10^{-7}$ rather than $10^{-12}$, but in the high-SNR limit the threshold reverts to the latter. [^20]: In particular, $\phi_0$ is strongly coupled to $t_0$, which produces frequency-dependent phase shifts through the exponential $\exp(2 \pi i f t_0)$. Adopting the new phase parameter $\phi'_0 = \phi_0 + 2 \pi f_0 t_0$, where $f_0$ is the *dominant* frequency at which $f^{-7/3} / S_n(f)$ is maximum, largely removes this coupling [@cpc2007]. [^21]: To see why this is the right thing to do, consider the integration of a function against the LSA distribution, and make a change of variables (with unit Jacobian) to eigenvalue components, and a second to rescaled components $\tilde{c}^{(i)} = A \sqrt{\lambda^{(i)}} c^{(i)}$: $$\int (\ldots) \, e^{-A^2 \theta_i \theta_j (\bar{h}_i,\bar{h}_j) / 2} d \theta = \int (\ldots) \, e^{-A^2 \sum_{(i)} \lambda^{(i)} [c^{(i)}]^2 / 2} d c \propto \int (\ldots) \, e^{- \sum_{(i)} [\tilde{c}^{(i)}]^2 / 2} d \tilde{c};$$ we see that the source parameters that correspond to $\tilde{c}$ lying on a sphere of fixed radius must lie on an isoprobability surface. To reassemble $\theta$ from the $\tilde{c}$, we need to *divide* the eigenvectors by $A \sqrt{\lambda^{(i)}}$. [^22]: Another way to organize this computation is to offset the integration variable $\bar{\theta}^j$ to $\bar{\theta}^j - (H^{jk})^{-1} n_k = \bar{\theta}^j - n^j$ in Eq. , obtaining $$\langle \bar{\theta}^{i_1} \cdots \bar{\theta}^{i_m} \rangle_p^{(0)} = \int (\bar{\theta}^{i_1} + n^{i_1}) \cdots (\bar{\theta}^{i_m} + n^{i_m}) e^{-H_{jk} \bar{\theta}^j \bar{\theta}^k / 2} d \theta \bigg/\!\! \int e^{-H_{jk} \bar{\theta}^j \bar{\theta}^k / 2} d \theta;$$ we can then expand the product in the integrand, bring the $n^{i_k}$ outside the integral, and apply Wick’s theorem \[Eq. \] to obtain each addend of the form $$n^{i_1} \cdots n^{i_{m-l}} \int \bar{\theta}^{i_1} \cdots \bar{\theta}^{i_l} e^{-H_{jk} \bar{\theta}^j \bar{\theta}^k / 2} d \theta \bigg/\!\! \int e^{-H_{jk} \bar{\theta}^j \bar{\theta}^k / 2} d \theta;$$ all integrals with odd $l$ are zero, while the integrals with even $l$ are given by the sum of all possible pairings of indices into products of $(H^{\cdots})^{-1}$.

ArXiv

--- abstract: 'It is often stated that heat baths with finite degrees of freedom i.e., finite baths, are sources of Tsallis distributions for the classical Hamiltonian systems. By using well-known fundamental statistical mechanical expressions, we rigorously show that Tsallis distributions with fat tails are possible *only* for finite baths with constant negative heat capacity while constant positive heat capacity finite baths yield decays with sharp cut-off with no fat tails. However, the correspondence between Tsallis distributions and finite baths holds at the expense of violating equipartition theorem for finite classical systems at equilibrium. Finally, we comment on the implications of the finite bath for the recent attempts towards a $q$-generalized central limit theorem.' author: - 'G. Baris Bagci' - Thomas Oikonomou title: 'Do Tsallis distributions really originate from the finite baths?' --- [^1] introduction ============ Recently, a considerably great deal of effort has been put into the construction of a non-extensive thermostatistics based on the Tsallis entropies $S_{q}(p)=\frac{\int d\Gamma p^{q}(\Gamma)-1}{1-q}$ where $\Gamma$ and $p(\Gamma)$ denote the phase space variables and the probability distribution, respectively [@Tsallis1]. The Tsallis $q$-distributions are obtained from the maximization of the Tsallis entropies either by the internal energy $E$ calculated from $E=\int d\Gamma p\left( \Gamma \right) H$ or $E=\frac{\int d\Gamma p^{q}\left( \Gamma \right) H}{\int d\Gamma p^{q}\left( \Gamma \right) }$ where $H$ is the Hamiltonian. The former ones are called as ordinary Tsallis distributions and of the form $1/\exp_q (\gamma_q H)$ ($\gamma_q$ being a positive constant) apart from normalization, where $\exp_q(x)=\big[1+(1-q)x\big]_+^{\frac{1}{1-q}}$ with $[x]_+=\max\{0,x\}$. The latter ones are escort Tsallis distributions of the form $\exp_q (- \gamma_q H)$ omitting the normalization constant. Despite these efforts, however, the true origin of Tsallis distributions in statistical mechanics is elusive. On the other hand, there seems to be an agreement on an important and very intuitive statistical mechanical source of Tsallis distributions, namely, heat baths with finite number of degrees of freedom, simply called finite baths [@Tsallis1; @Plastino1; @Almeida1; @Almeida2; @Campisi1; @Campisi2; @Biro]. According to this view, a system coupled to a finite heat bath attains an inverse power law distribution in the form of $q$-exponential decays for any of the two branches of the non-extensivity parameter $q$, $q>1$ and $q<1$. The heat capacity of the bath can be found as $\frac{1}{1-q}$ or $\frac{1}{q-1}$, depending on which Tsallis distributions are used i.e., ordinary or escort. Since the range of admissible $q$ values can be both above and below the value $q=1$, the heat bath can have both positive and negative heat capacity values. Despite seeming counterintuitive at first, the possibility of systems with negative heat capacity was first pointed out by Lynden-Bell and Wood [@Bell] , later investigated theoretically by Thirring who showed that microcanonical ensembles can in fact have negative specific heats [@Th]. Recently, negative heat capacity expressions have also been found in one dimensional evaporation models and long-range quantum spin systems in optical lattices treated as microcanonical ensembles [@Hilbert; @Kastner]. The experimental measurements of the negative microcanonical heat capacity have been carried out with small clusters of sodium atoms near the solid to liquid transition and liquid to gas transition for the cluster of hydrogen ions [@Schmidt; @Gobet]. As expected, when $q=1$, $q$-exponential decays become ordinary exponential distributions and simultaneously the heat bath attains infinite heat capacity [@Almeida1; @Almeida2]. In order to obtain the aforementioned results, one assumes that a subsystem (simply called system from now on) embedded in a finite heat bath interacts weakly, and together forms the total system. Then, the total system can be treated microcanonically. In treating subsystems in contact with an infinite reservoir, one can either choose the phase space surface $\Omega$ or volume $\Phi$ as the appropriate measure without loss of generality, since these two yield the same results in the thermodynamic limit [@Khinchin]. However, when dealing with systems of finite degrees of freedom, one should consider two phase space measures, separately, since these two measures might yield different results. Therefore, a rigorous study of the finite baths must take into account both ordinary/escort Tsallis distributions and phase space volume/surface demarcation into account. The paper is organized as follows: In the next section, we outline the general microcanonical approach that will be used throughout the manuscript. In Section III, we present the results by considering all possibilities i.e., ordinary/escort distributions as well as phase space volume/surface demarcations together with the positivity or negativity of the total energy. Finally, concluding remarks are presented. Microcanonical approach ======================= We begin by considering a system weakly coupled to another one acting as a (finite) bath so that the total Hamiltonian is assumed to be ergodic. Despite the presence of the interaction between the system and bath, the total system i.e., system plus bath, is assumed to be isolated. The Hamiltonian of the total system is given as $$\label{totalHam} H_\text{tot}(\x,\p,\X,\P)=H_{\text{\tiny{S}}}(\x,\p)+H_{\text{\tiny{B}}}(\X,\P) +h(\x,\X)\,,$$ where $H_\text{\tiny{S}}(\x, \p)$ and $H_\text{\tiny{B}}(\X,\P)$ denote the system and bath Hamiltonians, respectively, interacting with one another through the interaction term $h(\x,\X)$. The system and bath phase space coordinates are respectively given by $\x=\{\x_1,\ldots,\x_{N_\text{\tiny{S}}}\}$, $\p=\{\p_1,\ldots,\p_{N_\text{\tiny{S}}}\}$, $\X=\{\X_1,\ldots,\X_{N_\text{\tiny{B}}}\}$ $\P=\{\P_1,\ldots,\P_{N_\text{\tiny{B}}}\}$ where $\{\x_i,\p_i\}_{i=1}^{N_\text{\tiny{S}}},\{\X_i,\P_i\}_{i=1}^{N_\text{\tiny{B}}} \in\mathbb{R}^{D}$, $D$ being the dimensionality of the space. The system and bath Hamiltonians in particular read $$\label{HamS} H_\text{\tiny{S}}(\x,\p)=\sum_{i=1}^{N_\text{\tiny{S}}}\dfrac{\p_i^2}{2m}+ \mathcal{V}_\text{\tiny{S}}(\x)\,,\qquad\qquad H_\text{\tiny{B}}(\X,\P)=\sum_{i=1}^{N_\text{\tiny{B}}}\dfrac{\P_i^2}{2M}+ \mathcal{V}_\text{\tiny{B}}(\X)\,,$$ where $\mathcal{V}_\text{\tiny{S}}(\x)$ and $\mathcal{V}_\text{\tiny{B}}(\X)$ are the interactions within the system and bath. From here on, the positivity of the system energy $E_\text{\tiny{S}}$ is assumed. Moreover, the Boltzmann constant is set equal to the dimensionless unity so that the temperature has the same dimension as energy. The marginal probability distribution $ p(\x,\p)$ of finding the system $S$ in a particular state with positive energy $E_{\text{\tiny{S}}}$ reads [@Khinchin] $$\label{MargProb0} p( \x,\p) = \frac{\Omega _{\text{\tiny{B}}} \big( E_\text{tot}-H_{\text{\tiny{S}}}( \x, \p) \big) }{\Omega _\text{tot}( E_\text{tot} ) }=c\,\Omega _{\text{\tiny{B}}}\big( E_\text{tot}-H_{\text{\tiny{S}}}( \x, \p ) \big)\,,$$ where $c$ is the normalization constant given by the inverse of the density of states of the total system i.e. $$\label{tot} \Omega _\text{tot}\left( E_\text{tot}\right) =\int \text{d} \x\, \text{d} \p\, \text{d} \X \,\text{d} \P \,\delta \big( E_\text{tot}-H_\text{tot}\left( \x,\p, \X,\P \right)\big)\,.$$ $\delta$ denotes Dirac delta function. Similarly, the density of states of the bath is given by $$\label{bath} \Omega _{\text{\tiny{B}}}\left( E_{\text{\tiny{B}}} \right) =\int \text{d}\X\,\text{d}\P\, \delta \big( E_{\text{\tiny{B}}} -H_{\text{\tiny{B}}}\left( \X,\P \right) \big),$$ where $E_{\text{\tiny{B}}}$ is the energy of the finite bath. Before proceeding, we further assume that the finite bath has constant heat capacity i.e. $C_{\text{\tiny{B}}}= \frac{E_{\text{\tiny{B}}}}{T_{\text{\tiny{B}}}}$, implying in particular that the heat bath is either composed of finite number of non-interacting particles or particles interacting through linear harmonic potential. However, the heat capacity of the system, in full generality, is given by the expression $C_{\text{\tiny{S}}}(T_\text{\tiny{S}})=\frac{\partial E_{\text{\tiny{S}}}(T_\text{\tiny{S}})} {\partial T_{\text{\tiny{S}}}}$, $T_{\text{\tiny{S}}}$ being the temperature of the system. The system temperature becomes equal to that of the bath only when the heat capacity of the bath is infinite. The probability distribution $p(\x,\p)$ of the system in Eq. , due to the expression of the density of states of the bath in Eq. , involves Dirac delta function which is even, i.e., $\delta( x )=\delta ( - x )$. Therefore, Dirac delta function enforces two distinct cases: either $E_{\text{\tiny{tot}}}-H_{\text{\tiny{S}}}\geq0$ or $E_{\text{\tiny{tot}}}-H_{\text{\tiny{S}}}<0$. Since the marginal system distribution is finally obtained by identifying $E_{\text{\tiny{B}}} = E_{\text{\tiny{tot}}}-H_{\text{\tiny{S}}}$, one must consider both cases of the finite bath possessing constant positive and negative energies, or due to the relation $C_\text{\tiny{B}}= \frac{E_\text{\tiny{B}}}{T_{\text{\tiny{B}}}}$, constant positive or negative finite heat capacities. It is important to understand the constraints imposed on the system probability distribution due to these two possibilities: considering the case $E_{\text{\tiny{tot}}}-H_{\text{\tiny{S}}}\geq0$, one can see that the system energy at most can be equal to $E_{\text{\tiny{tot}}}$. This in turn implies that the probability distribution of the system in weak contact with a finite bath possessing positive heat capacity must have a cut-off at $H_\text{\tiny{S}}=E_{\text{\tiny{tot}}}$. This constraint excludes the possibility of system distribution having fat tails. Therefore, if fat power-law tails should ever emerge in this context, this must be the case when the constant heat capacity of the finite bath (or equivalently the heat bath energy $E_\text{B}$) is negative, i.e., $E_{\text{\tiny{tot}}}-H_{\text{\tiny{S}}}<0$. In order to proceed further, one must have an explicit expression for the density of states of the bath composed of finite classical particles. This expression can be shown to have the form $\Omega _{\text{\tiny{B}}}\left( E_\text{\tiny{B}} \right) \sim {\left| E_\text{\tiny{B}} \right|}^{k}$ with exponent $k$ apart from some multiplicative positive constant [@Almeida1; @Almeida2; @Campisi3]. The absolute value is needed to ensure the positivity of the density of states. Calculating the temperature of the finite bath with $T_{\text{\tiny{B}}}^{-1} = \frac{\partial \ln (\Omega _{\text{\tiny{B}}})}{\partial E_{\text{\tiny{B}}}}$ and comparing it with the expression $C_{\text{\tiny{B}}}= \frac{E_{\text{\tiny{B}}}}{T_{\text{\tiny{B}}}}$, we see that the exponent $k$ is equal to the finite heat capacity of the bath $C_{\text{\tiny{B}}}$ so that $$\label{particularbath} \Omega _{\text{\tiny{B}}}\left( E_{\text{\tiny{B}}} \right) \sim {\left| E_{\text{\tiny{B}}} \right|}^{C_{\text{\tiny{B}}}}\,,$$ where the finite heat bath capacity can be positive (i.e. $E_{\text{\tiny{tot}}}-H_{\text{\tiny{S}}}\geq0$) or negative (i.e. $E_{\text{\tiny{tot}}}-H_\text{\tiny{S}}<0$). At this point, we also remind the reader an important fact: one should consider both the density of states $\Omega$ and the volume of the phase space $\Phi$, since these two measures might yield different results when dealing with systems of finite degrees of freedom although they are equivalent in the thermodynamic limit [@Dunkel]. The two measures are related to one another by $$\label{twomeasures} \Omega \left( E\right) =\frac{\partial\, \Phi( E) }{\partial E}\,.$$ The equation above yields $\Phi_{\text{\tiny{B}}} \left( E_{\text{\tiny{B}}} \right) \sim {\left| E_{\text{\tiny{B}}} \right|}^{C_{\text{\tiny{B}}}+1} $ for a constant heat capacity bath. Therefore, by using $T_{\text{\tiny{B}}}^{-1}=\frac{\partial \ln (\Phi _{\text{\tiny{B}}})}{\partial E_{\text{\tiny{B}}}}$ and comparing with the expression $C_{\text{\tiny{B}}}= \frac{E_{\text{\tiny{B}}}}{T_{\text{\tiny{B}}}}$, we obtain an important relation $$\label{relation} C_{\text{\tiny{B}}}^{\Omega }+1=C_{\text{\tiny{B}}}^{\Phi },$$ which is valid as long as the heat capacity of the finite bath is constant. The superscripts denote the quantities calculated through the density of states $\Omega\equiv\Omega_{\text{\tiny{B}}}$ or phase space volume $\Phi\equiv\Phi_{\text{\tiny{B}}}$. Finite bath and Tsallis distributions ===================================== Having outlined the general microcanonical approach to the totality of the system plus bath, we now explore the constant positive and negative heat capacity possibilities distinctly: 0.2cm #### -0.3cm Case 0.1cm $E_{\text{\tiny{B}}}\geq0$ : This case i.e., $E_{\text{\tiny{tot}}}-H_\text{\tiny{S}}\geq0$, corresponds to the marginal system distribution stemming from the identification $E_\text{\tiny{B}} = E_{\text{\tiny{tot}}}-H_\text{\tiny{S}}\geq0$. In other words, the (finite) system is now coupled to a finite bath with positive energy and therefore constant positive heat capacity $C_\text{\tiny{B}}$. Therefore, the system energy can attain at most the value $E_\text{\tiny{tot}}$, for which a necessary cut-off condition has to be respected in the probability distribution of the system. In fact, using Eqs. and , we obtain the marginal probability distribution of the system $$\label{positive1} p( \x, \p ) = c \big( E_\text{\tiny{tot}}-H_\text{\tiny{S}}(\x,\p) \big)^{C_\text{\tiny{B}}^{\Omega }}\,,$$ where $c$ is the normalization constant. This distribution can be cast into the form of escort $q$-exponential by identifying $\alpha_q^{-1}:=(1-q)E_\text{\tiny{tot}}$ and $$\label{positive2} C_\text{\tiny{B}}^{\Omega }=\frac{1}{1-q}$$ so that $$\label{positive3} p( \x,\p) \sim \exp_q\big(-\alpha_q\, H_\text{\tiny{S}}( \x,\p) \big)$$ apart from the normalization. Considering consistently the conditions above together with $\alpha_q>0$, we obtain the following range of validity for the non-extensivity index $q$ $$\label{positive4} \left.\begin{array}{l} C_\text{\tiny{B}}^{\Omega }T^\Omega_\text{\tiny{B}}+E_\text{\tiny{S}}>0 \\ E_\text{\tiny{S}}\;\,>0 \\ C_\text{\tiny{B}}^{\Omega }T^\Omega_\text{\tiny{B}}\;>0 \\ \alpha_q\;\;\; >0 \\ \end{array}\right\}\qquad\Longrightarrow\qquad \left.\begin{array}{l} \frac{T^\Omega_\text{\tiny{B}}}{1-q}+E_\text{\tiny{S}}>0 \\ E_\text{\tiny{S}}\;\,>0 \\ \frac{T^\Omega_\text{\tiny{B}}}{1-q}\;>0 \\ \frac{1}{1-q} >0 \\ \end{array}\right\}\qquad\Longrightarrow\qquad \boxed{q<1}\,,$$ with $T^\Omega_\text{\tiny{B}}>0$. Accordingly, the distribution in Eq. for $q<1$ represents a sharp decay with a cut-off at $E_\text{\tiny{tot}}=H_\text{\tiny{S}}$ in the argument excluding the possibility of fat tails. On the other hand, in terms of the ordinary $q$-exponential distributions, Eq. can be rewritten as $$\label{positive5} p( \x, \p ) \sim \frac{1}{\exp_q\big(\epsilon_q\, H_\text{\tiny{S}}( \x, \p)\big)}$$ with $\epsilon_q^{-1}:=(q-1)E_\text{\tiny{tot}}$ and $$\label{positiveness1} C_\text{\tiny{B}}^{\Omega }=\frac{1}{q-1}.$$ With the condition $\epsilon_q>0$ now, the following range of validity for the index $q$ is found $$\label{positive6} \left.\begin{array}{l} C_\text{\tiny{B}}^{\Omega }T^\Omega_\text{\tiny{B}}+E_\text{\tiny{S}}>0 \\ E_\text{\tiny{S}}\;\,>0 \\ C_\text{\tiny{B}}^{\Omega }T^\Omega_\text{\tiny{B}}\;>0 \\ \epsilon_q\;\;\; >0 \\ \end{array}\right\}\qquad\Longrightarrow\qquad \left.\begin{array}{l} \frac{T^\Omega_\text{\tiny{B}}}{q-1}+E_\text{\tiny{S}}>0 \\ E_\text{\tiny{S}}\;\,>0 \\ \frac{T^\Omega_\text{\tiny{B}}}{q-1}\;>0 \\ \frac{1}{q-1} >0 \\ \end{array}\right\}\qquad\Longrightarrow\qquad \boxed{q>1}\,,$$ with $T^\Omega_\text{\tiny{B}}>0$. The distribution in Eq. for $q>1$ again represents a sharp decay with a cut-off at $H_\text{\tiny{S}}=E_\text{\tiny{tot}}$ in the argument excluding the possibility of fat tails. These results show that the adoption of ordinary or escort $q$-distributions does not change the form of the probability distribution of the system. The choice between the two aforementioned distributions is only related to the intervals of $q$ values, since they are related to one another through the relation $\left(2-q\right)$ (for more details on this issue, see Refs. [@us1; @us2]). Note also that the same calculations above can be redone in terms of $C_\text{\tiny{B}}^{\Phi }$ by using Eq. , but it can be observed that the adoption of $C_\text{\tiny{B}}^{\Phi }$ does not change the shape of the system distribution. Regarding the related intervals in Eqs. and , they change to $q<1 \wedge q>2$ and $q<0 \wedge q>1$, respectively. #### -0.3cm Case 0.1cm $E_{\text{\tiny{B}}}<0$ : This case implies a finite bath with constant negative heat capacity, since now $E_\text{\tiny{B}} = E_\text{\tiny{tot}}-H_\text{\tiny{S}}<0$. To ensure the negativity of the finite bath energy, $E_\text{\tiny{tot}}$ must always be negative so that the system distribution does not have a cut-off now. Using again Eqs. and , the marginal probability distribution of the system reads $$\label{negative1} p(\x, \p) = c\, \big( H_\text{\tiny{S}}(\x,\p) -E_\text{\tiny{tot}} \big)^{C_\text{\tiny{B}}^{\Omega }}\,,$$ where $c$ is the normalization constant. This distribution can be cast into the form of escort $q$-exponential by identifying $$\label{negative2} C_\text{\tiny{B}}^{\Omega }=\frac{1}{1-q}$$ so that $$\label{negative3} p(\x, \p) \sim \exp_q\big(-\alpha_q\, H_\text{\tiny{S}}( \x, \p) \big)$$ apart from the normalization. Considering consistently the conditions above together with $\alpha_q>0$, we obtain the following range of validity for the non-extensivity index $q$ $$\label{negative4} \left.\begin{array}{l} C_\text{\tiny{B}}^{\Omega }T^\Omega_\text{\tiny{B}}+E_\text{\tiny{S}}<0 \\ E_\text{\tiny{S}}\;\,>0 \\ C_\text{\tiny{B}}^{\Omega }T^\Omega_\text{\tiny{B}}\;<0 \\ \alpha_q\;\;\; >0 \\ \end{array}\right\}\qquad\Longrightarrow\qquad \left.\begin{array}{l} \frac{T^\Omega_\text{\tiny{B}}}{1-q}+E_\text{\tiny{S}}<0 \\ E_\text{\tiny{S}}\;\,>0 \\ \frac{T^\Omega_\text{\tiny{B}}}{1-q}\;<0 \\ \frac{1}{1-q} <0 \\ \end{array}\right\}\qquad\Longrightarrow\qquad \boxed{1<q<1+\Big(\frac{E_\text{\tiny{S}}}{T_\text{\tiny{B}}}\Big)^{-1}} \,,$$ with $T^\Omega_\text{\tiny{B}}>0$. As can be seen the term $\frac{E_\text{\tiny{S}}}{T_\text{\tiny{B}}}$ has the dimension of a heat capacity. Accordingly, the distribution in Eq. represents an inverse power-law decay with fat tails. This is the result if one agrees to obtain temperature and constant heat capacity of the finite bath in terms of the density of states $\Omega$. In terms of the phase space volume $\Phi$, using Eq. , we obtain $C^\Phi_\text{\tiny{B}}=\frac{2-q}{1-q}$ so that Eq. yields $1<q<1+(\frac{E_\text{\tiny{S}}}{T_\text{\tiny{B}}}+1)^{-1}$. This result again indicates that the system distribution is an inverse power-law with fat tails although in the thermodynamic limit ($E_\text{\tiny{S}}\rightarrow \infty$ so that $q$ is confined to the unique value of unity implying $C_\text{\tiny{B}}^{\Omega}\rightarrow - \infty$), these $q$-decays are replaced by the usual exponential distribution. It is worth noting that a particular case of the above general result was also obtained by Lutsko and Boon [@Lutsko] *solely* by considering the integration over momenta degrees of freedom so that $E_\text{\tiny{S}}/T_\text{\tiny{B}} = DN_\text{\tiny{S}} / 2$ (compare Eq. above to the one below Eq. (10) in Ref. [@Lutsko]) where $D$ and $N_\text{\tiny{S}}$ denote the dimensionality of the phase space and the number of particles in the system, respectively. It is indeed very remarkable that Lutsko and Boon obtained this particular result just by checking the integrability conditions of the concomitant Tsallis distributions [@Lutsko]. However, the results of Lutsko and Boon include neither the finiteness of the bath nor the necessity of its negative heat capacity for the escort $q$-distribution with fat tails to emerge. Due to the generality of the present calculations, one can also consider the influence of including other degrees of freedom on the interval of validity of the non-extensive parameter $q$: the more the degrees of freedom associated with the heat capacity of the system are, the more confined is the interval of the possible $q$ values. As such, the thermodynamic limit, for any system, corresponds to the unique value of the non-extensivity parameter $q$ i.e. unity, corresponding to the ordinary canonical case. On the other hand, in terms of the ordinary $q$-exponential distributions, Eq. can be rewritten as $$\label{negative6} p( \x, \p ) \sim \frac{1}{\exp_q\big(\epsilon_q\, H_\text{\tiny{S}}( \x,\p)\big)}$$ with $$\label{negativeness1} C_\text{\tiny{B}}^{\Omega }=\frac{1}{q-1}.$$ With the condition $\epsilon_q>0$ now, the following range of validity for the index $q$ is found $$\label{negative7} \left.\begin{array}{l} C_\text{\tiny{B}}^{\Omega }T^\Omega_\text{\tiny{B}}+E_\text{\tiny{S}}<0 \\ E_\text{\tiny{S}}\;\,>0 \\ C_\text{\tiny{B}}^{\Omega }T^\Omega_\text{\tiny{B}}\;<0 \\ \epsilon_q\;\;\; >0 \\ \end{array}\right\}\qquad\Longrightarrow\qquad \left.\begin{array}{l} \frac{T^\Omega_\text{\tiny{B}}}{q-1}+E_\text{\tiny{S}}<0 \\ E_\text{\tiny{S}}\;\,>0 \\ \frac{T^\Omega_\text{\tiny{B}}}{q-1}\;<0 \\ \frac{1}{q-1} <0 \\ \end{array}\right\}\qquad\Longrightarrow\qquad \boxed{1-\Big(\frac{E_\text{\tiny{S}}}{T_\text{\tiny{B}}}\Big)^{-1}<q<1}\,.$$ The distribution in Eq. represents an inverse power-law decay with fat tails by adopting the density of states $\Omega$. In terms of the phase space volume $\Phi$, using Eq. , we have $C_\text{\tiny{B}}^{\Phi }=\frac{q}{q-1}$ so that $1-(\frac{E_\text{\tiny{S}}}{T_\text{\tiny{B}}}+1)^{-1}<q<1$. As expected, when the thermodynamic limit is attained, i.e., $E_\text{\tiny{S}}\rightarrow \infty$, the value of $q$ assumes only unity, resulting an usual exponential decay. Finally, we note that all the results in this Section rely on one main ingredient: the dependence of the finite heat capacity of the bath on the non-extensivity parameter $q$ (check Eqs. , , and for example). In fact, this is the sole source for the emergence of the parameter $q$. However, the bath has constant (positive or negative) heat capacity which implies that it is composed of either finite number of non-interacting particles or particles coupled with linear harmonic interaction. In accordance with the equipartition theorem then, the constant heat capacity of the finite bath is only proportional to the degrees of freedom composing the bath with no explicit dependence on $q$. Considering further that the equipartition theorem is intact even for non-extensive systems [@Biro2], it is apparent that the non-extensivity parameter $q$ is merely originating from a substitution for the finite heat capacity of the bath as opposed to a possible genuine non-extensivity in the bath. In short, one indeed has inverse power-law distributions due to the finiteness of the heat capacity of the bath. However, these inverse power-law distributions are not Tsallis distributions.\ Conclusions =========== There has been a general consensus so far on relating finite baths to the Tsallis distributions. According to this view, the finite baths can have both positive and negative heat capacities depending on the use of ordinary and escort probability distributions [@Plastino1; @Almeida1; @Almeida2]. It is also held that one can have inverse power law distributions of Tsallis form with fat tails for all ranges of the nonextensivity parameter $q$. In order to shed light on all these issues, we have rigorously studied the probability distribution of the system through microcanonical approach and shown that it stems from the interplay between any arbitrary system and the constant heat capacity of the bath. Only when the bath has finite and constant negative heat capacity, the system attains an inverse power law distribution with fat tails. The finite baths with positive constant heat capacity lead to the system distributions with a well-determined cut-off condition leaving no possibility for the emergence of fat tails. Whether one adopts the ordinary or escort Tsallis distribution is found to be irrelevant, since the choice between the two does not change the nature of the distribution, but only serves for the same distribution to emerge in different intervals of the nonextensivity parameter $q$. Consider for example that one has a thermal bath with finite and constant positive heat capacity. Having this information suffices to determine the shape of the system probability distribution i.e. a sharp decay with a cut-off having no fat tails. However, the adoption of the ordinary Tsallis distribution for this particular case limits the exponent of the Tsallis distribution to the interval $q>1$ (see Eq. ) while the same physical case yields $q<1$ for the escort distributions (see Eq. ). Therefore, it is the feature of the finite bath which determines the shape of the system distribution. The most important question is finally to decide on whether the Tsallis distributions are indeed to emerge from the coupling of the physical system with a finite bath. The answer to this is no, since the emergence of the Tsallis distributions in finite bath scenario, be it ordinary or escort, requires the constant heat capacity of the finite heat bath to be $q$-dependent (see Eqs. , , and for a complete check). However, the bath, although finite because of consisting of finite number of non-interacting particles, should have a heat capacity such as $D N_B /2$, $D$ and $N_B$ being as usual the dimension of the space and the number of the particles in the bath, respectively. Therefore, its heat capacity must not depend on the non-extensivity parameter $q$. One might argue that the bath, being finite, might not be extensive so that its heat capacity can be $q$-dependent to account for the degree of non-extensivity. However, it can be rigorously shown that the equipartition theorem is intact despite the non-additivity of the Tsallis entropies so that even the heat capacity of a non-extensive classical Hamiltonian system must be independent of the non-extensivity parameter $q$ [@Biro2]. One might object to the above conclusion by stating that the interaction energy between the system and the bath has not been taken into account. However, if this is done, the probability distribution is found to be neither exponential nor $q$-exponential [@Tay] (see Eq. (18) therein). We also note that our results do not exclude the possibility of Tsallis distributions in non-ergodic systems such as composed of classical long-range interacting particles [@central1], since the ergodicity of the total system is assumed in the present work. Considered in the context of the recent attempts to $q$-generalized central limit theorems, our results on finite baths seem consistent. There could be no $q$-generalized central limit theorems if one could obtain Tsallis distributions from ordinary classical systems coupled to finite baths. In other words, if the finite baths were the source of genuine Tsallis distributions, one would expect them to emerge from the ordinary law of large numbers as an intermediate distribution despite the lack of correlation. This is apparently not the case [@central1; @central2]. [00]{} C. Tsallis, *Introduction to Nonextensive Statistical Mechanics*, (Springer) 2009. A.R. Plastino, A. Plastino, Phys. Lett. A 193 (1994) 140. M.P. Almeida, Physica A **300** (2001) 424. J.S. Andrade Jr., M.P. Almeida, A.A. Moreira and G.A. Farias, Phys. Rev. E **65** (2002) 036121. M. Campisi, Physica A **385** (2007) 501. M. Campisi, Phys. Lett. A **366** (2007) 335. T. S. Biró, arxiv: cond-mat 1211.5284 (2012). D. Lynden-Bell, Physica A **263** (1999) 293. W. Thirring, Z. Phys. B **235** (1970) 339. S. Hilbert and J. Dunkel, Phys. Rev. E **74** (2006) 011120. M. Kastner, Phys. Rev. Lett. **104** (2010) 240403. M. Schmidt *et al.*, Phys. Rev. Lett. **86** (2001) 1191. F. Gobet *et al.*, Phys. Rev. Lett. **89** (2002) 183403. A. Khinchin A, *Mathematical Foundations of Statistical Mechanics* (Dover, New York) 1949. M.Campisi, F. Zhan and P. Hänggi, EPL **99** (2012) 60004. J. Dunkel and S. Hilbert, Physica A Physica A **370** (2006) 390. T. Oikonomou and G. B. Bagci, J. Math. Phys. **50** (2009) 103301. T. Oikonomou and G. B. Bagci, Phys. Lett. A **374** (2010) 2225. J. F. Lutsko and J. P. Boon, EPL **95** (2011) 20006. T. S. Biró and P. Ván, Phys. Rev. E **83** (2011) 061147. J. Jiang *et al.*, Physica A **391** (2012) 3140. A. Pluchino, A. Rapisarda and C. Tsallis, EPL **80** (2007) 26002. O. Afsar and U. Tirnakli, Phys. Rev. E **82** (2010) 046210. [^1]: Corresponding Author

ArXiv

--- abstract: 'We present the correlation dimension of resolved young stars in four actively star-forming dwarf galaxies that are sufficiently resolved and transparent to be modeled as projections of three-dimensional point distributions. We use data in the Hubble Space Telescope archive; photometry for one of them, UGCA 292, is presented here for the first time. We find that there are statistically distinguishable differences in the nature of stellar clustering among the sample galaxies. The young stars of VII Zw 403, the brightest galaxy in the sample, have the highest value for the correlation dimension and also the most dramatic decrease with logarithmic scale, falling from $1.68\pm0.14$ to $0.10\pm0.05$ over less than a factor of ten in $r$. This decrease is consistent with the edge effect produced by a projected Poisson distribution within a 2:2:1 ellipsoid. The young stars in UGC 4483, the faintest galaxy in the sample, exhibit very different behavior, with a constant value of about 0.5 over this same range in $r$, extending nearly to the edge of the distribution. This behavior may indicate either a scale-free distribution with an unusually low correlation dimension, or a two-component (not scale-free) combination of cluster and field stars.' author: - Mary Crone Odekon title: The Correlation Dimension of Young Stars in Dwarf Galaxies --- Introduction ============ Two striking and controversial features in the distribution of gas and young stars in galaxies are their scale-free structure and their universality across different environments. Approximate forms of self-similarity are manifest through a variety of analysis techniques, including intensity power spectra, perimeter-area scaling, velocity-size relations, and autocorrelation functions, on scales ranging from several AU all the way up to entire galaxies. (See Elmegreen & Scalo 2004 for a review). One explanation is that scale-free structure is linked to the turbulent transfer of energy among scales, but the importance of physical effects like gas compressibility, magnetohydrodynamics, gravitational fragmentation, and energy output from massive stars is not clear. The identification of special scales where power laws change, and the identification of differences between the structure of gas and that of stars, may signal the regimes dominated by different physical processes (e.g. Larson 1995). Several authors specifically emphasize the universality of scaling laws across different environments. Heyer & Brunt (2004), for example, find similar stucture functions in 27 molecular clouds, and suggest that star formation is therefore dominated by a common mechanism like converging turbulent flows. Willett, Elmegreen, & Hunter (2005) emphasize the similarity in the power spectra of light in star forming galaxies over a large range of sizes. Another recent result supporting universality in stellar clustering is the similarity in clustering properties of OB stars in the Small Magellanic Cloud in clusters and the field (Oey et al. 2004). Previous studies of the projected distribution of starlight on galaxy-wide scales include Elmegreen & Elmegreen (2001), who conclude that the starlight in spiral galaxies is consistent with a fractal dimension of 1.35 (a typical value for gas clouds), and Parodi & Bingelli (2003) who find that star-forming complexes in dwarf irregular galaxies have cluster dimensions between 1.3 and 2. However, several difficulties complicate the interpretation and comparison of these studies. One is blending: large-scale clumps tend to hide the smaller-scale structure within them — indeed, the more highly clustered the distribution, the more likely that small-scale structure will be hidden within large-scale structure. A related, but distinct, problem is varying opacity. A region need not be visibly crowded to obscure starlight; the culprit may be dark dust. Another difficulty for observations based on optical light from stars is that clustering properties depend on the relative contributions from stars of different ages, older stars having had more time to relax into a smooth distribution. In this paper, we minimize these problems by selecting galaxies that are sufficiently resolved and instrinsically transparent to be modeled as projected three-dimensional distributions. In other words, blending and extinction are minimal for the stars and physical scales under consideration. In addition, we use resolved stars with ages we can estimate from their positions on a color-magnitude diagram. These constraints greatly simplify the quantification of uncertainties and comparison with models. Data and Photometry =================== Our choice of sample galaxies is driven by the goal of modeling the distribution of young stars as the projection of a three-dimensional distribution of points. We use very faint, low-metallicity dwarfs, for which extinction by dust is low and the line-of-sight depth is small (Table 1). The absolute magnitudes $M_B$ for our sample galaxies range from $-11.43$ to $-14.30$. For comparison, the 114 galaxies in the Palomar/Las Campanas Imaging Atlas of Blue Compact Dwarf Galaxies have a considerably brighter typical magnitude of $M_B = -16.1$ mag (Gil de Paz, Madore, & Pevunova 2003). Additional advantages of these galaxies are their small sizes on the sky and relatively high galactic latitudes, resulting in little contamination by galactic extinction or foreground stars. The four galaxies in our sample have deep Wide Field Planetary Camera 2 (WFPC2) images in the Hubble Space Telescope Archive. Single-star photometry for three of them has been published previously, while the photometry for UGCA 292 is published here for the first time. For this study, we performed the photometry for all four galaxies beginning with the original images, in order to assess completeness and blending. We examined archival data for several other dwarf galaxies and found them to be insufficient for our purposes because of shorter exposure times (snapshot survey galaxies), too much crowding and extinction (brighter galaxies like NGC 1549 and more distant galaxies like I Zw 18), or limited coverage of the star forming region (local group galaxies). Some of these galaxies are prime targets for future deep high-resolution images, and, indeed, observations of some of them are currently underway with the Advanced Camera for Surveys. Upon inspection, the images of the sample galaxies appear nearly transparent in the sense that distant background galaxies show through (Fig. 1; see also Fig. 2 in Izotov & Thuan 2002 and, for an especially obvious example, Fig. 1 in Crone et al. 2002). Further evidence for low extinction comes from the color-color diagrams for VII Zw 403 and UGCA 290, which show no evidence for internal reddening (Crone et al. 2002). There are a few small clusters, some with intense nebular emission, that are not transparent. For example, Izotov & Thuan find a patchy structure in the $V-I$ subtracted image of the emission nebula surrounding the dense cluster in UGC 4483, that they interpret as differential extinction. As discussed below, these small-scale instances of crowding and extinction are quantifiable, and do not preclude calculating the correlation dimension on sufficiently large scales. Data reduction is essentially the same for each of the four galaxies. Here, we describe the specific process for UGCA 292; details for the other galaxies, including errors and completeness, can be found in the references in Table 1. The observations for UGCA 292 use nine dither positions. We combined exposures at the same pointings using the CRREJ task within the STSDAS software package, and combined the different pointings using DRIZZLE onto a $1600\times1600$ grid. After masking out obvious background galaxies and foreground stars, we conducted single star photometry with DAOPHOT using zero points from the 1997 May SYNPHOT tables. To assess the completeness and accuracy of the photometry, we used ADDSTAR to superpose artificial stars of known magnitude on the inner 60 arcsecond by 40 arcsecond region of the galaxy (the rectangular region in Fig. 1). To avoid artifical overcrowding, we repeatedly performed the photometry with only ten additional stars at a time, for a total of 9000 stars over the magnitude range $21-30$ in each filter. The percentage of artificial stars recovered, as well as the difference between the input magnitude and the recovered magnitude, are in Fig. 2. We corrected for the small foreground extinction according to the values in Schlegel, Finkbeiner, & Davis (1998), and transformed the F555W and F814W magnitudes into V and I following Holtzman et al. (1995). The resulting CMD (Fig. 3) shows features characteristic of star forming dwarf galaxies: a plume of very blue ($V-I<0$) stars that generally includes blue supergiants, very blue main sequence stars, and blue Helium-burning stars; and a parallel plume of slightly red ($V-I\sim1$) stars that generally includes red supergiants, giants, and asymptotic giant branch stars. We estimate the distance to UGCA 292 using the tip of the red giant branch (TRGB), which occurs consistently at $M_I\sim-4.0$ for low metallicity stars (Lee, Freedman, & Madore 1993). The horizontal line in Fig. 3 indicates the most likely magnitude of the TRGB, based on the rise in the luminosity function of red ($V-I>0.5$) stars at $I\sim 24.5\pm0.2$; this edge in the luminosity function appears as a positive signal from a Sobel edge detector, shown as the dotted histogram in Fig. 3. Note the smaller rises at $I\sim 24.2$ and $I\sim 25.0$. Dwarf galaxies often exhibit a rise in the red luminosity function caused by asymptotic giant branch (AGB) stars that are slightly brighter than the TRGB, but is often clear from the CMD morphology that the larger peak is indeed the TRGB (see Fig. 4 and Schulte-Ladbeck, Crone, & Hopp 1998 for the case of VII Zw 403, which shows a red spray of AGB stars above the TRGB). In the case of UGCA 292, we take the largest peak to be the TRGB, with a statistical uncertainty of 0.2 mag, keeping in mind the possibility that the peak at 24.2 might be the actual TRGB. It is even possible that the rise at $I\sim 25.0$ is TRGB, but this would imply a strangely prominent AGB population; population synthesis modeling with very low metallicity evolutionary tracks might provide a solid test for this possibility. Photometric errors for the TRGB stars in UGCA 292 are small (nearly all less than 0.1 mag; see Fig. 2) and overall completeness is nearly 100%. This region of the CMD is also fairly well sampled, with over 300 red stars in the one-magnitude interval below the TRGB. There is, however, an uncertainty caused by the extremely low metallicity of this galaxy. As expected from its low nebular abundance (12+log(O/H)=7.32, van Zee & Haynes 2006), the red giants are abnormally blue. Following Salaris & Cassisi (1997), we can estimate the metallicity of red giant stars from the $(V-I)_o$ color of the RGB half a magnitude fainter than the TRGB. For UGCA 292, this color is in the range $1.0-1.1$ mag, corresponding to a metallicity just outside the range of their TRGB calibration ($1.2 < V-I < 2.0$ mag, corresponding to the metallicity range $-2.23 < [M/H] < -0.57$). We will still use the usual value for the TRGB magnitude, $M_I = -4.0 \pm 0.1$, because it holds over this entire range of metallicity. Comparing our tip determination to the absolute TRGB magnitude, we find a distance modulus of $28.5 \pm 0.24$ mag, corresponding to a distance of $5.0 \pm 0.4$ Mpc. If the rise at 24.2 is actually the TRGB, then UGCA 292 is a bit closer, at 4.0 Mpc. Either way, our distance estimate is larger than the 3.1 Mpc obtained using Virgocentric infall models (van Zee & Haynes 2006). In general, TRGB distance determinations for the other galaxies are more straightforward than that for UGCA 292, thanks to more obvious tip locations on the CMD and less extreme metallicity. For details, see the references in Table 1. We find distance moduli of $29.10 \pm 0.12$ for UGCA 290, $28.23 \pm 0.10$ for VII Zw 403, and $27.7 \pm 0.15$ for UGC 4483. Our distance determinations for UGCA 290 and VII Zw 403 are identical to those in the references in Table 1. Our distance determination for UGC 4483, $3.5 \pm 0.2$ Mpc, is consistent with Izotov & Thuan (2002), who find $3.4 \pm 0.2$ Mpc from the same data set. Fig. 4 shows the V, I color-magnitude diagrams (CMDs) in terms of absolute magnitude for all four galaxies. The selection criteria for young stars balance several considerations. There should be enough stars to produce statistically useful results. Photometry for the selected stars should be complete or nearly complete, requiring sufficiently bright stars. Finally, the stars should be sufficiently young that clusters on the scales under consideration have not dissipated, and we must be able to determine their ages from positions on the CMD. For age determination, we use the results of Crone et al. (2002; see in particular their Fig. 5), who modeled the star formation histories of UGCA 290 and VII Zw 403 with the Bologna synthetic code (Greggio et al. 1998). Note that synthetic modeling improves upon simple comparison with isochrones because it quantitatively estimates the degree to which each area of the CMD includes populations with a mix of different ages. They found that two sections of the CMD are limited to stars approximately in the range 0 to 20 Myr, with a few as old as 30 Myr: the brightest main sequence stars ($V-I<-0.2$, and $M_I<-4$) and the brightest supergiants ($M_I<-7.5$). For our sample galaxies, the number of stars in this population ranges from 32 for UGC 4483 to 125 for UGCA 290. We also consider a larger population including supergiants down to $M_I = -6$, corresponding to stellar ages up to about 100 Myr. For our sample galaxies, the number of stars in this population ranges from 47 to 164. Finally, for comparison, we also calculate the correlation dimension for all the resolved stars in each galaxy, including red giant stars billions of years old. Dissipation timescales for clusters on scales of tens of parsecs are likely to be hundreds of millions of years, so we do not expect the youngest stars to be significantly affected by dynamical relaxation on these scales. The crossing time $\tau_c$ for a 10 pc cluster with velocity dispersion 1 km s$^{-1}$ (typical for nearby clusters) is 10 Myr. According to the analytical approximation of Binney & Tremaine (1987), the dynamical relaxation time is $\tau_{rel} \sim (0.1N/\rm{ln}N)\tau_c$ , so a cluster with only $10^3$ stars has $\tau_{rel}$ as large as $\sim 300$ Myr. For the scales considered in this paper, both $N$ and cluster sizes are higher, producing still longer relaxation timescales. Velocity dispersions on these scales are very unlikely to be high enough to counter these effects: galaxy-wide gas velocity dispersions are only about 10 km s$^{-1}$ in galaxies this small (e.g. van Zee, Salzer, & Skillman 2001). It is also possible to estimate relaxation timescales through N-body simulations. Goodwin & Whitworth (2004) find that the structure in initially fractal star clusters is erased in one to several crossing times, depending on the initial velocity dispersion. In addition to these theoretical considerations, it should also be noted that many nearby galactic clusters with ages of hundreds of millions of years and sizes on the order of 10 pc are known to exist as discernable entities (e.g. Chen, Chen, & Shu 2004). Fig. 5 shows the spatial distribution of the stars in the second age group ($0-100$ Myr) for each galaxy, compared with R-band and B-band images from the online database of the Palomar/Las Campanas Imaging Atlas of Blue Compact Dwarf Galaxies. (For UGCA 292, which is not part of the Las Campanas Atlas, images are from the lower-resolution POSS-II survey.) The R-band isophotes approximate the extent of the older, dynamically relaxed stellar population, while bright clumps in the B band represent the extent of star-forming complexes. Comparison of these two populations shows that the WFPC2 field of view succeeds in covering most or all of the star-forming region in each galaxy. Note that, unlike the elliptical distribution of older stars, there is not generally a concentrated center in the distribution of young stars. These images also illustrate the range of galaxy morphologies in our sample. UGCA 290, with only a small elliptical distribution of older stars, is intermediate between a blue compact dwarf (BCD) and a dwarf irregular. VII Zw 403, meanwhile, is often cited as a classic type nE BCD, exhibiting central star formation within a large background of older stars (Isotov, Thuan, & Lipovetsky 1997). An interesting connection between these two galaxies is that despite their difference in morphology, the CMDs for their young stars (stars with $M_I < -4, M_V < -3.7$) are statistically indistinguishable (Crone et al. 2002). UGCA 292 is a dwarf irregular with the intriguing properties of especially low metallicity and especially high neutral hydrogen gas content $M_H/L_B = 6.9$ (van Zee 2000), suggesting that it is in a very early stage of development. Finally, UGC 4483 is a cometary-shaped galaxy sometimes cited as a BCD. However, based on its low peak surface brightness, Gil de Paz, Madore, and Pevunova (2003) conclude that it “should certainly not" be classified as a BCD but as a dwarf irregular. Many of the young stars in UGC 4483 are limited to a compact star-forming complex near its northern edge. As illustrated in Fig. 2 and the references in Table 1, artificial star tests that scatter stars at random over the face of each galaxy indicate nearly complete photometry for the young stellar populations we consider. However, if these stars are indeed clustered, stars may still be lost at a significant rate in a few small, crowded regions. In order to assess this issue, we determined whether unrecovered artificial stars are limited to small, obviously crowded regions, and if so, what the sizes of those regions are. We emphasize that this method relies on using stars for which photometry across most of the galaxy is nearly complete, and for which the few missing stars can be clearly identified with obviously crowded clumps. Otherwise, it is not possible to know the actual underlying distribution except through model-dependent consistency tests — the usual blending problem encountered in lower-resolution images. For each galaxy, we examined the locations of lost artificial stars for three populations: the main sequence at $-4.75 < M_I < -3.25$, supergiants at $-6.25 < M_I < -5.75$, and bright supergiants at $-8.25 < M_I < -7.75$. In other words, we considered the faint end of each region on the CMD that we use to select young stars. For all the galaxies, the very small percentage (0-1%) of lost bright supergiants are within two pixels (2 – 3 pc) of the center of very bright neighboring stars. The results are the same for the fainter supergiants, except in the case of VII Zw 403, where the completeness rate is down to 75% within a bright knot 20 pc across. For the main sequence stars, as well, the only stars lost are within obvious dense clumps, but these clumps are larger: 40 pc (UGCA 290), 50 pc (VII Zw 403), 23 pc (UGC 4483), and 36 pc (UGCA 292). In each of these cases, completeness for artificial stars randomly scattered *within the clumps* is $50\%-95\%$. Because the effective resolution is worse for the main sequence stars, we consider the supergiants separately, as well as augmented by the young main sequence stars, in the analysis that follows. The Correlation Dimension: Definitions ====================================== A self-similar distribution of points can be partially characterized by its correlation dimension $d_c$: $$N(r) \propto r^{d_c}$$ where the correlation integral $N(r)$ is the average number of particles in a region of radius $r$. For a random Poisson distribution in three-dimensional space, for example, $N$ simply scales as the volume, yielding a correlation dimension $d_c = 3$; the Poisson distributon fully samples the three-dimensional space. Various physical processes correspond to other predictions for $d_c$. The isothermal surfaces for simple Gauss Kolmogorov incompressible turbulence follow a fractal pattern with a dimension of 2.67 (Mandelbrot 1983), while other scaling laws obtain for turbulence models that include other, perhaps more realistic, physical processes (Elmegreen & Scalo 2004). Projection onto the two-dimensional plane of the sky reduces the expected value of $d_c$. Analytically, the projection of a mathematical fractal with dimension $d$ into two dimensions follows the rule $d_{proj} = d$ for $d < 2$, while $d_{proj} = 2$ for $d \geq 2$ (Falconer 1990). Sánchez, Alfaro, & Pérez (2005) tested this behavior for projected fractal clouds sampled by a finite number of points, and found that the correlation dimension roughly followed this behavior for $d$ less than about 1.5 and greater than about 1.9. At intermediate values, $d_{proj}$ was less than the analytic description — for example, $d=2$ produced $d_{proj} = 1.7$. Note that these results apply specifically to projections rather than slices; a two-dimensional slice, on the other hand, yields a dimension less by 1. Regardless of whether the distribution of stars is actually fractal, the correlation dimension as a function of scale $d_c(r)$, like any correlation function, provides a simple way to characterize a distribution that is hierarchically clustered. The correlation dimension, in particular, is used extensively to characterize time series data (Grassberger, Schreiber, & Schaffrath 1991). Its use for spatial distributions in astronomy includes application to young stars in Taurus (about 1.4, Larson 1995), water masers in galactic star-forming regions (between about 0.2 and 1.0, Strelnitski et al. 2002) and star forming complexes in dwarf galaxies (between about 1.2 and 1.8, Parodi & Binggeli 2004.) We calculated $N(r)$ by counting the average number of stars within a distance $r$ of each existing star. This is equivalent to the Grassberger & Procaccia (1983) pair-counting method, multiplied by the total number of particles. We then found $d_c$ by fitting a power-law function to $N(r)$ within logarithmic bins in $r$. Fig. 6 illustrates the ability of this technique to recover the dimension of a simple finite fractal randomly sampled by points. We use the Sierpinski triangle, a well-known fractal in the shape of a triangle composed of three smaller triangles half as large as the first, each of which is composed of three trianges half as large, and so on. This construction produces the fractal dimension log3/log2 $\sim 1.585$. As shown on the left panels in Fig. 6, our ability to recover the fractal dimension is limited by both the finite length of the fractal and the limited number of points. On large scales, edge effects become important, flattening $d_c$ to zero when $r$ becomes large enough to include all the particles. Regardless of the number of particles, the finite length of the fractal limits the scales over which $d=1.585$ is recovered to those less than about a quarter of the total length. On small scales, meanwhile, the ability to recover $d$ is limited by the number of particles available to sample it. For this particular fractal, it requires at least 50 particles to find 1.585 over an order of magnitude in $r$, and 500 particles to find 1.585 over two orders of magnitude. Limited spatial resolution is an additional concern. The panels on the right of Fig. 6 illustrate this problem, again using a Sierpinski triangle of length 1000, but this time with sharp resolution cutoffs at 1, 5, 10, and 20. Note that the steepening of $N(r)$ on small scales extends to scales considerably larger than the cutoff, about five times $r_c$. These sharp cutoffs more closely model the formal resolution scale of the images (e.g. the full width at half maximum of the point spread function), than the dense clumps, which are complete to better than 50%, but are a warning that resolution effects may creep in to length scales as large as five times the size of dense clumps. The models in Fig. 6 do not address the combination of edge and projection effects that apply to a finite distribution in three dimensions. In particular, the front and back edge of a three-dimensional distribution “appear" throughout an entire two-dimensional projection. A simple model that includes these effects is a random Poisson distribution bounded by a three-dimensional ellipsoid. We consider four specific models within this class: an edge-on oblate ellipsoid with axial ratio 1:2:2, a side-view prolate ellipsoid with axial ratio 1:1:2, a sphere, and a face-on disk (equivalent to an end-on cylinder). As expected, the pure Poisson dimension of two is obtained for small scales, but at larger scales the dimension gradually decreases to zero (Fig. 7). The edge effect is smallest for the two-dimensional disk, where the Poisson dimension $d_c = 2.0$ is recovered for scales up to about half of the semimajor axis. The problem is greatest for the ellipsoids, where projection carries the edge effect to scales as small as $r=250$, only one fourth of their semimajor axis. A further refinement of these projected Poisson models includes a decreasing volume number density. Fig. 8 shows results when the density decreases exponentially with ellipsoidal isodensity contours, such that the density falls by a factor of $e^{10}$ when the distance from the center along the major axis is 1000. The behavior of the cluster dimension is very similar to the constant density models over the scales where $d_c$ falls from 2.0 to about 0.5; for example, in both cases the oblate ellipsoids fall from 1.8 to 0.5 over a factor of five (a difference of 0.7 in log$r$). On larger scales, $d_c$ decreases more gradually for the exponential models, reflecting the smearing out of the edge. These models span a wide range of plausible shapes and orientations. For example, an end-on prolate ellipsoid has projection effects between those of a sphere and an end-on cylinder. Similarly, the behavior of a partially inclined, flat oblate ellipsoid is between that of the disk and the edge-on 2:2:1 ellipsoid. Distributions that are more extreme, such as a very flat, edge-on ellipsoid are not consistent with the observed projected distribution of stars (Fig. 5). The Correlation Dimension of the Sample Galaxies ================================================ Qualitatively, the correlation integrals for the actual stellar populations show the same behavior as the models, with approximate power laws that flatten on larger scales and steepen on smaller scales (Fig. 9). It is also apparent that the young population of UGC 4483 exhibits a considerably shallower slope, and that there are slight differences between the slopes for different ages and between young main sequence and young supergiants. But it is not clear from Fig. 9 alone whether these differences are a statistical effect. In particular, the young population in UGC 4483 is fairly small, at just 32 stars. (Note that the total number of stars in each population is indicated by the maximumum value of $N(r)$, and the spatial extent of the population by the value of $r$ where the slope flattens to zero.) In order to make quantitative comparisons, we fit $d_c$ within logarithmic bins in scale, and estimate uncertainties via bootstrap resampling. More specificially, we create a set of fifty random realizations of each galaxy by repeatedly selecting from the original distribution of stars, fit $d_c$ in logarithmic bins of $r$, and use the standard deviation in these values as our measure of uncertainty. Fig. 10 illustrates the usefulness of inspecting the full behavior of $d_c(r)$ rather than a single value. The correlation dimension for the young stars in VII Zw 403, in particular, is not resolved over a wide enough range of scales to reveal a constant power-law, dropping from $1.7 \pm 0.1$ to nearly zero over less than an order of magnitude in scale. The calculation of a single number for $d_c$ over these scales would be highly dependent on the range of scales chosen. For other populations, the correlation dimension changes more slowly. The youngest population in UGCA 290, for example, shifts from about 1.6 to 1.0 over about an order of magnitude in scale. Even more extreme, the young stars of UGC 4483 show a consistently low value for $d_c(r)$, at about 0.5. This galaxy has a dense young cluster on one side; perhaps this behavior reflects a non-scale free combination of two distinct components: a loose distribution of field stars and the compact cluster. As illustrated above, a sharp resolution cutoff steepens $d_c(r)$ on scales up to five times larger than the cutoff scale. For the supergiant populations, this does not pose a problem over the range in Fig. 9; their photometry is complete down to just a few parsecs. The photometry for the bright main sequence stars, on the other hand, is only 50-95% complete within a few specific clumps a few tens of parsecs across. A cautious estimate of largest scales that might be affected is five times the clump size: $\rm{log}r=2.3$ for UGCA 290, 2.4 for VII Zw 403, 2.25 for UGCA 292, and 2.1 for UGC 4483. Because the clump size is certainly not a sharp cutoff, this effect should be small for the scales in Fig. 10. The implication for interpreting the results in Fig. 10 is that the first two or three data points in bold for the $0-30$ Myr and $0-100$ Myr populations may be slightly too high. In fact, the supergiant and main sequence populations give statistically equivalent results. For comparison, we also calculate $d_c(r)$ for the entire resolved stellar population in each galaxy, a population that is both incomplete and heterogeneous in age. As expected for older stars with time to dynamically relax, $d_c(r)$ is very close to 2.0 on small scales and decreases in a manner similar to the exponential ellipsoid. Summary and Discussion ====================== Using high-resolution data for four very small star-forming galaxies, we are able to calculate the correlation dimension as a function of scale with quantified uncertainties. We find that there are statistically distinguishable differences in the nature of stellar clustering among the sample galaxies. The young stars of VII Zw 403, the brightest galaxy in the sample, show the highest dimension and also the most dramatic decrease with logarithmic scale, falling from $1.68\pm0.14$ to $0.10\pm0.05$ over less than a factor of ten in $r$. This decrease is consistent with the edge effect produced by a random distribution of points bounded by a 2:2:1 ellipsoid. The young stars in UGC 4483, the faintest galaxy in the sample, exhibit very different behavior, with a constant value of about $0.5$ over this same range in $r$, extending nearly to the edge of the distribution. This low, flat $d_c(r)$ is not consistent with a random distribution bounded by any ellipsoidal shape that fits the stellar distribution on the sky. The errors from bootstrap resampling are small enough to support the significance of a small dimension for UGC 4483. Further support that it is not a simple statistical effect comes from Fig. 6: while small populations produce larger uncertainties, there is no trend toward smaller dimension. As an additional test, we randomly subsampled the stars in UGC 290 to match the total number of stars in UGC 4483. The results were consistent with those for the full UGC 290 population, and not with UGC 4483. While our sample is small, it is worth noting that the brightest galaxy produces the highest dimension and the faintest galaxy the lowest. Parodi & Bingelli (2003) noted a slight but significant trend in this direction in their study of light from star-forming complexes. In our case, the small dimension of UGC 4483 may be linked to the fact that many of its young stars are in one compact clump that acts as a separate (not scale-free) component from the field distribution. Unfortunately, there are not enough stars to find statistically well-determined values for $d_c(r)$ for the field and cluster individually. A natural extension of our models would be projected, finite three-dimensional fractals. Sánchez et al. (2005) find that fractals with dimensions in the range $2.3-2.7$ yield projected dimensions higher than 1.85, rather close to the Poisson value of 2. This range of $2.3-2.7$ includes the values most often inferred from observations that are treated as slices rather than projections, and is also (not surprisingly) a range given much theoretical attention. From their correlation integrals $N(r)$, the models of Sánchez et al. appear to flatten on large scales in a manner similar to the projected ellipsoids. It would be interesting to examine $d_c(r)$ for these models, to verify if they really behave so similarly to the ellipsoids. An advantage of using $d_c(r)$ is that, like other correlation functions, it does not introduce any special scales; if there are special scales in the distribution, they will make themselves known as features in $d_c(r)$. Furthermore, the calculation of $d_c(r)$ does not depend on any independent characteristic of the distribution, like a determination of the center. The major limitations in our approach, where we use galaxies that are conducive to complete photometry but that are very small, are the range of accessible scales and the number of galaxies with images at sufficiently high resolution. These limitations are changing. High resolution images of dwarfs continue to be made by both HST and ground-based instruments, which will allow us to examine the clustering properties of interesting but fairly distant galaxies like I Zw 18, as well as galaxies near enough to require a larger field of view than that provided by the HST. Studies of nearby galaxies hold the potential to provide much deeper views of the stellar population (including lower-mass stars) and to resolve smaller scales. Finally, detailed modeling of the effects of extinction on $d_c(r)$ will open up the interpretation of resolved stellar population in larger galaxies where dust is more prevalent. In particular, it should be possible to come up with a reasonable range of possibilities that span projection and slicing, in order to model varying transparency in a large, dusty galaxy. We thank an anonymous referee for useful comments. This paper uses data products from the Hubble Space Telescope Archive and from the Palomar/Las Companas Imaging Atlas of Blue Compact Dwarf Galaxies. Chen, W. P., Chen, C. W., & Shu, C. G. 2004, , 128, 2306 Crone, M.M., Schulte-Ladbeck, R.E., Greggio, L., & Hopp, U. 2002, , 567, 258 Elmegreen, B.G., & Elmegreen, D. M. 2001, , 121, 1507 Elmegreen, B.G., & Scalo, J. 2004, , 42, 211 Gil de Paz, A., Madore, B.F., & Pevunova, O. 2003, , 147, 29 Goodwin, S. P., & Whitworth, A. P. 2004, Astronomy & Astrophysics, 413, 929 Grassberger, P., Schreiber, T., & Schaffrath, C. 1991, *Int. J. Bif. and Chaos*, 1:521-547 Grassberger, P., & Procaccia, I. 1983, , 50, 346 Hastings, H. M., & Sugihara, G., 1993, Fractals: A User’s Guide for the Natural Sciences, (Oxford: Oxford University Press) Heyer, M.H., & Brunt, C.M. 2004, , 615, L45 Holtzman, J. A., Burrows, C. J., Casertano, S., Hester, J. J., Trauger, J. T., Watson, A. M., & Worthey, G. 1995, PASP, 107, 1065 Izotov, Y.I., & Thuan, T.X. 2002, , 567, 875 Izotov, Y.I., Thuan, T.X. & Lipovetsky, V. A. 1997, , 108, 1 Larson, R.B. 1995, , 272, 213 Lee, M.G., Freedman, W.L., & Madore, B.F. 1993, , 417, 553 Lynds, R., Tolstoy, E., O’Neil, Jr,. E. J., Hunter, D. A. 1998, AJ, 116, 146 Mandelbrot, B. 1983, The Fractal Geometry of Nature, (San Francisco: W.H. Freeman and Comapany) Oey, M. S., King, N. L., Parker, J. W., Watson, A. M., Kern, K. M. 2004, in Gravitational Collapse: From Massive Stars to Planets, eds. G. García-Segura, G. Tenorio-Tagle, J. Franco, & H. W. Yorke) Revista Mexicana de Astronomía y Astrofísica (Serie de Conferencias) 22, 127 Parodi, B.R. & Binggeli, B. 2003, , 398, 501 Salaris, M., & Cassisi, S. 1997, , 289, 406 Sànchez, N., Alfaro, E.J., & Pèrez, E. 2005, , 625, 849 Schlegel D.J., Finkbeiner, D. P., & Davis, M. 1998, , 500, 525 Schulte-Ladbeck, R. E., Crone, M. M., & Hopp, U. 1998, , 493, L23 Schulte-Ladbeck, R. E., Hopp, U., Crone, M. M., & Greggio, L. 1999, , 525, 709 Sreenivasan, K. R. 1991, Ann. Rev. Fluid Mech., 23, 539 Strelnitski, V., Alexander, J., Gezari, S., Holder, B.P., Moran, J.M., & Reid, M.J. 2002, , 581, 1180 Theiler, J. 1986, Phys. Rev. A, 34, 2427. van Zee, L. 2000, , 119, 2757 van Zee, L. & Haynes, M.P. 2006, , 636, 214 van Zee, L., Salzer, J. J., & Skillman, E. D., 2001, , 122, 121 Willett, K.W., Elmegreen, B.G., & Hunter, D.A. 2005, , 129, 2186 [lrccclll]{} UGCA 290 &$-13.46$ &6.7 &7.80 &8122 &F555W &7800 &Crone et al. 2002\ & & & & &F814W &7800 &\ VII Zw 403 &$-14.30$ &4.4 &7.73 &6276 &F555W &4200 &Lynds et al. 1998, Schulte-Ladbeck et al 1999\ & & & & &F814W &4200 &\ UGCA 292 &$-11.43$ &5.0 &7.32 &9044 &F555W &13000 &\ & & & & &F814W &26000 &\ UGC 4833 &$-12.38$ &3.5 &7.56 &8769 &F555W &9500 &Izotov & Thuan 2002\ & & & & &F814W &6900 &\

ArXiv

--- abstract: 'Using a combination of computational simulations, atomic-scale resolution imaging and phenomenological modelling, we examine the underlying mechanism for nanodomain restructuring in lead zirconate titanate (PZT) nanodisks driven by electron beams. The observed subhertz nanodomain dynamics are identified with relaxation oscillations where the charging/discharging cycle time is determined by saturation of charge traps and nanodomain wall creep. These results are unusual in that they indicate very slow athermal dynamics in nanoscale systems.' author: - Nathaniel Ng - Rajeev Ahluwalia - Ashok Kumar - 'David J. Srolovitz' - Premala Chandra - 'James F. Scott' bibliography: - 'faceting7.bib' title: 'Electron-Beam Driven Relaxation Oscillations in Ferroelectric Nanodisks' --- The importance of finite-size boundaries for static polarization configurations has been emphasized in flux-closure [@Naumov04; @Aguado-pente08; @Jia11; @McGilly11; @Tang15] and faceted [@Ganpule02; @scott2008; @Lukyanchuk14] domain pattern studies in several ferroelectric materials. Recently it has become possible to reorient spatially domains and defects on nanoscales with controlled high resolution transmission electron microscopy (HRTEM) [@robertson; @Ahluwalia2013PRL; @kumar2014]. Nanodomain dynamics in ferroelectric nanodisks can also be probed with real-time HRTEM, and have been observed to be considerably slower (10 sec) [@kumar2014] than those in their nanomagnetic counterparts (50 nanoseconds)[@Bisig13]. Although in principle, such slow nanodomain reorientation could arise from thermal effects, it has been shown that such heating is negligible for our experimental conditions [@zheng]. Here we emphasize that the observed driven domain dynamics of such free-standing ferroelectric nanostructures are strongly influenced by electrostatic conditions. In a combined computational, experimental and phenomenological effort, we identify the charging/discharging mechanism underlying the observed oscillations between circular and hexagonal nanodomain patterns in lead zirconate titanate (PZT) nanodisks driven by electron beams. Using phase-field modelling [@Ahluwalia2013PRL; @Sriram2011ACSNano; @Ng2012Acta], we demonstrate the crucial influence of charge on the nanodomain patterns, and show that we can reproduce domain structures similar to those observed in experiment. The subhertz frequency response of the nanodomain reorientation results from relaxation oscillations [@Losev25; @vanderPol; @Pippard79] between threshhold charging of surface traps [@Li06] and subsequent discharging by nanodomain wall creep [@paruch], where quantitative agreement is made using prior measurements on PZT samples. Experimentally edge-supported 8-nm diameter disks of PZT were exposed to constant HRTEM beams with a $0.5\;\text{A/m}^2$ probe current density where further detail can be found elsewhere [@kumar2014]. Two distinct nanodomain patterns were observed: the nanodomain walls “flop” between being normal to the disk perimeter to being parallel to it as displayed in Figure \[fig:hrtem-phasefield\]a and \[fig:hrtem-phasefield\]c. Since HRTEM only reveals ferroelastic domain walls, modelling is needed to examine the full charging role played by the electron beam that drives the observed faceting oscillations and the simultaneous realignment of nanodomain walls. It is known that electron beams can lead to charging of insulating materials [@Sessler2004]. Furthermore in the context of ferroelectrics, it has been shown that electron beams can be used to switch the polarization direction [@bonnell] and can even stabilize unusual quadrant patterns in free standing nanodisks [@Ahluwalia2013PRL]. In order to investigate e-beam induced charging effects in ferroelectric nanodisks, we use a phase-field approach to simulate how domain patterns are influenced by radial fields generated by uniform free charge. The model, based on earlier studies [@Ahluwalia2009Nano; @Ng2012Acta; @Ahluwalia2013PRL], has been adapted to incorporate the circular geometry (see supplementary material [@supplementary1]) of the experiment; the goal is to determine whether the observed nanodomain pattern realignment is driven by depolarization fields in a charging/discharging cycle. ![HRTEM images for (a) circular (unfaceted) and (c) faceted 8-nm PZT dots; the ferroelastic domain walls are illustrated as purple lines, and the angles between them are indicated. Results from simulations (distribution of polarization vectors) for (b) uncharged and (d) charged nanodisks; red lines indicate the approximate positions of domain walls. Domain walls are *perpendicular* to the surface in the *unfaceted* cases (a. and b.), but *parallel* to the surface in the *faceted* cases (c. and d.)[]{data-label="fig:hrtem-phasefield"}](fig1.png){width="\columnwidth"} We first simulate the uncharged case by solving the phase field equations for a system quenched from the paraelectric state. A uniform charge density, $N_e = 4 \times 10^{26}\text{ m}^{-3}$, is then applied and the evolution of the domains is monitored. Fig \[fig:hrtem-phasefield\](b) shows the simulated pattern for the uncharged case with a flux closure domain pattern where the solid red lines indicate the positions of the ferroelastic domain walls, normal to the disk perimeter. We note the excellent agreement with the domain wall configuration in the HRTEM image in Fig. \[fig:hrtem-phasefield\](a), suggesting that it corresponds to an uncharged disk. Figure \[fig:hrtem-phasefield\](d) shows the nanodomain pattern when the nanocrystal is charged ($N_e = 4 \times 10^{26}\text{ m}^{-3}$), i.e. during the charging process, the flux closure domain pattern transforms to one where polarization vectors have a non-zero radial component at the perimeter and a polarization vortex type state, characterized by a vanishing radial component, in the center. This may be understood in terms of the radial electric field that vanishes at the center. This radial field arises due to the uniformly charge density [@Ahluwalia2013PRL] associated with the electron beam induced charging. Vortex formation at the center occurs to avoid head-to-head polarization, resulting in areas having $\mathbf{P \cdot n} = 0$ (center) and $\mathbf{P \cdot n} \ne 0$ along the perimeter. Here, a radial unit vector is introduced as $\mathbf{n}=\frac{\mathbf{r}-\mathbf{r_0}}{\left|\mathbf{r}-\mathbf{r_0}\right|}$, where $\mathbf{r_0}=\left(x_0,y_0\right)$ represents the coordinates of the centre of the disc. Regions with such distinct $\mathbf{P \cdot n}$ are separated by ferroelastic nanodomain walls that are accessible to the electron beam technique. The solid red lines in Fig \[fig:hrtem-phasefield\](d) indicate such ferroeleastic domain walls which are aligned along the perimeter of the nanocrystal. Thus this situation corresponds to the HRTEM image in \[fig:hrtem-phasefield\](c), indicating that it results from nanodisks with charged boundaries. While this charging/discharging scenario describes the realignment of domain walls, it does not directly model the observed kinetic faceting. We emphasize its dynamical nature since two-dimensional faceting cannot occur in thermal equilibrium at finite temperatures [@berge]. However, in Fig \[fig:hrtem-phasefield\](d), we see that the distribution of $\mathbf{P \cdot n}$ is anisotropic, namely strongly $\theta$-dependent, and this is further emphasized graphically in Figure \[fig:radial-polarizations\]. ![Distribution of the normalized radial component of the polarization, $\frac{1}{P_s}\mathbf{P \cdot n}$ corresponding to Fig 1d. The radial component of the polarization is zero in the center and nonzero towards the edges. Note that the distribution shows strong anisotropy (i.e. $\theta$-dependence)[]{data-label="fig:radial-polarizations"}](fig2.png){width="\columnwidth"} Indeed the distribution of elastic energy densities will also be orientation-dependent due to the underlying crystal axes of PZT. More generally kinetic faceting is dynamically selected to accommodate external anisotropies that could also include applied stress and substrate-related issues [@Ganpule02; @scott2008; @Lukyanchuk14] where details depend on the specifics of the experimental situation and the materials involved; indeed faceting nanodomain patterns were not observed in HRTEM experiments on PZT nanotargets with square and triangular boundaries [@kumar2014]. We emphasize that nanodomain realignment and faceting is observable only because our experiments uniquely involve free-standing, edge-supported nanodisks [@kumar2014], since substrate interactions would prevent such domain restructuring. Our phase-field simulations indicate that this nanodoman realignment occurs between charged and uncharged states of the PZT nanodisk. When the disk is uncharged, the usual flux closure is observed implying that the ferroelastic domain walls are normal to the surface. When it is charged, the polarization at the surface becomes radial and a vortex is formed at the center. In HRTEM, such a state is revealed by its nanodomains parallel to the surface. The faceting is selected to minimize both polarization and external anisotropies, principly due to strain fields at the perimeter due to the underlying (111) crystal axes of PZT. This electrostatic analysis accounts for the two observed nanodomain configurations, but suggests a frequency response on gigahertz scales in contrast to the subhertz frequencies observed [@kumar2014]. Because thermal effects have been ruled out for these experimental conditions [@zheng], this slow time-scale suggests extreme nonlinearity and possibly mechanical overshoot. Our phase-field modelling indicates a driven and reversible charging/discharging cycle that is naturally described as an effective relaxation oscillation circuit [@vanderPol; @Pippard79]. This is a system that is intrinsically out-of-equilibrium: it cycles between an unstable state and a frustrated attempt to reach equilibrium that is prevented by onset behavior of its nonlinear element that forces it back to its original unstable state. Electronically the nonlinear circuit element is often a neon light or a unijunction transitor, but its key feature is a threshhold voltage for negative resistance in parallel with a conducting channel that is usually a capacitor. To our knowledge the first report of self-sustaining subhertz oscillations in a physical system was in a wide-gap semiconductor ZnO [@Losev25], and we propose that the e-beam experiments on PZT nanodisks also present such an example. It is known that electron beam irradiation creates charge traps and defects in semiconducting targets. The time-scale for the charging cycle will then be set by the charge saturation of these surface states. This charge storage threshhold in PZT targets exposed to electron beams has been measured quantitatively to be $5 \text{C/m}^2$. For the experiments of interest [@kumar2014] with current density of $0.5 \text{A/m}^2$, this then corresponds to a charging time of $$\tau_{\text{charging}} = \frac{5 \text{C/m}^2}{0.5 \text{C/s-m}^2} = 10 \text{ sec}$$ that is the same order of magnitude as the observed faceting oscillations [@kumar2014]. We note that if we assume that these traps are singly charged, then there are $\frac{5 \text{C/m}^2}{1.6 \times 10^{-19} \text{C}/e} = 3 \times 10^{19} \rm{traps}/m^2$ that is consistent with concentrations reported for oxygen vacancies in commercial quality spin-on PZT films [@Mihara92]; this value is also in good agreement with the electron concentration used in our simulations ($N_e = 4 \times 10^{26} \text{m}^{-3}$) for 8 nm disks of thickness 100 nm. Finally we note that our system is a charge analogue of a seesaw with a faucet continuously dripping water into a container on one end, and a block on the other one. In this textbook mechanical relaxational oscillator [@Wang1999], the seesaw flips when the weight of the water container is greater than that of the block; however once the container touches the ground it empties, the seesaw goes back to its original position and the whole process starts again. In the context of our present work, the electron beam is analogous to the water flow from the faucet; the threshold is related to the Zener breakdown [@Zener1948] that occurs in pn-junctions at low fields ca. $E \ll 1 \frac{\text{V}}{\text{nm}}$ and involves excitation of electrons to the conduction band from the valence band or trap states within the bandgap. It is non-destructive and reversible, and thus differs from avalanche breakdown that occurs at higher fields. The charging of the nanocrystal creates depolarization fields that realign the domains and lead to the observed faceting (Fig 1). During the initial part of the cycle, the current from the electron beam is filling the surface states with charge and the nanodisk behaves like a capacitor. However once the charge saturation threshhold for these traps and defects in the PZT disk is reached [@Li06], the depolarization fields become strong enough to initiate the current flow in the nanodisk. This current leads to discharging and the subsequent decrease of the depolarization field that caused the faceting; consequently the domains realign back to their uncharged configurations. The presence of surface states and large electromechanical coupling means that this involves domain creep in a random elastic medium that has been discussed quantitatively and measured in PZT films [@paruch]. For the experimental parameters of interest, the domain wall creep in PZT epitaxial films is found to $v = 10^{-9}$ m/s [@paruch] so that for the dots of roughly 10 nm (of thickness 100 nm), the discharging time-scale is $$\tau_{\text{discharging}} = \frac{10 \text{nm}} {10^{-9} \text{m/s}} = 10 \text{ sec}$$ that is again in the same order of magnitude as the observed time-scale. This charging/discharging cycle is expected to continue as long as the electron beam is on and a there is a continuous flow of charge into the nanodisk. In contrast with many relaxation oscillators, in these PZT nanodisks there are slow time-scales associated with both the charging and the discharging cycles where the rate-limiting one is clearly the slower of the two. Thus the cycling times, achieved both by macroscopic (charging threshhold) and microscopic (discharging via domain creep) methods, is consistent with measurement. We note here that the two underlying phenomena, charge trapping and slow domain creep due to electromechanical coupling and mass, are both absent in magnetism which explains the many orders of magnitudes that separate the frequency response of field-driven ferroelectric and ferromagnetic nanodisks. In summary we have studied the mechanism underlying nanodomain restructuring in PZT nanodisks driven by electron beams. Our phase-field modelling identifies the two observed nanodomain patterns with uncharged and charged free-standing boundaries; furthermore anisotropy in the charged case selects faceting at the perimeter to minimize electrostatic and elastic energy costs. The observed subhertz frequency response, many orders of magnitude lower than expected from an electronic mechanism, is explained as relaxation oscillations where the charging/discharging times are determined by charge trap saturation and nanodomain wall creep. We predict that the oscillation time-scale can then be tuned by changing the surface trap density, either by varying the beam current or by annealing the PZT nanodisks in oxygen/ozone environments. Though this charging/discharging cycle here is driven by electron beams, similar behavior could be achieved by gating the PZT nanodisks to achieve a voltage-controlled source-gate-drain device. It would be unusual to have such small semiconductor triodes with such slow time-scales, and they could be important for biological applications like medical implants where time-scales are naturally of the order of seconds. This work was supported by the A\*STAR Computational Resource Centre through the use of its high performance computing facilities and by National Science Foundation grant NSF-DMR-1334428 (P. Chandra). PC is grateful for the hospitality of Trinity College, Cambridge where part of this work was performed. We thank Robert Laskowski for useful discussions.

ArXiv

-15mm \[par=1\] [**Noisy kink in microtubules**]{} H.C. Rosu$^{a,c}$ [^1], J.A. Tuszyński$^{b}$ [^2] and A. González$^{a}$ [^3] $^{a}$ [*Instituto de Física de la Universidad de Guanajuato, Apdo Postal E-143, Léon, Gto, México*]{} $^{b}$ [*Department of Physics, University of Alberta, Edmonton, AB, T6G 2J1, Canada*]{} $^{c}$ [*Institute of Gravitation and Space Sciences, Magurele-Bucharest, Romania*]{} ABSTRACT. We study the power spectrum of a class of noise effects generated by means of a digital-like disorder in the traveling variable of the conjectured Ginzburg-Landau-Montroll kink excitations moving along the walls of the microtubules. We have found a 1/f$^{\alpha}$ noise with $\alpha \in$ (1.82-2.04) on the time scales we have considered. PACS: 87.15.-v, 72.70+m, 87.10.+e One can find extensive, descriptive presentations of microtubules (MTs) in many biological papers. Here, we shall give only elementary definitions as follows. They are ubiquitous protein polymers in eukaryotic cells belonging to the category of [*biological filaments*]{} and making the most part of the cytoskeleton. They are hollow cylinders 25 nm in outer diameter and 17 nm in the inner one, with lengths ranging from nm to mm in some neuronal cells. The walls of the cylinders are usually made of 13 (the seventh Fibonacci number) protofilaments laterally associated. The surface structure of MT walls is very interesting [@mand]. Structurally, MTs are quasi one-dimensional chains of tubulin polar dimers (negative $\alpha$ and positive $\beta$ monomers, each of 4 nm in length) undergoing conformational changes induced by the guanosine triphosphate to diphosphate (GTP-GDP) hydrolysis. The whole assembly process of a MT is due to the hydrolyzation of GTP. The cation Mg$^{++}$ is essential to increase the affinity of tubulin for binding GTP and thus for generating MTs. Moreover, a unique dynamical property is the so-called [*dynamical instability*]{} [@-1] which is a random growth and shrinkage of the more active plus ends of MTs. It has raised much interest in recent years [@-2]. An energy-transfer mechanism in MTs by means of Ginzburg-Landau-Montroll (GLM) kinklike protofilament excitations has been discussed by Satarić, Tuszyński and Zakula [@1] in 1993. Also sine-Gordon (SG) solitons have been discussed by Chou, Zhang and Maggiora in 1994 [@czm]. We recall that various types of solitons have found interesting applications in biological physics (Davydov’s model [@dav], DNA/RNA [@dna]). Usually, in order to get nonlinear differential equations one performs a continuum limit for some lattice models in which discreteness effects are neglected. Some authors have shown that such effects might be important and suggested various ways of including them in continuous differential equations [@discr]. The digital disorder we shall comment on and use next is just a possible way to incorporate discreteness in a moving solitonic pattern. In 1993, Rosu and Canessa [@2] introduced a digital-like disorder in the Davydov $\beta$-kink leading to the 1/f$^{\alpha}$ noise, with $\alpha\approx1$, in the dynamics of that kink, and also commented on the multifractal features of the dynamics of the $\beta$-kink. The procedure is as follows. In the traveling variable $\xi=x-v_K t$ of the slowly moving kink (the acoustic Lorentz factor $\gamma _L \approx 1$) one puts $x_i=x_0+\Delta x_i$, where $x_0$ is the position of the center of mass of the kink, or its central position along the chain, and $\Delta x$ are small random displacements around $x_0$ (say, in the interval $\pm 1$). In the calculations, one can fix $x_0=0$. For each random position $\Delta x_i$ chosen from a uniform distribution, one calculates by means of a fast-Fourier-transform algorithm the [*noise power spectrum*]{} of the time series of the signal (considered to be the kink) in order to get the time correlations of the fluctuations of the signal, i.e., $$S_{K}(f)\propto\frac{1}{\tau}\left\langle|\int_0^{\tau}K(\xi)e^{2\pi ift}dt|^2 \right\rangle ~, \eqno(1)$$ where $0<t<\tau\approx 1/f$ and K($\xi$) is the kink function, and the brackets stand for averaging over ensembles. This approach to noise effects has been taken from the literature on the self-organized criticality (SOC) paradigm, see, e.g., [@vm]. On the other hand, the standard treatment of noise effects when they originate in thermal fluctuations is by means of the Langevin equation method. Valls and Lust [@vl] have studied the effect of thermal noise on the front propagation in the GL case. They have found a crossover between constant-velocity propagation at early times and diffusive behavior at late times. The purpose of the present work is to investigate the noise produced by the same type of disorder as in [@2] in the case of the GLM kink conjectured in MTs. The main assumption in [@1] is that the assembly of tubulin dimers/dipoles (${\cal D}_n$) form a quasi one-dimensional ferrodistortive system for which the double-well on-site potential model is a standard framework $$V({\cal D}_{n})=-\frac{1}{2}A {\cal D}_{n}^{2}+\frac{1}{4}B{\cal D}_{n}^{4}~. \eqno(2)$$ The variable ${\cal D}$ has been identified in [@1] with the amount of $\beta$-state distorsion vertically projected (the $\beta$-state is defined as having the mobile electron within the $\beta$-monomer). Moreover, in [@1] a GL hamiltonian/free-energy with intrinsic electric field and dissipation effects included led to the dimer Euler-Lagrange dimensionless equation of motion (EOM) in the traveling coordinate of the anharmonic oscillator form (with linear friction) $$%-\frac{Ma^2V_s^2}{\gamma _{L}^2} |A|{\cal D}^{''}-\gamma \alpha v_K{\cal D}^{'} -F({\cal D})=0~, \eqno(3)$$ where $\gamma$ is the friction coefficient, $\alpha=|A|\gamma ^2 _L/Mv_{sound}^2$ and $F({\cal D})=A{\cal D}-B{\cal D} ^3+qE$, with $q$ denoting the effective charge of a single dimer of mass M, and $E$ the magnitude of the intrinsic electric field. This EOM is known to have a unique kink solution given by the formula [@3] $$K (\xi)=a+\frac{b-a}{1+\exp (\beta \xi)}\equiv a+ \frac{\beta}{\sqrt{2}}(1-\tanh(\beta\xi /2))~, \eqno(4)$$ where $K={\cal D}/\sqrt{|A|/B}$ is a rescaled dipole variable, $\beta=(b-a)/\sqrt{2}$, whereas $a$ and $b$ are two of the solutions of the cubic equation $$F(K)\equiv(K -a)(K -b)(K -c)\equiv K^3 -K -\sigma=0~, \eqno(5)$$ where $\sigma =q\frac{\sqrt{B/|A|}}{|A|}E$ and $u_0$ units are used, where $u_0=\sqrt{|A|/B}\approx 1.4\cdot 10^{-11}$ m is the amplitude of the dimer displacement (shift of the double-well potential). Notice that the GLM kink is thin. Its width is $w_K=\frac{1}{\beta}\approx 0.7 u_0$. As we said, the type of digital disorder we consider here is very close to the ideas of the SOC paradigm that we understand in the broad sense of [*both*]{} spatial and temporal scaling of the dynamical state of the system [@soc]. The spatial scaling (self-similarity) is of the (multi)fractal type while the temporal scaling leads to $1/f^{\alpha}$ noises. A strong motivation for dealing with digital dynamics is the possibility of generating broken symmetries and therefore of having various types of dynamical phase transitions [@ch]. Thus, digital disorder, though might look a rather ad-hoc approach, focuses on both self-organized properties of MTs, i.e., to driven steady states with long-range spatio-temporal correlations, and to (dynamical) phase transitions, since digital dynamics allows for symmetry breaking. Our results are displayed in Figures 1 and 2 and show that the noise introduced by the digital disorder in the GLM kink variable is practically of the 1/f$^2$ (Brownian) type on the time scales we have considered. At low frequencies there is the known cross-over to a white noise due to the finite system size, which is moving to lower frequencies as the size of the system is increased [@fj]. On the other hand, the deviation from the power law at high frequencies is an artifact due to the so-called aliasing [@hjj]. The Brownian noise we have obtained is not unexpected since it is a common occurence in SOC models in any finite dimension [@cc], and only mean-field calculations reproduced the $1/f$ noise [@tb]. Finally, we would like to mention that if polarization does not exactly follow the displacement one needs a system of two coupled partial differential equations leading to two coupled traveling kink waves. Of course, the nonlinear models seem to be too simple-minded for the MT complexity. Nevertheless they provide guides for further insight and perhaps some partial answers. [**Acknowledgments**]{} This work was partially supported by CONACyT (Mexico) and by NSERC (Canada). [99]{} E.-M. Mandelkow et al., J. Cell Biology [**102**]{} (1986) 1067; E. Mandelkow et al., Neurobiology of Aging [**16**]{}(1995) 347. T. Mitchison and M. Kirschner, Nature [**312**]{} (1984) 232, 237. D. Kuchnir Fygenson, E. Braun, and A. Libchaber, Phys. Rev. E [**50**]{} (1994) 1579; H. Flyvbjerg, T.E. Holy and S. Leibler, Phys. Rev. Lett. [**73**]{} (1994) 2372; P.M. Bayley, M.J. Schilstra, and S.R. Martin, J. Cell Sci. [**95**]{} (1990) 33; D.J. Odde, L. Cassimeris, and H.M. Buettner, Biophys. J. [**69**]{} (1995) 796. M.V. Satarić, J.A. Tuszyński, R.B. Zakula, Phys. Rev. E [**48**]{} (1993) 589. See also: J.A. Tuszyński et al., J. Theor. Biol. [**174**]{} (1995) 371. K.C. Chou, C.T. Zhang, G.M. Maggiora, Biopolymers [**34**]{} (1994) 143. A.S. Davydov, [*Solitons in Molecular Systems*]{} (Reidel, Dordrecht, 1985). G. Gaeta, C. Reiss, M. Peyrard, T. Dauxois, Riv. Nuovo Cimento [**4**]{} (1994). J.C. Kimball, Phys. Rev. B [**21**]{} (1980) 2104; A. Sánchez and L. Vázquez, Phys. Lett. A [**152**]{} (1991) 184; J.A. Combs and S. Yip, Phys. Rev. B [**28**]{} (1983) 6873; M. Otwinowski, J.A. Tuszyński, J.M. Dixon, Phys. Rev. A [**45**]{} (1992) 7263. H. Rosu and E. Canessa, Phys. Rev. E [**47**]{} (1993) R3818. J. Vitting Andersen and O.G. Mouritsen, Phys. Rev. A [**45**]{}(1992) R5331. O.T. Valls and L.M. Lust, Phys. Rev. B [**44**]{} (1991) 4326. M.A. Collins, A. Blumen , J.F. Currie, and J. Ross, Phys. Rev. B [**19**]{} (1979) 3630. P. Bak, C. Tang and K. Wiesenfeld, Phys. Rev. Lett. [**59**]{} (1987) 381; Phys. Rev. A [**38**]{} (1988) 364. M.Y. Choi and B.A. Huberman, Phys. Rev. B [**28**]{} (1983) 2547. T. Fiig and H.J. Jensen, J. Stat. Phys. [**71**]{} (1993) 653; N.C. Pesheva, J.G. Brankov, and E. Canessa, Phys. Rev. E [**53**]{} (1996) 2099. H.J. Jensen, Mod. Phys. Lett. B [**5**]{} (1991) 625. H.J. Jensen, K. Christensen, and H.C. Fogedby, Phys. Rev. B [**40**]{} (1989) 7425; J. Kertész and L.B. Kiss, J. Phys. A [**23**]{} (1990) L433; H.F. Chau and K.S. Cheng, Phys. Rev. A [**46**]{} (1992) R2981. C. Tang and P. Bak, J. Stat. Phys. [**51**]{} (1988) 797. 4ex [\ Double logarithmic plot of the power spectrum of the digital noise perturbing the motion of the GLM kink moving along MTs at constant v$_K$ = 2m/s. The fitted slopes are as follows: (a) -2.032 (b) -2.0175 (c) -2.0164 for the time scales corresponding to that of the dimer, ten times bigger, and hundred times bigger, respectively. The errors in the slopes are at the level of 0.0003 for each case.]{} 2ex 4ex [\ The same plot as in Fig. 1 for v$_K$ = 100m/s. The fitted slopes are (a) -2.0083 (b) -1.8641 (c) -1.8205 (d) -2.0000 for temporal scales of 10$^{-2}$, 10$^{-1}$, 10$^{0}$, and 10$^{1}$ times that of the tubulin dimer, respectively. The level of the errors is the same as in Fig. 1.]{} [^1]: Electronic mail: rosu@ifug.ugto.mx [^2]: Electronic-mail:jtus@phys.ualberta.ca [^3]: Electronic mail: gonzalez@ifug.ugto.mx

ArXiv

--- abstract: 'Atomic metallic hydrogen has been produced in the laboratory at high pressure and low temperature, prompting further investigations of its different properties. However, purely experimental approaches are infeasible because of the extreme requirements in producing and preserving the metastable phase. Here we perform a systematic investigation of the electronic, optical, mechanical and thermal properties of $I4_1/amd$ hydrogen at 495 GPa using first-principles calculations. We calculate the electronic structure and dielectric function to verify the metallic behaviour of $I4_1/amd$ hydrogen. The calculated total plasma frequency from both intraband and interband transitions, 33.40 eV, agrees well with the experimental result. The mechanical properties including elastic stability and sound velocity are also investigated. The mechanical stability of $I4_1/amd$ hydrogen is limited by shear modulus other than bulk modulus, and the high Young’s modulus indicates that $I4_1/amd$ hydrogen is a stiff material. After investigating the lattice vibrational properties, we study the thermodynamical properties and lattice anharmonicity to understand thermal behaviours in metallic hydrogen. Finally, the lattice thermal conductivity of $I4_1/amd$ hydrogen is calculated to be 194.72 W/mK and 172.96 W/mK along the $x$ and $z$ directions, respectively. Using metallic hydrogen as an example, we demonstrate that first-principles calculations can be a game-changing solution to understand a variety of material properties in extreme conditions.' author: - 'Bo Peng$^1$, Ke Xu$^1$, Hao Zhang$^{1*}$, Hezhu Shao$^{2}$, Gang Ni$^1$, Jing Li$^1$, Liangcai Wu$^3$, Hongliang Lu$^4$, Qingyuan Jin$^1$, and Heyuan Zhu$^1$' bibliography: - 'new.bib' title: 'Theoretical investigation of novel electronic, optical, mechanical and thermal properties of metallic hydrogen at 495 GPa' --- Introduction ============ In 1935, Wigner and Huntington predicted that, hydrogen molecules would become a metallic solid similar to the alkalis at high pressure [@Wigner1935]. Metallic hydrogen is expected to be a wonder material: it may be a room-temperature superconductor and have significant applications in energy and rocketry [@Ashcroft1968]. However, producing metallic hydrogen has been a great challenge. Although a lot of theoretical investigations have been performed on metallic hydrogen [@Cudazzo2010a; @McMahon2011; @McMahon2011a; @McMahon2012; @Azadi2014; @McMinis2015; @Borinaga2016; @Borinaga2018], not until recently has atomic metallic hydrogen been synthesized in the laboratory at high pressure and low temperature [@Dias2017]. At a pressure of 495 GPa, hydrogen becomes metallic with a high reflectivity, and the metallic phase is metastable when the pressure is released. This prompts a systematic study of its electronic, optical, mechanical and thermal properties for further applications. However, the challenge is to produce and preserve such metastable phase for experimental investigations. Several months after the synthesis, the world’s only metallic hydrogen sample has disappeared. Even if the reproducibility of the sample is improved, characterizing metallic hydrogen under such pressure is extremely difficult due to the limitations of conventional techniques. For instance, for hydrogen samples in a diamond anvil cell at extremely high pressure, neutron scattering and X-ray diffraction experiments are unavailable. Thus it is imperative to develop alternative techniques for exhaustive characterization. Theoretical estimation, as a reference to the experiments, is indeed another option. Recently, computational material techniques have been developed to perform an *ab initio* study on materials in extreme conditions [@Burakovsky2010; @Gao2008; @Alfe2002; @Zhu2017; @Moustafa2017; @Ma2009a]. In addition, state-of-the-art techniques can be used to calculate accurately the ground- and excited-state [@Shishkin2006; @Shishkin2007; @Fuchs2007], mechanical [@LePage2002; @Wu2005], and phononic properties [@ShengBTE; @Li2012a; @Li2012; @Li2013a], enabling the feasibility of theoretical predictions in a variety of different properties of metallic hydrogen. Here, we perform an exhaustive study of metallic hydrogen at 495 GPa. Our fully first-principles calculations shed light into its electronic, optical, mechanical and thermal properties. First we focus on ground- and excited-state properties such as electronic band structure, dielectric function and plasma frequency. Then, mechanical properties including elastic anisotropy and sound velocity are investigated. After that, we discuss the lattice vibrational properties as well as lattice anharmonicity. Finally we study the phonon transport properties of metallic hydrogen. Our results provide clear means of characterizing metallic hydrogen via these distinct features, strongly calling for experimental verifications. Ground- and excited-state properties ==================================== Crystal structure ----------------- ![Conventional cell of $I4_1/amd$ hydrogen at 495 GPa, and 3D electron localization function (isosurface=0.5) of its 3$\times$3$\times$2 supercell. At the boundary, increasing electron localization from 0.5 to 1 is plotted with colours from green to red.[]{data-label="f1"}](H-Figure1.pdf){width="0.45\linewidth"} We start by structural optimization using the Vienna *ab-initio* simulation package (VASP) based on density functional theory (DFT) [@Kresse1996; @Kresse1996a] under the generalized gradient approximation (GGA) expressed by the Perdew-Burke-Ernzerhof (PBE) functional [@Perdew1996]. A plane-wave basis set with kinetic energy cutoff of 800 eV is employed with 51$\times$51$\times$21 (41$\times$41$\times$41) **k**-mesh for conventional (primitive) cell at a pressure of 495 GPa during structural relaxation, until the energy differences are converged within 10$^{-6}$ eV, with a Hellman-Feynman force convergence threshold of 10$^{-4}$ eV/Å. No. $hkl$ 2$\theta$ ($^{\circ}$) $d$ (Å) FWHM ($^{\circ}$) ----- ------- ------------------------ --------- ------------------- -- -- -- 1 001 28.6 3.1222 0.075 2 002 59.2 1.5611 0.081 3 100 79.0 1.2114 0.091 4 011 86.1 1.1294 0.096 5 003 95.6 1.0407 0.104 6 110 107.6 0.9571 0.118 : Calculated X-ray diffraction patterns for the conventional cell of $I4_1/amd$ hydrogen. \[t1\] The conventional cell of the $I4_1/amd$ phase at 495 GPa is shown in Figure \[f1\]. The calculated lattice parameters are $a=b=1.2114$ Å and $c=3.1222$ Å, respectively, which are consistent with other theoretical studies [@McMahon2011; @Borinaga2016]. The electron localization function maps of $I4_1/amd$ hydrogen at 495 GPa show a typical metallic behavior: Regions with constant 0.5 (corresponding to the electron-gas like pair probability) are connected one to the other by channels, forming infinite three-dimensional networks [@Silvi2000]. In order to confirm the structure of the experimentally reported metallic hydrogen, we simulate the X-ray diffraction patterns of the conventional cell [@vesta]. The results are shown in Table \[t1\]. Electronic structure -------------------- ![(a) Electronic structure and (b) imaginary part of dielectric function for $I4_1/amd$ hydrogen at 495 GPa.[]{data-label="2"}](H-Figure2.pdf){width="\linewidth"} Figure \[2\](a) shows the electronic band structure of $I4_1/amd$ hydrogen at 495 GPa. The Fermi level is crossed by five different bands, confirming the metallic behaviour. The huge dispersion is associated to the dominating kinetic term in the energies of the electronic states. Different from the free-electron approximation [@Borinaga2016], the band gap open across the Brillouin zone at about 20 eV due to the interaction of the electrons with the proton lattice. Borinaga *et al* has pointed out that the electron-electron interactions such as Hartree, exchange and correlation effects give no significant contribution to the band structure [@Borinaga2016]. However, after photo-excitation, an electron is excited from valence band into conduction band, leaving behind a positively charged hole. This single-particle excitation cannot be described by non-interacting particles with infinite lifetime. Due to interactions with other particles, the excited electrons and holes become quasiparticles with finite lifetimes. This acquires a self-energy to account for all the electron-electron interactions [@Louie2006]. In addition, the negatively charged quasielectron is attracted by the quasihole, stabilizing the electron-hole pair and forming a new quasiparticle, an exciton. Apart from this attraction, a possible electron-hole exchange is also an important many-body effect. Moreover, the electron-hole pairs are weakened by the screening of the electronic system. Therefore, self-energy, excitonic, and other many-electron interaction effects may strongly influence the excited-state properties of metallic hydrogen [@Bechstedt2016]. Optical properties ------------------ Subsequent to the standard-DFT results, self-consistent GW$_0$ corrections are undertaken [@Shishkin2006; @Shishkin2007; @Fuchs2007] with eight iterations of $G$. The energy cutoff for the response function is set to be 300 eV. A total of 20 (valence and conduction) bands are used with a **k**-point sampling of $27\times 27\times 27$. The Bethe-Salpeter equation (BSE) is carried out on top of GW$_0$ calculation with the Tamm-Dancoff approximation [@Albrecht1998; @Rohlfing1998; @Sander2015]. No profound effects on the optical properties are found by including the excitonic effects. Thus we present the GW$_0$ results hereafter. direction $\omega_{p}^{intra}$ (eV) $\omega_{p}^{inter}$ (eV) $\sigma$ (kS/m) ----------- --------------------------- --------------------------- ----------------- $x$ 23.17 24.06 1097 $z$ 54.61 17.20 6094 \[plasma\] The calculated imaginary part of dielectric function of $I4_1/amd$ hydrogen at 495 GPa is present in Figure \[2\](b). Here we use current-current correlation function instead of density-density correlation function to account for the contribution of Drude terms in metals [@Sangalli2012]. The dielectric function diverges at zero frequency because of the free-electron contribution. We also list the plasma frequency from intraband and interband transitions, as well as the electrical conductivity $\sigma$ in Table \[plasma\]. The calculated total plasma frequency along the $x$ direction is 33.40 eV, agreeing well with the experimental value of 32.5$\pm$2.1 eV [@Dias2017]. Mechanical properties ===================== Mechanical stability -------------------- To evaluate the mechanical stability of $I4_1/amd$ hydrogen at 495 GPa, we calculate the elastic tensor coefficients of the primitive cell including ionic relaxations using the finite differences method [@LePage2002; @Wu2005] with a $\Gamma$-centered 34$\times$34$\times$34 **k**-mesh. The elastic coefficients with the contributions for distortions with rigid ions, the contributions from the ionic relaxations and including both, are present in Table \[t2\]. $C_{11}$ (GPa) $C_{33}$ (GPa) $C_{44}$ (GPa) $C_{66}$ (GPa) $C_{12}$ (GPa) $C_{13}$ (GPa) ------------------- ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- -- Rigid ions 1974.22 1249.60 111.33 18.54 -317.16 598.73 Ionic relaxations -13.74 0.0 -45.57 0.0 13.74 0.0 Total 1960.48 1249.60 65.76 18.54 -303.42 598.73 \[t2\] $I4_1/amd$ hydrogen is tetragonal crystal system with $a=b\neq c$. According to Born-Huang’s lattice dynamical theory [@Born1954; @Wu2007], the mechanical stability conditions for tetragonal phase are given by $C_{11} > 0,\ C_{33} > 0,\ C_{44} > 0,\ C_{66} > 0,\ (C_{11} - C_{12}) > 0,\ (C_{11} + C_{33} - 2C_{12}) > 0,\ 2(C_{11} + C_{12}) + C_{33} + 4C_{13} > 0$. The calculated elastic constants of metallic hydrogen satisfy the corresponding Born stability criteria, indicating the $I4_1/amd$ phase is mechanically stable. Bulk modulus and shear modulus ------------------------------ The bulk modulus $B$ describes the material’s response to uniform hydrostatic pressure. For tetragonal structure, the Voigt and Reuss methods are used to evaluate $B$ [@Wu2007; @Bannikov2012; @Yu2015a] $$B_V = \frac{2(C_{11} + C_{12}) + C_{33} + 4C_{13}}{9},$$ $$B_R = \frac{(C_{11} + C_{12}) C_{33} - 2C_{13}^2}{C_{11} + C_{12} + 2C_{33} - 4C_{13}}.$$ The shear modulus $G$ the material’s response to shear stress, and can be given by $$G_V = \frac{4C_{11} - 2C_{12} + 2C_{33} - 4C_{13} + 12C_{44} + 6C_{66}}{30},$$ $$G_R = \frac{15}{18B_V/[(C_{11} + C_{12}) C_{33} - 2C_{13}^2] + 6/(C_{11}-C_{12}) + 6/C_{44} + 3/C_{66}}.$$ The Hill method calculations are as follows: $$B_H = \frac{1}{2} (B_V + B_R),$$ $$G_H = \frac{1}{2} (G_V + G_R).$$ The calculated bulk modulus and shear modulus are shown in Table \[t3\]. $B_V$ (GPa) $B_R$ (GPa) $B_H$ (GPa) $G_V$ (GPa) $G_R$ (GPa) $G_H$ (GPa) $B_H/G_H$ $E$ (GPa) $\nu$ $A^U$ $A^B$ $A^G$ ------------- ------------- ------------- ------------- ------------- ------------- ----------- ----------- ------- ------- ------- -------- 773.18 768.56 770.87 315.11 56.39 185.75 4.15 515.83 0.39 22.94 0.30% 69.64% \[t3\] The bulk modulus of metallic hydrogen is even higher than diamond (443 GPa) and cubic C$_3$N$_4$ (496 GPa). This result can be easily explained by the extreme high synthesis pressure. The shear modulus, which represents the resistance to shear deformation against external forces, is much lower than the bulk modulus. This implies that the shear modulus limits the mechanical stability of hydrogen. The $B/G$ ratio measures the malleability of materials. A high value represents ductility, while a low value is associated with brittleness. The critical value separating ductile and brittle materials is approximately 1.75 [@Yu2015a]. Our calculated $B/G$ ratio is 4.15, revealing that metallic hydrogen is ductile. Young’s modulus and Poisson’s ratio ----------------------------------- The Young’s modulus can be evatulated from $B_H$ and $G_H$ [@Bannikov2012] $$E = \frac{9B_H G_H}{3 B_H + G_H},$$ and the Poisson’s ratio is $$\nu = \frac{3B_H - 2G_H}{2 (3 B_H + G_H)}.$$ The calculated Young’s modulus and Poisson’s ratios are present in Table \[t3\]. Young’s modulus is a mechanical property of linear elastic solid materials, and measures the stiffness of a solid material. A material with a higher Young’s modulus is stiffer, which needs more force to deform compared to a soft material. The high Young’s modulus indicates that metallic hydrogen is a stiff material. The Poisson’s ratio describes the response in the direction orthogonal to uniaxial strain. Poisson’s ratio close to 0 indicates very little lateral expansion when compressed, while a Poisson’s ratio of exactly 0.5 represents a perfectly incompressible material deformed elastically at small strains. For $I4_1/amd$ hydrogen, a Poisson’s ratio of 0.39 suggests it is less compressible. Elastic anisotropy ------------------ In most single crystals, the elastic response is usually anisotropic. Elastic anisotropy exhibits a different bonding character in different directions and is important in diverse applications such as phase transformations, dislocation dynamics. The universal anisotropy index, which represents a universal measure to quantify the single crystal elastic anisotropy, is defined as [@Bannikov2012; @Ranganathan2008] $$A^U = 5\frac{G_V}{G_R} + \frac{B_V}{B_R} - 6,$$ $A^U$ is identically zero for locally isotropic single crystals. The departure of $A^U$ from zero defines the extent of single crystal anisotropy. Another way implies the estimation in compressibility and shear $$A^B = \frac{B_V-B_R}{B_V+B_R},$$ $$A^G = \frac{G_V-G_R}{G_V+G_R}.$$ A value of zero denotes elastic isotropy and a value of 100% represents the largest anisotropy. The calculated anisotropy indexes are obtained in Table \[t3\]. According to universal anisotropy index $A^U$, the average elastic anisotropy is more than two times higher than Li [@Ranganathan2008]. Comparing $A^B$ and $A^G$, there is much more anisotropy in shear than in compressibility. Sound velocity and Debye temperature ------------------------------------ ------------ --------------- --------------- ------------ --------------- --------------- ------------ --------------- --------------------- -- -- -- (km/s) \[001\] (km/s) \[110\] (km/s) $[100]v_l$ $[001]v_{t1}$ $[010]v_{t2}$ $[001]v_l$ $[100]v_{t1}$ $[010]v_{t2}$ $[110]v_l$ $[001]v_{t1}$ $[1\bar{1}0]v_{t2}$ 36.65 6.71 3.56 29.26 3.56 3.56 24.09 6.71 27.85 ------------ --------------- --------------- ------------ --------------- --------------- ------------ --------------- --------------------- -- -- -- \[t4\] The sound velocity is determined by the symmetry of the crystal and the propagation direction. The tetragonal symmetry dictates that pure transverse and longitudinal modes can only exist for \[100\], \[001\] and \[110\] directions. In all other directions the propagating waves are either quasi-transverse or quasi-longitudinal. In the principal directions the acoustic velocities can be simply written as [@Feng2012] For \[100\] direction: $$[100]v_l = \sqrt{C_{11}/\rho},\ \ [001] v_{t1} = \sqrt{C_{44}/\rho},\ \ [010] v_{t2} = \sqrt{C_{66}/\rho};$$ For \[001\] direction: $$[001]v_l = \sqrt{C_{33}/\rho},\ \ [100] v_{t1} = [010] v_{t2} = \sqrt{C_{66}/\rho};$$ For \[110\] direction: $$[110]v_l = \sqrt{(C_{11}+C_{12}+2C_{66})/\rho},\ \ [001] v_{t1} = \sqrt{C_{44}/\rho},\ \ [1\bar{1}0] v_{t2} = \sqrt{(C_{11}-C_{12})/2\rho}.$$ The calculated sound velocities are presented in Table \[t4\]. $v_l$ (km/s) $v_t$ (km/s) $v_s$ (km/s) $\theta_D$ (K) -------------- -------------- -------------- ---------------- -- -- -- -- -- 26.42 11.28 12.75 3626.64 : Calculated average longitudinal sound velocity $v_l$, transverse sound velocity $v_t$, average sound velocity $v_s$ and Debye temperature $\theta_D$ for the primitive cell of $I4_1/amd$ hydrogen. \[t5\] The Debye temperature $\theta_D$ can be calculated as follows [@Shao2016] $$\label{Debye-stiffness} \theta_D=\frac{h}{k_B} \bigg( \frac{3N}{4\pi V} \bigg)^{1/3} v_s,$$ where $h$ is the Planck constant, $k_B$ is the Boltzmann constant, $N$ is the number of atoms in the cell, $V$ is the volume of the unit cell, and $v_s$ is the average sound velocity given by $$v_s=\bigg[\frac{1}{3} \big( \frac{1}{v_l^3} + \frac{2}{v_t^3} \big) \bigg]^{-1/3},$$ The average sound velocity in crystals can be determined from $B$ and $G$ [@Yu2015a], $$v_l=\sqrt{\frac{3B+4G}{3\rho}},$$ $$v_t=\sqrt{\frac{G}{\rho}}.$$ The calculated sound velocities and Debye temperature are shown in Table \[t5\]. The speed of sound can be used to measure the speed of phonons propagating through the lattice. The ultrahigh longitudinal sound velocity of metallic hydrogen indicates high phonon velocity. The Debye temperature measures the temperature above which all modes begin to be excited [@Nakashima1992]. Thus a high $\theta_D$ indicates weak three-phonon scattering and hence a high lattice thermal conductivity. Phonon properties ================= Dynamical stability ------------------- The Born-Huang mechanical stability criteria provide a necessary condition for the dynamical stability, but not a sufficient one [@Zhou2014a]. Therefore we need to examine the dynamical stability of $I4_1/amd$ hydrogen at 495 GPa by calculating the phonon dispersion of the primitive cell. The harmonic interatomic force constants are obtained using density functional perturbation theory (DFPT) within a supercell approach [@DFPT]. A 7$\times$7$\times$7 supercell with 5$\times$5$\times$5 **q**-mesh is used. The phonon dispersion and thermodynamical properties are calculated from the interatomic force constants using the PHONOPY code [@Togo2008; @Togo2015]. {width="\linewidth"} As shown in Figure \[f2\](a), no imaginary frequencies exist in the whole Brillouin zone, indicating dynamical stability at 0 K. The calculated Raman active modes are 1200.24 cm$^{-1}$ ($E_g$) and 2692.70 cm$^{-1}$ ($B_{1g}$) respectively, which can be used as a reference for characterization of $I4_1/amd$ hydrogen at 495 GPa. The calculated phonon dispersion agrees well with previous result [@McMahon2011]. Thermodynamical properties -------------------------- For light elements such as hydrogen, phonons play an important role in determining the thermodynamical properties of crystals both at 0 K and at finite temperatures [@Setten2007]. Using phonon frequencies in the Brillouin zone, we further examine the thermodynamical properties of metallic hydrogen by calculating the Helmholtz free energy $F$ [@Setten2007], $$F = E_{tot} + \frac{1}{2} \sum\limits_{\textbf{q}j}\hbar\omega_{\textbf{q}j}+k_BT\sum\limits_{\textbf{q}j}\ln[1-\exp(-\hbar \omega_{\textbf{q}j}/k_BT)],$$ where $E_{tot}$ is the total energy of the crystal, $\hbar$ is the reduced Planck constant, $\omega_{\textbf{q}j}$ is the phonon frequency of the $j$-th branch with wave vector **q**, $T$ is the temperature, and the summation term is the Helmholtz free energy for phonons [@Togo2008; @Togo2015]. The first summation term is a temperature-free term corresponding to the zero point energy of phonons; and the second summation term is a temperature-dependent term referring to the thermally induced occupation of the phonon modes. The calculated zero point energy of $I4_1/amd$ hydrogen at 495 GPa is 0.295 eV/atom, which is the Helmholtz free energies of phonons at 0 K. Temperature is also an important thermodynamic variable for determining the stability of materials. The Helmholtz free energies $F$ as a function of temperature are shown in Figure \[f2\](b). At higher temperature, the phonon modes are occupied according to Bose-Einstein statistics, and the free energy further decreases. From the Helmholtz free energy, the other thermodynamical behavior can be deduced [@Born1954]. The entropy is $$S=-\frac{\partial F}{\partial T}.$$ The calculated entropies for the four structures as a function of temperature are shown in Fig. \[f2\](c). The difference between the entropies of different phases can be used to determine the relative stability [@Pavone1998; @McMahon2011a]. The isometric heat capacity can be calculated as $${C_V}=\sum_{\textbf{q}j}k_B\left(\frac{\hbar\omega_{\textbf{q}j}}{k_BT}\right)^2\frac{\exp(\hbar\omega_{\textbf{q}j}/k_BT)}{[\exp(\hbar\omega_{\textbf{q}j}/k_BT)-1]^2}.$$ The calculated heat capacities per atom are shown in Figure \[f2\](d), which approach their Dulong-Petit classical limit (2$\times$24.94 J/K$mol$) at high temperatures. Lattice anharmonicity --------------------- Accurate simulation of anharmonic properties such as thermomechanics and thermal expansion is important for understanding thermal behaviours in solids and their realistic applications. The anharmonic properties can be derived from the volume dependences of phonon dispersion using the quasiharmonic approximation method [@Togo2008; @Togo2015], and temperature is assumed to indirectly affect vibrational properties via thermal expansion. Using the same supercell and **q**-mesh in previous phonon calculations, we calculate the phonon spectra of ten volumes, and the thermal expansion and thermomechanics are derived by fitting the free energy-volume relationship. {width="\linewidth"} Gibbs free energy $G$ at given temperature $T$ and pressure $p$ is obtained from Helmholtz free energy $F$ via finding a minimum value by changing volume $V$ [@Togo2015], $$G(T,p)=\textrm{Min}_V[F(T;V)+pV],$$ where $F$ is the sum of electronic internal energy and phonon Helmholtz free energy. The calculated $G$ is depicted in Figure \[f3\](a). The minimum value is obtained by the fits to third-order Birch-Murnaghan equation of states. The calculated isothermal bulk modulus is present in Figure \[f3\](b). An anticipated thermal softening is observed in the temperature dependence of $B_T$. From 0 to 1000 K, the softening of the isothermal bulk modulus for metallic hydrogen is 157.98 GPa, which is much larger than those in C and Al$_2$O$_3$ [@Huang2016]. The Grüneisen parameter $\gamma$ describes the effect of thermal expansion on vibrational properties, and provides information on the anharmonic interactions. A larger Grüuneisen parameter indicates stronger anharmonic vibrations. As shown in Figure \[f3\](c), the calculated $\gamma$ at room temperature is 0.84. The volume thermal expansion coefficient $\beta$ can be calculated from the Grüneisen parameter [@Grimvall] $$\beta = \frac{\gamma C_V}{B_T V}.$$ Another approach is to obtain $\beta$ from the calculated equilibrium volumes $V$ at temperature $T$ $$\beta = \frac{1}{V(T)}\frac{\partial V(T)}{\partial T}.$$ As shown in Figure \[f3\](d), the $\beta$ curves from these methods are in excellent agreement with each other. It should be noticed that it is challenging to measure $\beta$ accurately in experiment because thermal expansion varies with the crystalline orientation and there may be some metastable phases of metallic hydrogen at relative high temperatures. Phonon transport ---------------- The lattice thermal conductivity $\kappa$ can be calculated using the self-consistent iterative approach [@Omini1996; @ShengBTE; @Li2012a; @Li2012; @Li2013a] as a sum of contribution of all the phonon modes $\lambda$, $$\label{kappa-eq} \kappa^{\alpha\alpha}=\frac{1}{V}\sum_{\textbf{q}j}C_{\textbf{q}j} \tau_{\textbf{q}j} {v_{\textbf{q}j}^{\alpha}}^2,$$ where $C_{\textbf{q}j}$ is the heat capacity per mode, $\tau_{\textbf{q}j}$ is the mode relaxation time, and $v_{\textbf{q}j}^{\alpha}$ is the group velocity along $\alpha$ direction, respectively. The lattice dynamical properties $C_{\textbf{q}j}$ and $v_{\textbf{q}j}^{\alpha}$ in Eq. (\[kappa-eq\]) can be obtained by the phonon dispersion relation with harmonic interatomic force constants as input, while the $\tau_{\textbf{q}j}$ provides information about the anharmonic interactions and can be obtained using the anharmonic interatomic force constants. The anharmonic interatomic force constants are calculated using a supercell-based, finite-difference method [@Li2012], and a 7$\times$7$\times$7 supercell is used with a 5$\times$5$\times$5 **q**-mesh. We include the interactions with the 40th nearest-neighbour atoms. A discretizationa of the Brillouin zone into a $\Gamma$-centered regular grid of 32$\times$32$\times$32 $\textbf{q}$ points is introduced with scale parameter for broadening chosen as 0.1. When the phonon $\textbf{q}$-points mesh increases from 28$\times$28$\times$28 to 32$\times$32$\times$32, the difference of $\kappa$ between the two grids is less than 1%, verifying the convergence of our calculations. {width="\linewidth"} Figure \[f4\](a) presents the phonon group velocities calculated from the phonon spectrum within the whole Brillouin zone. The group velocity shows strong anisotropy along the $x$ and $z$ directions. The phonon group velocities in long-wavelength limit for the acoustic modes agree well with the sound velocities calculated from elastic coefficients. The phonon scattering depends on two factors: the number of channels available for a phonon to get scattered, and the strength of each scattering channel. The former factor is determined by whether there exist three phonon groups that can satisfy both energy and quasi-momentum conservations. The latter factor depends on the anharmonicity of a phonon mode, which is described by the mode Grüneisen parameter. The mode Grüneisen parameters provide information about anharmonic interactions, and can be calculated directly from the anharmonic interatomic force constants. Figure \[f4\](b) shows the mode Grüneisen parameters. Strong anisotropy is found in the Grüneisen parameters at low frequencies. To provide more physical insight, we investigate the phonon relaxation time of each phonon mode at 300 K in Figure \[f4\](c). The relaxation times are highly anisotropic at low frequencies as well. The thermal conductivity for $I4_1/amd$ hydrogen at 495 GPa at 300 K are higher than 170 W/mK along the $x$ and $z$ directions, as listed in Table \[thermal\]. We extract the contributions of different phonon branches to $\kappa$ along the $x$ and $z$ directions, as shown in Table \[thermal\]. Here the different modes are simply distinguished by frequency [@Peng2016]. For phonon transport along $x$ direction, the contribution of TA$_2$ phonons is largest among all phonon modes. For phonon transport along $z$ direction, TA$_1$ phonons contribute the most to $\kappa$. For both directions, the acoustic phonons dominate the heat transport. direction $\kappa$ (W/mK) TA$_1$ (%) TA$_2$ (%) LA (%) Optical (%) ----------- ----------------- ------------ ------------ -------- ------------- $x$ 194.72 28.6 46.8 24.3 0.3 $z$ 172.96 55.6 22.7 21.4 0.3 \[thermal\] The intrinsic lattice thermal conductivity $\kappa$ of $I4_1/amd$ hydrogen at 495 GPa is present in Figure \[f4\](d). The $\kappa$ is weakly anisotropic ($\kappa_{xx}=\kappa_{yy}\neq\kappa_{zz}$). The anisotropy in thermal transport can be attributed to the anisotropic natures of phonon dispersion of acoustic modes, which results in the orientation-dependent group velocities and Grüneisen parameters as mentioned above. The orientation-dependence of thermal transport properties provides a way to determine the optimized transport directions for potential applications. Conclusion ========== State-of-the-art computational material techniques enable a systematic investigation of different properties of metallic hydrogen. The calculated electronic structure and dielectric function confirm the metallic behaviour of $I4_1/amd$ hydrogen at 495 GPa. The calculated total plasma frequency from intraband and interband transitions of 33.40 eV agrees well with the experimental value. The phase is mechanical stable, and the stability is limited by shear modulus other than bulk modulus. The high Young’s modulus indicates that $I4_1/amd$ hydrogen is a stiff material, while the Poisson’s ratio suggests it is less compressible. The elastic anisotropy, sound velocity and Debye temperature are also investigated in detail. The dynamical stability of $I4_1/amd$ hydrogen is confirmed by phonon dispersion. The thermodynamical properties, as well as lattice anharmonicity are also calculated to understand thermal behaviours in metallic hydrogen. The lattice thermal conductivity of $I4_1/amd$ hydrogen is 194.72 W/mK and 172.96 W/mK along the $x$ and $z$ directions, respectively, where acoustic phonons dominate heat transport. Our results show the power of first-principles calculations in predicting a variety of material properties in extreme conditions where purely experimental approaches are impractical, and pave way for potential applications of metallic hydrogen at extremely high pressure and high temperature. Acknowledgement {#acknowledgement .unnumbered} =============== This work is supported by the National Natural Science Foundation of China under Grants No. 11374063 and 11404348, and the National Basic Research Program of China (973 Program) under Grant No. 2013CBA01505. References {#references .unnumbered} ==========

ArXiv

--- abstract: 'In these lectures I want to discuss how the structure functions in deep inelastic scattering relate to quark and gluon correlation functions. In particular we will consider the issue of intrinsic transverse momenta of quarks, which becomes important in processes like 1-particle inclusive leptoproduction. Some examples of cross sections and asymmetries, in particular in polarized scattering processes are discussed.' author: - 'P.J. MULDERS' title: STRUCTURE OF HADRONS IN HARD PROCESSES --- / Introduction ============ The central point of these lectures is the availability of a field theoretical framework for the strong interactions. It is the nonabelian gauge theory based on the color symmetry group $SU(N_c)$ with $N_c$ = 3. This theory, Quantum Chromodynamics (QCD) underlies the strong interactions of which the existence is known since the thirties. At that time the goal was to understand the forces in the atomic nucleus. The experimental results of many experiments revealed the existence of a rich spectrum of hadrons, baryons and mesons. The presence of these excitation spectra and also explicit measurements of the charge and current distributions of the hadrons revealed the composite nature of hadrons. The quark model brought some order into this situation, describing baryons as composite systems of three valence quarks and mesons as a composite system of a valence quark and antiquark. Valence refers here to the fact that these quarks are the contributors to the quantum numbers (upness, downness, strangeness, etc.) of the hadrons. The symmetry considerations based on flavor $SU(3)$ and flavor-spin $SU(6)$ are the basis of the success of the quark model. Color was introduced in a rather early stage to solve a number of problems such as the fermion statistics of the quarks. The experimental results of the SLAC-MIT deep inelastic electron scattering experiments, which indicated the existence of hard ’point-like’ constituents in the nucleon and the field-theoretical developments in nonabelian gauge theories, specifically the notion of asymptotic freedom and the proof of renormalizability, led to the natural emergence of QCD as the theory for the strong interactions between the quarks and gluons. The important feature of QCD is the fact that the force becomes weaker at short distances. This anti-screening behavior or asymptotic freedom is a unique feature of non-abelian gauge theories. QCD is part of the Standard Model that describes the strong and electroweak forces. All forces in the Standard Model are described within the framework of nonabelian gauge theories, based on a gauge symmetry. For testing directly QCD, one preferably avoids the presence of hadrons, as these constitute complex bound state systems. Useful scattering processes for this purpose are aaaaaaaaaaaaaaaaaaaaaaaaaabbbbbbbb $e^+ e^- \longrightarrow X$ (inclusive annihilation),\ $e^+ e^- \longrightarrow$ jets (jet production),\ $\ell H \longrightarrow \ell^\prime X$ (inclusive leptoproduction). Analyzing jets produced in $e^+e^-$ scattering represent, modulo complications arising from hadronization, hard scattering processes $e^+e^-$ $\rightarrow$ $q \bar q$, $e^+e^-$ $\rightarrow$ $q \bar q g$, etc. For these processes perturbative QCD can be used to do reliable calculations. Similarly one can use perturbative QCD to study the $\ln Q^2$ dependence of the elementary $\ell q$ $\rightarrow$ $\ell^\prime q$ cross sections that are important in leptoproduction. The properties of hadrons are poorly described directly starting from QCD. They require a nonperturbative approach in quantum field theory. This has led to the use of models that incorporate some features like confinement and asymptotic freedom and symmetries of the underlying theory, such as quark potential models, bag models or chiral models. Most promising from a fundamental point of view are lattice gauge approaches. Controlling and selectively probing the nonperturbative regime in high energy scattering processes is the key to study the structure of hadrons in the context of QCD. The control parameters for the target and the probe are the spin and flavor, which in combination with the kinematical flexibility in scattering processes is used to select the observable and its gluonic or quarkic nature. Examples are aaaaaaaaaaaaaaaa = bbbbbbbbbbbbbbbbbbbbbb = cccccccccccc $\ell H \longrightarrow \ell^\prime H$ (elastic leptoproduction) (spacelike) form factors,\ $\ell H \longrightarrow \ell^\prime X$ (inclusive leptoproduction) distribution functions,\ $\ell H \longrightarrow \ell^\prime h X$ (1-particle inclusive distribution and\ leptoproduction) fragmentation functions,\ $e^+ e^- \longrightarrow h \bar h$ (annihilation into $h\bar h$) (timelike) form factors,\ $e^+ e^- \longrightarrow h X$ (1-particle inclusive fragmentation functions,\ annihilation)\ $H_1 H_2 \longrightarrow \mu^+ \mu^- X$ (Drell-Yan scattering) distribution functions. These notes consist of two parts. After this introduction we discuss the basic formalism to deal with the beforementioned processes, introducing the hadron tensor, form factors, structure functions. These are needed to write down general expressions for the electroweak cross sections. The second part reviews the introduction of parton distribution and fragmentation functions that enable a systematic treatment of the structure functions at high energies and/or momentum transfer. Inclusive Leptoproduction ========================= The hadron tensor ----------------- For the process $\ell + H \rightarrow \ell^\prime + X$ (see Fig. \[fig0\]), the cross section can be separated into a lepton and hadron part. Although the lepton part is simpler, let us start with the hadron part, &&2MW\_\^[(H)]{}( q; [P S]{} ) = \_X (2)\^4 \^4 (q + P - P\_X)\ && |[J\_(0)]{}|P\_X P\_X |[J\_(0)]{}|[P S]{} , $$\begin{aligned} &&\ell H \longrightarrow \ell^\prime X\end{aligned}$$ $$\begin{aligned} &&{x_{_B}}= \frac{Q^2}{2P\cdot q} \\ &&y = \frac{P\cdot q}{P\cdot k}\end{aligned}$$ The simplest thing to do is to parametrize this tensor in standard tensors and structure functions. Instead of the traditional choice [@Roberts] for these tensors, $g_{\mu\nu}$ and $P_\mu P_\nu$ and structure functions $W_1$ and $W_2$, we immediately go to a dimensionless representation, using a natural space-like momentum (defined by $q$) and a time-like momentum constructed from $P$ and $q$, z\^= -q\^&=& -,\ t\^&=& = (P\^- q\^) = . Using hermiticity for the currents, parity invariance and current conservation one obtains as the most general form the symmetric tensor MW\_S\^([q,P]{}) = \_[-g\_\^]{}[F\_1]{} + t\^t\^\_[F\_L]{} , \[param\] where the structure functions $F_1$ and $F_2$ or the transverse and longitudinal structure functions, $F_T = F_1$ and $F_L$, depend only on the for the hadron part relevant invariants $Q^2$ and ${x_{_B}}$. In all equations given here we have omitted target mass effects of order $M^2/Q^2$. The lepton tensor ----------------- In order to write down the cross section one needs to include the necessary phase space factors and include the lepton part given by the tensor L\_\^[(H)]{} (k ; k\^\^) = 2 k\_k\^\_ + 2 k\_k\^\_ - Q\^2 g\_ +2i\_e \_ q\^k\^ . We have included here the (longitudinal) lepton polarization ($\lambda_e = \pm 1$). For later convenience it is useful to rewrite this tensor also in terms of the space-like and time-like vectors $\hat q$ and $\hat t$. It is a straightforward exercise to get k\^= q\^+ t\^+ \^, where $\hat \ell$ is the perpendicular direction defining the lepton scattering plane (see Fig. \[fig1\]). This perpendicular direction becomes relevant only if other vectors than $P$ and $q$ are present, e.g. a spin direction in polarized scattering or the momentum of a produced hadron in 1-particle inclusive processes. The lepton tensor becomes L\^\_[(H)]{} & = & , \[leptontensor\] where $\epsilon_\perp^{\mu\nu} \equiv \epsilon^{\mu\nu\rho\sigma}\hat t_\rho \hat q_\sigma$. The inclusive cross section --------------------------- The cross section for unpolarized lepton and hadron only involves the first two (symmetric) terms in the lepton tensor and one obtains & = & { ( 1-y + y\^2) [F\_T([x\_[\_B]{}]{},Q\^2)]{}\ && + ( 1 -y) [F\_L([x\_[\_B]{}]{},Q\^2)]{}}. As soon as the exchange of a $Z^0$ boson becomes important the hadron tensor is no longer constrained by parity invariance and a third structure function $F_3$ becomes important. Target polarization ------------------- The use of polarization in leptoproduction provides new ways to probe the hadron target. For a spin 1/2 particle the initial state is described by a 2-dimensional spin density matrix $\rho = \sum_\alpha \vert \alpha\rangle p_\alpha \langle \alpha\vert$ describing the probabilities $p_\alpha$ for a variety of spin possibilities. This density matrix is hermitean with Tr$\,\rho$ = 1. It can in the target rest frame be expanded \_[ss\^]{} = ( 1 + S\_[ss\^]{}), where ${\mbox{\boldmath $S$}}$ is the spin vector. When $\vert {\mbox{\boldmath $S$}}\vert = 1$ one has a pure state (only one state $\vert \alpha\rangle$ and $\rho^2 = \rho$), when $\vert {\mbox{\boldmath $S$}}\vert \le 1$ one has an ensemble of states. For the case $\vert {\mbox{\boldmath $S$}}\vert = 0$ one has simply an averaging over spins, corresponding to an unpolarized ensemble. To include spin one could generalize the hadron tensor to a matrix in spin space, $\tilde W_{s^\prime s}^{\mu\nu}(q,P)$ $\propto$ $\langle P, s^\prime\vert J^\mu\vert X\rangle\langle X\vert J^\nu\vert P, s>$ depending only on the momenta or one can look at the tensor $\sum_\alpha p_\alpha \tilde W_{\alpha\alpha}^{\mu\nu}(q,P)$, which is given by W\^(q,P,S) = ( (P,S) W\^(q,P) ), \[defS\] with the spacelike spin vector $S$ appearing [*linearly*]{} and in an arbitrary frame satisfying $P\cdot S = 0$. It has invariant length $-1 \le S^2 \le 0$. It is convenient to write the spin vector as S\^= ( P\^- q\^) + S\_\^, \[inclspin\] with $\lambda = M(S\cdot q)/(P\cdot q)$. For a pure state one has $\lambda^2 + {\mbox{\boldmath $S$}}_\perp^2 = 1$. Using symmetry constraints one obtains for electromagnetic interactions (parity conservation) an antisymmetric part in the hadron tensor, MW\_A\^([q,P,S]{}) = \_[-i\_\^]{} [g\_1]{} + i t\_\^[\[]{}\_\^[\]]{} [S\_]{} [g\_T]{}. \[wanti\] The polarized part of the cross section becomes & = & \_e { ( 1- )[g\_1([x\_[\_B]{}]{},Q\^2)]{}\ && -\_S\^ [g\_T([x\_[\_B]{}]{},Q\^2)]{}}. Semi-inclusive leptoproduction ============================== The hadron tensor ----------------- More flexibility in probing new aspects of hadron structure is achieved in semi-inclusive scattering processes. For instance in 1-particle inclusive measurements one can measure azimuthal dependences in the cross sections. The central object of interest for 1-particle inclusive leptoproduction, the hadron tensor, is given by &&2M[W]{}\_\^[(H)]{}( q; [P S; P\_h S\_h]{} ) = \_X \^4 (q + P - P\_X - P\_h)\ && |[J\_(0)]{}|P\_X; [P\_h S\_h]{} P\_X; [P\_h S\_h]{} |[J\_(0)]{}|[P S]{} , where $P,\ S$ and $P_h,\ S_h$ are the momenta and spin vectors of target hadron and produced hadron, $q$ is the (space-like) momentum transfer with $-q^2$ = $Q^2$ sufficiently large. The kinematics is illustrated in Fig. \[fig1\], where also the scaling variables are introduced. $$\begin{aligned} &&\ell H \longrightarrow \ell^\prime h X\end{aligned}$$ $$\begin{aligned} &&{x_{_B}}= \frac{Q^2}{2P\cdot q} \\ &&y = \frac{P\cdot q}{P\cdot k}\\ &&z_h = \frac{P\cdot P_h}{P\cdot q}\\ &&Q_\st = \frac{\vert P_{h\perp}\vert}{z_h}\end{aligned}$$ For the parametrization of the hadron tensor in terms of structure functions it is useful to introduce the directions $\hat q$ and $\hat t$ as before and using the vector $P_h$ to construct a vector that is orthogonal to these vectors. For the situation that $P\cdot P_h$ is ${\cal O}(Q^2)$ (current fragmentation!) one finds that q\_\^= q\^+[x\_[\_B]{}]{}P\^- = - -Q\_Th\^, is such a vector. This vector is proportional to the transverse momentum of the outgoing hadron with respect ot $P$ and $q$. It can also be considered as the transverse momentum of the photon with respect to the hadron momenta $P$ and $P_h$. For an unpolarized (or spin 0) hadron in the final state the symmetric part of the tensor is given by M[W]{}\_S\^([q,P,P\_h]{}) &=& - g\_\^[H]{}\_T + t\^t\^[H]{}\_L\ && + t\^[{]{}h\^[}]{}[H]{}\_[LT]{} + 2h\^h\^+ g\_\^\_[TT]{} . Noteworthy is that also an antisymmetric term in the tensor is allowed, M[W]{}\_A\^([q,P,P\_h]{}) = - it\^[\[]{}h\^[\]]{}[H]{}\^\_[LT]{}. \[sidiswanti\] The semi-inclusive cross section -------------------------------- Clearly the lepton tensor in Eq. \[leptontensor\] is able to distinguish all the structures in the semi-inclusive hadron tensor. The symmetric part gives the cross section for unpolarized leptons, & = & [x\_[\_B]{}]{}[z\_h]{}{ ( 1-y+y\^2 )[H]{}\_T + (1-y)[H]{}\_L\ && - (2-y)\_h\^[H]{}\_[LT]{}\ && + (1-y)2\_h\^[H]{}\_[TT]{} } while the antisymmetric part gives the cross section for a polarized lepton (note the target is not polarized!) = \_e[z\_h]{}\_h\^[H]{}\^\_[LT]{} . Of course many more structure functions appear for polarized targets or if one considers polarimetry in the final state. Form factors ============ A special case is the situation in which the final state is identical to the initial state, elastic scattering. In that case the final state four momentum is $P^\prime = P+q$ and is fixed to be $(P+q)^2 = M^2$, i.e. ${x_{_B}}= 1$. We can still use the formalism for inclusive leptoproduction but the hadron tensor becomes becomes 2MW\_(q,P) = \_[H\_(P;P\^)]{} (1-[x\_[\_B]{}]{}). One needs the current matrix elements of the electromagnetic current, which using hermiticity, parity and current conservation can be parametrized in terms of (real) form factors, for a spin 1/2 particle P\^, S\^J\_(x) P,S = e\^[iqx]{}|U\_[S\^]{} (P\^) \_[\_(P,P\^)]{} U\_S(P), where $U_S(P)$ are the standard Dirac spinors. In order to interpret the form factors for spacelike $q$, note that in leptoproduction there always exist a frame in which $q$ is purely spacelike ($q^0$ = 0), the socalled brick-wall or Breit frame. Working out the current expression for the nucleon in this frame ($P^{\prime 0} = P^0$, ${\mbox{\boldmath $P$}} = -{\mbox{\boldmath $q$}}/2$, ${\mbox{\boldmath $P$}}^\prime = +{\mbox{\boldmath $q$}}/2$), gives P\^, S\^J\_0\^[em]{} P,S & = & 2Me\^[iqx]{}\ & & 2M G\_E(Q\^2) e\^[iqx]{},\ P\^, S\^\^[em]{} P,S & = & (i\_N q) e\^[iqx]{}\ & & G\_M(Q\^2)(i\_N q) e\^[iqx]{}, where $\chi_{S^\prime}^\dagger\,{{\mbox{\boldmath $\sigma$}}}\,\chi_S$ $\equiv$ ${{\mbox{\boldmath $\sigma$}}}_N$. These expressions show the relevance of the Sachs form factors $G_E$ and $G_M$ being the Fourier transfer of the spatial charge and current distribution. The quantity $e\,G_E(0)$ is the charge of the nucleon, $e\,G_M(0)/M$ is the magnetic moment of the nucleon. The quantity $\kappa$ = $F_2(0)$ = $G_M(0)-G_E(0)$ is the anomalous magnetic moment. > One has for a point-particle (e.g. electron or muon) (in lowest order in $\alpha$) $$F_1(0) = G_E(0) = 1 , \qquad F_2(0) = 0, \qquad G_M(0) = 1,$$ and no $Q^2$-dependence, while for a composite particle like the nucleon one has a combination of quark form factors weighted by the charges, giving $$\begin{aligned} > && F_1^p(0) = G_E^p(0) = 1, \qquad F_2^p(0) = \kappa_p \approx 1.79, > \quad \qquad G_M^p = \mu_p \approx 2.79,\\ > && F_1^n(0) = G_E^n(0) = 0, \qquad F_2^n(0) = \kappa_n \approx -1.91, > \qquad G_M^n = \mu_n \approx -1.91.\end{aligned}$$ The $Q^2$-dependence for the electromagnetic form factors of the nucleon approximately is given by $$G_E^p(Q^2) \approx \frac{G_M^p(Q^2)}{\mu_p} \approx > \frac{G_M^n(Q^2)}{\mu_n} > \approx \frac{1}{\left( 1 + Q^2/0.69\ \mbox{GeV}^2\right)^2}$$ (dipole form factors). For the tensor $H_{\mu \nu}$ one obtains the result $$\begin{aligned} H_{\mu \nu}(P,P^\prime) & = & \frac{1}{2} \sum_{S,S^\prime} \langle P,S \vert J_\mu (0) \vert P^\prime,S^\prime \rangle \langle P^\prime,S^\prime \vert J_\nu(0) \vert P,S \rangle \nonumber \\ & = & \frac{1}{2}\,Tr\ \left[\Gamma_\mu (\slash P^\prime + M) \Gamma_\nu (\slash P + M) \right] \nonumber \\ & = & \left( \frac{q_\mu q_\nu}{q^2} - g_{\mu \nu} \right) \,Q^2 (F_1 + F_2)^2 + 4\,\tilde P_\mu \tilde P_\nu\,\left( F_1^2 + \frac{Q^2}{4M^2}\,F_2^2 \right) \nonumber \\ & = & - g_{\perp\,\mu \nu}\,Q^2 G_M^2 + \hat t_\mu \hat t_\nu\,4M^2\,G_E^2\end{aligned}$$ One thus sees the following [*elastic*]{} contribution in the structure functions &&F\_T([x\_[\_B]{}]{},Q\^2) = G\_M\^2(Q\^2)(1-[x\_[\_B]{}]{}),\ &&F\_L([x\_[\_B]{}]{},Q\^2) = G\_E\^2(Q\^2)(1-[x\_[\_B]{}]{}). In particular when one is also considering other currents than the electromagnetic case, it is useful to realize that the currents for the $\gamma$, $Z$- or $W$-particles of course all are known in terms of quark vector and axial vector currents, e.g. &&J\_\^[()]{}(x) = \_[q]{}QV\_\^[(q)]{}(x)\ &&J\_\^[(Z)]{}(x) = \_[q]{}(I\_W\^3 - 2Q\^2\_W ) V\_\^[(q)]{}(x) - I\_W\^3A\_\^[(q)]{}(x),\ &&J\_\^[(W\^)]{}(x) = \_[q]{} I\_W\^(V\_\^[(q)]{}(x) - A\_\^[(q)]{}(x)), where the vector and axial vector quark currents are &&V\_\^[(q)]{}(x) = (x)\_(x),\ &&A\_\^[(q)]{}(x) = (x)\_\_5 (x), One can of course also consider a parametrization of these currents in terms of form factors for a particular flavor, again using hermiticity, parity and the conservation of the vector current. For a spin 1/2 particle (e.g. the nucleon) one obtains $$\begin{aligned} &&\langle P^\prime, S^\prime \vert V_\mu (0) \vert P,S \rangle = \bar U_{S^\prime} (P^\prime) \underbrace{\left[ \gamma_\mu\, F^q_1(Q^2) + \frac{i\sigma_{\mu\nu}q^\nu}{2M}\, F^q_2(Q^2) \right]}_{\Gamma_\mu^V(P,P^\prime)} \,U_S(P), \nonumber \\ &&\\ &&\langle P^\prime, S^\prime \vert A_\mu (0) \vert P,S \rangle = \bar u_{S^\prime} (P^\prime) \underbrace{\left[ \gamma_\mu \gamma_5\, G^q_A(Q^2) + \gamma_5\,q_\mu\, G^q_P(Q^2) \right]}_{\Gamma_\mu^A(P,P^\prime)} \,U_S(P). \nonumber \\ &&\end{aligned}$$ Exclusive processes in principle then offer possibilities to measure vector or axial vector form factors for various flavor currents. One can for instance use the $(\gamma Z)$ interference term in electron-nucleon scattering or $\nu_e p \rightarrow e^+n$ processes to separate the currents for different flavors. > Note that the normalizations of the densities for a given flavor imply for the form factor at $Q^2 = 0$, $$F^q_1(0) = G_E^q(0) = n_q ,$$ and one has e.g. $F_1(0) = \sum_q e_q\,n_q = Q$. Note that $n_q$ is the number of quarks minus antiquarks. Other form factors at zero momentum transfer just define some numbers, e.g. $$F^q_2(0) = \kappa_q, \qquad G^q_M(0) = \mu_q, \qquad > G_A^q(0) = g_A^q.$$ One has e.g. $G_M(0) = \sum_q e_q\,\mu_q$. For the axial currents one finds from $\beta-decay$ a proton-neutron transition element of the axial current, that using isospin symmetry can be simply converted into G\_A\^[pn]{}(0) = g\_A\^u - g\_A\^d = 1.26. \[axform\] Quark correlation functions in leptoproduction ============================================== Within the framework of QCD and knowing that the photon or $Z^0$ current couples to the quarks, it is possible to write down a diagrammatic expansion for leptoproduction, with in the deep inelastic limit ($Q^2 \rightarrow \infty$) as relevant diagrams only the ones given in Fig. \[fig2\] for inclusive and 1-particle inclusive scattering respectively. The expression for ${\cal W}_{\mu\nu}$ can be rewritten as a nonlocal product of currents and it is a straightforward exercise to show by inserting the currents $j_\mu(x) = :\overline \psi(x) \gamma_\mu \psi(x):$ that for 1-particle inclusive scattering one obtains in tree approximation && 2M[W]{}\_( q; P S; P\_h S\_h )\ & & = d\^4x e\^[iqx]{} P S |:\_j (x) (\_)\_[jk]{} \_k(x): \_X |X; P\_h S\_h\ &&X; P\_h S\_h | :\_l(0) (\_)\_[li]{} \_i(0):|P S\ & & = d\^4x e\^[iqx]{} P S |\_j (x) \_i(0) P S (\_)\_[jk]{}\ && 0 \_k(x) \_X |X; P\_h S\_h X; P\_h S\_h | \_l(0) 0 (\_)\_[li]{}\ && + d\^4x e\^[iqx]{} P S |\_k (x) \_l(0) P S (\_)\_[li]{}\ &&0 \_j(x) \_X |X; P\_h S\_h X; P\_h S\_h | \_i(0) 0 (\_)\_[jk]{},\ & & = d\^4p d\^4k\^4(p+q-k)  ( (p) \_(k) \_) + { [c]{} q -q\ },\ && \[basic\] where && \_[ij]{}(p) = d\^4 e\^[ip]{} P S |\_j (0) \_i()|P S ,\ && \_[kl]{}(k) = d\^4 e\^[ik]{} 0 \_k() \_X |X; P\_h S\_h X; P\_h S\_h | \_l(0) 0 . Note that in $\Phi$ (quark production) a summation over colors is assumed, while in $\Delta$ (quark decay) an averaging over colors is assumed. The quantities $\Phi$ and $\Delta$ correspond to the blobs in Fig. \[fig2\] and parametrize the soft physics. Soft refers to all invariants of momenta being small as compared to the hard scale, i.e. for $\Phi(p)$ one has $p^2 \sim p\cdot P \sim P^2 = M^2 \ll Q^2$. In general many more diagrams have to be considered in evaluating the hadron tensors, but in the deep inelastic limit they can be neglected or considered as corrections to the soft blobs. We return to this later. As mentioned above, the relevant structural information for the hadrons is contained in soft parts (the blobs in Fig. \[fig2\]) which represent specific matrix elements of quark fields. The form of $\Phi$ is constrained by hermiticity, parity and time-reversal invariance. The quantity depends besides the quark momentum $p$ on the target momentum $P$ and the spin vector $S$ and one must have $$\begin{aligned} && \mbox{[Hermiticity]} \quad \Rightarrow \quad \Phi^\dagger (p,P,S) = \gamma_0 \,\Phi(p,P,S)\,\gamma_0 , \\ && \mbox{[Parity]} \quad \Rightarrow \quad \Phi(p,P,S) = \gamma_0 \,\Phi(\bar p,\bar P,-\bar S)\,\gamma_0 , \\ &&\mbox{[Time\ reversal]} \ \Rightarrow \ \Phi^\ast(p,P,S) = (-i\gamma_5 C)\,\Phi(\bar p,\bar P, \bar S)\,(-i\gamma_5 C) ,\end{aligned}$$ where $C$ = $i\gamma^2 \gamma_0$, $-i\gamma_5 C$= $i\gamma^1\gamma^3$ and $\bar p$ = $(p^0,-{\mbox{\boldmath $p$}})$. The most general way to parametrize $\Phi$ using only the constraints from hermiticity and parity invariance, is [@RS79; @TM95] (p,P,S) & = & MA\_1 + A\_2/P + A\_3 /p + iA\_4\ & & + iA\_[5]{}(pS) \_5 + MA\_6 /S \_5 + A\_7/P \_5\ & & + A\_8/p \_5 + A\_9\_5 + A\_[10]{}\_5\ & & + A\_[11]{}\_5 + A\_[12]{}, \[lorentz\] where the first four terms do not involve the hadron polarization vector. Hermiticity requires all the amplitudes $A_i$ = $A_i(p\cdot P, p^2)$ to be real. The amplitudes $A_4$, $A_5$ and $A_{12}$ vanish when also time reversal invariance applies. Inclusive scattering ==================== The relevant soft parts ----------------------- In order to find out which information in the soft parts is important in a hard process one needs to realize that the hard scale $Q$ leads in a natural way to the use of lightlike vectors $n_+$ and $n_-$ satisfying $n_+^2 = n_-^2 = 0$ and $n_+\cdot n_-$ = 1. For inclusive scattering one parametrizes the momenta $$\left. \begin{array}{l} q^2 = -Q^2 \\ P^2 = M^2\\ 2\,P\cdot q = \frac{Q^2}{{x_{_B}}} \\ \end{array} \right\} \longleftrightarrow \left\{ \begin{array}{l} q =\ \frac{Q}{\sqrt{2}}\,n_- \ - \ \frac{Q}{\sqrt{2}}\,n_+ \\ \mbox{} \\ P = \frac{{x_{_B}}M^2}{Q\sqrt{2}}\,n_- + \frac{Q}{{x_{_B}}\sqrt{2}}\,n_+ \end{array} \right.$$ The above are the external momenta. Next turn to the internal momenta, looking at the left diagram in Fig. \[fig2\]. In the soft part actually [*all*]{} momenta, that is $p$ [*and*]{} $P$ have a minus component that can be neglected compared to that in the hard part, since otherwise $p\cdot P$ would be hard. Thus because $p$ must have only a hard plus component, $q$ has two hard components and $k$ being the current jet also must be soft, i.e. only can have one large lightcone component, one must have $$\begin{aligned} p &=& \quad\ldots\quad + \frac{Q}{\sqrt{2}}\,n_+ , \\ q &=& \frac{Q}{\sqrt{2}}\,n_- - \frac{Q}{\sqrt{2}}\,n_+ , \\ p+q = k &=& \frac{Q}{\sqrt{2}}\,n_- + \quad\ldots\ .\end{aligned}$$ where the …parts indicate (negligible) $1/Q$ terms. Also the transverse component is not relevant for the hard part. One thus sees that for inclusive scattering the only relevant dependence of the soft part is the $p^+$ dependence. Moreover, the above requirements on the internal momenta already indicate that the lightcone fraction $x = p^+/P^+$ must be equal to ${x_{_B}}$. This will come out when we do the actual calculation in one of the next sections. The minus component $p^- \equiv p\cdot n_+$ and transverse components thus can be integrated over restricting the nonlocality in $\Phi(p)$. The relevant soft part then is some Dirac trace of the quantity [@Soper77; @Jaffe83] \_[ij]{}(x) & = & dp\^-d\^2p\_ \_[ij]{}(p,P,S)\ & = & .  e\^[ip]{} P,S\_j(0) \_i() P,S|\_[\^+ = \_= 0]{}, depending on the lightcone fraction $x = p^+/P^+$. To be precise one puts in the full form for the quark momentum, p = xP\^+n\_+ + n\_- + p\_, \[quarkmom\] and performs the integration over $\Phi(p)$ using dp\^-d\^2p\_…= d(pP)dp\^2 … . When one wants to calculate the leading order in $1/Q$ for a hard process, one only needs to look at leading parts in $M/P^+$ because $P^+ \propto Q$ (see opening paragraph of this section) [@JJ92]. In this case that turns out to be the part proportional to $(M/P^+)^0$, (x) = { f\_1(x)[0.2 em n-0.4em /]{}\_+ + g\_1(x) \_5[0.2 em n-0.4em /]{}\_+ + h\_1(x)} + [O]{}() The precise expression of the functions $f_1(x)$, etc. as integrals over the amplitudes can be easily written down. Calculating the inclusive cross section --------------------------------------- Using field theoretical methods the left diagram in Fig. \[fig2\] can now be calculated. Omitting the sum over flavors ($\sum_a$), the quark charges $e_a^2$ and the $(q\leftrightarrow -q, \mu\leftrightarrow \nu)$ ’antiquark’ diagram, the symmetric part of the hadron tensor the result is 2MW\^(P,q) = dp\^-dp\^+d\^2p\_ ((p) \^(p+q) \^), where (k) = (/k + m)(k\^2-m\^2) (k\^+) , and in the approximation anything proportional to $1/Q^2$ has been neglected. One obtains $$\begin{aligned} 2M\,W_S^{\mu \nu}(P,q) &=& \int dp^-\,dp^+\,d^2p_\perp \ \frac{1}{2}\,\mbox{Tr}\left(\Phi(p) \,\gamma^\mu \,\gamma^+\,\gamma^\nu\right) \, \delta ( p^+ + q^+) \nonumber \\ &= & -g_\perp^{\mu\nu}\ \left. \mbox{Tr}\left(\gamma^+\,\Phi(x) \right) \right|_{x = {x_{_B}}} \nonumber \\ & = & -g_\perp^{\mu\nu}\,f_1({x_{_B}}). \label{inclcalc}\end{aligned}$$ Antiquarks arise from the diagram with opposite fermion flow, proportional to $\mbox{Tr}\left(\overline\Phi(p) \,\gamma^\nu \overline\Delta(k) \gamma^\mu\right)$ with \_[ij]{}(p) = d\^4 e\^[-ip]{} P S |\_i () \_j(0)|P S . The [*proper*]{} definition of antiquark distributions starts from $\Phi^c(x)$ containing antiquark distributions $\bar f_1(x)$, etc. The quantity $\Phi^c(p)$ is obtained from $\Phi(p)$ after the replacement of $\psi$ by $\psi^C = C\overline\psi^T$. One then finds $\overline \Phi(p)$ = $-C(\Phi^c)^TC^\dagger$, i.e. one has to be aware of sign differences. Symmetry relations between quark and antiquark relations can be obtained using the anticommutation relations for fermions, giving $\overline \Phi_{ij}(p) = - \Phi_{ij}(-p)$. One finds that $\bar f_1(x) = -f_1(-x)$, $\bar g_1(x) = g_1(-x)$, and $\bar h_1(x) = -h_1(-x)$. Finally, after including the flavor summation and the quark charges squared one can compare the result with Eq. \[param\] to obtain for the structure function 2F\_1([x\_[\_B]{}]{}) = \_a e\_a\^2(f\_1\^a([x\_[\_B]{}]{}) + f\_1\^[|a]{}([x\_[\_B]{}]{})), while $F_L({x_{_B}}) = 0$ (Callan-Gross relation). The antisymmetric part of $W^{\mu\nu}$ in the above calculation is left as an exercise. The answer is $$\begin{aligned} 2M\,W_A^{\mu \nu}(P,q) &=& i\,\epsilon_\perp^{\mu\nu}\,g_1({x_{_B}}),\end{aligned}$$ which after inclusion of antiquarks, flavor summation gives after comparison with Eq. \[wanti\] 2g\_1([x\_[\_B]{}]{}) = \_a e\_a\^2(g\_1\^a([x\_[\_B]{}]{}) + g\_1\^[|a]{}([x\_[\_B]{}]{})). Interpretation of the functions ------------------------------- The functions $f_1$, $g_1$ and $h_1$ can be obtained from the correlator $\Phi(x)$ after tracing with the appropriate Dirac matrix, f\_1(x) & = & .  e\^[ip]{} P,S(0) \^+ () P,S|\_[\^+ = \_= 0]{},\ g\_1(x) & = & .  e\^[ip]{} P,S(0) \^+\_5 () P,S|\_[\^+ = \_= 0]{},\ S\_\^ih\_1(x) & = & .  e\^[ip]{} P,S(0) i\^[i+]{}\_5() P,S|\_[\^+ = \_= 0]{}, By introducing [*good*]{} and [*bad*]{} fields $\psi_\pm \equiv \frac{1}{2}\gamma^\mp\gamma^\pm \psi$, one sees that $f_1$ can be rewritten as f\_1(x) & = & .  e\^[ip]{} P,S\^\_+(0) \_+() P,S|\_[\^+ = \_= 0]{}\ & = & \_n | P\_n \_+P|\^2(P\_n\^+ - (1-x)P\^+) , i.e. it is a quark lightcone momentum distribution. For the functions $g_1$ and $h_1$ one needs in addition the projectors on quark chirality states, $P_{R/L} = \frac{1}{2}(1\pm \gamma_5)$, and on quark transverse spin states [@Artru; @JJ92], $P_{\uparrow/\downarrow} = \frac{1}{2}(1\pm \gamma^i \gamma_5)$ to see that && f\_1(x) = f\_[1R]{}(x) + f\_[1L]{}(x) = f\_[1]{}(x) + f\_[1]{}(x),\ && g\_1(x) = f\_[1R]{}(x) - f\_[1L]{}(x),\ && h\_1(x) = f\_[1]{}(x) - f\_[1]{}(x). One sees some trivial bounds such as $f_1(x) \ge 0$ and $\vert g_1(x)\vert \le f_1(x)$. Since $P_n^+ \le 0$ and sees $x \le 1$. From the antiquark distribution $\bar f_1(x)$ and its relation to $f_1(x)$ one obtains $x \ge -1$, thus the support of the functions is $-1 \le x \le 1$. Bounds on the distribution functions ------------------------------------ The trivial bounds on the distribution functions ($\vert h_1(x)\vert \le f_1(x)$ and $\vert g_1(x)\vert \le f_1(x)$) can be sharpened. For instance one can look explicitly at the structure in Dirac space of the correlation function $\Phi_{ij}$. Actually, we will look at the correlation functions $(\Phi\,\gamma_0)_{ij}$, which involves at leading order matrix elements $\psi_{+j}^\dagger (0)\psi_{+i}(\xi)$. One has in Weyl representation ($\gamma^0 = \rho^1$, $\gamma^i = -i\rho^2\sigma^i$, $\gamma_5 = i\gamma^0\gamma^1\gamma^2\gamma^3 = \rho^3$) the matrices $$\begin{aligned} P_+ & = & \left\lgroup \begin{array}{rrrr} 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 1 \end{array}\right\rgroup, \quad \\ P_+\gamma_5 & = & \left\lgroup \begin{array}{rrrr} 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & -1 \end{array}\right\rgroup, \quad P_+\gamma^1\gamma_5 = \left\lgroup \begin{array}{rrrr} 0 & 0 & 0 & 1 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \end{array}\right\rgroup .\end{aligned}$$ The good projector only leaves two (independent) Dirac spinors, one righthanded (R), one lefthanded (L). On this basis of good R and L spinors the for hard scattering processes relevant matrix $(\Phi\slash n_-)$ is given by (/n\_-)\_[ij]{}(x) = [cc]{} f\_1 + g\_1 & (S\_\^1-iS\_\^2)h\_1\ &\ (S\_\^1+iS\_\^2)h\_1 & f\_1 - g\_1 One can also turn the $S$-dependent correlation function $\Phi$ defined in analogy with $W(q,P,S)$ in Eq. \[defS\] into a matrix in the nucleon spin space. If (x; P,S) &=& \_O + \_L + S\_\^1\_\^1 + S\_\^2\_\^2 , then one has on the basis of spin 1/2 target states with $\lambda = +1$ and $\lambda = -1$ respectively \_[ss\^]{}(x) = [cc]{} \_O + \_L & \_\^1 - i\_\^2\ &\ \_\^1 + i\_\^2 & \_O - \_L\ The matrix relevant for bounds is the matrix $M$ = $(\Phi \slash n_-)^T$ (for this matrix one has $v^\dagger M v \ge 0$ for any direction $v$). On the basis $+R$, $-R$, $+L$ and $-L$ it becomes ((x)/n\_-)\^T = [cccc]{} f\_1 + g\_1 & 0 & 0 & 2h\_1\ &\ 0 & f\_1 - g\_1 & 0 & 0\ &\ 0 & 0 & f\_1 - g\_1 & 0\ &\ 2h\_1 & 0 & 0 & f\_1 + g\_1 . Of this matrix any diagonal matrix element must always be positive, hence the eigenvalues must be positive, which gives a bound on the distribution functions stronger than the trivial bounds, namely h\_1(x)( f\_1(x) + g\_1(x)) known as the Soffer bound [@Soffer]. Sum rules --------- For the functions appearing in the soft parts, and thus also for the structure functions, one can derive sum rules. Starting with the traces defining the quark distributions, $$\begin{aligned} f_1(x) & = & \left. \int \frac{d\xi^-}{4\pi}\ e^{ip\cdot \xi} \,\langle P,S\vert \overline \psi(0) \gamma^+ \psi(\xi) \vert P,S\rangle \right|_{\xi^+ = \xi_\st = 0}, \\ g_1(x) & = & \left. \int \frac{d\xi^-}{4\pi}\ e^{ip\cdot \xi} \,\langle P,S\vert \overline \psi(0) \gamma^+\gamma_5 \psi(\xi) \vert P,S\rangle \right|_{\xi^+ = \xi_\st = 0},\end{aligned}$$ and integrating over $x = p^+/P^+$ one obtains (using symmetry relation as indicated above to eliminate antiquarks $\bar f_1$), \_0\^1 dx( f\_1(x) - |f\_1(x)) = \_[-1]{}\^1 dx f\_1(x) = , which as we have seen in the section on elastic scattering is nothing else than a form factor at zero momentum transfer, i.e. the number of quarks of that particular flavor. Similarly one finds the sum rule \_0\^1 dx( g\_1(x) + |g\_1(x)) = \_[-1]{}\^1 dx g\_1(x) = , \[axsumr\] which precisely is the axial charge $g_A$ for a particular quark flavor. These sum rules for the quark distributions underly the sum rules for the structure functions, e.g. the Bjorken sum rule following from Eq. \[axsumr\] and Eq. \[axform\]. \_0\^1 d[x\_[\_B]{}]{} (g\_1\^p([x\_[\_B]{}]{},Q\^2)-g\_1\^n([x\_[\_B]{}]{},Q\^2)) = (g\_A\^u - g\_A\^d) = G\_A\^[pn]{}(0). 1-particle inclusive scattering =============================== The relevant distribution functions ----------------------------------- For 1-particle inclusive scattering one parametrizes the momenta $$\left. \begin{array}{l} q^2 = -Q^2 \\ P^2 = M^2\\ P_h^2 = M_h^2 \\ 2\,P\cdot q = \frac{Q^2}{{x_{_B}}} \\ 2\,P_h\cdot q = -z_h\,Q^2 \end{array} \right\} \longleftrightarrow \left\{ \begin{array}{l} P_h = \frac{z_h\,Q}{\sqrt{2}}\,n_- + \frac{M_h^2}{z_h\,Q\sqrt{2}}\,n_+ \\ \mbox{} \\ q =\ \frac{Q}{\sqrt{2}}\,n_- \ - \ \frac{Q}{\sqrt{2}}\,n_+\ +\ q_T \\ \mbox{} \\ P = \frac{{x_{_B}}M^2}{Q\sqrt{2}}\,n_- + \frac{Q}{{x_{_B}}\sqrt{2}}\,n_+ \end{array} \right.$$ Note that this works for socalled current fragmentation, in which case the produced hadron is [*hard*]{} with respect to the target momentum, i.e. $P\cdot P_h \sim Q^2$. The minus component $p^-$ is irrelevant in the lower soft part, while the plus component $k^+$ is irrelevant in the upper soft part. Note that after the choice of $P$ and $P_h$ one can no longer omit a transverse component in the other vector, in this case the momentum transfer $q$. This is precisely the vector $q_T$ introduced earlier in the discussion of the structure functions for 1-particle inclusive leptoproduction. One immediately sees that one can no longer simply integrate over the transverse component of the quark momentum, defined in Eq. \[quarkmom\]. At this point it turns out that the most convenient way to describe the spin vector of the target is via an expansion of the form S\^= -n\_- + n\_+ + S\_. One has up to ${\cal O}(1/Q^2)$ corrections $\lambda \approx M\,(S\cdot q)/(P\cdot q)$ and $S_\st \approx S_\perp$. For a pure state one has $\lambda^2 + {\mbox{\boldmath $S$}}_\st^2 = 1$, in general this quantity being less or equal than one. The soft part to look at is (x,p\_T) = .  e\^[ip]{} P,S(0) () P,S|\_[\^+ = 0]{}. For the leading order results, it is parametrized as (x,p\_) = \_O(x,p\_) + \_L(x,p\_) + \_T(x,p\_), with the parts involving unpolarized targets (O), longitudinally polarized targets (L) and transversely polarized targets (T) up to parts proportional to $M/P^+$ given by \_O(x,p\_) & = & { f\_1(x,p\_)[0.2 em n-0.4em /]{}\_+ + h\_1\^(x,p\_) }\ \_L(x,p\_) & = & { g\_[1L]{}(x,p\_)\_5[0.2 em n-0.4em /]{}\_+ + h\_[1L]{}\^(x,p\_) }\ \_T(x,p\_) & = & { f\_[1T]{}\^(x,p\_)\ & & + g\_[1T]{}(x,p\_) \_5[0.2 em n-0.4em /]{}\_+ + h\_[1T]{}(x,p\_)\ & & + h\_[1T]{}\^(x,p\_) }. All functions appearing here have a natural interpretation as densities. This is seen as discussed before for the ${\mbox{\boldmath $p$}}_\st$-integrated functions. Now it includes densities such as the density of longitudinally polarized quarks in a transversely polarized nucleon ($g_{1T}$) and the density of transversely polarized quarks in a longitudinally polarized nucleon ($h_{1L}^\perp$). The interpretation of all functions is illustrated in Fig. \[fig3\]. Several functions vanish from the soft part upon integration over $p_\st$. Actually we will find that particularly interesting functions survive when one integrates over ${\mbox{\boldmath $p$}}_\st$ weighting with $p_\st^\alpha$, e.g. \_\^(x) && d\^2 p\_ (x,p\_)\ & = & { -g\_[1T]{}\^[(1)]{}(x)S\_\^/n\_+\_5 -h\_[1L]{}\^[(1)]{}(x)\ && -[f\_[1T]{}\^[(1)]{}]{} \^\_[  ]{}\^n\_-\^ - [h\_1\^[(1)]{}]{} }, \[Phid\] where we define ${\mbox{\boldmath $p$}}_\st^2/2M^2$-moments as g\_[1T]{}\^[(1)]{}(x) = d\^2p\_  g\_[1T]{}(x,p\_), and similarly the other functions. The functions $h_1^\perp$ and $f_{1T}^\perp$ are T-odd, vanishing if T-reversal invariance can be applied to the matrix element. For $p_\st$-dependent correlation functions, matrix elements involving gluon fields at infinity (gluonic poles [@bmt]) can for instance prevent application of T-reversal invariance. The functions describe the possible appearance of unpolarized quarks in a transversely polarized nucleon ($f_{1T}^\perp$) or transversely polarized quarks in an unpolarized hadron ($h_1^\perp$) and lead to single-spin asymmetries in various processes [@Sivers90; @Anselmino95]. The interpretation of these functions is also illustrated in Fig. \[fig3\]. \ \ \ Of course just integrating $\Phi(x,p_\st)$ over $p_\st$ gives the result used in inclusive scattering with $f_1(x)$ = $\int d^2p_\st \ f_1(x,p_\st)$, $g_1(x)$ = $g_{1L}(x)$ and $h_1(x)$ = $h_{1T}(x) + h_{1T}^{\perp (1)}(x)$. We note that the function $h_{1T}^{\perp (2)}$ appears after weighting with $p_\st^\alpha p_\st^\beta$. The relevant fragmentation functions ------------------------------------ Just as for the distribution functions one can perform an analysis of the soft part describing the quark fragmentation. One needs [@CS82] \_[ij]{}(z,k) = . \_X e\^[ik]{} Tr 0 \_i () P\_h,XP\_h,X\_j(0) 0 |\_[\^- = 0]{}. For the production of unpolarized hadrons $h$ in hard processes one needs to leading order in $1/Q$ the correlation function, \_O(z,k\_) = zD\_1(z,k\^\_)[0.2 em n-0.4em /]{}\_- + zH\_1\^(z,k\^\_) + [O]{}(). when we limit ourselves to an unpolarized or spin 0 final state hadron. The arguments of the fragmentation functions $D_1$ and $H_1^\perp$ are $z$ = $P_h^-/k^-$ and ${\mbox{\boldmath $k$}}^\prime_\st$ = $-z{{{\mbox{\boldmath $k$}}}_{_T}}$. The first is the (lightcone) momentum fraction of the produced hadron, the second is the transverse momentum of the produced hadron with respect to the quark. The fragmentation function $D_1$ is the equivalent of the distribution function $f_1$. It can be interpreted as the probability of finding a hadron $h$ in a quark. The function $H_1^\perp$, interpretable as the difference in production probabilities of unpolarized hadrons from a transversely polarized quark depending on transverse momentum, is allowed because of the non-applicability of time reversal invariance [@Collins93]. This is natural for the fragmentation functions [@HHK83; @JJ93] because of the appearance of out-states $\vert P_h, X\rangle$ in the definition of $\Delta$, in contrast to the plane wave states appearing in $\Phi$. After ${{{\mbox{\boldmath $k$}}}_{_T}}$-averaging one is left with the functions $D_1(z)$ and the ${\mbox{\boldmath $k$}}_\st^2/2M^2$-weighted result $H_1^{\perp (1)}(z)$. The semi-inclusive cross section -------------------------------- After the analysis of the soft parts, the next step is to find out how one obtains the information on the various correlation functions from experiments, in this particular case in lepton-hadron scattering via one-photon exchange as discussed before. To get the leading order result for semi-inclusive scattering it is sufficient to compute the diagram in Fig. \[fig2\] (right) by using QCD and QED Feynman rules in the hard part and the matrix elements $\Phi$ and $\Delta$ for the soft parts, parametrized in terms of distribution and fragmentation functions. The most well-known results for leptoproduction are: && = \_[a,|a]{} e\_a\^2 1 + (1-y)\^2[x\_[\_B]{}]{}[f\^a\_1]{}([x\_[\_B]{}]{})[ D\^a\_1]{}(z\_h)\ && = [\_e]{} \_[a,|a]{} e\_a\^2  y (2-y)  [x\_[\_B]{}]{}[g\^a\_1]{}([x\_[\_B]{}]{})[D\^a\_1]{}(z\_h) The indices attached to the cross section refer to polarization of lepton (O is unpolarized, L is longitudinally polarized) and hadron (O is unpolarized, L is longitudinally polarized, T is transversely polarized). Note that the result is a weighted sum over quarks and antiquarks involving the charge $e_a$ squared. Comparing with well-known formal expansions of the cross section in terms of structure functions one can simply identify these. For instance the above result for unpolarized scattering (OO) shows that after averaging over azimuthal angles, only one structure function survives if we work at order $\alpha_s^0$ and at leading order in $1/Q$. As we have seen, in 1-particle inclusive unpolarized leptoproduction in principle four structure functions appear, two of them containing azimuthal dependence of the form $\cos (\phi_h^\ell)$ and $\cos (2\phi_h^\ell)$. The first one only appears at order $1/Q$ [@LM94], the second one even at leading order but only in the case of the existence of nonvanishing T-odd distribution functions. To be specific if we define weighted cross section such as d\^2q\_[T]{} (2\_h\^) \_[OO]{} we obtain the following asymmetry,. \_[OO]{} = (1-y)\_[a,|a]{} e\_a\^2 [x\_[\_B]{}]{}[h\_[1]{}\^[(1)a]{}]{}([x\_[\_B]{}]{}) H\_1\^[(1)a]{}. In lepton-hadron scattering this asymmetry requires T-odd distribution functions and therefore most likely is absent or very small. In $e^+e^-$ annihilation, however, a $\cos 2\phi$ asymmetry between produced particles (e.g. pions) in opposite jets involves two very likely nonvanishing fragmentation functions $H_1^\perp$ and $\overline H_1^\perp$. Indications for the presence of these fragmentation functions have been found in LEP data[@Efremov]. For polarized targets, several azimuthal asymmetries arise already at leading order. For example the following possibilities were investigated in Refs [@KM96; @Collins93; @Kotzinian95; @TM95b]. && \_[LT]{} =\ && [\_eS ]{} y(2-y)\_[a,|a]{} e\_a\^2 [x\_[\_B]{}]{}[g\_[1T]{}\^[(1)a]{}]{}([x\_[\_B]{}]{}) [D\^a\_1]{}(z\_h), \[asbas\]\ && \_[OL]{} =\ && - (1-y)\_[a,|a]{} e\_a\^2 [x\_[\_B]{}]{}[h\_[1L]{}\^[(1)a]{}]{}([x\_[\_B]{}]{}) [H\_1\^[(1)a]{}]{}(z\_h), \[as2\]\ && \_[OT]{} =\ && [S ]{} (1-y)\_[a,|a]{} e\_a\^2 [x\_[\_B]{}]{}[h\_1\^a]{}([x\_[\_B]{}]{}) [H\_1\^[(1)a]{}]{}(z\_h). \[finalstate\] The latter two are single spin asymmetries involving the fragmentation function $H_1^{\perp (1)}$. The last one was the asymmetry proposed by Collins [@Collins93] as a way to access the transverse spin distribution function $h_1$ in pion production. Note, however, that in using the azimuthal dependence one needs to be very careful. For instance, besides the $<\sin (\phi_h^\ell + \phi_S^\ell)>_{OT}$, one also finds at leading order a $<\sin (3\phi_h^\ell - \phi_S^\ell)>_{OT}$ asymmetry which is proportional to $h_{1T}^{\perp (2)}\,H_1^{\perp (1)}$ [@TM95b]. Inclusion of subleading contributions ===================================== Subleading inclusive leptoproduction ------------------------------------ If one proceeds up to order $1/Q$ one also needs terms in the parametrization of the soft part proportional to $M/P^+$. Limiting ourselves to the ${\mbox{\boldmath $p$}}_\st$-integrated correlations one needs $$\begin{aligned} \Phi(x) & = & \frac{1}{2}\,\Biggl\{ f_1(x)\,{\kern 0.2 em n\kern -0.4em /}_+ + \lambda\,g_1(x)\, \gamma_5\,{\kern 0.2 em n\kern -0.4em /}_+ + h_1(x)\,\frac{\gamma_5\,[{\kern 0.2 em S\kern -0.56em \raisebox{0.3ex}{/}}_\st,{\kern 0.2 em n\kern -0.4em /}_+]}{2}\Biggr\} \nonumber \\ & + & \frac{M}{2P^+}\Biggl\{ e(x) + g_T(x)\,\gamma_5\,{\kern 0.2 em S\kern -0.56em \raisebox{0.3ex}{/}}_\st + \lambda\,h_L(x)\,\frac{\gamma_5\,[{\kern 0.2 em n\kern -0.4em /}_+,{\kern 0.2 em n\kern -0.4em /}_-]}{2} \Biggr\} \nonumber \\ & + & \frac{M}{2P^+}\Biggl\{ -\lambda\,e_L(x)\,i\gamma_5 - f_T(x)\,\epsilon_\st^{\rho\sigma}\gamma_\rho S_{\st\sigma} + h(x)\,\frac{i\,[{\kern 0.2 em n\kern -0.4em /}_+,{\kern 0.2 em n\kern -0.4em /}_-]}{2} \Biggr\}.\end{aligned}$$ The last set of three terms proportional to $M/P^+$ vanish when time-reversal invariance applies. Actually in the calculation of the cross section one has to be careful. Let us use inclusive scattering off a transversely polarized nucleon (transverse means $\vert {\mbox{\boldmath $S$}}_\perp\vert = 1$ in Eq. \[inclspin\]) as an example. The hadronic tensor is zero in leading order in $1/Q$. At order $1/Q$ one obtains from the handbag diagram a contribution 2MW\_[A (a)]{}\^([q,P,S\_]{}) = i t\_\^[\[]{}\_\^[\]]{} [S\_]{} ( g\_[1T]{}\^[(1)]{}([x\_[\_B]{}]{}) - h\_1([x\_[\_B]{}]{})). It shows that one must be very careful with the integration over $p_\st$. There is a second contribution at order $1/Q$ coming from diagrams as the one shown in Fig. \[figdis1\]. For these gluon diagrams one needs matrix elements containing $\overline \psi(0)\,gA_\st^\alpha(\eta)\,\psi(\xi)$. At order $1/Q$ one only needs the matrix element of the bilocal combinations $\overline \psi(0)\,gA_\st^\alpha(\xi)\,\psi(\xi)$ and $\overline \psi(0)\,gA_\st^\alpha(0)\,\psi(\xi)$ These soft parts have a structure quite similar to $\Phi_\partial^\alpha$ and are parametrized as \_A\^(x) &= & { -xg\_[T]{}(x)S\_\^/n\_+\_5 -xh\_[L]{}(x)\ && -xf\_T(x) \^\_[  ]{}\^n\_-\^ - xh(x) }. \[phiA\] This contributes also to $W_A^{\mu\nu}$, 2MW\_[A (b)]{}\^([q,P,S\_]{}) = i t\_\^[\[]{}\_\^[\]]{} [S\_]{} g\_T([x\_[\_B]{}]{}). Using the QCD equations of motion, however, these functions can be related to the functions appearing in $\Phi$. To be precise one combines $i\partial$ in $\Phi_\partial$ (see Eq. \[Phid\]) and $A_\mu$ in $\Phi_A$ to $\Phi_D$ containing $iD_\mu = i\partial_\mu + g\,A_\mu$ for which one has via the equations of motion \_D\^(x) &= & { -(xg\_[T]{}- h\_1)S\_\^/n\_+\_5\ && -(xh\_[L]{}-g\_1)\ && -xf\_T(x) \^\_[  ]{}\^n\_-\^ - xh(x) }. Hence one obtains && xg\_T = xg\_T - g\_[1T]{}\^[(1)]{} - h\_1,\ && xh\_L = xh\_L - h\_[1L]{}\^[(1)]{} - g\_1,\ && xf\_T = xf\_T + f\_[1T]{}\^[(1)]{},\ && xh = xh + 2h\_[1]{}\^[(1)]{}. and one obtains the full contribution 2MW\_[A]{}\^([q,P,S\_]{}) = i t\_\^[\[]{}\_\^[\]]{} [S\_]{} g\_T([x\_[\_B]{}]{}), leading for the structure function ${\it g}_T({x_{_B}},Q^2)$ defined in Eq. \[wanti\] to the result \_T([x\_[\_B]{}]{},Q\^2) = \_a e\_a\^2 ( g\_T\^a([x\_[\_B]{}]{}) + g\_T\^[|a]{}([x\_[\_B]{}]{})). From Lorentz invariance one obtains, furthermore, some interesting relations between the subleading functions and the $k_\st$-dependent leading functions [@BKL84; @MT96; @BM98]. Just by using the expressions for the functions in terms of the amplitudes $A_i$ in Eq. \[lorentz\] one finds &&g\_T = g\_1 + g\_[1T]{}\^[(1)]{}, \[gTrel\]\ &&h\_L = h\_1 - h\_[1L]{}\^[(1)]{}, \[hLrel\]\ &&f\_T = - f\_[1T]{}\^[(1)]{},\ &&h = - h\_[1]{}\^[(1)]{}. \[rel4\] As an application, one can eliminate $g_{1T}^{(1)}$ using Eq. \[gTrel\] and obtain (assuming sufficient neat behavior of the functions) for $g_2 = g_T - g_1$ g\_2(x) & = & -+\ && + . One can use this to obtain for each quark flavor $\int dx\,g^a_2(x) = 0$, the Burkhardt-Cottingham sumrule [@BC]. Neglecting the interaction-dependent part one obtains the Wandzura-Wilczek approximation [@WW] for $g_2$, which in particular when one neglects the quark mass term provides a simple and often used estimate for $g_2$. It has become the standard with which experimentalists compare the results for $g_2$. Actually the SLAC results for $g_2$ can also be used to estimate the function $g_{1T}^{(1)}$ and the resulting asymmetries, e.g. the one in Eq. \[asbas\]. For this one needs the exact relation in Eq. \[gTrel\]. Results can be found in Refs [@KM96] and [@BM99]. Subleading 1-particle inclusive leptoproduction ----------------------------------------------- Also for the transverse momentum dependent functions dependent distribution and fragmentation functions one can proceed to subleading order [@MT96]. We will not discuss these functions here. In semi-inclusive cross sections one also needs fragmentation functions, for which similar relations exist, e.g. the relation in Eq. \[rel4\] for distribution functions has an analog for fragmentation functions, relating $H_1^{\perp (1)}$ (appearing in Eqs \[as2\] and \[finalstate\]) and an at subleading order appearing function $H(z)$, = z\^2 (). \[frag1\] An interesting subleading asymmetry involving $H_1^\perp$ is a $\sin(\phi_h^\ell)$ single spin asymmetry appearing as the structure functions ${\cal H}_{LT}^\prime$ in Eq. [sidiswanti]{} for a polarized lepton but unpolarized target [@LM94], &&\_[LO]{} =\ &&[\_e]{} y [x\_[\_B]{}]{}\^2 [e\^a]{}([x\_[\_B]{}]{})[H\_1\^[(1)a]{}]{}(z\_h) \[as1\] where $\tilde e^a(x) = e^a(x) - (m_a/M)\,(f_1^a(x)/x)$. This cross section involves, besides the time-reversal odd fragmentation function $H_1^\perp$, the distribution function $e$. The tilde function that appear in the cross sections is in fact the socalled interaction dependent part of the twist three functions. It would vanish in any naive parton model calculation in which cross sections are obtained by folding electron-parton cross sections with parton densities. Considering the relation for $\tilde e$ one can state it as $x\,e(x)$ = $(m/M)\,f_1(x)$ in the absence of quark-quark-gluon correlations. The inclusion of the latter also requires diagrams dressed with gluons as shown in Fig. \[figdis1\]. Color gauge invariance ====================== We have sofar neglected two problems. The first problem is that the correlation function $\Phi$ discussed in previous sections involve two quark fields at different space-time points and hence are not color gauge invariant. The second problem comes from the gluonic diagrams similar as the ones we have discussed in the previous section (see Fig. \[figdis1\]) We note that diagrams involving matrix elements with longitudinal ($A^+$) gluon fields, $$\overline \psi_j(0)\,gA^+(\eta)\,\psi_i(\xi)$$ do not lead to any suppression. The reason is that because of the $+$-index in the gluon field the matrix element is proportional to $P^+$, $p^+$ or $M\,S^+$ rather than the proportionality to $M\,S_\st^\alpha$ or $p_\st^\alpha$ that we have seen in Eq. \[phiA\] for a gluonic matrix element with transverse gluons. A straightforward calculation, however, shows that the gluonic diagrams with one or more longitudinal gluons involve matrix elements (soft parts) of operators $\overline \psi \psi$, $\overline \psi\,A^+\,\psi$, $\overline \psi\,A^+A^+\,\psi$, etc. that can be resummed into a correlation function \_[ij]{}(x) = .  e\^[ip]{} P,S\_j(0)[U]{}(0,)\_i() P,S|\_[\^+ = \_= 0]{}, where ${\cal U}$ is a gauge link operator (0,) = [P]{}(-i\_0\^[\^-]{} d\^-A\^+()) (path-ordered exponential with path along $-$-direction). Et voila, the unsuppressed gluonic diagrams combine into a color gauge invariant correlation function. We note that at the level of operators, one expands (0)() = \_n (0)\_[\_1]{}…\_[\_n]{}(0), in a set of local operators, but only the expansion of the nonlocal combination with a gauge link (0)() = \_n (0)D\_[\_1]{}…D\_[\_n]{}(0), is an expansion in terms of local gauge invariant operators. The latter operators are precisely the local (quark) operators that appear in the operator product expansion applied to inclusive deep inelastic scattering. For the $p_\st$-dependent functions, one finds that inclusion of $A^+$ gluonic diagrams leads to a color gauge invariant matrix element with links running via $\xi^= = \pm \infty$. For instance in lepton-hadron scattering one finds (x,p\_T) = .  e\^[ip]{} P,S(0)[U]{}(0,) [U]{}(,)() P,S|\_[\^+ = 0]{}, where the gauge links are at constant $\xi_\st$. One can multiply this correlator with $p_\st^\alpha$ and make this into a derivative $\partial_\alpha$. Including the links one finds the color gauge invariant result &&p\_\^\_[ij]{}(x,p\_) = (\_\^)\_[ij]{}(x,p\_)\ && =  e\^[ip]{} { P,S\_j(0) [U]{}(0,) iD\_\^\_i() P,S|\_[\^+=0]{}\ && - P,S\_j(0)[U]{}(0,) \_\^[\^-]{}d\^- [U]{}(,)\ && gG\^[+]{}() [U]{}(,)\_i() P,S|\_[\^+=0]{}}, which gives after integration over $p_\st$ the expected result $\Phi_\partial^\alpha(x) = \Phi_D^\alpha(x) - \Phi_A^\alpha(x)$. Note that in $A^+ = 0$ gauge all the gauge links disappear, while one has $G^{+\alpha} = \partial^+A^\alpha$, but there presence is essential to perform the above differentiations. Evolution ========= The explicit treatment of transverse momenta provides also a transparent way to include the evolution equations for quark distribution and fragmentation functions. Remember that we have assumed that soft parts vanish sufficiently fast as a function of the invariants $p\cdot P$ and $p^2$, which at constant $x$ implies a sufficiently fast vanishing as a function of ${\mbox{\boldmath $p$}}_\st^2$. This simply turns out not to be true. Assuming that the result for ${\mbox{\boldmath $p$}}_\st^2 \ge \mu^2$ is given by the diagram shown in Fig. \[figdis2\] one finds f\_1(x,p\_\^2) & = & (\^2-p\_\^2)f\_1(x,p\_\^2)\ & + & (p\_\^2-\^2) \_x\^1 P\_[qq]{}()f\_1(y;\^2), where $f_1(x;\mu^2) = \pi \int_0^{\mu^2} d{\mbox{\boldmath $p$}}_\st^2\ f_1(x,{\mbox{\boldmath $p$}}_\st^2)$ and the splitting function is given by P\_[qq]{}(z) = C\_F, with $\int dz\,f(z)/(1-z)_+ \equiv \int dz\,(f(z)-f(1))/(1-z)$ and the color factor $C_F$ = 4/3 for $SU(3)$. With the introduction of the scale in $f_1(x;\mu^2)$ one sees that the scale dependence satisfies = \_x\^1 P\_[qq]{}()f\_1(y;\^2). This is the standard [@Roberts] nonsinglet evolution equation for the valence quark distribution function. For the flavor singlet combination of quark distributions or the sea distributions one also needs to take into account contributions as shown in Fig. \[figdis2\] (right) involving the gluon distribution functions related to matrix elements with gluon fields $F_{\mu\nu}(\xi)$ but otherwise proceeding along analogous lines. The $\delta$-function contribution can be explicitly calculated by including vertex corrections (socalled virtual diagrams), but it is easier to derive them by requiring that the sum rules for $f_1$ remain valid under evolution, which requires that $\int_0^1 dz\ P_{qq}(z) = 0$. > Except for logarithmic contributions also finite $\alpha_s$ contributions show up in deep inelastic scattering [@Roberts]. For instance in inclusive scattering one finds that the lowest order result for $F_L$ is of this type, $$\begin{aligned} > F_L({x_{_B}},Q^2) & = & \frac{\alpha_s(Q^2)}{4\pi} \Biggl[ C_F \int_{{x_{_B}}}^1 > \frac{dy}{y}\,\left(\frac{2{x_{_B}}}{y}\right)^2 y\,f_1(y;Q^2) \nonumber \\ > &&\mbox{} + \left( 2\sum_q e_q^2 \right) \int_{{x_{_B}}}^1 > \frac{dy}{y} \,\left( \frac{2{x_{_B}}}{y} \right)^2 \left( 1 - \frac{{x_{_B}}}{y} > \right) \,y\,G(y;Q^2) \Biggr], > \nonumber \\ &&\end{aligned}$$ the second term involving the gluon distribution function $G(x)$. Concluding remarks ================== In these lectures I have discussed aspects of hard scattering processes, in particular inclusive and 1-particle inclusive lepton-hadron scattering. The goal is the study of the quark and gluon structure of hadrons. For example, by considering polarized targets or particle production one can measure spin and azimuthal asymmetries and use them to obtain information on specific correlations between spin and momenta of the partons. The reason why this is a promising route is the existence of a field theoretical framework that allows a clean study of the observables as well-defined hadronic matrix elements. Acknowledgements {#acknowledgements .unnumbered} ================ I acknowledge many discussions with colleagues, in particular Elliot Leader, Stan Brodsky. [99]{} R.G. Roberts, [*The structure of the proton*]{}, Cambridge University Press 1990 J.P. Ralston and D.E. Soper, Nucl. Phys. [**B152 (1979)**]{} 109. R.D. Tangerman and P.J. Mulders, Phys. Rev. [**D51**]{} (1995) 3357. D.E. Soper, Phys. Rev. [**D 15**]{} (1977) 1141; Phys. Rev. Lett. 43 (1979) 1847. R.L. Jaffe, Nucl. Phys. [**B 229**]{} (1983) 205. R.L. Jaffe and X. Ji, Nucl. Phys. [**B 375**]{} (1992) 527. X. Artru and M. Mekhfi, Z. Phys. [**45**]{} (1990) 669, J.L. Cortes, B. Pire and J.P. Ralston, Z. phys. [**C55**]{} (1992) 409. J. Soffer and D. Wray, Phys. Lett. [**43B**]{} (1973) 514. N. Hammon, O. Teryaev and A. Schäfer, Phys. Lett. [**B390**]{} (1997) 409; D. Boer, P.J. Mulders and O.V. Teryaev, D. Sivers, Phys. Rev. [**D41**]{} (1990) 83 and Phys. Rev. [**D43**]{} (1991) 261. M. Anselmino, M. Boglione and F. Murgia, Phys. Lett. [**B362**]{} (1995) 164; M. Anselmino and F. Murgia, Phys. Lett. [**B442**]{} (1998) 470. J.C. Collins and D.E. Soper, Nucl. Phys. [**B 194**]{} (1982) 445. J. Collins, Nucl. Phys. [**B396**]{} (1993) 161. K. Hagiwara, K. Hikasa and N. Kai, Phys. Rev. [**D27**]{} (1983) 84. R.L. Jaffe and X. Ji, Phys. Rev. Lett. [**71**]{} (1993) 2547. Phys. Rev. [**D57**]{} (1998) 3057. J. Levelt and P.J. Mulders, Phys. Rev. [**D 49**]{} (1994) 96; Phys. Lett. [**B 338**]{} (1994) 357. A.V. Efremov, O.G. Smirnova, L.G. Tkachev, in proceedings of the 13. International Symposium on High-Energy Spin Physics (SPIN98), Protvino, Russia, 8-12 Sep 1998, Nucl. Phys. Suppl. [**74**]{} (1999) 49. A. Kotzinian and P.J. Mulders, Phys. Rev. [**D54**]{} (1996) 1229. A. Kotzinian, Nucl. Phys. [**B 441**]{} (1995) 234. R.D. Tangerman and P.J. Mulders, Phys. Lett. [**B352**]{} (1995) 129. A.P. Bukhvostov, E.A. Kuraev and L.N. Lipatov, Sov. Phys. [**JETP 60**]{} (1984) 22. P.J. Mulders and R.D. Tangerman, Nucl. Phys. [**B461**]{} (1996) 197. D. Boer and P.J. Mulders, Phys. Rev. [**D 57**]{} (1998) 5780. H. Burkhardt and W.N. Cottingham, Ann. Phys. (N.Y.) [**56**]{} (1976) 453. S. Wandzura and F. Wilczek, Phys. Rev. [**D16**]{} (1977) 707. M. Boglione and P.J. Mulders, Phys. Rev. [**D60**]{} (1999) 054007.

ArXiv

--- abstract: 'We derive a functional limit theorem for the partial maxima process based on a long memory stationary $\alpha$-stable process. The length of memory in the stable process is parameterized by a certain ergodic theoretical parameter in an integral representation of the process. The limiting process is no longer a classical extremal Fréchet process. It is a self-similar process with $\alpha$-Fréchet marginals, and it has the property of stationary max-increments, which we introduce in this paper. The functional limit theorem is established in the space $D[0,\infty)$ equipped with the Skorohod $M_1$-topology; in certain special cases the topology can be strengthened to the Skorohod $J_1$-topology.' address: - | School of Operations Research and Information Engineering\ Cornell University\ Ithaca, NY 14853 - | School of Operations Research and Information Engineering\ and Department of Statistical Science\ Cornell University\ Ithaca, NY 14853 author: - Takashi Owada - Gennady Samorodnitsky title: 'Maxima of long memory stationary symmetric $\alpha$-stable processes, and self-similar processes with stationary max-increments' --- [^1] Introduction {#sec:intro} ============ The asymptotic behavior of the partial maxima sequence $M_n=\max_{1 \leq k \leq n} X_k$, $n=1,2,\dots$ for an i.i.d. sequence $(X_1,X_2,\ldots)$ of random variables is the subject of the classical extreme value theory, dating back to [@fisher:tippett:1928]. The basic result of this theory says that only three one-dimensional distributions, the Fréchet distribution, the Weibull distribution and the Gumbel distribution, have a max-domain of attraction. If $Y$ has one of these three distributions, then for a distribution in its domain of attraction, and a sequence of i.i.d. random variables with that distribution, $$\label{e:1din.conv} \frac{M_n-b_n}{a_n} \Rightarrow Y$$ for properly chosen sequences $(a_n)$, $(b_n)$; see e.g. Chapter 1 in [@resnick:1987] or Section 1.2 in [@dehaan:ferreira:2006]. Under the same max-domain of attraction assumption, a functional version of was established in [@lamperti:1964]: with the same sequences $(a_n)$, $(b_n)$ as in , $$\label{e:funct.conv} \left( \frac{M_{\lfloor nt\rfloor}-b_n}{a_n}, \ t\geq 0\right) \Rightarrow (Y(t), \, t\geq 0)$$ for a nondecreasing right continuous process $(Y(t), \, t\geq 0)$, and the convergence is weak convergence in the Skorohod $J_1$ topology on $D[0,\infty)$. The limiting process is often called [*the extremal process*]{}; its properties were established in [@dwass:1964; @dwass:1966] and [@resnick:rubinovitch:1973]. Much of the more recent research in extreme value theory concentrated on the case when the underlying sequence $(X_1,X_2,\ldots)$ is stationary, but may be dependent. In this case the extrema of the sequence may cluster, and it is natural to expect that the limiting results and will, in general, have to be different. The extremes of moving average processes have received special attention; see e.g. [@rootzen:1978], [@davis:resnick:1985] and [@fasen:2005]. The extremes of the GARCH(1,1) process were investigated in [@mikosch:starica:2000b]. The classical work on this subject of the extremes of dependent sequences is [@leadbetter:lindgren:rootzen:1983]; in some cases this clustering of the extremes can be characterized through the [*extremal index*]{} (introduced, originally, in [@leadbetter:1983]). The latter is a number $0\leq \theta\leq 1$. Suppose that a stationary sequence $(X_1,X_2,\ldots)$ has this index, and let $(\tilde X_1,\tilde X_2,\ldots)$ be an i.i.d. sequence with the same one-dimensional marginal distributions as $(X_1,X_2,\ldots)$. If and hold for the i.i.d. sequence, then the corresponding limits will satisfy $\tilde Y{\stackrel{d}{=}}\tilde Y(1)$, but the limit in for the dependent sequence $(X_1,X_2,\ldots)$ will satisfy $Y {\stackrel{d}{=}}\tilde Y(\theta)$. In particular, the limit will be equal to zero if the extremal index is equal to zero. This case can be viewed as that of long range dependence in the extremes, and it has been mostly neglected by the extreme value community. Long range dependence is, however, an important phenomenon in its own right, and in this paper we take a step towards understanding how long range dependence affects extremes. A random variable $X$ is said to have a regularly varying tail with index $-\alpha$ for $\alpha>0$ if $$P(X>x) = x^{-\alpha} L(x), \ x>0\,,$$ where $L$ is a slowly varying at infinity function, and the distribution of any such random variable is in the max-domain of attraction of the Fréchet distribution with the same parameter $\alpha$; see e.g. [@resnick:1987]. Recall that the Fréchet law $F_{\alpha,\sigma}$ on $(0,\infty)$ with the tail index $\alpha$ and scale $\sigma>0$ satisfies $$\label{e:frechet} F_{\alpha,\sigma}(x) = \exp\bigl\{ -\sigma^\alpha x^{-\alpha}\bigr\}, \ x>0\,.$$ Sometimes the term [*$\alpha$-Fréchet*]{} is used. In this paper we discuss the case of regularly varying tails and the resulting limits in . The limits obtained in this paper belong to the family of the so-called [*Fréchet processes*]{}, defined below. We would like to emphasize that, even for stationary sequences with regularly varying tails, non-Fréchet limits may appear in . We are postponing a detailed discussion of this point to a future publication. A stochastic process $(Y(t), \, t \in T)$ (on an arbitrary parameter space $T$) is called a Fréchet process if for all $n \geq 1$, $a_1, \dots, a_n >0$ and $t_1, \dots, t_n \in T$, the weighted maximum $\max_{1 \leq j \leq n} a_j Y(t_j)$ follows a Fréchet law as in . The best known Fréchet process is the extremal Fréchet process obtained in the scheme starting with an i.i.d. sequence with regularly varying tails. The extremal Fréchet process $\bigl( Y(t),\, t\geq 0\bigr)$ has finite dimensional distributions defined by $$\label{e:extreme.frechet} \bigl( Y(t_1),Y(t_2),\ldots, Y(t_n)\bigr) {\stackrel{d}{=}}\biggl( X^{(1)}_{\alpha,t_1^{1/\alpha}}, \ \max \Bigl( X^{(1)}_{\alpha,t_1^{1/\alpha}}, \, X^{(2)}_{\alpha,(t_2-t_1)^{1/\alpha}}\Bigr), \, \ldots \,$$ $$\max \Bigl( X^{(1)}_{\alpha,t_1^{1/\alpha}}, \, X^{(2)}_{\alpha,(t_2-t_1)^{1/\alpha}}, \ldots, X^{(n)}_{\alpha,(t_n-t_{n-1})^{1/\alpha}}\Bigr)\biggr)$$ for all $n$ and $0\leq t_1<t_2:\ldots <t_n$. The different random variables in the right hand side of are independent, with $X^{(k)}_{\alpha,\sigma}$ having the Fréchet law $F_{\alpha,\sigma}$ in , for any $k=1,\ldots, n$. This independence makes the extremal Fréchet processes the extremal Fréchet versions of the better known Lévy processes. The structure of general Fréchet processes has been extensively studied in the last several years. These processes were introduced in [@stoev:taqqu:2005], and their representations (as a part of a much more general context) were studied in [@kabluchko:stoev:2012]. Stationary Fréchet processes (in particular, their ergodicity and mixing) were discussed in [@stoev:2008], [@kabluchko:schlather:dehaan:2009] and [@wang:stoev:2010]. In this paper we concentrate on the maxima of stationary $\alpha$-stable processes with $0<\alpha<2$. Recall that a random vector ${{\bf X}}$ in ${{{\mathbb R}}}^d$ is called $\alpha$-stable if for any $A$ and $B>0$ we have $$A{{\bf X}}^{(1)} + B{{\bf X}}^{(2)} {\stackrel{d}{=}}(A^\alpha+B^\alpha)^{1/\alpha} {{\bf X}}+ {{\bf y}}\,,$$ where ${{\bf X}}^{(1)}$ and ${{\bf X}}^{(2)}$ are i.i.d. copies of ${{\bf X}}$, and ${{\bf y}}$ is a deterministic vector (unless ${{\bf X}}$ is deterministic, necessarily, $0<\alpha\leq 2$). A stochastic process $(X(t), \, t \in T)$ is called $\alpha$-stable if all of its finite-dimensional distributions are $\alpha$-stable. We refer the reader to [@samorodnitsky:taqqu:1994] for information on $\alpha$-stable processes. When $\alpha=2$, an $\alpha$-stable process is Gaussian, while in the case $0<\alpha<2$, both the left and the right tails of a (nondegenerate) $\alpha$-stable random variable $X$ are (generally) regularly varying with exponent $\alpha$. That is, $$P(X>x) \sim c_+ x^{-\alpha}, \ \ P(X<-x) \sim c_- x^{-\alpha} \ \ \text{as $x\to\infty$}$$ for some $c_+, \, c_-\geq 0$, $c_++c_->0$. That is, if $(X_1,X_2,\ldots)$ is an i.i.d. sequence of $\alpha$-stable random variables, then the i.i.d. sequence $(|X_1|,|X_2|,\ldots)$ satisfies and with $a_n=n^{1/\alpha}$ (and $b_n=0$), $n\geq 1$. Of course, we are not planning to study the extrema of an i.i.d. $\alpha$-stable sequence. Instead, we will study the maxima of (the absolute values of) a stationary $\alpha$-stable process. The reason we have chosen to work with stationary $\alpha$-stable processes is that their structure is very rich, and is also relatively well understood. This will allow us to study the effect of that structure on the limit theorems and . We are specifically interested in the long range dependent case, corresponding to the zero value of the extremal index. The structure of stationary symmetric $\alpha$-stable ([S$\alpha$S]{}) processes has been clarified in the last several years in the works of Jan Rosiński; see e.g. [@rosinski:1995; @rosinski:2006]. The integral representation of such a process can be chosen to have a very special form. The class of stationary [S$\alpha$S]{} processes we will investigate requires a representation slightly more restrictive than the one generally allowed. Specifically, we will consider processes of the form $$\label{e:underlying.proc} X_n = \int_E f \circ T^n(x) \, dM(x), \ \ \ n=1,2,\dots\,,$$ where $M$ is a [S$\alpha$S]{} random measure on a measurable space $(E,\mathcal{E})$ with a $\sigma$-finite [*infinite*]{} control measure $\mu$. The map $T:E \to E$ is a measurable map that preserves the measure $\mu$. Further, $f \in L^{\alpha}(\mu)$. See [@samorodnitsky:taqqu:1994] for details on $\alpha$-stable random measures and integrals with respect to these measures. It is elementary to check that a process with a representation is, automatically, stationary. Recall that any stationary [S$\alpha$S]{} process has a representation of the form: $$\label{e:general.integral} X_n = \int_E f_n(x) dM(x), \ \ \ n=1,2,\dots\,,$$ with $$\label{e:general.rosinski} f_n(x) = a_n(x) \left( \frac{d\mu \circ T^{n}}{d\mu}(x) \right)^{1/\alpha} f\circ T^{n}(x), \ \ \ x \in E$$ for $n=1,2,\dots$, where $T:E \to E$ is a one-to-one map with both $T$ and $T^{-1}$ measurable, mapping the control measure $\mu$ into an equivalent measure, and $$a_n(x) = \prod_{j=0}^{n-1} u \circ T^j(x), \ \ \ x \in E$$ for $n=1,2,\dots$, with $u:E \to \{-1,1 \}$ a measurable function. Here $M$ is [S$\alpha$S]{} (and $f \in L^{\alpha}(\mu)$). See [@rosinski:1995]. The main restriction of is, therefore, to a measure preserving map $T$. It has been observed that the ergodic-theoretical properties of the map $T$, either in or in , have a major impact on the memory of a stationary $\alpha$-stable process. See e.g. [@surgailis:rosinski:mandrekar:cambanis:1993], [@samorodnitsky:2004a; @samorodnitsky:2005], [@roy:2008], [@resnick:samorodnitsky:2004], [@owada:samorodnitsky:2012], [@owada:2013]. The most relevant for this work is the result of [@samorodnitsky:2004a], who proved that, if the map $T$ in or in is conservative, then using the normalization $a_n=n^{1/\alpha}$ ($b_n=0$) in , as indicated by the marginal tails, produces the zero limit, so the partial maxima grow, in this case, strictly slower than at the rate of $n^{1/\alpha}$. On the other hand, if the map $T$ is not conservative, then the normalization $a_n=n^{1/\alpha}$ in is the correct one, and it leads to a Fréchet limit (we will survey the ergodic-theoretical notions in the next section). Therefore, the extrema of [S$\alpha$S]{} processes corresponding to conservative flows cluster so much that the sequence of the partial maxima grows at a slower rate than that indicated by the marginal tails. This case can be thought of as indicating long range dependence. It is, clearly, inconsistent with a positive extremal index. The Fréchet limit obtained in by [@samorodnitsky:2004a] remains valid when the map $T$ is conservative (but with the normalization of a smaller order than $n^{1/\alpha}$), as long as the map $T$ satisfies a certain additional assumption. If one views the stationary $\alpha$-stable process as a natural function of the Poisson points forming the random measure $M$ in then, informally, this assumption guarantees that only the largest Poisson point contributes, distributionally, to the asymptotic behavior of the partial maxima of the process. In this paper we restrict ourselves to this situation as well. However, we will look at the limits obtained in the much more informative functional scheme . In this paper the assumption on the map $T$ will be expressed in terms of the rate of growth of the so-called return sequence, which we define in the sequel. We would like to emphasize that, when this return sequence grows at a rate slower than the one assumed in this paper, new phenomena seem to arise. Multiple Poisson points may contribute to the asymptotic distribution of the partial maxima, and non-Fréchet limit may appear in . We leave a detailed study of this to a subsequent work. In the next section we provide the elements of the infinite ergodic theory needed for the rest of the paper. In Section \[sec:lim.processes\] we introduce a new notion, that of a process with stationary max-increments. It turns out that the possible limits in the functional maxima scheme (with $b_n=0$) are self-similar with stationary max-increments. We discuss the general properties of such processes and then specialize to the concrete limiting process we obtain in the main result of the paper, stated and proved in Section \[sec:FLTPM\]. Ergodic Theoretical Notions {#sec:ergodic} =========================== In this section we present some basic notation and notions of, mostly infinite, ergodic theory used in the sequel. The main references are [@krengel:1985], [@aaronson:1997], and [@zweimuller:2009]. Let $(E,\mathcal{E},\mu)$ be a $\sigma$-finite, infinite measure space. We will say that $A = B$ mod $\mu$ if $A, B \in \mathcal{E}$ and $\mu(A\triangle B)=0$. For $f\in L^1(\mu)$ we will often write $\mu(f)$ for the integral $\int f\, d\mu$. Let $T:E \to E$ be a measurable map preserving the measure $\mu$. The sequence $(T^n)$ of iterates of $T$ is called a [*flow*]{}, and the ergodic-theoretical properties of the map and the flow are identified. A map $T$ is called [*ergodic*]{} if any $T$-invariant set $A$ (i.e., a set such that $T^{-1}A = A$ mod $\mu$) satisfies $\mu(A)=0$ or $\mu(A^c)=0$. A map $T$ is said to be [*conservative*]{} if $$\sum_{n=1}^{\infty} {{\bf 1}}_A \circ T^n = \infty \ \ \text{a.e. on } A$$ for any $A \in \mathcal{E}$, $0 < \mu(A) < \infty$; if $T$ is also ergodic, then the restriction “[*on*]{} $A$” is not needed. The [*conservative part*]{} of a measure-preserving $T$ is the largest $T$-invariant subset $C$ of $E$ such that the restriction of $T$ to $C$ is conservative. The set $D=E \setminus C$ is the [ *dissipative part*]{} of $T$ (and the decomposition $E=C\cup D$ is called [*the Hopf decomposition*]{} of $T$). The [*dual operator*]{} $\widehat{T}: L^1(\mu) \to L^1(\mu)$ is defined by $$\label{e:define.dual} \widehat{T} f = \frac{d \bigl(\nu_f \circ T^{-1}\bigr)}{d \mu}, \ f\in L^1(\mu)\,,$$ where $\nu_f$ is the signed measure $\nu_f(A) = \int_{A} f d\mu$, $A \in \mathcal{E}$. The dual operator satisfies the duality relation $$\label{e:dual.rel} \int_E \widehat{T} f\cdot g\, d\mu = \int_E f\cdot g \circ T\, d\mu$$ for $f\in L^1(\mu), \, g\in L^\infty(\mu)$. Note that makes sense for any nonnegative measurable function $f$ on $E$, and the resulting $\widehat{T} f$ is again a nonnegative measurable function. Furthermore, holds for arbitrary nonnegative measurable functions $f$ and $g$. A conservative, ergodic and measure preserving map $T$ is said to be [*pointwise dual ergodic*]{}, if there exists a normalizing sequence $a_n \nearrow \infty$ such that $$\label{e:pde} \frac{1}{a_n} \sum_{k=1}^n \widehat{T}^k f \to \mu(f) \ \ \ \text{a.e.} \; \text{for every } f \in L^1(\mu)\,.$$ The property of pointwise dual ergodicity rules out invertibility of the map $T$. Sometimes we require that for some functions the above convergence takes place uniformly on a certain set. A set $A \in \mathcal{E}$ with $0 < \mu(A) < \infty$ is said to be a [ *uniform set*]{} for a conservative, ergodic and measure preserving map $T$, if there exist a normalizing sequence $a_n \nearrow \infty$ and a nontrivial nonnegative measurable function $f \in L^1(\mu)$ (nontriviality means that $f$ is different from zero on a set of positive measure) such that $$\label{e:uniform_set} \frac{1}{a_n} \sum_{k=1}^n \widehat{T}^k f \to \mu(f) \ \ \ \text{uniformly, a.e. on } A\,.$$ If holds for $f={{\bf 1}}_A$, the set $A$ is called [*a Darling-Kac set*]{}. A conservative, ergodic and measure preserving map $T$ is pointwise dual ergodic if and only if $T$ admits a uniform set; see Proposition 3.7.5 in [@aaronson:1997]. In particular, it is legitimate to use the same normalizing sequence $(a_n)$ both in (\[e:pde\]) and (\[e:uniform\_set\]). Let $A \in \mathcal{E}$ with $0 < \mu(A) < \infty$. The frequency of visits to the set $A$ along the trajectory $( T^n x), \, x\in E$, is naturally related to the [*wandering rate*]{} sequence $$\label{e:wanderingrate} w_n = \mu \Bigl(\bigcup_{k=0}^{n-1} T^{-k}A\Bigr)\,.$$ If we define the first entrance time to $A$ by $$\varphi_A(x) = \min \{ n \geq 1: T^n x \in A \}$$ (notice that $\varphi_A < \infty$ a.e. on $E$ since $T$ is conservative and ergodic), then $w_n \sim \mu(\varphi_A < n)$ as $n\to\infty$. Since $T$ is also measure preserving, we have $\mu(A \cap \{ \varphi_A > k \}) = \mu(A^c \cap \{ \varphi_A = k \})$ for $k \geq 1$. Therefore, alternative expressions for the wandering rate sequence are $$w_n = \mu(A) + \sum_{k=1}^{n-1} \mu (A^c \cap \{ \varphi_A = k \}) = \sum_{k=0}^{n-1} \mu(A \cap \{ \varphi_A > k \})\,.$$ Suppose now that $T$ is a pointwise dual ergodic map, and let $A$ be a uniform set for $T$. It turns out that, under an assumption of regular variation, there is a precise connection between the wandering rate sequence $(w_n)$ and the normalizing sequence $(a_n)$ in and . Specifically, let $RV_{\gamma}$ represent the class of regularly varying at infinity sequences (or functions, depending on the context) of index $\gamma$. If either $(w_n) \in RV_{\beta}$ or $(a_n) \in RV_{1-\beta}$ for some $\beta \in [0,1]$, then $$\label{e:prop3.8.7} a_n \sim \frac{1}{\Gamma(2-\beta) \Gamma(1+\beta)} \frac{n}{w_n} \ \ \ \text{as } n \to \infty\,.$$ Proposition 3.8.7 in [@aaronson:1997] gives one direction of this statement, but the argument is easily reversed. We finish this section with a statement on distributional convergence of the partial maxima for pointwise dual ergodic flows. It will be used repeatedly in the proof of the main theorem. For a measurable function $f$ on $E$ define $$M_n(f)(x) = \max_{1 \leq k \leq n}|f \circ T^k(x)|\,, \ \ x \in E, \ n \geq 1\,.$$ In the sequel we will use the convention $\max_{k\in K}b_k = 0$ for a nonnegative sequence $(b_n)$, if $K=\emptyset$. \[p:max.ergodic\] Let $T$ be a pointwise dual ergodic map on a $\sigma$-finite, infinite, measure space $(E,\mathcal{E},\mu)$. We assume that the normalizing sequence $(a_n)$ is regularly varying with exponent $1-\beta$ for some $0 < \beta \leq 1$. Let $A \in \mathcal{E}$, $0 < \mu(A) < \infty$, be a uniform set for $T$. Define a probability measure on $E$ by $\mu_n (\cdot) = \mu(\cdot \cap \{\varphi_A \leq n\}) / \mu(\{\varphi_A \leq n \})$. Let $f:E \to {{{\mathbb R}}}$ be a measurable bounded function supported by the set $A$. Let $\| f \|_{\infty} = \inf \{M : |f(x)| \leq M \ \text{a.e. on } A \}$. Then $$\left( M_{\lfloor nt\rfloor}(f), \ 0\leq t\leq 1 \right) \Rightarrow \|f\|_\infty \left( {{\bf 1}}_{\{ V_{\beta} \leq t \}}, \ 0\leq t\leq 1 \right) \ \ \ \text{in the $M_1$ topology on $D[0,1] $,}$$ where the law of the left hand side is computed with respect to $\mu_n$, and $V_{\beta}$ is a random variable defined on a probability space $(\Omega^{\prime},\mathcal{F}^{\prime},P^{\prime})$ with $P^{\prime}(V_{\beta} \leq x) = x^{\beta}$, $0 < x \leq 1$. For $0<{\varepsilon}<1$ let $A_{\varepsilon}= \{ x\in A:\, |f(x)|\geq (1-{\varepsilon})\| f\|_\infty\}$. Note that each $A_{\varepsilon}$ is uniform since $A$ is uniform. Clearly, $$(1-{\varepsilon})\| f\|_\infty {{\bf 1}}_{ \{\varphi_{A_{\varepsilon}}(x)\leq nt \} } \leq M_{\lfloor nt\rfloor}(f)(x)\leq \| f\|_\infty {{\bf 1}}_{ \{\varphi_A(x)\leq nt \} } \ \ \ \mu \text{-a.e.}$$ for all $n\geq 1$ and $0\leq t\leq 1$. Since for monotone functions weak convergence in the $M_1$ topology is implied by convergence in finite dimensional distributions, we can use Theorem 3.2 in [@billingsley:1999] in a finite-dimensional situation. The statement of the proposition will follow once we show that, for a uniform set $B$ (which could be either $A$ or $A_{\varepsilon}$) the law of $\varphi_B/n$ under $\mu_n$ converges to the law of $V_{\beta}$. Let $(w_n^{(B)})$ be the corresponding wandering rate sequence. Since holds for $(w_n^{(B)})$ with the same normalizing constants $a_n$, $$\mu_n \left( \frac{\varphi_{B}}{n} \leq x \right) = \frac{\mu(\varphi_{B} \leq \lfloor nx \rfloor )}{\mu(\varphi_A \leq n)} \sim \frac{w_{\lfloor nx \rfloor}^{(B)}}{w_n} \to x^{\beta}$$ for all $0 < x \leq 1$, because the wandering rate sequence is regularly varying with index $\beta$ by . Self-similar processes with stationary max-increments {#sec:lim.processes} ===================================================== The limiting process obtained in the next section shares with any possible limits in the functional maxima scheme (with $b_n=0$) two very specific properties, one of which is classical, and the other is less so. Recall that a stochastic process $\bigl( Y(t), \, t\geq 0\bigr)$ is called self-similar with exponent $H$ of self-similarity if for any $c>0$ $$\bigl( Y(ct), \, t\geq 0\bigr){\stackrel{d}{=}}\bigl( c^HY(t), \, t\geq 0\bigr)$$ in the sense of equality of all finite-dimensional distributions. The best known classes of self-similar processes arise in various versions of a functional central limit theorem for stationary processes, and they have an additional property of stationary increments. Recall that a stochastic process $\bigl( Y(t), \, t\geq 0\bigr)$ is said to have stationary increments if for any $r\geq 0$ $$\label{e:stat.incr} \bigl( Y(t+r)-Y(r), \, t\geq 0\bigr) {\stackrel{d}{=}}\bigl( Y(t)-Y(0), \, t\geq 0\bigr) \,;$$ see e.g. [@embrechts:maejima:2002] and [@samorodnitsky:2006LRD]. In the context of the functional limit theorem for the maxima , a different property appears. \[d:max.stat.def\] A stochastic process $( Y(t), \, t \geq 0 )$ is said to have stationary max-increments if for every $r \geq 0$, there exists, perhaps on an enlarged probability space, a stochastic process $\bigl( Y^{(r)}(t), \, t \geq 0 \bigr)$ such that $$\begin{aligned} \label{e:def.statmaxi} \bigl( Y^{(r)}(t), \, t \geq 0 \bigr) &\stackrel{d}{=} ( Y(t), \, t \geq 0 )\,, \nonumber \\ ( Y(t+r), \, t \geq 0 ) &\stackrel{d}{=} \bigl( Y(r) \vee Y^{(r)}(t), \, t \geq 0 \bigr)\,.\end{aligned}$$ Notice the analogy between the definition of stationary increments (when $Y(0)=0$) and Definition \[d:max.stat.def\]. Since the operations of taking the maximum is not invertible (unlike summation), the latter definition, by necessity, is stated in terms of existence of the max-increment process $\bigl( Y^{(r)}(t), \, t \geq 0 \bigr)$. \[t:max.lamperti\] Let $(X_1,X_2,\ldots)$ be a stationary sequence. Assume that for some sequence $a_n\to\infty$, and a stochastic process $(Y(t), \, t\geq 0)$ such that, for every $t>0$, $Y(t)$ is not a constant, $$\left( \frac{1}{a_n} M_{\lfloor nt\rfloor}, \ t\geq 0\right) \Rightarrow (Y(t), \, t\geq 0)$$ in terms of convergence of finite-dimensional distributions. Then $(Y(t), \, t\geq 0)$ is self-similar with exponent $H>0$ of self-similarity, and has stationary max-increments. Furthermore, the sequence $(a_n)$ is regularly varying with index $H$. The facts that the limiting process $(Y(t), \, t\geq 0)$ is self-similar with exponent $H>0$ of self-similarity, and that the sequence $(a_n)$ is regularly varying with index $H$, follow from the Lamperti theorem; see [@lamperti:1962], or Theorem 2.1.1 in [@embrechts:maejima:2002]. Note that the distinction between convergence along a discrete sequence $n\to\infty$, or along a continuous sequence $\lambda\to\infty$, disappears for maxima of stationary processes. We check now the stationarity of the max-increments of the limiting process. Let $r>0$, and $t_i>0, \, i=1,\ldots, k$, some $k\geq 1$. Write $$\label{e:decompose} \frac{1}{a_n} M_{\lfloor n(t_i+r)\rfloor} = \frac{1}{a_n} M_{\lfloor nr\rfloor} \bigvee \frac{1}{a_n} \max_{nr<j\leq n(t_i+r)} X_j, \ i=1,\ldots, k\,.$$ By the assumption of the theorem and stationarity of the process $(X_1,X_2,\ldots)$, $$\frac{1}{a_n} M_{\lfloor nr\rfloor} \Rightarrow Y(r), \ \ \Bigl( \frac{1}{a_n} \max_{nr<j\leq n(t_i+r)} X_j, \ i=1,\ldots, k\Bigr) \Rightarrow \bigl( Y(t_1),\ldots, Y(t_k)\bigr)$$ as $n\to\infty$. Using the standard tightness argument, we conclude that for every sequence $n_m\to\infty$ there is a subsequence $n_{m(l)}\to\infty$ such that $$\biggl( \frac{1}{a_{n_{m(l)}}} M_{\lfloor {n_{m(l)}}r\rfloor}, \ \Bigl( \frac{1}{a_{n_{m(l)}}} \max_{n_{m(l)}r<j\leq n_{m(l)}(t_i+r)} X_j, \ i=1,\ldots, k\Bigr)\biggr) \Rightarrow \Bigl( Y(r), \, \bigl( Y^{(r)}(t_1),\ldots, Y^{(r)}(t_k)\bigr)\Bigr)$$ as $l\to\infty$, where $\bigl( Y^{(r)}(t_1),\ldots, Y^{(r)}(t_k)\bigr){\stackrel{d}{=}}\bigl( Y(t_1),\ldots, Y(t_k)\bigr)$. Let now $\tau_i, \, i=1,2,\ldots$ be an enumeration of the rational numbers in $[0,\infty)$. A diagonalization argument shows that there is a sequence $n_m\to\infty$ and a stochastic process $\bigl( Y^{(r)}(\tau_i), \, i=1,2,\dots \bigr)$ with $\bigl( Y^{(r)}(\tau_i), \, i=1,2,\dots \bigr){\stackrel{d}{=}}\bigl( Y(\tau_i), \, i=1,2,\dots \bigr)$ such that $$\label{e:rational.conv} \biggl( \frac{1}{a_{n_{m}}} M_{\lfloor {n_{m}}r\rfloor}, \ \Bigl( \frac{1}{a_{n_{m}}} \max_{n_{m}r<j\leq n_{m}(\tau_i+r)} X_j, \ i=1,2,\ldots \Bigr)\biggr) \Rightarrow \Bigl( Y(r), \, \bigl( Y^{(r)}(\tau_i), \ i=1,2,\ldots\bigr)\Bigr)$$ in finite-dimensional distributions, as $m\to\infty$. We extend the process $Y^{(r)}$ to the entire positive half-line by setting $$Y^{(r)}(t) = \frac12 \Bigl( \lim_{\tau\uparrow t, \ \text{rational}} Y^{(r)}(\tau) + \lim_{\tau\downarrow t, \ \text{rational}} Y^{(r)}(\tau) \Bigr), \ \ t\geq 0\,.$$ Then and monotonicity imply that as $m \to \infty$, $$\label{e:full.conv} \biggl( \frac{1}{a_{n_{m}}} M_{\lfloor {n_{m}}r\rfloor}, \ \Bigl( \frac{1}{a_{n_{m}}} \max_{n_{m}r<j\leq n_{m}(t+r)} X_j, \ t\geq 0 \Bigr)\biggr) \Rightarrow \Bigl( Y(r), \, \bigl( Y^{(r)}(t), \ t\geq 0\bigr)\Bigr)$$ in finite-dimensional distributions. Now the stationarity of max-increments follows from , and continuous mapping theorem. \[rk:sup.measures\] [Self-similar processes with stationary max-increments arising in a functional maxima scheme are close in spirit to the stationary self-similar extremal processes of [@obrien:torfs:vervaat:1990], while extremal processes themselves are defined as random sup measures. A random sup measure is, as its name implies, indexed by sets. They also arise in a limiting maxima scheme similar to , but with a stronger notion of convergence. Every stationary self-similar extremal processes trivially produces a self-similar process with stationary max-increments via restriction to sets of the type $[0,t]$ for $t\geq 0$, but the connection between the two objects remains unclear. Our limiting process in Theorem \[t:main.max\] below can be extended to a stationary self-similar extremal processes, but the extension is highly nontrivial, and will not be pursued here. ]{} It is not our goal in this paper to study in details the properties of self-similar processes with stationary max-increments, so we restrict ourselves to the following basic result, \[pr:sssmaxi\] Let $\bigl( Y(t), \, t \geq 0 \bigr)$ be a nonnegative self-similar process with stationary max-increments, and exponent $H$ of self-similarity. Suppose $\bigl(Y(t), \, t \geq 0 \bigr)$ is not identically zero. Then $H\geq 0$, and the following statements hold. [(a)]{}  If $H=0$, then $Y(t)=Y(1)$ a.s. for every $t>0$. [(b)]{}  If $0<EY(1)^p<\infty$ for some $p>0$, then $H\leq 1/p$. [(c)]{} If $H>0$, $\bigl( Y(t), \, t \geq 0 \bigr)$ is continuous in probability. By the stationarity of max-increments, $Y(t)$ is stochastically increasing with $t$. This implies that $H\geq 0$. If $H=0$, then $Y(n){\stackrel{d}{=}}Y(1)$ for each $n=1,2\ldots$. We use with $r=1$. Using $t=1$ we see that, in the right hand side of , $Y(1)=Y^{(1)}(1)$ a.s. Since $Y^{(1)}(n)\geq Y^{(1)}(1)$ a.s., we conclude, using $t=n$ in the right hand side of , that $Y(1)=Y^{(1)}(n)$ a.s for each $n=1,2,\ldots$. By monotonicity, we conclude that the process $\bigl( Y^{(1)}(t), \, t \geq 0 \bigr)$, hence also the process $\bigl( Y(t), \, t \geq 0 \bigr)$, is a.s. constant on $[1,\infty)$ and then, by self-similarity, also on $(0,\infty)$. Next, let $p>0$ be such that $0<EY(1)^p<\infty$. It follows from with $r=1$ that $$2^H Y(1){\stackrel{d}{=}}Y(2){\stackrel{d}{=}}\max\bigl( Y(1), \, Y^{(1)}(1)\bigr)\,.$$ Therefore, $$2^{pH} EY(1)^p = E Y(2)^p = E \bigl[ Y(1)^p \vee Y^{(1)}(1)^p \bigr] \leq 2 EY(1)^p$$ This means that $pH\leq 1$. Finally, we take arbitrary $0<s<t$. We use with $r=s$. For every $\eta > 0$, $$P \bigl( Y(t) - Y(s) > \eta \bigr) = P \bigl( Y(s) \vee Y^{(s)}(t-s) - Y(s) > \eta \bigr)$$ $$\leq P \bigl( Y^{(s)}(t-s) > \eta \bigr) = P \bigl( (t-s)^H Y(1) > \eta \bigr).$$ Hence continuity in probability. We now define the limiting process obtained in the main limit theorem of Section \[sec:FLTPM\], and place it in the general framework introduced earlier in this section. Let $\alpha>0$, and consider the extremal Fréchet process $Z_{\alpha}(t), \, t\geq 0$, defined in , with the scale $\sigma=1$. For $0<\beta<1$ we define a new stochastic process by $$Z_{\alpha, \beta}(t) = Z_{\alpha}(t^\beta), \ t\geq 0\,.$$ Since the extremal Fréchet process is self-similar with $H=1/\alpha$, it is immediately seen that the process $Z_{\alpha, \beta}$ is self-similar with $H=\beta/\alpha$. We claim that the process $Z_{\alpha, \beta}$ has stationary max-increments as well. To show this, we start with a useful representation of the extremal Fréchet process $Z_{\alpha}(t), \, t\geq 0$ in terms of the points of a Poisson random measure. Let $\bigl( (j_k,s_k)\bigr)$ be the points of a Poisson random measure on ${{{\mathbb R}}}_+^2$ with mean measure $\rho_{\alpha} \times \lambda$, where $\rho_{\alpha}(x,\infty) = x^{-\alpha}$, $x>0$ and $\lambda$ is the Lebesgue measure on ${{{\mathbb R}}}_+$. Then an elementary calculation shows that $$\bigl( Z_{\alpha}(t), \, t\geq 0\bigr) {\stackrel{d}{=}}\Bigl( \sup\bigl\{ j_k:\, s_k\leq t\bigr\}, \, t\geq 0\Bigr)\,.$$ Therefore, $\bigl( Z_{\alpha,\beta}(t), \, t\geq 0\bigr) {\stackrel{d}{=}}\bigl( U_{\alpha,\beta}(t), \, t\geq 0\bigr) $, where $$\label{e:represent.U} U_{\alpha,\beta}(t) = \sup \, \{ \, j_k : s_k \leq t^{\beta} \}\,, \ \ t \geq 0\,.$$ Given $r > 0$, we define $$U_{\alpha,\beta}^{(r)}(t) = \sup \, \{ \, j_k : (t+r)^{\beta}-t^{\beta} \leq s_k \leq (t+r)^{\beta} \}\,.$$ Since $$\bigl( (t_1+r)^{\beta}-t_1^{\beta}, \, (t_1+r)^{\beta}\bigr) \subset \bigl( (t_2+r)^{\beta}-t_2^{\beta}, \, (t_2+r)^{\beta}\bigr)$$ for $0\leq t_1<t_2$, it follows that $$\bigl( U_{\alpha,\beta}^{(r)}(t), \, t \geq 0 \bigr) \stackrel{d}{=} \bigl( U_{\alpha,\beta}(t), \, t \geq 0 \bigr)\,.$$ Furthermore, since $(t+r)^{\beta}-t^{\beta}\leq r^\beta$, we see that $$U_{\alpha,\beta}(t+r) = U_{\alpha,\beta}(r) \vee U_{\alpha,\beta}^{(r)}(t) \ \ \text{for all } t \geq 0\,.$$ This means that the process $U_{\alpha,\beta}$ has stationary max-increments and, hence, so does the process $Z_{\alpha,\beta}$. It is interesting to note that, by part (b) of Proposition \[pr:sssmaxi\], any $H$-self-similar process with stationary max-increments and $\alpha$-Fréchet marginals, must satisfy $H\leq 1/\alpha$. The exponent $H=\beta/\alpha$ with $0<\beta\leq 1$ of the process $Z_{\alpha,\beta}$ (with $\beta=1$ corresponding to the extremal Fréchet process $Z_{\alpha}$) covers the entire interval $(0,1/\alpha]$. Therefore, the upper bound of part (b) of Proposition \[pr:sssmaxi\] is, in general, the best possible. As a general phenomenon, the stationary increments of a self-similar stochastic process arising in a functional central limit theorem (for partial sums) exhibit dependence reflecting, at least partially the dependence in the original stationary process. This can be seen, for example, in [@owada:samorodnitsky:2012] that starts with a stationary process similar in its nature to the one considered in this paper. Do the stationary max-increments of a process arising in the functional maxima scheme reflect similarly the dependence structure in the original process? The fact the operation of taking partial maxima tends to reduce the memory in the original process has been noticed before; for example, for stationary Gaussian processes partial maxima converge, with the same normalization, to the extremal Gumbel process, as long as the correlations decay more rapidly than $(\log n)^{-1}$; see [@berman:1964]. Within that range of correlation decay, on the other hand, one would observe that the partial sums stop converging to the Brownian motion, and start converging to the Fractional Brownian motion, and the normalization will change as well. A related example with power tails is given by the infinite moving average models of the form $X_n = \sum_{j=0}^{\infty} c_j Z_{n-j}$, where $(Z_n)$ are i.i.d. random variables, belonging to the max-domain of attraction of an $\alpha$-Fréchet law, and $(c_j)$ is a deterministic sequence satisfying certain conditions for the process to be well-defined. In this case it was shown by [@davis:resnick:1985] that the normalized partial maxima of $(X_n)$ converge in the functional sense to an extremal $\alpha$-Fréchet process; this paper assumes absolute summability of $(c_j)$, but it is only needed for the tail estimate (2.7), which is valid without the absolute summability, see [@mikosch:samorodnitsky:2000]. In contrast, partial sums of the infinite moving average processes can converge, if $0<\alpha<2$, in certain cases (some of which prevent lack of summability of the coefficients), to the linear fractional $\alpha$-stable motions, and not to the $\alpha$-stable Lévy motions; see e.g. Section 4.7 in [@whitt:2002]. With the discussion in view, we investigate the dependence properties of the stationary max-increments of the process $Z_{\alpha,\beta}$ defined above. This process will appear, in the case $0< \alpha < 2$ and $1/2 < \beta < 1$, in the main limit theorem, Theorem \[t:main.max\] below. Equivalently, we will look at the max-increments of the process $U_{\alpha,\beta}$ in . Let $$\label{e:max.increment.2} V_n^{(\alpha,\beta)} = \sup \bigl\{ \, j_k : n^{\beta}-1 < s_k \leq n^{\beta} \bigr\}\,, \ \ n=1,2,\dots\,.$$ An important observation is that this sequence is independent only when $\beta=1$, that is, when $Z_{\alpha,\beta}$ and $U_{\alpha,\beta}$ coincide with the extremal $\alpha$-Fréchet process. In general, we have the following statement. The stationary process $(V_n^{(\alpha,\beta)}, \, n \geq 1)$ is mixing and is generated by a dissipative flow. Notice that $$(V_n^{(\alpha,\beta)}, \, n \geq 1) \stackrel{d}{=} \bigl( M_{\alpha}((n^{\beta}-1, n^{\beta}]), \, n \geq 1\bigr)\,,$$ where $M_{\alpha}$ is an independently scattered $\alpha$-Fréchet sup-measure on ${{{\mathbb R}}}_+$ with the Lebesgue control measure. See [@stoev:taqqu:2005]. It follows by Theorem 3.4 in [@stoev:2008] that $(V_n^{(\alpha,\beta)}, \, n \geq 1)$ is mixing. Further, $\sum_{n=1}^{\infty} {{\bf 1}}_{(n^{\beta}-1, n^{\beta}] }(x) < \infty$ a.e., and Theorem 5.2 of [@wang:stoev:2010] shows that $V_n^{(\alpha,\beta)}$ is generated by a dissipative flow. The main result in Theorem \[t:main.max\] below deals with the partial maxima of stationary S$\alpha$S processes for which the underlying flows are conservative; these are associated with long memory in the process. The resulting limiting process is not an extremal process, hence its max-increments are dependent. The memory in the original S$\alpha$S is long enough to guarantee that. On the other hand, the max-increments of the latter are generated by a dissipative flow, which is associated with shorter memory. We observe, therefore, another example of “memory reduction” by the operation of taking partial maxima. We finish this section by mentioning that an immediate conclusion from is the following representation of the process $Z_{\alpha, \beta}$ on the interval $[0,1]$: $$\label{e:represent.Z} \left( Z_{\alpha, \beta}(t), \, 0 \leq t \leq 1 \right) \stackrel{d}{=} \left( \bigvee_{j=1}^{\infty} \Gamma_j^{-1/\alpha} {{\bf 1}}_{\{ V_j \leq t \}}, \, 0 \leq t \leq 1 \right)\,,$$ where $\Gamma_j$, $j=1,2,\dots$, are arrival times of a unit rate Poisson process on $(0,\infty)$, and $(V_j)$ are i.i.d. random variables with $P(V_1 \leq x) = x^{\beta}$, $0 < x \leq 1$, independent of $(\Gamma_j)$. A Functional Limit Theorem for Partial Maxima {#sec:FLTPM} ============================================= In this section we state and prove our main result, a functional limit theorem for the partial maxima of the process ${{\bf X}}=(X_1,X_2,\dots)$ given in (\[e:underlying.proc\]). Recall that $T$ is a conservative, ergodic and measure preserving map on a $\sigma$-finite, infinite, measure space $(E,\mathcal{E},\mu)$. We will assume that $T$ is a pointwise dual ergodic map with normalizing sequence $(a_n)$ that is regularly varying with exponent $1-\beta$; equivalently, the wandering sequence $(w_n)$ in is assumed to be regularly varying with exponent $\beta$. Crucially, we will assume that $1/2 < \beta < 1$. See Remark \[rk:other.beta\] after the statement of Theorem \[t:main.max\] below. Define $$\label{e:b.n} b_n = \left( \int_E \max_{1 \leq k \leq n}\bigl| f\circ T^n(x)\bigr|^{\alpha} \mu(dx) \right)^{1/\alpha}, \ \ \ n=1,2,\dots\,.$$ The sequence $(b_n)$ is known to play an important role in the rate of growth of partial maxima of an $\alpha$-stable process of the type . It also turns out to be a proper normalizing sequence for our functional limit theorem. In [@samorodnitsky:2004a] it was shown that, for a canonical kernel (\[e:general.rosinski\]), if the map $T$ is conservative, then the sequence $(b_n)$ grows at a rate strictly slower than $n^{1/\alpha}$. The extra assumptions imposed in this paper will guarantee a more precise statement: $(b_n) \in RV_{\beta/\alpha}$. Specifically, $$\label{e:RV.exp.bn} \lim_{n\to\infty}\frac{b_n^\alpha}{ w_n}=\| f\|_\infty$$ (where $(w_n)$ is the wandering sequence). This fact has an interesting message, because it explicitly shows that the rate of growth of the partial maxima is determined both by the heaviness of the marginal tails (through $\alpha$) and by the length of memory (through $\beta$). In contrast, if the map $T$ has a non-trivial dissipative component, then the sequence $(b_n)$ grows at the rate $n^{1/\alpha} $, and so do the partial maxima of a stationary [S$\alpha$S]{} process; see [@samorodnitsky:2004a]. This is the limiting case of the setup in the present paper, as $\beta$ gets closer to $1$. Intuitively, the smaller is $\beta$, the longer is the memory in the process. The functional limit theorem in Theorem \[t:main.max\] below involves weak convergence in the space $D[0,\infty)$ equipped with two different topologies, the Skorohod $J_1$-topology and the Skorohod $M_1$-topology, introduced in [@skorohod:1956]. The details could be found, for instance, in [@billingsley:1999] (for the $J_1$-topology), and in [@whitt:2002] (for the $M_1$-topology). See Remark \[rk:topologies\]. We recall the tail constant of an $\alpha$-stable random variable given by $$C_{\alpha} = \left( \int_0^{\infty} x^{-\alpha} \sin x \, dx \right)^{-1} = \begin{cases} (1-\alpha) / \bigl(\Gamma(2-\alpha) \cos(\pi \alpha / 2)\bigr) & \ \ \text{if } \alpha \neq 1, \\ 2 / \pi & \ \ \text{if } \alpha=1; \end{cases}$$ see [@samorodnitsky:taqqu:1994]. \[t:main.max\] Let $T$ be a conservative, ergodic and measure preserving map on a $\sigma$-finite infinite measure space $(E,\mathcal{E},\mu)$. Assume that $T$ is a pointwise dual ergodic map with normalizing sequence $(a_n) \in RV_{1-\beta}$, $0\leq \beta\leq 1$. Let $f\in L^{\alpha}(\mu)\cap L^{\infty}(\mu)$, and assume that $f$ is supported by a uniform set $A$ for $T$. Let $\alpha>0$. Then the sequence $(b_n)$ in satisfies . Assume now that $0<\alpha<2$ and $1/2<\beta<1$. If $M$ is a S$\alpha$S random measure on $(E,\mathcal{E})$ with control measure $\mu$, then the stationary S$\alpha$S process ${{\bf X}}$ given in (\[e:underlying.proc\]) satisfies $$\label{e:weak.conv} \Bigl( \frac{1}{b_n} \max_{1 \leq k \leq \lfloor nt \rfloor}|X_k|, \, t\geq 0\Bigr) \Rightarrow \bigl( C_{\alpha}^{1/\alpha} Z_{\alpha, \beta}(t), \, t\geq 0\bigr) \ \ \ \text{in } D[0,\infty)$$ in the Skorohod $M_1$-topology. Moreover, if $f = {{\bf 1}}_A$, then the above convergence occurs in the Skorohod $J_1$-topology as well. \[rk:other.beta\] [Convergence to a Fréchet limit in no longer holds in the range $0<\beta<1/2$. We will discuss elsewhere what non-Fréchet limiting processes may appear in that case. In the boundary case $\beta=1/2$, however, the statement may still hold. This will be the case when the “one jump” property below is satisfied. This is the case, for example, for the Markov shift operator $T$ presented at the end of the paper. See also [@samorodnitsky:2004a]. ]{} \[rk:topologies\] [It is not difficult to see why the weak convergence in holds in the $J_1$-topology for indicator functions, but only in the $M_1$-topology in general. Indeed, for a general function $f$ the values of the process $M_{\lfloor n\cdot\rfloor}(f)$ in Proposition \[p:max.ergodic\] may have multiple jumps on the time scale $o(n)$ before reaching the limiting value $\|f\|_\infty$. Since the limiting process has a single jump, one does not expect the $J_1$-convergence. On the other hand, if $f = {{\bf 1}}_A$, then the value $\|f\|_\infty=1$ is reached in a single jump, matching the single jump in the limiting process. ]{} We start with verifying . Obviously, $$b_n^{\alpha} \leq \, \| f \|_{\infty} \mu (\varphi_A \leq n)\,,$$ and the definition of the wandering rate sequence gives us the upper bound $$\limsup_{n\to\infty}\frac{b_n^\alpha}{ w_n}\leq \| f\|_\infty\,.$$ On the other hand, take an arbitrary $\epsilon \in \bigl( 0,\| f \|_{\infty}\bigr)$. The set $$B_{\epsilon} = \bigl\{x \in A: |f(x)| \geq \, \| f \|_{\infty} - \epsilon \bigr\}\,.$$ is a uniform set for $T$. Let $\bigl( w_n^{({\varepsilon})}\bigr)$ be the corresponding wandering rate sequence. Then a lower bound for $b_n^{\alpha}$ is obtained by $$b_n^{\alpha} \geq \bigl( \| f \|_{\infty} - \epsilon \, \bigr) \, \mu \Bigl( \bigcup_{j=1}^n T^{-j} B_{\epsilon} \Bigr)\,,$$ so $$\liminf_{n\to\infty}\frac{b_n^\alpha}{ w_n} = \liminf_{n\to\infty}\frac{b_n^\alpha}{ w_n^{({\varepsilon})}} \geq \| f \|_{\infty} - \epsilon\,.$$ Letting ${\varepsilon}\to 0$, we obtain . Suppose now that $0<\alpha<2$ and $1/2<\beta<1$. We continue with proving convergence in the finite dimensional distributions in . Since for random elements in $D[0,\infty)$ with nondecreasing sample paths, weak convergence in the $M_1$-topology is implied by the finite-dimensional weak convergence, this will also establish in the sense of weak convergence in the $M_1$-topology. Fix $0 = t_0< t_1 < \dots < t_d$, $d \geq 1$. We may and will assume that $t_d \leq 1$. We use a series representation of the random vector $(X_1, \dots, X_n)$: with $f_k=f\circ T^k$, $k=1,2,\ldots$, $$\label{e:series.max} (X_k, \, k=1,\dots,n) \stackrel{d}{=} \left( b_n \, C_{\alpha}^{1/\alpha} \sum_{j=1}^{\infty} \epsilon_j \Gamma_j^{-1/\alpha} \frac{f_k(U_j^{(n)})}{\max_{1 \leq i \leq n} |f_i(U_j^{(n)})|}, \, k=1,\dots,n \right)\,.$$ Here $(\epsilon_j)$ are i.i.d. Rademacher random variables (symmetric random variables with values $\pm 1$), $(\Gamma_j)$ are the arrival times of a unit rate Poisson process on $(0,\infty)$, and $(U_j^{(n)})$ are i.i.d. $E$-valued random variables with the common law $\eta_n$ defined by $$\label{e:eta.n} \frac{d\eta_n}{d\mu}(x) = \frac{1}{b_n^{\alpha}} \max_{1 \leq k \leq n} |f_k(x)|^{\alpha}\,, \ \ x \in E\,.$$ The sequences $(\epsilon_j)$, $(\Gamma_j)$, and $(U_j^{(n)})$ are taken to be independent. We refer to Section 3.10 of [@samorodnitsky:taqqu:1994] for series representations of $\alpha$-stable random vectors. The representation was also used in [@samorodnitsky:2004a], and the argument below is structured similarly to the corresponding argument [*ibid.*]{} The crucial consequence of the assumption $1/2<\beta<1$ is that, in the series representation (\[e:series.max\]), only the largest Poisson jump will play an important role. It is shown in [@samorodnitsky:2004a] that, under the assumptions of Theorem \[t:main.max\], for every $\eta>0$, $$\label{e:single.poisson.jump} \varphi_n(\eta) \equiv P \left( \bigcup_{k=1}^n \left\{ \Gamma_j^{-1/\alpha} \frac{|f_k(U_j^{(n)})|}{\max_{1 \leq i \leq n}|f_i(U_j^{(n)})|} > \eta \ \ \text{for at least 2 different } j=1,2,\dots \right\} \right) \to 0$$ as $n \to \infty$. We will proceed in two steps. First, we will prove that $$\label{e:fidi.Malphabeta} \left( \bigvee_{j=1}^{\infty} \Gamma_j^{-1/\alpha} \frac{\max_{1 \leq k \leq \lfloor nt_i \rfloor} |f_k(U_j^{(n)})|}{\max_{1 \leq k \leq n} |f_k(U_j^{(n)})|}, \, i=1,\dots,d \right) \Rightarrow \left( Z_{\alpha, \beta}(t_i), \, i=1,\dots,d \right) \ \ \ \text{in } {{{\mathbb R}}}_+^d\,.$$ Next, we will prove that, for fixed $\lambda_1, \dots, \lambda_d > 0$, for every $0<\delta<1$, $$P \bigl( b_n^{-1} \max_{1 \leq k \leq \lfloor nt_i \rfloor} |X_k| > \lambda_i, \ \ i=1,\dots,d \bigr)$$ $$\label{e:upp.bdd.max} \leq P \left( C_{\alpha}^{1/\alpha} \bigvee_{j=1}^{\infty} \Gamma_j^{-1/\alpha} \frac{\max_{1 \leq k \leq \lfloor nt_i \rfloor}|f_k(U_j^{(n)})|}{\max_{1 \leq k \leq n}|f_k(U_j^{(n)})|} > \lambda_i(1-\delta), \ \ i=1,\dots,d \right) + o(1)$$ and that $$P \bigl( b_n^{-1} \max_{1 \leq k \leq \lfloor nt_i \rfloor} |X_k| > \lambda_i, \ \ i=1,\dots,d \bigr)$$ $$\label{e:lower.bdd.max} \geq P \left( C_{\alpha}^{1/\alpha} \bigvee_{j=1}^{\infty} \Gamma_j^{-1/\alpha} \frac{\max_{1 \leq k \leq \lfloor nt_i \rfloor}|f_k(U_j^{(n)})|}{\max_{1 \leq k \leq n}|f_k(U_j^{(n)})|} > \lambda_i(1+\delta), \ \ i=1,\dots,d \right) + o(1)\,.$$ Since the Fréchet distribution is continuous, the weak convergence $$\Bigl( b_n^{-1} \max_{1 \leq k \leq \lfloor nt_i \rfloor} |X_k|, \ \ i=1,\dots,d \Bigr) \Rightarrow \left( Z_{\alpha, \beta}(t_i), \, i=1,\dots,d \right) \ \ \ \text{in } {{{\mathbb R}}}_+^d$$ will follow by taking $\delta$ arbitrarily small. We start with proving . For $n=1,2,\ldots$, $N_n=\sum_{j=1}^{\infty} \delta_{(\Gamma_j, U_j^{(n)})}$ is a Poisson random measure on $(0,\infty)\times \cup_{k=1}^{n} T^{-k}A$ with mean measure $\lambda \times \eta_n$. Define a map $S_n:{{{\mathbb R}}}_+ \times \cup_{k=1}^{n} T^{-k}A \to {{{\mathbb R}}}_+^d$ by $$S_n(r,x) = r^{-1/\alpha} \bigl( M_n(f)(x) \bigr)^{-1} \bigl( M_{\lfloor nt_1 \rfloor}(f)(x), \dots, M_{\lfloor nt_d \rfloor}(f)(x) \bigr)\,, \ \ r > 0, \ x \in \cup_{k=1}^{n} T^{-k}A\,.$$ Then, for $\lambda_1, \dots, \lambda_d > 0$, $$P \left( \bigvee_{j=1}^{\infty} \Gamma_j^{-1/\alpha} \frac{\max_{1 \leq k \leq \lfloor nt_i \rfloor} |f_k(U_j^{(n)})|}{\max_{1 \leq k \leq n} |f_k(U_j^{(n)})|} \leq \lambda_i\,, \ \ i=1,\dots,d \right)$$ $$=P \left[ N_n\left( S_n^{-1} \Bigl( (0,\lambda_1] \times \dots \times (0,\lambda_d] \Bigr)^c\right) = 0 \right]$$ $$=\exp \left\{ - (\lambda \times \eta_n) \left( S_n^{-1} \Bigl( (0,\lambda_1] \times \dots \times (0,\lambda_d] \Bigr)^c \right)\right\}$$ $$=\exp \left\{ - (\lambda \times \eta_n) \left\{(r,x): \bigvee_{j=1}^d \lambda_j^{-\alpha} \frac{\bigl(M_{\lfloor nt_j \rfloor}(f)(x)\bigr)^{\alpha}}{\bigl(M_n(f)(x)\bigr)^{\alpha}} > r \right\} \right\}$$ $$= \exp \left\{ -b_n^{-\alpha} \int_E \bigvee_{j=1}^d \lambda_j^{-\alpha} M_{\lfloor nt_j \rfloor}(f)^{\alpha} d\mu \right\}\,.$$ We use and the weak convergence in Proposition \[p:max.ergodic\] to obtain $$b_n^{-\alpha} \int_E \bigvee_{j=1}^d \lambda_j^{-\alpha} M_{\lfloor nt_j \rfloor}(f)^{\alpha} d\mu \sim \| f \|_{\infty}^{-1} \int_E \bigvee_{j=1}^d \lambda_j^{-\alpha} M_{\lfloor nt_j \rfloor}(f)^{\alpha} d\mu_n$$ $$\to \int_{\Omega^{\prime}} \bigvee_{j=1}^d \lambda_j^{-\alpha} {{\bf 1}}_{\{ V_{\beta} \leq t_j \}} dP^{\prime} = \sum_{i=1}^d (t_i^{\beta} - t_{i-1}^{\beta}) \left( \bigwedge_{j=i}^d \lambda_j \right)^{-\alpha}.$$ Therefore, $$P \left( \bigvee_{j=1}^{\infty} \Gamma_j^{-1/\alpha} \frac{\max_{1 \leq k \leq \lfloor nt_i \rfloor} |f_k(U_j^{(n)})|}{\max_{1 \leq k \leq n} |f_k(U_j^{(n)})|} \leq \lambda_i\,, \ \ i=1,\dots,d \right)$$ $$\to \exp \left\{ - \sum_{i=1}^d (t_i^{\beta} - t_{i-1}^{\beta}) \left( \bigwedge_{j=i}^d \lambda_j \right)^{-\alpha} \right\} = P \bigl( Z_{\alpha, \beta}(t_i) \leq \lambda_i, \ \ i=1,\dots,d \bigr) \,.$$ The claim (\[e:fidi.Malphabeta\]) has, consequently, been proved. We continue with the statements (\[e:upp.bdd.max\]) and (\[e:lower.bdd.max\]). Since the arguments are very similar, we only prove (\[e:upp.bdd.max\]). Let $K \in {{\mathbb N}}$ and $0 < \epsilon < 1$ be constants so that $$K+1 > \frac{4}{\alpha} \ \ \ \text{and } \ \ \delta - \epsilon K > 0\,.$$ Then $$P \bigl( b_n^{-1} \max_{1 \leq k \leq \lfloor nt_i \rfloor} |X_k| > \lambda_i, \ \ i=1,\dots,d \bigr)$$ $$\leq P \left( C_{\alpha}^{1/\alpha} \bigvee_{j=1}^{\infty} \Gamma_j^{-1/\alpha} \frac{\max_{1 \leq k \leq \lfloor nt_i \rfloor}|f_k(U_j^{(n)})|}{\max_{1 \leq k \leq n}|f_k(U_j^{(n)})|} > \lambda_i(1-\delta), \ \ i=1,\dots,d \right) + \varphi_n \bigl( C_{\alpha}^{-1/\alpha} \epsilon \min_{1 \leq i \leq d} \lambda_i \bigr)$$ $$+ \sum_{i=1}^d P \Biggl( C_{\alpha}^{1/\alpha} \max_{1 \leq k \leq \lfloor nt_i \rfloor} \left| \sum_{j=1}^{\infty} \epsilon_j \Gamma_j^{-1/\alpha} \frac{f_k(U_j^{n})}{\max_{1 \leq l \leq n}|f_l(U_j^{(n)})|} \right| > \lambda_i\,,$$ $$C_{\alpha}^{1/\alpha} \bigvee_{j=1}^{\infty} \Gamma_j^{-1/\alpha} \frac{\max_{1 \leq k \leq \lfloor nt_i \rfloor}|f_k(U_j^{(n)})|}{\max_{1 \leq k \leq n}|f_k(U_j^{(n)})|} \leq \lambda_i(1-\delta)\,, \ \ \text{and for each } m=1,\dots,n,$$ $$C_{\alpha}^{1/\alpha} \Gamma_j^{-1/\alpha} \frac{|f_m(U_j^{(n)})|}{\max_{1 \leq l \leq n}|f_l(U_j^{(n)})|} > \epsilon \min_{1 \leq l \leq d}\lambda_l \ \ \text{for at most one } j=1,2,\dots \Biggr)\,.$$ By (\[e:single.poisson.jump\]), it is enough to show that for all $\lambda>0$ and $0 \leq t \leq 1$, $$\label{e:neg.term} P \Biggl( C_{\alpha}^{1/\alpha} \max_{1 \leq k \leq \lfloor nt \rfloor} \left| \sum_{j=1}^{\infty} \epsilon_j \Gamma_j^{-1/\alpha} \frac{f_k(U_j^{(n)})}{\max_{1 \leq i \leq n}|f_i(U_j^{(n)})|} \right| > \lambda\,,$$ $$C_{\alpha}^{1/\alpha} \bigvee_{j=1}^{\infty} \Gamma_j^{-1/\alpha} \frac{\max_{1 \leq k \leq \lfloor nt \rfloor}|f_k(U_j^{(n)})|}{\max_{1 \leq k \leq n}|f_k(U_j^{(n)})|} \leq \lambda(1-\delta)\,, \ \ \text{and for each } m=1,\dots,n,$$ $$C_{\alpha}^{1/\alpha} \Gamma_j^{-1/\alpha} \frac{|f_m(U_j^{(n)})|}{\max_{1 \leq i \leq n}|f_i(U_j^{(n)})|} > \epsilon \lambda \ \ \text{for at most one } j=1,2,\dots \Biggr) \to 0\,.$$ For every $k=1,2,\dots,n$, the Poisson random measure represented by the points $$\bigl( \epsilon_j \Gamma_j^{-1/\alpha} f_k(U_j^{(n)}) \bigl( \max_{1 \leq i \leq n}|f_i(U_j^{(n)})| \bigr)^{-1}, \, j=1,2,\dots \bigr)$$ has the same mean measure as that represented by the points $$\bigl( \epsilon_j \Gamma_j^{-1/\alpha} \| f \|_{\alpha} b_n^{-1}, \, j=1,2,\dots \bigr)\,.$$ Hence, these two Poisson random measures coincide distributionally. We conclude that the probability in (\[e:neg.term\]) is bounded by $$\sum_{k=1}^{\lfloor nt \rfloor} P \Biggl( C_{\alpha}^{1/\alpha} \left| \sum_{j=1}^{\infty} \epsilon_j \Gamma_j^{-1/\alpha} \frac{f_k(U_j^{(n)})}{\max_{1 \leq i \leq n}|f_i(U_j^{(n)})|} \right| > \lambda\,, \ \ C_{\alpha}^{1/\alpha} \bigvee_{j=1}^{\infty} \Gamma_j^{-1/\alpha} \frac{f_k(U_j^{(n)})}{\max_{1 \leq i \leq n}|f_i(U_j^{(n)})|} \leq \lambda(1-\delta)\,,$$ $$C_{\alpha}^{1/\alpha} \Gamma_j^{-1/\alpha} \frac{|f_k(U_j^{(n)})|}{\max_{1 \leq i \leq n}|f_i(U_j^{(n)})|} > \epsilon \lambda \ \ \text{for at most one } j=1,2,\dots \Biggr)$$ $$= \lfloor nt \rfloor P \Biggl( C_{\alpha}^{1/\alpha} \left| \sum_{j=1}^{\infty} \epsilon_j \Gamma_j^{-1/\alpha} \right| > \lambda \| f \|_{\alpha}^{-1} b_n\,, \ \ C_{\alpha}^{1/\alpha} \bigvee_{j=1}^{\infty} \Gamma_j^{-1/\alpha} \leq \lambda (1-\delta) \| f \|_{\alpha}^{-1} b_n\,,$$ $$C_{\alpha}^{1/\alpha} \Gamma_j^{-1/\alpha} > \epsilon \lambda \| f \|_{\alpha}^{-1} b_n \ \ \text{for at most one } j=1,2,\dots \Biggr)$$ $$\leq n P \left( C_{\alpha}^{1/\alpha} \left| \sum_{j=K+1}^{\infty} \epsilon_j \Gamma_j^{-1/\alpha} \right| > (\delta - \epsilon K) \lambda \| f \|_{\alpha}^{-1} b_n \right)$$ $$\leq \frac{n \| f \|_{\alpha}^4 C_{\alpha}^{4/\alpha} }{(\delta - \epsilon K)^4 \lambda^4 b_n^4} \ E \left| \sum_{j=K+1}^{\infty} \epsilon_j \Gamma_j^{-1/\alpha} \right|^4\,.$$ Due to the choice $K+1>4/\alpha$, $$E \left| \sum_{j=K+1}^{\infty} \epsilon_j \Gamma_j^{-1/\alpha} \right|^4 < \infty;$$ see [@samorodnitsky:2004a] for a detailed proof. Since $n/b_n^4 \to 0$ as $n \to \infty$, (\[e:neg.term\]) follows. Suppose now that $f = {{\bf 1}}_A$. In that case the probability measure $\eta_n$ defined in coincides with the probability measure $\mu_n$ of Proposition \[p:max.ergodic\]. In order to prove weak convergence in the $J_1$-topology, we will use a truncation argument. First of all, we may and will restrict ourselves to the space $D[0,1]$. Let $K=1,2,\ldots$. First of all, we show, in the notation of , the convergence $$\label{e:fidi.truncate} \left( C_{\alpha}^{1/\alpha} \max_{1 \leq k \leq \lfloor nt \rfloor} \left| \sum_{j=1}^K \epsilon_j \Gamma_j^{-1/\alpha} {{\bf 1}}_A \circ T^k(U_j^{(n)}) \right|, \ 0\leq t\leq 1 \right) \Rightarrow \left( C_{\alpha}^{1/\alpha} \bigvee_{j=1}^K \Gamma_j^{-1/\alpha} {{\bf 1}}_{\{ V_j \leq t \}}, \ 0\leq t\leq 1 \right)$$ in the $J_1$-topology on $D[0,1]$. Indeed, by , outside of an event of asymptotically vanishing probability, the process in the left hand side of is $$\label{e:fidi.truncate1} \left( C_{\alpha}^{1/\alpha} \bigvee_{j=1}^K \Gamma_j^{-1/\alpha} \max_{1 \leq k \leq \lfloor nt \rfloor} {{\bf 1}}_A \circ T^k(U_j^{(n)}), \ 0\leq t\leq 1 \right) \,.$$ By Proposition \[p:max.ergodic\], we can put all the random variables involved on the same probability space so that the time of the single step in the $j$th term in converges a.s. for each $j=1,\ldots, K$ to $V_j$. Then, trivially, the process in converges a.s. in the $J_1$-topology on $D[0,1]$ to the process in the right hand side of . Therefore, the weak convergence in follows. Next, we note that in the $J_1$-topology on the space $D[0,1]$, $$\left( C_{\alpha}^{1/\alpha} \bigvee_{j=1}^K \Gamma_j^{-1/\alpha} {{\bf 1}}_{\{ V_j \leq t \}}, \ 0\leq t\leq 1\right) \to \left( C_{\alpha}^{1/\alpha} \bigvee_{j=1}^{\infty} \Gamma_j^{-1/\alpha} {{\bf 1}}_{\{ V_j \leq t_i \}} \ 0\leq t\leq 1\right) \ \ \ \text{as } K \to \infty \ \ \ \text{a.s.}$$ This is so because, as $K \to \infty$, $$\sup_{0 \leq t \leq 1} \left( \bigvee_{j=1}^{\infty} \Gamma_j^{-1/\alpha} {{\bf 1}}_{\{ V_j \leq t \}} - \bigvee_{j=1}^K \Gamma_j^{-1/\alpha} {{\bf 1}}_{\{ V_j \leq t \}} \right) \leq \Gamma_{K+1}^{-1/\alpha} \to 0 \ \ \ \text{a.s.}.$$ According to Theorem 3.2 in [@billingsley:1999], the $J_1$-convergence in will follow once we show that $$\lim_{K \to \infty} \limsup_{n \to \infty} P \left( \max_{1 \leq k \leq n} \left| \sum_{j=K+1}^{\infty} \epsilon_j \Gamma_j^{-1/\alpha} {{\bf 1}}_A \circ T^k(U_j^{(n)}) \right| > \epsilon \right) = 0$$ for every $\epsilon>0$. Write $$P \left( \max_{1 \leq k \leq n} \left| \sum_{j=K+1}^{\infty} \epsilon_j \Gamma_j^{-1/\alpha} {{\bf 1}}_A \circ T^k(U_j^{(n)}) \right| > \epsilon \right)$$ $$\leq \int_0^{(\epsilon/2)^{-\alpha}} e^{-x} \frac{x^{K-1}}{(K-1)!} dx + \int_{(\epsilon/2)^{-\alpha}}^{\infty} e^{-x} \frac{x^{K-1}}{(K-1)!} \, P \left( \max_{1 \leq k \leq n} \left| \sum_{j=1}^{\infty} \epsilon_j \bigl( \Gamma_j + x \bigr)^{-1/\alpha} {{\bf 1}}_A \circ T^k(U_j^{(n)}) \right| > \epsilon \right) dx \,.$$ Clearly, the first term vanishes when $K \to \infty$. Therefore, it is sufficient to show that for every $x \geq (\epsilon/2)^{-\alpha}$, $$\label{e:upper.app.Gamma1} P \left( \max_{1 \leq k \leq n} \left| \sum_{j=1}^{\infty} \epsilon_j \bigl( \Gamma_j + x \bigr)^{-1/\alpha} {{\bf 1}}_A \circ T^k(U_j^{(n)}) \right| > \epsilon \right) \to 0$$ as $n \to \infty$.\ To this end, choose $L \in {{\mathbb N}}$ and $0 < \xi < 1/2$ so that $$\label{e:const.rest2} L+1 > \frac{4}{\alpha} \ \ \ \text{and } \ \ \frac{1}{2} - \xi L > 0\,.$$ By we can write $$\label{e:one.term.vanish} P \left( \max_{1 \leq k \leq n} \left| \sum_{j=1}^{\infty} \epsilon_j \bigl( \Gamma_j + x \bigr)^{-1/\alpha} {{\bf 1}}_A \circ T^k(U_j^{(n)}) \right| > \epsilon \right)$$ $$\leq P \Biggl( \max_{1 \leq k \leq n} \left| \sum_{j=1}^{\infty} \epsilon_j \bigl( \Gamma_j + x \bigr)^{-1/\alpha} {{\bf 1}}_A \circ T^k(U_j^{(n)}) \right| > \epsilon\,, \ \ \text{and for each } m = 1,\dots, n,$$ $$\bigl( \Gamma_j + x \bigr)^{-1/\alpha} {{\bf 1}}_A \circ T^m(U_j^{(n)}) > \xi \epsilon \ \ \text{for at most one } j =1,2,\dots \Biggr) + o(1)\,.$$ Notice that for every $k=1,\dots, n$, the Poisson random measure represented by the points $$\bigl( \epsilon_j \bigl( \Gamma_j + x \bigr)^{-1/\alpha} {{\bf 1}}_A \circ T^k(U_j^{(n)}), \, j=1,2,\dots \bigr)$$ is distributionally equal to the Poisson random measure represented by the points $$\bigl( \epsilon_j \bigl( b_n^{\alpha} \mu(A)^{-1} \Gamma_j +x \bigr)^{-1/\alpha}, \, j=1,2,\dots \bigr)\,.$$ Therefore, the first term on the right hand side of (\[e:one.term.vanish\]) can be bounded by $$\sum_{k=1}^n P \Biggl( \left| \sum_{j=1}^{\infty} \epsilon_j \bigl( \Gamma_j + x \bigr)^{-1/\alpha} {{\bf 1}}_A \circ T^k(U_j^{(n)}) \right| > \epsilon\,,$$ $$\bigl( \Gamma_j + x \bigr)^{-1/\alpha} {{\bf 1}}_A \circ T^k(U_j^{(n)}) > \xi \epsilon \ \ \text{for at most one } j=1,2,\dots \Biggr)$$ $$= nP \Biggl( \left| \sum_{j=1}^{\infty} \epsilon_j \bigl( b_n^{\alpha} \mu(A)^{-1} \Gamma_j + x \bigr)^{-1/\alpha} \right| > \epsilon\,,$$ $$\bigl( b_n^{\alpha} \mu(A)^{-1} \Gamma_j + x \bigr)^{-1/\alpha} > \xi \epsilon \ \ \text{for at most one } j = 1,2,\dots \Biggr)$$ $$\leq n P \left( \left| \sum_{j=L+1}^{\infty} \epsilon_j \bigl( b_n^{\alpha} \mu(A)^{-1} \Gamma_j + x \bigr)^{-1/\alpha} \right| > \left( \frac{1}{2} - \xi L \right)\epsilon \right)\,.$$ In the last step we used the fact that, for $x \geq (\epsilon/2)^{-\alpha}$, the magnitude of each term in the infinite sum does not exceed $\epsilon/2$. By the contraction inequality for Rademacher series (see e.g. Proposition 1.2.1 of [@kwapien:woyczynski:1992]), $$nP \left( \left| \sum_{j=L+1}^{\infty} \epsilon_j \bigl( b_n^{\alpha} \mu(A)^{-1} \Gamma_j + x \bigr)^{-1/\alpha} \right| > \left( \frac{1}{2} - \xi L \right)\epsilon \right)$$ $$\leq 2 n P \left( \left| \sum_{j=L+1}^{\infty} \epsilon_j \Gamma_j^{-1/\alpha} \right| > \left( \frac{1}{2} - \xi L \right)\epsilon \mu(A)^{-1/\alpha} b_n \right)\,.$$ As before, by Markov’s inequality and using the constraints of the constants $L \in {{\mathbb N}}$ and $0 < \xi < 1/2$ given in (\[e:const.rest2\]), $$2n P \left( \left| \sum_{j=L+1}^{\infty} \epsilon_j \Gamma_j^{-1/\alpha} \right| > \left( \frac{1}{2} - \xi L \right)\epsilon \mu(A)^{-1/\alpha} b_n \right) \leq \frac{2n\mu(A)^{4/\alpha}}{(2^{-1}-\xi L)^4 \epsilon^4 b_n^4} \ E \left| \sum_{j=L+1}^{\infty} \epsilon_j \Gamma_j^{-1/\alpha} \right|^4 \to 0$$ as $n \to \infty$ and, hence, (\[e:upper.app.Gamma1\]) follows. [There is not doubt that the convergence result in Theorem \[t:main.max\] can be extended to more general infinitely divisible random measures $M$ in , under appropriate assumptions of regular variation of the Lévy measure of $M$ and integrability of the function $f$. In particular, processes $Z_{\alpha, \beta}$ with any $\alpha>0$ are likely to appear in the limit in (and not only with $0<\alpha<2$ allowed by the assumption of stability). Furthermore, the symmetry of the process ${{\bf X}}$ has very little to do with the limiting distribution of the partial maxima. For example, a straightforward symmetrization argument allows one to extend to skewed $\alpha$-stable processes, at least in the sense of convergence of finite-dimensional distributions. The reason we decided to restrict the presentation to the symmetric stable case had to do with a particularly simple form of the series representation available in this case. This has allowed us to avoid certain technicalities that might have otherwise blurred the main message, which is the effect of memory on the functional limit theorem for the partial maxima. ]{} We conclude by mentioning that the result of Theorem \[t:main.max\] applies, among others, to the two examples in [@owada:samorodnitsky:2012], that of a flow generated by a null recurrent Markov chain, and that of an AFN-system. We only remind the reader of the setup of the former example, since it also appears in [@samorodnitsky:2004a]. Consider an irreducible null recurrent Markov chain $(x_n, \, n \geq 0)$ defined on an infinite countable state space ${\protect{\mathbb S}}$ with the transition matrix $(p_{ij})$. Let $(\pi_i, \, i \in {\protect{\mathbb S}})$ be its unique (up to constant multiplication) invariant measure with $\pi_0=1$. Note that $(\pi_i)$ is necessarily an infinite measure. Define a $\sigma$-finite and infinite measure on $(E,\mathcal{E}) = ({\protect{\mathbb S}}^{{{\mathbb N}}}, \mathcal{B}({\protect{\mathbb S}}^{{{\mathbb N}}}))$ by $$\mu(B) = \sum_{i \in {\protect{\mathbb S}}} \pi_i P_i(B), \ \ B \subseteq {\protect{\mathbb S}}^{{{\mathbb N}}}\,,$$ where $P_i(\cdot)$ denotes the probability law of $(x_n)$ starting in state $i \in {\protect{\mathbb S}}$. Let $$T(x_0, x_1, \dots) = (x_1,x_2,\dots)$$ be the usual left shift operator on ${\protect{\mathbb S}}^{{{\mathbb N}}}$. Then $T$ preserves $\mu$. Since the Markov chain is irreducible and null recurrent, $T$ is conservative and ergodic (see [@harris:robbins:1953]). Let $A = \{ x \in {\protect{\mathbb S}}^{{{\mathbb N}}}: x_0=i_0 \}$ for a fixed state $i_0\in {\protect{\mathbb S}}$, and $$\varphi_A(x) = \min \{ n \geq 1: x_n \in A \}\,, \ \ x \in {\protect{\mathbb S}}^{{{\mathbb N}}}$$ be the first entrance time. Assume that $$\sum_{k=1}^n P_0(\varphi_A \geq k) \in RV_{\beta}$$ for some $\beta \in (1/2, 1)$. Then all of the assumptions of Theorem \[t:main.max\] are satisfied for any $f\in L^{\alpha}(\mu)\cap L^{\infty}(\mu)$, supported by $A$; see [@aaronson:1997]. [43]{} natexlab\#1[\#1]{} <span style="font-variant:small-caps;">Aaronson, J.</span> (1997): *An Introduction to Infinite Ergodic Theory*, vol. 50 of *Mathematical Surveys and Monographs*, Providence: American Mathematical Society. <span style="font-variant:small-caps;">Berman, S.</span> (1964): “Limit theorems for the maximum term in stationary sequences,” *Annals of Mathematical Statistics*, 35, 502–516. <span style="font-variant:small-caps;">Billingsley, P.</span> (1999): *Convergence of Probability Measures*, New York: Wiley, second ed. <span style="font-variant:small-caps;">Cohen, S. and G. Samorodnitsky</span> (2006): “Random rewards, Fractional Brownian local times and stable self-similar processes,” *Annals of Applied Probability*, 16, 1432–1461. <span style="font-variant:small-caps;">Davis, R. and S. Resnick</span> (1985): “Limit Theory for Moving Averages of Random Variables with Regularly Varying Tail Probabilities,” *The Annals of Probability*, 13, 179–195. <span style="font-variant:small-caps;">de Haan, L. and A. Ferreira</span> (2006): *Extreme Value Theory: An Introduction*, New York: Springer. <span style="font-variant:small-caps;">Dwass, M.</span> (1964): “Extremal processes,” *The Annals of Mathematical Statistics*, 35, 1718–1725. ——— (1966): “Extremal processes, II,” *The Illinois Journal of Mathematics*, 10, 381–391. <span style="font-variant:small-caps;">Embrechts, P. and M. Maejima</span> (2002): *Selfsimilar Processes*, Princeton and Oxford: Princeton University Press. <span style="font-variant:small-caps;">Fasen, V.</span> (2005): “Extremes of regularly varying Lévy driven mixed moving average processes,” *Advances in Applied Probability*, 37, 993–1014. <span style="font-variant:small-caps;">Fisher, R. A. and L. Tippett</span> (1928): “Limiting forms of the frequency distributions of the largest or smallest member of a sample,” *Proceedings of Cambridge Philosophical Society*, 24, 180–190. <span style="font-variant:small-caps;">Harris, T. and H. Robbins</span> (1953): “Ergodic theory of Markov chains admitting an infinite invariant measure,” *Proc. Natl. Acad. Sci.*, 39, 860–864. <span style="font-variant:small-caps;">Kabluchko, Z., M. Schlather, and L. de Haan</span> (2009): “Stationary max-stable fields associated to negative definite functions,” *Annals of Probability*, 37, 2042–2065. <span style="font-variant:small-caps;">Kabluchko, Z. and S. Stoev</span> (2012): “Minimal spectral representations of infinitely divisible and max-infinitely divisible processes,” Preprint. Available at: [ http://arxiv.org/pdf/1207.4983v1.pdf]{}. <span style="font-variant:small-caps;">Krengel, U.</span> (1985): *Ergodic Theorems*, Berlin, New York: De Gruyter. <span style="font-variant:small-caps;">Kwapień, S. and N. Woyczyński</span> (1992): *Random Series and Stochastic Integrals: Single and Multiple*, Boston: Birkh[ä]{}user. <span style="font-variant:small-caps;">Lamperti, J.</span> (1962): “Semi-stable stochastic processes,” *Transaction of the American Mathematical Society*, 104, 62–78. ——— (1964): “On extreme order statistics,” *The Annals of Mathematical Statistics*, 35, 1726–1737. <span style="font-variant:small-caps;">Leadbetter, M.</span> (1983): “Extremes and local dependence of stationary sequences,” *Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete*, 65,, 291–306. <span style="font-variant:small-caps;">Leadbetter, M., G. Lindgren, and H. Rootz[é]{}n</span> (1983): *Extremes and Related Properties of Random Sequences and Processes*, New York: Springer Verlag. <span style="font-variant:small-caps;">Mikosch, T. and G. Samorodnitsky</span> (2000): “The supremum of a negative drift random walk with dependent heavy–tailed steps,” *Annals of Applied Probability*, 10, 1025–1064. <span style="font-variant:small-caps;">Mikosch, T. and C. Stărică</span> (2000): “Limit theory for the sample autocorrelations and extremes of a GARCH(1,1) process,” *The Annals of Statistics*, 28, 1427–1451. <span style="font-variant:small-caps;">O[’]{}Brien, G., P. Torfs, and W. Vervaat</span> (1990): “Stationary self-similar extremal processes,” *Probability Theory and Related Fields*, 87, 97–119. <span style="font-variant:small-caps;">Owada, T.</span> (2013): “Limit theory for the sample autocovariance for heavy tailed stationary infinitely divisible processes generated by conservative flows,” Technical Report. Available at: [ http://arxiv.org/pdf/1302.0058v1.pdf]{}. <span style="font-variant:small-caps;">Owada, T. and G. Samorodnitsky</span> (2012): “Functional Central Limit Theorem for heavy tailed stationary infinitely divisible processes generated by conservative flows,” Technical Report. <span style="font-variant:small-caps;">Resnick, S.</span> (1987): *Extreme Values, Regular Variation and Point Processes*, New York: Springer-Verlag. <span style="font-variant:small-caps;">Resnick, S. and M. Rubinovitch</span> (1973): “The structure of extremal processes,” *Advances in Applied Probability*, 5, 287–307. <span style="font-variant:small-caps;">Resnick, S. and G. Samorodnitsky</span> (2004): “Point processes associated with stationary stable processes,” *Stochastic Processes and Their Applications*, 114, 191–210. <span style="font-variant:small-caps;">Rootzén, H.</span> (1978): “Extremes of moving averages of stable processes,” *The Annals of Probability*, 6, 847–869. <span style="font-variant:small-caps;">Rosiński, J.</span> (1995): “On the structure of stationary stable processes,” *The Annals of Probability*, 23, 1163–1187. ——— (2006): “Minimal integral representations of stable processes,” *Probability and Mathematical Statistics*, 26, 121–142. <span style="font-variant:small-caps;">Roy, E.</span> (2008): “Ergodic properties of Poissonian ID processes,” *The Annals of Probability*, 35, 551–576. <span style="font-variant:small-caps;">Samorodnitsky, G.</span> (2004): “Extreme value theory, ergodic theory, and the boundary between short memory and long memory for stationary stable processes,” *Annals of Probability*, 32, 1438–1468. ——— (2005): “Null flows, positive flows and the structure of stationary symmetric stable processes,” *Annals of Probability*, 33, 1782–1803. ——— (2006): *Long Range Dependence*, vol. 1:3 of *Foundations and Trends in Stochastic Systems*, Boston: Now Publishers. <span style="font-variant:small-caps;">Samorodnitsky, G. and M. Taqqu</span> (1994): *[S]{}table [N]{}on-[G]{}aussian [R]{}andom [P]{}rocesses*, New York: Chapman and Hall. <span style="font-variant:small-caps;">Skorohod, A.</span> (1956): “Limit theorems for stochastic processes,” *Theor. Probab. Appl.*, 1, 261–290. <span style="font-variant:small-caps;">Stoev, S.</span> (2008): “On the ergodicity and mixing of max-stable processes,” *Stochastic Processes and their Applications*, 118, 1679–1705. <span style="font-variant:small-caps;">Stoev, S. and M. Taqqu</span> (2005): “Extremal stochastic integrals: a parallel between max-stable processes and $\alpha$-stable processes,” *Extremes*, 8, 237–266. <span style="font-variant:small-caps;">Surgailis, D., J. Rosiński, V. Mandrekar, and S. Cambanis</span> (1993): “Stable mixed moving averages,” *Probab. Theory Related Fields*, 97, 543–558. <span style="font-variant:small-caps;">Wang, Y. and S. Stoev</span> (2010): “On the structure and representations of max-stable processes,” *Advances in Applied Probability*, 42, 855–877. <span style="font-variant:small-caps;">Whitt, W.</span> (2002): *Stochastic-Process Limits. An Introduction to Stochastic-Process Limits and Their Applications to Queues*, New York: Springer. <span style="font-variant:small-caps;">Zweimüller, R.</span> (2009): “[*Surrey [N]{}otes on [I]{}nfinite [E]{}rgodic [T]{}heory*]{},” Lecture notes, Surrey University. [^1]: This research was partially supported by the ARO grants W911NF-07-1-0078 and W911NF-12-10385, NSF grant DMS-1005903 and NSA grant H98230-11-1-0154 at Cornell University.

ArXiv

--- abstract: 'Finding the clique of maximum cardinality in an arbitrary graph is an NP-Hard problem that has many applications, which has motivated studies to solve it exactly despite its difficulty. The great majority of algorithms proposed in the literature are based on the Branch and Bound method. In this paper, we propose an exact algorithm for the maximum clique problem based on the Russian Dolls Search method. When compared to Branch and Bound, the main difference of the Russian Dolls method is that the nodes of its search tree correspond to decision subproblems, instead of the optimization subproblems of the Branch and Bound method. In comparison to a first implementation of this Russian Dolls method from the literature, several improvements are presented. Some of them are adaptations of techniques already employed successfully in Branch and Bound algorithms, like the use of approximate coloring for pruning purposes and bit-parallel operations. Two different coloring heuristics are tested: the standard greedy and the greedy with recoloring. Other improvements are directly related to the Russian Dolls scheme: the adoption of recursive calls where each subproblem (doll) is solved itself via the same principles than the Russian Dolls Search and the application of an elimination rule allowing not to generate a significant number of dolls. Results of computational experiments show that the algorithm outperforms the best exact combinatorial algorithms in the literature for the great majority of the dense graphs tested, being more than twice faster in several cases.' author: - | \ [Universidade Federal Rural do Rio de Janeiro,]{}\ [Departamento de Ciência da Computação,]{}\ [Av. Governador Roberto Silveira S/N,]{}\ [26020-740 Nova Iguaçu - RJ, Brazil]{} - | \ [Université d’Avignon et des Pays du Vaucluse,]{}\ [Laboratoire d’Informatique d’Avignon,]{}\ [F-84911 Avignon, Cedex 9, France]{} - | ,\ [Université de Versailles Saint Quentin,]{}\ [45 Avenue des Etats Unis,]{}\ [78035 Versailles, France]{} - | \ [Sciences Institute,]{}\ [National University of General Sarmiento,]{}\ [J. M. Gutiérrez 1150, Malvinas Argentinas,]{}\ [(1613) Buenos Aires, Argentina]{} bibliography: - './clqdolls.bib' title: 'A Bit-Parallel Russian Dolls Search for a Maximum Cardinality Clique in a Graph[^1]' --- [^1]: This work has been partially supported by the Stic/AmSud joint program by CAPES (Brazil), CNRS and MAE (France), CONICYT (Chile) and MINCYT (Argentina) – project STABLE – and the Pronem program by FUNCAP/CNPq (Brazil) – project ParGO. The first author was with Universidade Federal do Ceará, Departamento de Computação, Brazil, when this work was done. The second author has been partially granted by the “Pesquisador Visitante Especial” (CNPq program) process 313831/2013-0.

ArXiv

--- author: - Arthur Genthon - David Lacoste bibliography: - 'biblio.bib' title: Fluctuation relations and fitness landscapes of growing cell populations --- Introduction ============ While the growth of cell populations appears deterministic, many processes occurring at the single cell level are stochastic. Among many possibilities, stochasticity at the single cell level can arise from stochasticity in the generation times [@sandler_lineage_2015], from stochasticity in the partition at division [@hosoda_origin_2011; @thomas_making_2017], or from the stochasticity of single cell growth rates, which are usually linked to stochastic gene expression [@elowitz_stochastic_2002]. Ideally one would like be able to disentangle the various sources of stochasticity present in experimental data [@barizien_growing_2019]. This would allow to understand and predict how the various sources of stochasticity affect macroscopic parameters of the cell population, such as the Malthusian population growth rate [@powell_growth_1956; @olivier_how_2017]. Beyond this specific question, research in this field attempts to elucidate the fundamental physical constraints which control growth and divisions in cell populations. With the advances in single cell experiments, where the growth and divisions of thousand of individual cells can be tracked, robust statistics can be acquired. New theoretical methods are needed to exploit this kind of data and to relate experiments carried out at the population level with experiments carried out at the single cell level. For instance, one would like to relate single-cell time-lapse videomicroscopy experiments of growing cell populations [@kiviet_stochasticity_2014], which provide information on all the lineages in the branched tree, with experiments carried out with the mother machine configuration, which provide information on single lineages [@taheri-araghi_cell-size_2015; @wang_robust_2010]. Here, we develop a theoretical framework to relate observables measured at the single cell level and at the population level, building on a number of theoretical works [@jafarpour_bridging_2018; @lin_effects_2017; @thomas_making_2017; @thomas_single-cell_2017] and on our own previous work on this topic [@garcia-garcia_linking_2019]. Following Nozoe et al. [@nozoe_inferring_2017], we introduce two different ways to sample lineages, namely the forward (chronological) and backward (retrospective) samplings. We show that the statistical bias present in the backward sampling with respect to the forward sampling is captured by a general relation called fluctuation relation [@thomas_making_2017; @hoffmann_nonparametric_2016; @baake_mutation_2007]. This fluctuation relation depends only on the structure of the branched tree but not on the class of dynamical models defined on it. This relation has analogies with the fluctuation relations well known in Stochastic Thermodynamics [@seifert_stochastic_2012], as first noted in [@sughiyama_pathwise_2015; @kobayashi_fluctuation_2015], which we further discuss here. Three classical models of cell size control have been proposed in the literature: the ‘sizer’ in which the division is controlled by the size of the cell, the ‘timer’ in which the division depends on the age of cell, and the ‘adder’ for which the cell divides after adding a constant volume to its birth volume [@jun_fundamental_2018; @amir_cell_2014; @osella_concerted_2014; @tzur_cell_2009]. We study size models, age models and also the general case of mixed models, in which the division rate is controlled by both the size and the age of the cell. Mixed models include the three policies mentioned above, for instance the added volume appearing in the adder policy can be expressed as a function of the size and the age of the cell. We develop a framework based on fluctuation relations for these models and we explore some consequences. In section \[app\_path\_op\], we introduce a framework based either on path integrals or on operators, which characterize the symmetry contained in fluctuation relations for these models. For these specific models and for key phenotypic variables such as the size and the age, we study in section \[sec:fitness\] a function called fitness landscape [@nozoe_inferring_2017], which informs whether a specific phenotypic variable affects the division rate of the cell population. The backward and forward processes {#sec_back_for} ---------------------------------- Let us consider a branched tree, starting with $N_0$ cells at time $t=0$ and ending with $N(t)$ cells at time $t$ as shown on \[fig\_back\_for\]. We assume that all lineages survive up to time $t$, and therefore the final number $N(t)$ of cells correspond to the number of lineages in the tree. The most natural way to sample the lineages is to put uniform weight on all of them. This sampling is called the backward, (or retrospective) because at the end of the experiment one randomly chooses one lineage among the $N(t)$ with a uniform probability and then one traces the history of the lineage backward in time from time $t$ to $0$, until reaching the ancestor population. The backward weight associated with a lineage $l$ is defined as $$\label{eq_back_wei} \omega_{{\text{back}}}(l)=N(t)^{-1} \,.$$ In a tree, some lineages divide more often than others, which results in an over-representation of lineages that have divided more often than the average. Therefore by choosing a lineage with uniform distribution, we are more likely to choose a lineage with more divisions than the average number of divisions in the tree. The other way of sampling a tree is the forward (or chronological) one and consists in putting the weight $$\label{eq_for_wei} \omega_{\text{for}}(l)= N_0^{-1} m^{-K(l)} \,,$$ on a lineage $l$ with $K(l)$ divisions, where $m$ is the number of offsprings at division. This choice of weights is called forward because one starts at time $0$ by uniformly choosing one cell among the $N_0$ initial cells, and one goes forward in time up to time $t$, by choosing one of the $m$ offsprings with equal weight $1/m$ at each division. ![Example of a tree with $N_0=1$ and $N(t)=10$ lineages at time $t$. Two lineages are highlighted, the first in blue with $2$ divisions and the second in orange with $5$ divisions. The forward sampling is represented with the green right arrows: it starts at time $t=0$ and goes forward in time by choosing one of the two daughters lineages at each division with probability $1/2$. The backward sampling is pictured by the left purple arrows: starting from time $t$ with uniform weight on the $10$ lineages it goes backward in time down to time $t=0$.[]{data-label="fig_back_for"}](for_back_article_test.pdf){width="\linewidth"} The backward and forward weights are properly normalized probabilities, defined on the $N(t)$ lineages in the tree at time $t$: $\sum_{i=1}^{N(t)} \omega_{{\text{back}}}(l_i) = \sum_{i=1}^{N(t)} \omega_{{\text{for}}}(l_i) =1$. Single lineage experiments are precisely described by a forward process since experimentally, at each division, only one of the two daughter cells is conserved while the other is eliminated (for instance flushed away in a microfluidic channel [@taheri-araghi_cell-size_2015; @wang_robust_2010]). In these experiments, a tree is generated but at each division only one of the two lineages is conserved, with probability $1/2$, while the rest of the tree is eliminated. This means that single lineage observables can be measured without single lineage experiments, provided population experiments are analyzed with the correct weights on lineages. Link with the population growth rate ------------------------------------ Since the backward weight put on a lineage depends on the number of cells at time $t$, it takes into account the reproductive performance of the colony but it is unaffected by the reproductive performance of the lineage considered. On the contrary, the forward weight put on a specific lineage depends on the number of divisions of that lineage but is unaffected by the reproductive performance of other lineages in the tree. Therefore, the difference between the values of the two weights for a particular lineage informs on the difference between the reproductive performance of the lineage with respect to the colony. We now introduce the population growth rate: $$\label{eq_def_lambda} \Lambda_t=\frac{1}{t} \ln \frac{N(t)}{N_0} \,,$$ which is linked to forward weights by the relation $$\label{eq_w_for_N} \frac{N(t)}{N_0}=\sum_{i=1}^{N(t)} m^{K_i} \ \omega_{\text{for}}(l_i) = \langle m^K \rangle_{\text{for}} \,,$$ where $\langle \cdot \rangle_{\text{for}}$ is the average over the lineages weighted by $\omega_{\text{for}}$, and $K_i=K(l_i)$. Combining the two equations above, we obtain [@levien_large_2020]: $$\label{eq_w_for_lambda} \Lambda_t=\frac{1}{t} \ln \langle m^K \rangle_{\text{for}} \,,$$ which allows an experimental estimation of the population growth rate from the knowledge of the forward statistics only. can also be re-written to express the bias between the forward and backward weights of the same lineage $$\label{eq_fr_wei_1} \frac{\omega_{{\text{back}}}(l)}{\omega_{\text{for}}(l)}=\frac{m^{K(l)}}{\langle m^K \rangle_{\text{for}}} \,,$$ which is the reproductive performance of the lineage divided by its average in the colony with respect to $\omega_{\text{for}}$. A similar relation is derived using the relation $$\label{eq_w_back_N} \frac{N_0}{N(t)}=\sum_{i=1}^{N(t)} m^{-K_i} \ \omega_{\text{back}}(l_i) = \langle m^{-K} \rangle_{\text{back}} \,.$$ Combining \[eq\_w\_for\_lambda,eq\_w\_back\_N\] we obtain: $$\label{eq_w_for_lambda_2} \Lambda_t= - \frac{1}{t} \ln \langle m^{-K} \rangle_{\text{back}} \,.$$ As for \[eq\_w\_for\_N\], \[eq\_w\_back\_N\] is reshaped to show the bias between the forward and backward weights of the same lineage: $$\label{eq_fr_wei_2} \frac{\omega_{{\text{back}}}(l)}{\omega_{\text{for}}(l)}=\frac{\langle m^{-K} \rangle_{\text{back}}}{m^{-K(l)}} \,.$$ Combining \[eq\_for\_wei,eq\_back\_wei,eq\_def\_lambda\], we obtain the fluctuation relation [@garcia-garcia_linking_2019; @nozoe_inferring_2017]: $$\label{eq_fr_wei} \omega_{{\text{back}}}(l)= \omega_{\text{for}}(l) \ e^{K(l) \ln m - t \Lambda_t} \,.$$ If we now introduce the probability distribution of the number of divisions for the forward sampling $p_{\text{for}}(K)=\sum_l \delta(K-K(l)) \omega_{\text{for}}(l)$ and similarly for the backward sampling, we can also recast the above relation as a fluctuation relation for the distribution of the number of divisions: $$\label{eq_fr_K} p_{{\text{back}}} (K,t)=p_{{\text{for}}} (K,t) \ e^{K \ln m - t \Lambda_t} \,.$$ Let us now introduce the Kullback-Leibler divergence between two probability distributions $p$ and $q$, which is the non-negative number: $$\mathcal{D}_{{\text{KL}}}(p||q)=\int {\mathrm{d}}x \, p(x) \ln \frac{p(x)}{q(x)} \geq 0 \,.$$ Using \[eq\_fr\_wei\], we obtain $$\mathcal{D}_{{\text{KL}}}(\omega_{{\text{back}}}|| \omega_{\text{for}}) = \langle K \rangle_{{\text{back}}} \ln m - t \Lambda_t \geq 0 \,.$$ A similar inequality follows by considering $\mathcal{D}_{{\text{KL}}}(\omega_{{\text{for}}}|| \omega_{\text{back}})$. Finally we obtain $$\label{eq_ineq_K} \frac{t}{\langle K \rangle_{{\text{back}}}} \leq \frac{\ln m}{\Lambda_t} \leq \frac{t}{\langle K \rangle_{\text{for}}} \,.$$ In the long time limit, $ \lim\limits_{t \rightarrow + \infty} t/\langle K \rangle_{{\text{back}}} = \langle \tau \rangle_{{\text{back}}}$, where $\tau$ is the inter-division time, or generation time, defined as the time between two consecutive divisions on a lineage. The same argument goes for the forward average. In the case of cell division where each cell only gives birth to two daughter cells ($m=2$), the center term in the inequality tends to the population doubling time $T_d$. Therefore, this inequality reads in the long time limit: $$\label{eq_ineq_tau} \langle \tau \rangle_{{\text{back}}} \leq T_d \leq \langle \tau \rangle_{\text{for}} \,.$$ Let us now mention a minor but subtle point related to this long time limit. For a lineage with $K$ divisions up to time $t$, we can write $t=a + \sum_{i=1}^{K} \tau_i$, where $a$ is the age of the cell at time $t$ and where $\tau_i$ is the generation time associated with the $i^{\rm{th}}$ division. Then $t/ K= \tau_m + a/K$, where $\tau_m$ is the mean generation time along the lineage. For finite times, all we can deduce is $t/ K \geq \tau_m $. Therefore the left inequality of \[eq\_ineq\_tau\] always holds $$\langle \tau \rangle_{{\text{back}}} \leq \frac{t}{\langle K \rangle_{{\text{back}}}} \leq \frac{\ln m}{\Lambda_t} \,,$$ while the right inequality does not necessarily hold at finite time. The inequalities of \[eq\_ineq\_tau\] have been theoretically derived in [@hashimoto_noise-driven_2016] for age models and these authors have also experimentally verified them using experimental data. In our previous work [@garcia-garcia_linking_2019], we have replotted their experimental data for clarity and we have shown theoretically that the same inequalities should also hold for size models. In fact, as the present derivation shows, the relation \[eq\_ineq\_K\] is very general and only depends on the branching structure of the tree, while the relation \[eq\_ineq\_tau\] requires in addition the existence of a steady state. These inequalities and \[eq\_fr\_K\] express fundamental constraints between division and growth, which should hold for any model of this type (size model, age model or mixed size-age model), irrespective of the precise form of the division rate or on the partition at cell division. Stochastic thermodynamic interpretation --------------------------------------- The results derived above have a structure similar to that found for non-equilibrium systems in Stochastic Thermodynamics [@seifert_stochastic_2012]. Indeed, \[eq\_w\_for\_lambda\] is the analog of the Jarzynski relation while \[eq\_fr\_K\] is similar to the Crooks fluctuation relation, with the number of divisions $K$ the analog of the work, and the population growth rate the analog of the free energy. Given that the Jarzynski or Crooks fluctuation relations have been used to infer free energies from non-equilibrium measurements, we could similarly use \[eq\_w\_for\_lambda\] or \[eq\_fr\_K\] to determine the population growth rate from the statistics of the number of divisions within the appropriate sampling of lineages. Relations of this type show the central role played by the population growth rate in these models. In this context, the inequalities \[eq\_ineq\_K\] are expressing a constraint equivalent to the second law of thermodynamics, which classically follows from the Jarzynski or Crooks fluctuation relations. Fluctuation relations also take a slightly different form when expressed at finite time or at steady state, which is indeed the case here when comparing \[eq\_ineq\_K\] with \[eq\_ineq\_tau\]. A difference between the two sets of results is that the work fluctuation relations involve two specific dynamics which are related by time-reversal symmetry, whereas no dynamics is needed here to derive \[eq\_w\_for\_lambda\] or \[eq\_fr\_K\]. We now precisely introduce dynamical variables on this tree, which will provide us with the equivalent of the non-equilibrium trajectories of Stochastic Thermodynamics. Let us introduce $M$ variables labeled $(y_1,y_2, \ldots ,y_M)$ to describe a dynamical state of the system, then a path is fully determined by the values of these variables at division, and the times of each division. We call $\boldsymbol{y}(t)=(y_1(t),y_2(t), \ldots ,y_M(t))$ a vector state at time $t$ and $\{\boldsymbol{\boldsymbol{y}}\}=\{\boldsymbol{y}(t_j)\}_{j=1}^{K}$ a path with $K$ divisions. The probability $\mathcal{P}$ of path $\{\boldsymbol{y}\}$ is defined as the sum over all lineages of the weights of the lineages that follow the path $\{\boldsymbol{y}\}$: $$\label{eq_def_path_pb} \mathcal{P}(\{\boldsymbol{y}\},K,t)=\sum_{i=1}^{N(t)} \omega(l_i) \, \delta (K-K_i) \delta (\{\boldsymbol{y}\} - \{\boldsymbol{y}\}_i) \,,$$ where $\{\boldsymbol{y}\}_i$ is the path followed by lineage $l_i$. Using the normalization of the weights $\omega$ on the lineages, we show that $\mathcal{P}$ is properly normalized: $\int {\mathrm{d}}\{\boldsymbol{y}\} \sum_K \mathcal{P}(\{\boldsymbol{y}\},K,t) = 1$. We then define the number $n(\{\boldsymbol{y}\},K,t)$ of lineages in the tree at time $t$ that follow the path $\{\boldsymbol{y}\}$ with $K$ divisions: $$\label{eq_def_n} n(\{\boldsymbol{y}\},K,t)=\sum_{i=1}^{N(t)} \delta (K-K_i) \delta (\{\boldsymbol{y}\} - \{\boldsymbol{y}\}_i) \,.$$ This number of lineages is normalized as $\int {\mathrm{d}}\{\boldsymbol{y}\} \sum_K n(\{\boldsymbol{y}\},K,t) = N(t)$. Then, the path probability can be re-written as $$\label{eq_def_path_pb-v2} \mathcal{P}(\{\boldsymbol{y}\},K,t) = n(\{\boldsymbol{y}\},K,t) \cdot \omega(l) \,.$$ Since $n(\{\boldsymbol{y}\},K,t)$ is independent of a particular choice of lineage weighting, we obtain $$\label{eq_fr_path} \frac{\mathcal{P}_{{\text{back}}}(\{\boldsymbol{y}\},K,t)}{\mathcal{P}_{\text{for}}(\{\boldsymbol{y}\},K,t)}=\frac{\omega_{{\text{back}}}(l)}{\omega_{\text{for}}(l)}= \ e^{K \ln m - t \Lambda_t} \, ,$$ which generalizes \[eq\_fr\_K\]. In our previous work [@garcia-garcia_linking_2019], we have derived this relation for size models with individual growth rate fluctuations (i.e. $\boldsymbol{y}=(x,\nu)$) but we were not aware of the weighting method introduced by [@nozoe_inferring_2017], and for this reason, we used the term ‘tree’ to denote the backward sampling, and the term ‘lineage’ to denote the forward sampling. Mixed age-size controlled models ================================ Dynamics at the population level -------------------------------- The state of a cell is described by its size $x$, its age $a$ and its individual growth rate $\nu$, with $\boldsymbol{y}=(x,a,\nu)$. The evolution of the number of cells $n(\boldsymbol{y},K,t)$ in the state $\boldsymbol{y}$ at time $t$, that belong to a lineage with $K$ divisions up to time $t$ is governed by the equation $$\begin{gathered} \label{eq_mix_mod} {\left(}\partial_t + \partial_a {\right)}n(\boldsymbol{y},K,t) +\partial_x {\left[}\nu x n(\boldsymbol{y},K,t) {\right]}\\ + B(\boldsymbol{y})n(\boldsymbol{y},K,t) =0 \,,\end{gathered}$$ and the boundary condition $$\begin{gathered} \label{eq_mix_model_bc} n(x,a=0,\nu,K,t)= \\ m \int {\mathrm{d}}\boldsymbol{y}' B(\boldsymbol{y}') \Sigma(\boldsymbol{y}|\boldsymbol{y}') n(\boldsymbol{y}',K-1,t) \,,\end{gathered}$$ where $B(\boldsymbol{y})$ is the division rate and $\Sigma(\boldsymbol{y}|\boldsymbol{y}')$ is the conditional probability (also called division kernel) for a newborn cell to be in state $\boldsymbol{y}$ knowing its mother divided while in state $\boldsymbol{y}'$, normalized as $\int \Sigma(\boldsymbol{y}|\boldsymbol{y}') {\mathrm{d}}\boldsymbol{y} =1$, for any $\boldsymbol{y}'$. Dynamics at the probability level {#sec_edp_proba} --------------------------------- While $n(\boldsymbol{y},K,t)$ in \[eq\_mix\_mod\] is independent of the choice of weights put on the lineages, we now turn to a description in terms of the probability $p(\boldsymbol{y},K,t)$ for a cell to be in state $(\boldsymbol{y},K)$ at time $t$ if chosen randomly among the $N(t)$ cells in the tree at that time. To do so, one has to choose how to weight each cell in the colony, which is equivalent to weight each lineage, since at time $t$ each cell is the ending point of one lineage. The first possibility is the backward sampling, for which each lineage is weighted uniformly. In this case, we define $p_{{\text{back}}}$ as $$p_{{\text{back}}}(\boldsymbol{y},K,t)=\frac{n(\boldsymbol{y},K,t)}{N(t)} \,.$$ Dividing \[eq\_mix\_mod\] and the boundary condition \[eq\_mix\_model\_bc\] by $N(t)$ we obtain $$\begin{gathered} \label{eq_mix_mod_p} {\left(}\partial_t + \partial_a {\right)}p_{{\text{back}}}(\boldsymbol{y},K,t) +\partial_x {\left[}\nu x p_{{\text{back}}}(\boldsymbol{y},K,t) {\right]}\\ + {\left[}B(\boldsymbol{y}) + \Lambda_p(t) {\right]}p_{{\text{back}}}(\boldsymbol{y},K,t) =0 \,, \end{gathered}$$ and $$\begin{gathered} \label{eq_mix_mod_p_bc} p_{{\text{back}}}(x,a=0,\nu,K,t)= \\ m \int {\mathrm{d}}\boldsymbol{y}' B(\boldsymbol{y}') \Sigma(\boldsymbol{y}|\boldsymbol{y}') p_{{\text{back}}}(\boldsymbol{y}',K-1,t) \,, \end{gathered}$$ where we defined the instantaneous population growth rate as $$\label{eq_def_lambda_inst} \Lambda_p(t)=\frac{\dot{N}}{N} \,.$$ The instantaneous population growth rate and the population growth rate defined in \[eq\_def\_lambda\] are related by: $$\Lambda_t=\frac{1}{t} \int_{0}^{t} \Lambda_p(t') {\mathrm{d}}t' \,.$$ In the long-time limit, $N$ grows exponentially with constant rate $\Lambda_p$, and thus $\Lambda_t=\Lambda_p=\Lambda$. The other possibility is to use the forward statistics, in which case we define the probability $p_{{\text{for}}}$, as $$p_{{\text{for}}}(\boldsymbol{y},K,t)=\frac{n(\boldsymbol{y},K,t)}{m^K} \,.$$ Dividing \[eq\_mix\_mod\] and the boundary condition \[eq\_mix\_model\_bc\] by $m^K$ we obtain $$\begin{gathered} \label{eq_mix_mod_p_sc} {\left(}\partial_t + \partial_a {\right)}p_{{\text{for}}}(\boldsymbol{y},K,t) +\partial_x {\left[}\nu x p_{{\text{for}}}(\boldsymbol{y},K,t) {\right]}\\ + B(\boldsymbol{y}) p_{{\text{for}}}(\boldsymbol{y},K,t) =0 \,, \end{gathered}$$ and $$\begin{gathered} \label{eq_mix_mod_p_sc_bc} p_{{\text{for}}}(x,a=0,\nu,K,t)= \\ \int {\mathrm{d}}\boldsymbol{y}' B(\boldsymbol{y}') \Sigma(\boldsymbol{y}|\boldsymbol{y}') p_{{\text{for}}}(\boldsymbol{y}',K-1,t) \,. \end{gathered}$$ One can notice that the backward statistics is well suited to study the population, while the forward statistics reproduce the behaviour of single lineage experiments. Indeed, by taking \[eq\_mix\_mod\_p\] for the population/backward probability $p_{{\text{back}}}$, and choosing $\Lambda_p(t)=0$ and $m=1$ we recover \[eq\_mix\_mod\_p\_sc\]. This equation is then a population equation in which we follow only one cell, so that $\Lambda_p(t)=0$ and $m=1$, which we call single lineage experiment. Volume conservation ------------------- Let us explain why the conservation of volume between the mother cell and the two daughter cells at division is compatible with \[eq\_mix\_mod,eq\_mix\_model\_bc\]. To do so, let us impose the volume conservation and show that the resulting dynamical equations reduce to \[eq\_mix\_mod,eq\_mix\_model\_bc\]. For simplicity we consider the particular case of a size-control mechanism, described in \[app\_size\_mod\], with constant individual growth rate, which means that $\boldsymbol{y}=x$. Our derivation would easily extend to mixed models since age does not play any role in the conservation of volume. One can consider that when a cell of size $x_0$ divides, the volume of one of the two daughter cells is randomly distributed according to the kernel $\Sigma(x_1 | x_0)$, and the volume of the second daughter is imposed by the conservation of volume: $x_2=x_0-x_1$. Thus, considering the evolution of $n(x_1,K,t)$, the source term reads $$\begin{gathered} S = \int {\mathrm{d}}x_0 \Sigma (x_1 | x_0) B(x_0) n(x_0,t) \\ + \int {\mathrm{d}}x_0 {\mathrm{d}}x_2 \Sigma (x_2|x_0) B(x_0) n(x_0,t) \delta(x_0-x_1-x_2) \,,\end{gathered}$$ where the first term accounts for cells of size $x_0$ randomly dividing into cells of size $x_1$, and where the second term describes the divisions of cells of size $x_0$ randomly dividing into cells of size $x_2$, thus giving birth to cells of size $x_0-x_2$ by conservation of the volume. This result can be reshaped as $$S = \int {\mathrm{d}}x_0 B(x_0) n(x_0,t) {\left[}\Sigma (x_1 | x_0) + \Sigma (x_0-x_1|x_0) {\right]}\,.$$ We now define a modified division kernel $\tilde{\Sigma}$, as $\tilde{\Sigma}(x_1|x_0)={\left[}\Sigma (x_1 | x_0) + \Sigma (x_0-x_1|x_0) {\right]}/2$, leading the source term to have the same form as in \[eq\_size\_mod\], but for the modified kernel. Moreover if $\Sigma (x | x')$ is symmetrical around $x=x'/2$, then we have the exact correspondence $\tilde{\Sigma}(x|x')=\Sigma (x | x')$. Test of the fluctuation relation -------------------------------- We simulated the time evolution of colonies of cells, obeying \[eq\_mix\_mod,eq\_mix\_model\_bc\], for age and size models in order to test the fluctuation relation. Since results are very similar -as expected- for age models, we restrict ourselves to size models. We tested two results: the fluctuation relation for the number of divisions \[eq\_fr\_K\] and one of its consequences: the inequality for the mean number of divisions \[eq\_ineq\_K\]. All simulations for size models were conducted with the division rate $B(x,\nu)=\nu x^{\alpha}$, where $\alpha$ is the strength of the control and $x$ is the dimensionless size. Power law were found to be good approximations for empirical division rates $B(x)$ as function of size [@hosoda_origin_2011; @robert_division_2014; @osella_concerted_2014], and for $B(a)$ as function of age [@robert_division_2014]. The factor $\nu$, being the only time scale for size models, give $B(x)$ its proper dimension. For age models, we chose $B(a,\nu)=\nu f(a \nu)$, where $a$ is the age of the cell and with a power law dependence on age: $f(u)=u^{\alpha}$. The constant $\nu^{\alpha+1}$ is redefined as $\nu$ since it does not play a role, so that $B(a,\nu)=\nu a^{\alpha}$. On \[fig\_test\_fr\], the backward and forward probability distributions of the number of divisions are shown for a size model. The two distributions intersect at the number of divisions $K=t \Lambda_t / \ln 2$. The inset of \[fig\_test\_fr\] shows the logarithm of the ratio $q(K,t)=p_{{\text{back}}} (K,t)/p_{{\text{for}}} (K,t)$ of the two distributions, which is as expected a straight line of slope $ - \ln 2$ when plotted against the number of divisions. For convenience and for \[fig\_test\_fr\] only, noise in the volume partition at division has been introduced, by choosing for the conditional probability $\Sigma(x | x')$ a uniform distribution between sizes $x=0$ and $x=x'$. This has the effect of broadening the distributions $P(K)$ with respect to the case of deterministic symmetrical volume partition. ![Distributions of the number of divisions in orange filled histogram for the forward statistics, and with the blue empty histogram for the backward statistics, for a size model with division rate $B(x)=\nu x^{\alpha}$, constant growth rate $\nu=1$, $\alpha=2$ and $t=7$. The vertical dashed line at $K=t \Lambda_t / \ln 2$ is the theoretical value at which the two distributions should intersect. The inset shows the logarithm of the ratio $q(K)$ of the forward to backward probabilities (purple crosses), and the theoretical result $t \Lambda_t-K \ln 2$ (green line).[]{data-label="fig_test_fr"}](distrib_K_inset_size.pdf){width="\linewidth"} Then, we tested the inequality on the mean numbers of divisions by varying the strength of the size-control $\alpha$. Results are shown on \[fig\_test\_fr\_ineq\]. One one hand, we see that the less control on size, the more discrepancy between the two determinations $\langle K \rangle_{{\text{back}}}$ and $\langle K \rangle_{{\text{for}}}$. On the other hand, when increasing the control, the two determinations converge to the population doubling time, where no stochasticity in the number of divisions is left, and every lineage carries the same number of divisions, leading to the equality of the backward and forward statistics. ![Test of the inequality on the mean number of divisions measured with the backward and forward statistics against $\alpha$. The quantity $t/\langle K \rangle$, re-scaled by the population doubling time $T_d=\ln 2 / \Lambda_t$ is shown with orange diamonds (resp. blue circles) for the forward (resp. backward) statistics. The division rate is given by $B(x)=\nu x^{\alpha}$ where $\nu=1$, $\alpha$ is varied from $1$ to $15$ and $t=6$. The volume repartition between the two daughter cells at division is symmetrical, so that $\Sigma(x | x')=\delta(x-x'/2)$.[]{data-label="fig_test_fr_ineq"}](size_ineq_test.pdf){width="\linewidth"} Path integral and operator formalisms {#app_path_op} ===================================== Path integral formalism ----------------------- Using a path integral approach, \[eq\_mix\_mod\] can be solved exactly even for mixed models. A path is defined as the ensemble of values $\{\boldsymbol{y}\}$ taken by $\boldsymbol{y}$ from $0$ to time $t$. Since the evolution of the size between two divisions is deterministic, the path is fully characterized by only a finite number of variables which are: the number of divisions $K$, $\forall k \in [0,K]$, the size $x_k$ at birth and the individual growth rate $\nu_k$ of the cycle, $\forall k \in [0,K-1]$ the age $a_k$ at division, and the final size $x$ and age $a$. Initial conditions are then given by $\boldsymbol{y_0}=(x_0,a=0,\nu_0)$. The number of paths $n(\{\boldsymbol{y}\},K,t)$ is given by $$\begin{gathered} \label{eq_path_density} n(\{\boldsymbol{y}\},K,t)= m^K \exp {\left[}-\int_{0}^{t} {\mathrm{d}}\tau B {\left(}\boldsymbol{y}(\tau) {\right)}{\right]}n_0(x_0,\nu_0) \\ \times \prod_{k=1}^K \mathbb{T} {\left(}x_k,\nu_k \big| x_{k-1}e^{\nu_{k-1} a_{k-1}},\nu_{k-1}, a_{k-1} {\right)}\,.\end{gathered}$$ The proof of this equality and the definition of the transition matrix $\mathbb{T}$ are given in \[app\_mix\_sol\]. Recovering Wakamoto’s relation for age models --------------------------------------------- For age models, the division rate $B(a)$ only depends on the age $a$ of the cell. In this case the theoretical distribution of generation times for the forward statistics is given by $$\label{eq_gen_age} f_{\text{for}}(\tau)=B(\tau) e^{-\int_{0}^{\tau} {\mathrm{d}}t B(t)},$$ which is related to the distribution of generation times for the backward statistics $f_{\text{back}}(\tau)$ by [@hashimoto_noise-driven_2016] : $$\label{eq_waka_age} f_{{\text{back}}}(\tau)=2 f_{\text{for}}(\tau) e^{- \tau \Lambda_t}.$$ This relation can be viewed as a consequence of the general fluctuation relation at the level of path probabilities as explained below. In the case of age models, we show in \[app\_mix\_sol\] that the path probability can be reduced to $$\label{eq_path_age_for} \mathcal{P}_{\text{for}}(\{a_k\},a,K,t) = g_{\text{for}}(a) \prod_{k=0}^{K-1} f_{\text{for}}(a_k) \,,$$ where $g_{\text{for}}(a)$ is the probability for the cell not to divide up to age $a$, between the last division at time $t-a$ and final time $t$; and where $a_k$ is the age of the cell at the $(k+1)^{\rm{th}}$ division. We define the same quantities $f_{{\text{back}}}(\tau)$ and $g_{{\text{back}}}(a)$ at the backward level, obeying $$\label{eq_path_age_back} \mathcal{P}_{{\text{back}}}(\{a_k\},a,K,t) = g_{{\text{back}}}(a) \prod_{k=0}^{K-1} f_{{\text{back}}}(a_k) \,.$$ Using the fluctuation relation at the path level, namely \[eq\_fr\_path\], and making a special choice for $a_k$: $\forall k \in [0,K-1], \ a_k=\tau=(t-a)/K$, we obtain $$\label{eq_waka_age_t} f_{{\text{back}}}(\tau)=m f_{\text{for}}(\tau) e^{- \tau \Lambda_t} \, e^{-a \Lambda/K} {\left[}\frac{g_{\text{for}}(a)}{g_{{\text{back}}}(a)} {\right]}^{1/K} \,.$$ In a steady state, when $t \rightarrow \infty$ and $K \rightarrow \infty$, $e^{-a \Lambda/K} {\left[}g_{\text{for}}(a)/g_{{\text{back}}}(a) {\right]}^{1/K}$ tends to $1$ and \[eq\_waka\_age\_t\] gives back Wakamoto’s relation namely \[eq\_waka\_age\], in the case $m=2$. Operator formalism ------------------ Here, we present an operator-based framework, which provides an alternate route to \[eq\_w\_for\_lambda\], \[eq\_w\_for\_lambda\_2\] and \[eq\_fr\_K\], that completely avoids path integrals. For simplicity, let us illustrate this formalism on the case of the size model, discussed in \[app\_size\_mod\]. The case of the general mixed model, which includes both age and size control can be treated along the same lines. For the backward sampling, we define the generating function $f_{{\text{back}}}(\boldsymbol{y},\lambda,t)$ : $$\label{eq_def_gen_fun} f_{{\text{back}}}(\boldsymbol{y},\lambda,t)=\sum_{K=0}^{\infty} e^{-\lambda K} p_{{\text{back}}}(\boldsymbol{y},K,t) \,,$$ We then multiply \[eq\_size\_mod\_back\] by $e^{-\lambda K}$ and sum over $K$ to obtain $$\label{eq_eq_diff_f} \partial_t f_{{\text{back}}}(\boldsymbol{y},\lambda,t) = \mathcal{L}_{{\text{back}}} (\lambda )\ f_{{\text{back}}}(\boldsymbol{y},\lambda,t) \,.$$ The linear operator $\mathcal{L}_{{\text{back}}} (\lambda )$, acting on $f_{{\text{back}}}(\boldsymbol{y},\lambda,t)$, is defined on a test function $f$ as $$\begin{gathered} \mathcal{L}_{{\text{back}}}( \lambda) f = - \partial_x [\nu x \, f] - (B(\boldsymbol{y}) + \Lambda_p(t)) f \\ + 2e^{-\lambda} \int {\mathrm{d}}\boldsymbol{y}' B(\boldsymbol{y}')\Sigma(\boldsymbol{y}|\boldsymbol{y}') f \,.\end{gathered}$$ Although this operator explicitly depends on the state $\boldsymbol{y}$, we choose not to write this dependency explicitly to ease the reading. By the same method, we obtain the operator at the forward level: $$\begin{gathered} \mathcal{L}_{{\text{for}}}( \lambda) f = - \partial_x [\nu x \, f] - B(\boldsymbol{y}) f \\ + e^{-\lambda} \int {\mathrm{d}}\boldsymbol{y}' B(\boldsymbol{y}')\Sigma(\boldsymbol{y}|\boldsymbol{y}') f \,.\end{gathered}$$ By direct comparison, we obtain the fluctuation relation at the level of operators $$\label{eq_fr_op_size} \mathcal{L}_{{\text{back}}} (\lambda)=\mathcal{L}_{\text{for}} (\lambda - \ln(2)) - \Lambda_p(t) \mathds{1} \,.$$ This equality between two operators implies relations between the eigenvalues and the eigenvectors as well. Let us call $\chi_{{\text{back}}}(\boldsymbol{y},\lambda)$ (resp. $\chi_{{\text{for}}}(\boldsymbol{y},\lambda)$) an eigenvalue of the operator $\mathcal{L}_{{\text{back}}} (\lambda)$ (resp. $\mathcal{L}_{{\text{for}}} (\lambda)$), and $f_{{\text{back}}}^{\chi}(\boldsymbol{y},\lambda)$ (resp. $f_{{\text{for}}}^{\chi}(\boldsymbol{y},\lambda)$) the associated eigenvector. Then the fluctuation relation at the operator level \[eq\_fr\_op\_size\] gives $$\begin{aligned} \label{eq_fr_eva_size} \chi_{{\text{back}}}(\boldsymbol{y},\lambda)&=\chi_{{\text{for}}}(\boldsymbol{y},\lambda-\ln 2) - \Lambda_p(t) \,, \\ \label{eq_fr_eve_size} f_{{\text{back}}}^{\chi}(\boldsymbol{y},\lambda)&=f_{{\text{for}}}^{\chi}(\boldsymbol{y},\lambda- \ln 2) \,.\end{aligned}$$ When solving \[eq\_eq\_diff\_f\], the long-time behavior of $f_{{\text{back}}}(\boldsymbol{y},\lambda,t)$ is controlled by the largest eigenvalue of $\mathcal{L}_{{\text{back}}} (\lambda)$, which we call $\mu_{{\text{back}}}(\boldsymbol{y},\lambda)$, and reads $$\label{eq_gen_fun_long_t} f_{{\text{back}}}(\boldsymbol{y},\lambda,t) \underset{t \rightarrow \infty}{\sim} C_{{\text{back}}}^{\mu} (\boldsymbol{y},\lambda) f_{{\text{back}}}^{\mu}(\boldsymbol{y},\lambda) \, e^{\mu_{{\text{back}}}(\boldsymbol{y},\lambda) t} \,,$$ where $C_{{\text{back}}}^{\mu}(\boldsymbol{y},\lambda)$ is the constant coefficient of the eigenvector $f_{{\text{back}}}^{\mu}(\boldsymbol{y},\lambda)$ associated with the largest eigenvalue, in the decomposition of the initial condition $f_{{\text{back}}}(\boldsymbol{\boldsymbol{y}},\lambda,t=0)$ on the set of the eigenvectors of $\mathcal{L}_{{\text{back}}} (\lambda)$. We investigate the particular case $\lambda=0$. On the one hand, using the definition \[eq\_def\_gen\_fun\] we obtain $$f_{{\text{back}}}(\boldsymbol{y},0,t)=p_{{\text{back}}}(\boldsymbol{y},t) \,,$$ and thus the normalization of the probability gives $$\label{eq_cond1} \int {\mathrm{d}}\boldsymbol{y} f_{{\text{back}}}(\boldsymbol{y},0,t) =1 \,.$$ On the other hand, integrating the long time behaviour \[eq\_gen\_fun\_long\_t\] over $\boldsymbol{y}$, we get $$\begin{gathered} \label{eq_cond2} \int {\mathrm{d}}\boldsymbol{y} f_{{\text{back}}}(\boldsymbol{y},0,t) \\ \underset{t \rightarrow \infty}{\sim} \int {\mathrm{d}}\boldsymbol{y} \, C_{{\text{back}}}^{\mu} (\boldsymbol{y},0) f_{{\text{back}}}^{\mu}(\boldsymbol{y},0) \, e^{\mu_{{\text{back}}}(\boldsymbol{y},0) t} \,.\end{gathered}$$ The only solution to satisfy both conditions \[eq\_cond1,eq\_cond2\] is $$\forall \boldsymbol{y}, \, \mu_{{\text{back}}}(\boldsymbol{y},0)=0 \,,$$ $$\label{eq_ope_int_1} \int {\mathrm{d}}\boldsymbol{y} \, C_{{\text{back}}}^{\mu} (\boldsymbol{y},0) f_{{\text{back}}}^{\mu}(\boldsymbol{y},0)=1 \,.$$ Since the largest backward eigenvalue is independent of the state $\boldsymbol{y}$ for $\lambda=0$, we define the state-independent backward eigenvalue $\hat{\mu}_{{\text{back}}}(\lambda=0)=0$. Reporting $\hat{\mu}_{{\text{back}}}(\lambda=0)$ in \[eq\_fr\_eva\_size\], the right hand side of the equation has to be independent of the state $\boldsymbol{y}$ as well, so we define the state-dependent forward eigenvalue: $\hat{\mu}_{{\text{for}}}(\lambda=-\ln 2)= {\mu}_{{\text{for}}}(\boldsymbol{y},\lambda=-\ln 2)$ for any $\boldsymbol{y}$. Finally, \[eq\_fr\_eva\_size\] gives $$\label{eq_eva_for} \hat{\mu}_{{\text{for}}}(\lambda=-\ln 2)= \Lambda \,,$$ where $\Lambda$ is the steady-state value of the population growth rate $\Lambda_p(t)$. We proved that the steady-state population growth rate $\Lambda$ is the largest eigenvalue of the forward (resp. backward) operator $\mathcal{L}_{{\text{for}}} (\lambda)$ (resp. $\mathcal{L}_{{\text{back}}} (\lambda)$) for the specific choice $\lambda=- \ln 2$ (resp. $\lambda= 0$).\ This choice of $\lambda$ can be understood as follows. Let us consider the steady-state version of equation \[eq\_eq\_diff\_f\] for $\lambda=0$: $$\begin{gathered} \label{eq_def_malthus_ep} - \partial_x [\nu x \, f_{{\text{back}}}(\boldsymbol{y},0)] - B(\boldsymbol{y}) f_{{\text{back}}}(\boldsymbol{y},0) \\ + 2\int {\mathrm{d}}\boldsymbol{y}' B(\boldsymbol{y}')\Sigma(\boldsymbol{y}|\boldsymbol{y}') f_{{\text{back}}}(\boldsymbol{y},0) = \Lambda f_{{\text{back}}}(\boldsymbol{y},0) \,.\end{gathered}$$ This equation, together with the normalization condition of $f_{{\text{back}}}(\boldsymbol{y},0)$ that we already noted in \[eq\_cond1\], form an eigenproblem whose unique eigenvalue is called the Malthus parameter [@olivier_how_2017], which is indeed equal to the steady-state population growth rate. Finally, the operator acting on the left hand side of \[eq\_def\_malthus\_ep\] differs from $\mathcal{L}_{{\text{back}}} (\lambda=0)$, only by a term $- \Lambda \mathds{1}$, which gives back the value $0$ for the eigenvalue of operator $\mathcal{L}_{{\text{back}}} (\lambda=0)$.\ We now use the long-time behaviour and the computed value of the largest forward eigenvalue to propose a second derivation of the link between the population growth rate and the forward statistics for the number of divisions \[eq\_w\_for\_lambda\]. On the one hand, using the definition of the generation function we get $$\label{eq_cond_ope_for_1} \int {\mathrm{d}}\, \boldsymbol{y} f_{{\text{for}}}(\boldsymbol{y},-\ln 2,t) = \sum_K 2^K p_{{\text{for}}}(K,t) \,.$$ On the other hand, $$\begin{gathered} \label{eq_cond_ope_for_2} f_{{\text{for}}}(\boldsymbol{y},- \ln 2,t) \\ \underset{t \rightarrow \infty}{\sim} C_{{\text{for}}}^{\mu} (\boldsymbol{y},- \ln 2) f_{{\text{for}}}^{\mu}(\boldsymbol{y},- \ln 2) \, e^{\hat{\mu}_{{\text{for}}}(- \ln 2) t} \,.\end{gathered}$$ Using \[eq\_fr\_eve\_size,eq\_ope\_int\_1\] and the fact that the initial condition is the same for both backward and forward samplings: $f_{{\text{for}}}(\boldsymbol{y},\lambda,t=0)=f_{{\text{back}}}(\boldsymbol{y},\lambda,t=0)$ for any $\boldsymbol{y}$ and $\lambda$, we prove that $$\label{eq_int_1_for} \int {\mathrm{d}}\boldsymbol{y} \, C_{{\text{for}}}^{\mu} (\boldsymbol{y},- \ln 2) f_{{\text{for}}}^{\mu}(\boldsymbol{y},- \ln 2)=1 \,.$$ Finally, integrating \[eq\_cond\_ope\_for\_2\] over $\boldsymbol{y} $ and combining \[eq\_eva\_for,eq\_cond\_ope\_for\_1,eq\_int\_1\_for\], we recover $$\Lambda=\frac{1}{t} \ln \sum_K 2^K p_{{\text{for}}}(K,t) \,,$$ which is \[eq\_w\_for\_lambda\] for $m=2$. Phenotypic fitness landscapes {#sec:fitness} ============================= The fitness of a phenotypic trait $s$ is a measure of the reproductive success of individuals carrying it. It is usually defined as the number of offsprings of one individual with a given value of the trait and is quite difficult to evaluate. Nozoe et al. suggested that one way to measure it could be to compare the chronological and retrospective marginal probabilities [@nozoe_inferring_2017] and accordingly defined it as: $$\label{eq_def_h} h(s)=\Lambda_t + \frac{1}{t} \ln {\left[}\frac{P_{{\text{back}}}(s)}{P_{{\text{for}}}(s)} {\right]}\,,$$ so that $$P_{{\text{back}}}(s)=P_{{\text{for}}}(s) \exp {\left[}(h(s)-\Lambda_t)t {\right]}\,.$$ This has again the form of a fluctuation relation similar to \[eq\_fr\_K\], except for the replacement of the factor $K \ln 2 /t $ by the function $h(s)$. This suggests that the fitness landscape $h(s)$ plays a role similar to that of an effective division rate, which depends on the trait $s$. In line with this interpretation, in the particular case where $s=K$, \[eq\_fr\_K\] leads to $\tilde{h}(K)=K \ln 2 /t$, where the fitness landscape for trait $K$ is called the lineage fitness and is written $\tilde{h}$. Indeed, in a branched tree, lineages with a large number of divisions $K$ are exponentially over-represented in the population as compared to lineages with small numbers of divisions. This means that lineages with large $K$ have a larger fitness than the ones with a small $K$, which is coherent with $\tilde{h}(K)$ being an increasing function of $K$. Nozoe et al. also introduced a selection pressure, which measures how the population growth rate changes when a trait $s$ changes and which is defined as [@nozoe_inferring_2017] $$\Pi {\left[}s {\right]}= \langle h(s) \rangle_{{\text{back}}} - \langle h(s) \rangle_{{\text{for}}} \,.$$ This may be written as $$\begin{aligned} t \, \Pi {\left[}s {\right]}&= \int ds \left( P_{{\text{back}}}(s) - P_{{\text{for}}}(s) \right) \ln \frac{P_{{\text{back}}}(s)}{P_{{\text{for}}}(s)} {\nonumber}\\ &= \mathcal{D}_{{\text{KL}}}(P_{{\text{back}}} || P_{{\text{for}}} ) + \mathcal{D}_{{\text{KL}}}(P_{{\text{for}}} || P_{{\text{back}}} ) {\nonumber}\\ &= \mathcal{J}_{{\text{KL}}}(P_{{\text{back}}} , P_{{\text{for}}} ) \ge 0 \,,\end{aligned}$$ where in the last step, we introduced the Jeffreys divergence. Note that when $\Pi {\left[}s {\right]}$ is large, the trait $s$ is strongly correlated with the lineage fitness. In the following, we rewrite the definition of $h(s)$ in a slightly different way using $$\begin{aligned} \label{eq_FR_s} P_{{\text{back}}}(s) &=\sum_K P_{{\text{back}}}(s,K) \nonumber\\ &=e^{-t \Lambda_t}\sum_K 2^K P_{{\text{for}}}(s,K) \nonumber\\ &=e^{-t \Lambda_t} P_{{\text{for}}}(s) \sum_K 2^K R_{{\text{for}}}(K|s) \nonumber\\ &= e^{[h(s)-\Lambda_t]t}P_{{\text{for}}}(s) \,,\end{aligned}$$ where we have introduced the probability of the number of division events conditioned on trait $s$ at the forward level, $R_{{\text{for}}}(K|s)$. Lastly, the fitness landscape reads $$\label{eq_def2_h} h(s)=\frac{1}{t}\ln {\left[}\sum_K 2^K R_{{\text{for}}}(K|s) {\right]}\,.$$ An increasing or decreasing fitness landscape means a positive or negative correlation of the trait value with the capacity to divide, whereas a constant fitness landscape means that the trait is not correlated with the number of divisions. Indeed, if we consider a trait $s$ which does not affect the number $K$ of divisions, then $R_{{\text{for}}}(K|s)=\mathcal{P}_{{\text{for}}}(K)$ and \[eq\_def2\_h\] reads $ h(s)=\ln {\left[}\sum_K 2^K \mathcal{P}_{{\text{for}}}(K) {\right]}/t $, which is equal to $\Lambda_t$ according to \[eq\_w\_for\_lambda\]. In that case, we find that the backward and forward probabilities for that trait $s$ are equal. In the next sections, we evaluate the relevance of the key variables from our model, namely the size and the age by evaluating their fitness landscape in size and age models. Size models ----------- We start with a case where the fitness landscape is fully solvable namely a size model with no individual growth rate fluctuations and with symmetric division. These hypotheses may apply to E. Coli, which is known to show small variability in single cell growth rates, and to divide approximately symmetrically. Let us consider a colony starting with one ancestor cell of size $x_0$. Then, the available sizes at time $t$ are discrete and given by $x=x_0 \exp [\nu t] / 2^K$ where $K$ is the number of divisions undergone by the cell. Therefore a particular size $x$ can be reached only if there is an integer $K$ satisfying this relation, and this integer is unique, leading to $$\label{eq_def_R_fit_land} R_{{\text{for}}}(K|x)=\delta {\left(}K - \frac{\ln {\left[}\frac{x_0 e^{\nu t}}{x} {\right]}}{\ln 2} {\right)}\,.$$ Using this relation in \[eq\_def2\_h\], one finds $$\label{eq_h_size} h(x)= \nu + \frac{1}{t} \ln {\left(}\frac{x_0}{x} {\right)}\,.$$ This result was tested numerically and illustrated in \[fig\_hx\]. The fitness landscape of the size is a decreasing function, which implies that cells with a small size are over-represented in the population. This is coherent with the over-representation of cells that divided a lot, since these cells are more likely to be small due to the numerous divisions. ![Size fitness landscape for size models with $B(x)=\nu x^{\alpha}$, constant $\nu=1$, $\alpha=1$, $t=5$, an initial cell of size $x_0=1$ and symmetrical division. The simulated purple dots are positioned at discrete values of $x$ and the green curve is the theoretical prediction. The black horizontal dashed line represents the population growth rate $\Lambda$.[]{data-label="fig_hx"}](hx.pdf){width="\linewidth"} Reporting this result in \[eq\_FR\_s\], we obtain a fluctuation relation for the size $$\label{eq_FR_x} P_{{\text{back}}}(x) = e^{{\left(}\nu - \Lambda_t {\right)}t} \frac{x_0}{x} P_{{\text{for}}}(x) \,,$$ which in the long time limit becomes $$\label{eq_FR_x-ss} P_{{\text{back}}}(x) = \frac{x_0}{x} P_{{\text{for}}}(x) \, ,$$ where we used the property that in a steady state, the population growth rate and the individual growth rate are equal when there is no individual growth rate variability. In some setups, experiments do not start with a unique ancestor cell but with $N_0 > 1$ initial cells, with possibly heterogeneous sizes. We describe this heterogeneity by the average initial size $\langle x_0 \rangle$ and the standard deviation $\sigma_{x_0}$. In this case, accessible sizes are still discrete but depend on both the number of divisions and the initial cell that started the lineage, and are expressed $ x_0^i \exp [\nu t] /2^K$, where $K$ takes integer values from $0$ to $\infty$ and where $x_0^i \in \mathcal{X}_0$, with $\mathcal{X}_0$ the set of initial sizes. Consequently, a final size $x$ can possibly be reached by different couples $(K_i,x_0^i)$. In order to go further, we now introduce explicitly the initial sizes $x_0^i$ in \[eq\_def2\_h\] as $$\begin{aligned} h(x)&=\frac{1}{t}\ln {\left[}\sum_K \sum_{i} 2^K R_{{\text{for}}}(K,x_0^i|x) {\right]}{\nonumber}\\ \label{eq_def2_h_xi} &= \frac{1}{t}\ln {\left[}\sum_K \sum_{i} 2^K R_{{\text{for}}}(K|x,x_0^i) R_{{\text{for}}}(x_0^i|x){\right]}\,.\end{aligned}$$ When conditioning on the initial size $x_0^i$, there is only one possible number of divisions $K$ to reach size $x$, so that $R_{{\text{for}}}(K|x,x_0^i)$ obeys an equation similar to \[eq\_def\_R\_fit\_land\]: $$\label{eq_def_R_fit_land_xi} R_{{\text{for}}}(K|x,x_0^i)=\delta {\left(}K - \frac{\ln {\left[}\frac{x_0^i e^{\nu t}}{x} {\right]}}{\ln 2} {\right)}\,.$$ Let us examine two limit cases: (i) small variability in the initial sizes and (ii) large variability in the initial sizes. [0.49]{} {width="\linewidth"} [0.49]{} {width="\linewidth"} Case (i) is characterized by a small number $N_0$ of initial cells and a small coefficient of variation $\sigma_{x_0}/\langle x_0 \rangle$. In this case, it is realistic to say that a final size $x$ can only be reached by one couple $(K^*,x^*)$, because the sets of accessible sizes generated by each initial cell do not overlap. Therefore, $R_{{\text{for}}}(x_0^i|x)=\delta(x_0^i-x^*)$ and so for any final size $x$, only one initial size $x^*$ survives in the sum, so that \[eq\_def2\_h\_xi\] reads $$\begin{aligned} h(x)&= \nu + \frac{1}{t} \ln {\left(}\frac{x^*}{x} {\right)}{\nonumber}\\ &=\nu + \frac{1}{t} \ln {\left(}\frac{x^*}{x^* \exp [\nu t]/2^K} {\right)}{\nonumber}\\ &=\frac{K \ln 2}{t} {\nonumber}\\ &=\tilde{h}(K) \,.\end{aligned}$$ We learn from this formula that cells that come from lineages with the same number of divisions $K$ have the same fitness landscape value $h(x)$ for the size, regardless of the size $x^*$ of the initial cell of their lineages, and this value is the lineage fitness $\tilde{h}(K)$. Thus, available values for $h(x)$ are quantified by $K$ and form plateaus, where points representing cells coming from different ancestors but with the same number of divisions accumulate, as shown in \[fig\_hx\_N\_ini\_small\].\ Case (ii) is characterized by a large number $N_0$ of initial cells and a large coefficient of variation $\sigma_{x_0}/\langle x_0 \rangle$. Unlike in case (i), the sets of accessible sizes generated by each initial cell have many overlaps, so that a final size $x$ can be reached by many different couples $(K_i,x_0^i)$. We make the hypothesis that a final size $x$ can be reached by any initial cell with uniform probability, so that $R_{{\text{for}}}(x_0^i|x)=1/N_0$. Therefore, \[eq\_def2\_h\_xi\] becomes $$\begin{aligned} h(x)&=\frac{1}{t}\ln {\left[}\frac{1}{N_0} \sum_{i} \frac{x_0^i e^{\nu t}}{x} {\right]}{\nonumber}\\ &=\nu + \frac{1}{t}\ln \frac{\langle x_0 \rangle }{x} \,.\end{aligned}$$ This behavior was tested numerically and the result plotted on \[fig\_hx\_N\_ini\_large\] confirms that the plateaus observed in case (i) are replaced by a smooth curve depending on the mean initial size. We observe the same effect, namely the loss of the plateaus, when introducing fluctuations in individual growth rates. Age Models ---------- ### Constant individual growth rate We consider the case where the individual growth rate is constant and equal to $\nu$. In steady-state, the forward age distribution \[eq\_sol\_age\_for\] reads (see \[app\_age\_mod\] for details) $$\label{eq_sol_pa_for} p_{{\text{for}}}(a)=p_{{\text{for}}}(0) \, \exp {\left[}-\int_{0}^{a} B(a') {\mathrm{d}}a' {\right]}\,.$$ To find the integration constant $p_{{\text{for}}}(0)$, we use the normalization of probability $p_{{\text{for}}}$: $$\label{eq_def_Z} Z=p_{{\text{for}}}(0)^{-1}=\int_{0}^{\infty} {\mathrm{d}}a \exp {\left[}-\int_{0}^{a} B(a') {\mathrm{d}}a' {\right]}\,.$$ Similarly, the steady-state backward distribution of ages \[eq\_sol\_age\_back\] reads $$\label{eq_sol_pa_back} p_{{\text{back}}}(a)=p_{{\text{back}}}(0) \, \exp {\left[}-\Lambda a -\int_{0}^{a} B(a') {\mathrm{d}}a' {\right]}\,.$$ In this case, the integration constant $p_{{\text{back}}}(0)$ can be expressed both using the normalization of $p_{{\text{back}}}(a)$, as done for the forward case, or using \[eq\_age\_mod\_lambda\] which leads to $p_{{\text{back}}}(0)=2 \Lambda$. Therefore, the ratio of the age distributions using the backward and forward statistics reads $$\label{eq_age_fr} \frac{p_{{\text{back}}}(a)}{p_{{\text{for}}}(a)}= 2 Z \Lambda e^{-\Lambda a} \,,$$ where $Z$ is defined in \[eq\_def\_Z\] and only depends on the division rate $B(a)$. This relation has a similar form as the relation derived by Hashimoto et al. [@hashimoto_noise-driven_2016] for the distributions of generation times \[eq\_waka\_age\], except for the extra age-independent factor $Z \Lambda$. Finally, the fitness landscape reads $$\label{eq_age_fit_land} h(a)=\frac{1}{t} {\left[}\Lambda (t-a) + \ln (2 Z \Lambda) {\right]}\,.$$ For the same reason as for $h(x)$ in size models, $h(a)$ in age models is a decreasing function of $a$ because lineages that divided a lot are over-represented in the population and are therefore more likely to contain young cells at time $t$. The initial condition does not play any role in this derivation, therefore, unlike size models, the results obtained are unchanged for any number $N_0$ of initial cells with heterogeneous initial ages.\ The above calculation is general because we did not put any constraint on $B(a)$. Let us now go into more details by choosing a power law for the division rate: $B(a)=\nu a^{\alpha}$. In this case, the integral of \[eq\_def\_Z\] is solvable and gives $$\label{eq_Z_pow_law} Z=\frac{1}{\alpha +1} {\left(}\frac{\alpha+1}{\nu} {\right)}^{\frac{1}{\alpha+1}} \Gamma {\left(}\frac{1}{\alpha+1} {\right)}\,,$$ where the Gamma function is defined as $$\Gamma(u)=\int_{0}^{\infty} e^{-x} x^{u-1} \, {\mathrm{d}}x \,.$$ Results are plotted on \[fig\_age\_fit\_land\], which shows that theoretical predictions for the backward and forward age distributions are in good agreement with the numerical histograms. The inset plot shows the age fitness landscape, which follows the linear behavior predicted by \[eq\_age\_fit\_land\]. ![Age distributions with the forward statistics (orange filled histogram) and with the backward statistics (blue unfilled histogram), and the corresponding theoretical curves in red and green. The inset plot shows the age fitness landscape (purple crosses) and the theoretical linear law (green). The horizontal black dashed-line represents the population growth rate $\Lambda$. Simulations were conducted with $B(a)=\nu a^{\alpha}$, constant $\nu=1$, $\alpha=2$, $t=12$ and $N_0=1$.[]{data-label="fig_age_fit_land"}](age_inset_ha.pdf){width="\linewidth"} Let us examine the particular case of uncontrolled models, for which the division rate is constant: $B=\nu$. This corresponds to the case $\alpha=0$ in the power law analysis conducted above. Replacing $\alpha$ by $0$ in \[eq\_Z\_pow\_law\] leads to $Z=1/\nu$; moreover in steady state $\Lambda=\nu$, so that $$p_{{\text{back}}}(a) = 2 \, p_{{\text{for}}}(a) \, e^{-\Lambda a} \,.$$ The extra factor $Z \Lambda$ cancels and we find the same relation between the backward and forward age distributions as Wakamoto’s one between generation times distributions. Moreover, the distributions themselves are greatly simplified and read $$\begin{aligned} p_{{\text{for}}}(a)&=\nu e^{ - \nu a} \,, \\ p_{{\text{back}}}(a)&=2 \nu e^{ - 2 \nu a} \,,\end{aligned}$$ which shows that in this special case the age distributions are themselves identical with the generation time distributions. ### Fluctuating individual growth rates We now introduce the possibility for the individual growth rate $\nu$ to be randomly re-defined at each division, so that the sate of a cell is now determined by two traits: $\boldsymbol{y}=(a,\nu)$. In this case, the steady-state joint distributions of ages and individual growth rates are given by \[eq\_sol\_age\_for,eq\_sol\_age\_back\]. To obtain the steady-state age distributions one needs to integrate these joint distributions on individual growth rates: $$\label{eq_pa_for_nu_fluct} p_{{\text{for}}}(a)= \int {\mathrm{d}}\nu \, p_{{\text{for}}}(0,\nu) \, \exp {\left[}-\int_{0}^{a} B(a',\nu) {\mathrm{d}}a' {\right]}\,,$$ $$\label{eq_pa_back_nu_fluct} p_{{\text{back}}}(a)= e^{-\Lambda a} \int {\mathrm{d}}\nu \, p_{{\text{back}}}(0,\nu) \, \exp {\left[}-\int_{0}^{a} B(a',\nu) {\mathrm{d}}a' {\right]}\,,$$ where $p_{{\text{for}}}(0,\nu)$ and $p_{{\text{back}}}(0,\nu)$ are given by the boundary terms \[eq\_age\_model\_for\_bc,eq\_age\_mod\_back\_bc\]: $$p_{{\text{for}}}(0,\nu)= \int {\mathrm{d}}a {\mathrm{d}}\nu' B(a,\nu') \Sigma {\left(}\nu | \nu' {\right)}p_{{\text{for}}}(a,\nu') \,,$$ $$p_{{\text{back}}}(0,\nu)=2 \int {\mathrm{d}}a {\mathrm{d}}\nu' B(a,\nu') \Sigma {\left(}\nu | \nu' {\right)}p_{{\text{back}}}(a,\nu') \,.$$ In the absence of mother-daughter correlations for the individual growth rate, then $\Sigma {\left(}\nu | \nu' {\right)}= \hat{\Sigma} {\left(}\nu {\right)}$, which implies that $p_{{\text{for}}}(0,\nu)$ and $p_{{\text{back}}}(0,\nu)$ have the same dependency in $\nu$: $$\begin{aligned} p_{{\text{for}}}(0,\nu)&= \hat{\Sigma} {\left(}\nu {\right)}\int {\mathrm{d}}a {\mathrm{d}}\nu' B(a,\nu') p_{{\text{for}}}(a,\nu') \,, \\ &=\hat{\Sigma}{\left(}\nu {\right)}\, \hat{Z}^{-1} \\ p_{{\text{back}}}(0,\nu)&=2 \hat{\Sigma} \int {\mathrm{d}}a {\mathrm{d}}\nu' B(a,\nu') p_{{\text{back}}}(a,\nu') \\ &= 2 \hat{\Sigma} {\left(}\nu {\right)}\, \Lambda \,.\end{aligned}$$ Finally, the fluctuation relation for the age reads $$\frac{p_{{\text{back}}}(a)}{p_{{\text{for}}}(a)}= 2 \hat{Z} \Lambda e^{-\Lambda a} \,,$$ which is the equivalent of \[eq\_age\_fr\] for fluctuating growth rates without mother-daughter correlations. Therefore, the age fitness landscape features the same linear dependency in age with a slope $- \Lambda$ as in the case of constant individual growth rate. In the general case with mother-daughter correlations, this statement is not necessarily true though. Indeed, $p_{{\text{for}}}(0,\nu)$ and $p_{{\text{back}}}(0,\nu)$ do not have in general the same dependency in $\nu$ and therefore the integrals on $\nu$ in \[eq\_pa\_for\_nu\_fluct,eq\_pa\_back\_nu\_fluct\] do not have the same dependency in $a$, thus adding an extra age-dependent term in the age fitness landscape. Consequently, looking at the slope of the age fitness landscape informs on the presence of mother-daughter correlations. We checked numerically that the age fitness landscape without mother-daughter correlations aligns with the theoretical prediction of slope $- \Lambda$; while the same function with mother-daughter correlations presents a non-linear age dependency. Discussion ========== We have studied two different methods to sample lineages in a branched tree: one sampling called backward or retrospective presents a statistical biais with respect to the forward or chronological sampling, an observation which is important to relate experiments carried out at the population level with the ones carried out at the single lineage level. This statistical bias can be rationalized by a set of fluctuation relations, which relate probability distributions in the two ensembles and which are similar to fluctuation relations known in Stochastic Thermodynamics. This analogy suggests new methods to infer the population growth rate based on statistics sampled only in the forward or backward direction. Distributions of the number of divisions are also constrained by these fluctuation relations, which can also be expressed at the level of path probabilities or at the operator level. Interesting inequalities between the mean number of divisions or the mean generation times follow from these fluctuation relations. One inequality concerning the mean generation time has already been verified experimentally by Hashimoto et al. [@hashimoto_noise-driven_2016] for various strains of E Coli. It would be interesting to perform more experimental investigations of this kind for other cell types or under different conditions, since such tests are very important to test our framework. By measuring the difference in the two samplings, namely the forward and the backward one, for a specific trait, one can detect whether that particular trait is or not correlated with the division rate. The fitness landscape, related to this difference, has been introduced by Nozoe et al. [@nozoe_inferring_2017]. While these authors have applied that concept to variables which are not reset or redistributed at division in their work, in the present paper, we used the concept of fitness landscape for variables like the size and the age, which precisely undergo a reset at division in size and age models. We derived expressions for these fitness landscapes, which agree with the statistical biais which we expect when measuring size or age distributions in cell populations. In addition, we also find that the precise form of the age fitness function appears to inform whether or not mother-daughter correlations are present in age models. In the future, it would be interesting to further investigate models with mother-daughter correlations or with correlated single cell growth rates. It would be also valuable to extend the calculations of fitness landscapes to include other important phenotypic state variables besides size or age, as done in [@nakashima_lineage_2020]. We hope that our work has contributed to clarifying the connection between single lineage and population statistics and to understanding the fundamental constraints which cell growth and division must obey.\ \ **Acknowledgments**: The authors acknowledge R. García-García for a previous collaboration, which made possible the present work. We would also like to thank L. Robert, P. Gaspard and J. Unterberger for stimulating discussions.\ **Funding**: This work was partially funded by the Labex CelTisPhysBio (ANR-10-LBX-0038). Age and size controlled models {#app_simpl_mod} ============================== Age models {#app_age_mod} ---------- For models where the division is only controlled by the age $a$ of the cell and its growth rate $\nu$ (so that $\boldsymbol{y}=(a,\nu)$), through a division rate $B(a,\nu)$, \[eq\_mix\_mod\] and \[eq\_mix\_model\_bc\] simplify as $$\label{eq_age_mod} {\left(}\partial_t + \partial_a {\right)}n(a,\nu,K,t) + B(a,\nu)n(a,\nu,K,t) =0 \,,$$ $$n(a=0,\nu,K,t)= 2 \int {\mathrm{d}}a' {\mathrm{d}}\nu' B(a',\nu') \Sigma(\nu|\nu') n(a',\nu',K-1,t) \,.$$ As discussed in \[sec\_edp\_proba\], the dynamics can also be expressed at the level of probabilities instead of populations. Then, the equation and boundary term for the forward probability, summed over $K$, read $$\label{eq_age_mod_for} {\left(}\partial_t + \partial_a {\right)}p_{{\text{for}}}(a,\nu,t) + B(a,\nu)p_{{\text{for}}}(a,\nu,t) =0 \,,$$ $$\label{eq_age_model_for_bc} p_{{\text{for}}}(0,\nu,t)=\int {\mathrm{d}}a {\mathrm{d}}\nu' B(a,\nu') \Sigma(\nu|\nu') p_{{\text{for}}}(a,\nu',t) \,.$$ Similarly, for the backward probability, we obtain $$\label{eq_age_mod_back} {\left(}\partial_t + \partial_a {\right)}p_{{\text{back}}}(a,\nu,t) + {\left[}B(a,\nu) + \Lambda_p(t) {\right]}p_{{\text{back}}}(a,\nu,t) =0 \,,$$ $$\label{eq_age_mod_back_bc} p_{{\text{back}}}(0,\nu,t)=2 \int {\mathrm{d}}a {\mathrm{d}}\nu' \, B(a,\nu') \Sigma(\nu|\nu') p_{{\text{back}}}(a,\nu',t) \,.$$ We can relate the instantaneous population growth rate $\Lambda_p(t)$ to the backward averaged division rate by integrating \[eq\_age\_mod\_back\] over $a$ and $\nu$ and using the boundary condition \[eq\_age\_mod\_back\_bc\] and the normalization of $p_{{\text{back}}}$: $$\label{eq_age_mod_lambda} \Lambda_p(t) = \int {\mathrm{d}}a {\mathrm{d}}\nu \, B(a,\nu) p_{{\text{back}}}(a,\nu,t) = \frac{1}{2} \, \int {\mathrm{d}}\nu \, p_{{\text{back}}}(0,\nu,t) \,.$$ In steady-state, \[eq\_age\_mod\_for,eq\_age\_mod\_back\] can be solved to give the joint distribution of age and individual growth rate: $$\begin{aligned} \label{eq_sol_age_for} p_{{\text{for}}}(a,\nu)&=p_{{\text{for}}}(0,\nu) \, \exp {\left[}-\int_{0}^{a} B(a',\nu) {\mathrm{d}}a' {\right]}\,, \\ \label{eq_sol_age_back} p_{{\text{back}}}(a,\nu)&=p_{{\text{back}}}(0,\nu) \, \exp {\left[}-\Lambda a -\int_{0}^{a} B(a',\nu) {\mathrm{d}}a' {\right]}\,.\end{aligned}$$ Size models {#app_size_mod} ----------- For models where the division is only controlled by the size $x$ of the cell and its growth rate $\nu$ (so that $\boldsymbol{y}=(x,\nu)$), through a division rate $B(x,\nu)$, \[eq\_mix\_mod\] and the boundary condition \[eq\_mix\_model\_bc\] merge as $$\begin{aligned} \label{eq_size_mod} \partial_t n(x,\nu,K,t) &+\partial_x {\left[}\nu x n(x,\nu,K,t) {\right]}+ B(x,\nu)n(x,\nu,K,t) {\nonumber}\\ &- 2 \int {\mathrm{d}}x' {\mathrm{d}}\nu' B(x',\nu') \Sigma(x,\nu|x',\nu') n(x',\nu',K-1,t)=0 \,.\end{aligned}$$ As for age models, the equation for the forward probability reads $$\begin{aligned} \label{eq_size_mod_for} \partial_t p_{{\text{for}}}(x,\nu,K,t) &+\partial_x {\left[}\nu x p_{{\text{for}}}(x,\nu,K,t) {\right]}+ B(x,\nu)p_{{\text{for}}}(x,\nu,K,t) {\nonumber}\\ &- \int {\mathrm{d}}x' {\mathrm{d}}\nu' \, B(x',\nu') \Sigma(x,\nu|x',\nu') p_{{\text{for}}}(x',\nu',K-1,t)=0 \,.\end{aligned}$$ Similarly, for the backward probability, we obtain $$\begin{aligned} \label{eq_size_mod_back} \partial_t p_{{\text{back}}}(x,\nu,K,t) &+\partial_x {\left[}\nu x p_{{\text{back}}}(x,\nu,K,t) {\right]}+ {\left[}B(x,\nu) + \Lambda_p(t) {\right]}p_{{\text{back}}}(x,\nu,K,t) {\nonumber}\\ &- 2\int {\mathrm{d}}x' {\mathrm{d}}\nu' \, B(x',\nu') \Sigma(x,\nu|x',\nu') p_{{\text{back}}}(x',\nu',K-1,t)=0 \,.\end{aligned}$$ Integrating \[eq\_size\_mod\_back\] over $x$ and $\nu$, summing over $K$ and using the normalization of $p_{{\text{back}}}$ we obtain $$\label{eq_size_mod_lambda} \Lambda_p(t) = \int {\mathrm{d}}x {\mathrm{d}}\nu \, B(x,\nu) p_{{\text{back}}}(x,\nu,t) \,.$$ Path integral solution {#app_mix_sol} ====================== General solution ---------------- We start by giving a recursive solution to \[eq\_mix\_mod\] $$\begin{aligned} \label{eq_mix_mod_rec_sol} n(\boldsymbol{y},K,t) &= m \int {\mathrm{d}}x' {\mathrm{d}}x_K {\mathrm{d}}\nu_K {\mathrm{d}}a_{K-1} {\mathrm{d}}\nu_{K-1} \delta(\nu-\nu_K) \delta {\left(}x-x_Ke^{\nu_K a} {\right)}\exp {\left[}-\int_{0}^{a} {\mathrm{d}}\tau B {\left(}x_K e^{\nu_K \tau},\tau,\nu {\right)}{\right]}{\nonumber}\\ &\times B {\left(}x',a_{K-1},\nu_{K-1} {\right)}\Sigma {\left(}x_K,\nu_K \big| x',\nu_{K-1} {\right)}n(x',a_{K-1},\nu_{K-1},t-a,K-1) \,.\end{aligned}$$ One can compute the cell number iteratively until reaching $K=0$, which represents cells that have not divided up to time $t$. This boundary term reads $$\label{eq_mix_k0} n(\boldsymbol{y},K=0,t) = \int {\mathrm{d}}x_0 {\mathrm{d}}\nu_0 \delta (t-a) \delta(\nu-\nu_0) \delta {\left(}x-x_0e^{\nu_0 a} {\right)}\exp {\left[}-\int_{0}^{a} {\mathrm{d}}\tau B {\left(}x_0 e^{\nu_0 \tau},\tau,\nu_0 {\right)}{\right]}n_0(x_0,\nu_0) \,.$$ The final result then reads $$\begin{aligned} \label{eq_mix_mod_fin_sol} n(\boldsymbol{y},K,t) &= m^K\int {\mathrm{d}}x_K {\mathrm{d}}\nu_K \prod_{k=0}^{K-1} {\mathrm{d}}\boldsymbol{y}_k \delta {\left(}t-a-\sum_{k=0}^{K-1} a_k {\right)}\delta(\nu-\nu_K) \delta {\left(}x-x_Ke^{\nu_K a} {\right)}{\nonumber}\\ &\times\exp {\left[}-\int_{0}^{t} {\mathrm{d}}\tau B {\left(}\boldsymbol{y}(\tau) {\right)}{\right]}\prod_{k=1}^K \mathbb{T} {\left(}x_k,\nu_k \big| x_{k-1}e^{\nu_{k-1} a_{k-1}},\nu_{k-1}, a_{k-1} {\right)}n_0(x_0,\nu_0) \,,\end{aligned}$$ where trajectories $\boldsymbol{y}(\tau)$ explicitly appearing in the exponential are given by $\nu(\tau)=\nu_k$, $x(\tau)=x_k\exp(\nu_k(\tau-t_k))$ and $a(\tau)=\tau-t_k$ for $\tau \in [t_k, t_{k+1}[$. The $t_k$ are the division times, given by $t_k=\sum_{k=0}^{k-1} a_k $ for $k \in [1,K-1]$ and $t_0=0$, $t_{K+1}=t$. In addition, the transition matrix $\mathbb{T}$ is defined as $\mathbb{T}(x,\nu|x',\nu',a')=\Sigma(x,\nu|x',\nu')B(x',a',\nu')$. Lastly, we can define the number of paths $n(\{\boldsymbol{y}\},K,t)$ as the number of lineages in the tree that followed the path $\{\boldsymbol{y}\}$ with $K$ divisions: $$\begin{aligned} n(\{x_k,a_k,\nu_k\},K,t)= & m^K \exp {\left[}-\int_{0}^{t} {\mathrm{d}}\tau B {\left(}\boldsymbol{y}(\tau) {\right)}{\right]}n_0(x_0,\nu_0) {\nonumber}\\ & \times \prod_{k=1}^K \mathbb{T} {\left(}x_k,\nu_k \big| x_{k-1}e^{\nu_{k-1} a_{k-1}},\nu_{k-1}, a_{k-1} {\right)}\,.\end{aligned}$$ Age models {#age-models-1} ---------- In the case of age models with constant individual growth rate $\nu$, then $\boldsymbol{y}=y=a$, and $\Sigma(a|a')=\delta(a)$ because new cells always appear at age $0$. Therefore \[eq\_mix\_mod\_fin\_sol\] reads $$\label{eq_mix_mod_fin_sol_age} n(a,K,t) = m^K\int \prod_{k=0}^{K-1} {\mathrm{d}}a_k \, \delta {\left(}t-a-\sum_{k=0}^{K-1} a_k {\right)}\exp {\left[}-\int_{0}^{t} {\mathrm{d}}\tau B {\left(}a(\tau) {\right)}{\right]}\prod_{k=1}^K B(a_{k-1}) \,.$$ The number $n(\{a_k\},K,t)$ of paths then reads $$\label{eq_path_density_age} n(\{a_k\},a,K,t)= m^K \exp {\left[}-\int_{0}^{a} {\mathrm{d}}\tau B {\left(}\tau {\right)}{\right]}\prod_{k=1}^K B(a_{k-1}) \exp {\left[}-\int_{0}^{a_{k-1}} {\mathrm{d}}\tau B {\left(}\tau {\right)}{\right]}\,.$$ Using the definition of the forward path probability $\mathcal{P}_{\text{for}}(\{a_k\},a,K,t)=n(\{a_k\},a,K,t) / m^{K}$ we obtain $$\label{eq_path_number_age} \mathcal{P}_{\text{for}}(\{a_k\},a,K,t) = g_{\text{for}}(a) \prod_{k=0}^{K-1} f_{\text{for}}(a_k) \,,$$ where $g_{\text{for}}(a)=\exp {\left[}-\int_{0}^{a} {\mathrm{d}}\tau B {\left(}\tau {\right)}{\right]}$ is the probability for the cell not to divide from age $0$ at time $t-a$ to age $a$ at finale time $t$; and where $f_{\text{for}}$ is the forward generation time distribution $$f_{\text{for}}(\tau)=B(\tau) e^{-\int_{0}^{\tau} {\mathrm{d}}t B(t)} \,.$$ We define the equivalent quantities $f_{{\text{back}}}(\tau)$ and $g_{{\text{back}}}(a)$ at the backward level: $$\mathcal{P}_{{\text{back}}}(\{a_k\},a,K,t) = g_{{\text{back}}}(a) \prod_{k=0}^{K-1} f_{{\text{back}}}(a_k) \,.$$ Using the fluctuation relation at the path level, namely \[eq\_fr\_path\], and making the choice $\forall k \in [0,K-1], \ a_k=\tau=(t-a)/K$, we obtain $$f_{{\text{back}}}(\tau)=m f_{\text{for}}(\tau) e^{- \tau \Lambda_t} \, e^{-a \Lambda/K} {\left[}\frac{g_{\text{for}}(a)}{g_{{\text{back}}}(a)} {\right]}^{1/K} \,.$$ In a steady state, when $t \rightarrow \infty$ and $K \rightarrow \infty$, then $e^{-a \Lambda/K} {\left[}g_{\text{for}}(a)/g_{{\text{back}}}(a) {\right]}^{1/K}$ $\rightarrow 1$ so the steady-state backward distribution of generation times is given by $$f_{{\text{back}}}(\tau)=m f_{\text{for}}(\tau) e^{- \tau \Lambda_t} \,.$$

ArXiv

--- abstract: | Quadratic variations of Gaussian processes play important role in both stochastic analysis and in applications such as estimation of model parameters, and for this reason the topic has been extensively studied in the literature. In this article we study the problem for general Gaussian processes and we provide sufficient and necessary conditions for different types of convergence which include convergence in probability, almost sure convergence, $L^p$-convergence as well as convergence in law. Furthermore, we study general Gaussian vectors from which different interesting cases including first or second order quadratic variations can be studied by appropriate choice of the underlying vector. Finally, we provide a practical and simple approach to attack the problem which simplifies the existing methodology considerably. [*Keywords: Quadratic variations; Gaussian vectors; Gaussian processes; Convergence in probability; Strong convergence; Convergence in $L^p$; Central limit theorem*]{} skip [*2010 AMS subject classification: 60F05, 60F15, 60F25, 60G15*]{} author: - Lauri Viitasaari - Lauri Viitasaari bibliography: - 'bibli.bib' title: '**Sufficient and Necessary Conditions for Limit Theorems for Quadratic Variations of Gaussian Sequences** ' --- Introduction ============ Quadratic variation of a stochastic process $X$ plays an important role in different applications. For example, the concept is important is one is interested to develop stochastic calculus with respect to $X$. Furthermore, quadratic variations can be used to build estimators for the model parameters such as self-similarity index or parameter describing long range dependence which have important applications in all fields of science such as hydrology, chemistry, physics, and finance to simply name a few. For such applications one is interested to study the convergence of the quadratic variation. Furthermore, a wanted feature is to obtain a central limit theorem which allows one to apply statistical tools developed for normal random variables. For Gaussian processes the study of quadratic variation goes back to Lévy who studied standard Brownian motion and showed the almost sure convergence $$\lim_{n\to\infty}\sum_{k=1}^{2^n} \left[W_{\frac{k}{2^n}}-W_{\frac{k-1}{2^n}}\right]^2 =1.$$ Later this result was extended to cover more general Gaussian processes in Baxter [@Bax] and in Gladyshev [@glady] for uniformly divided partitions. General subdivisions were studied in Dudley [@dud] and Klein & Gine [@klei-gine] where the optimal condition $o\left(\frac{1}{\log n}\right)$ for the mesh of the partition was obtained in order to obtain almost sure convergence. It is also known that for the standard Brownian motion the condition $o\left(\frac{1}{\log n}\right)$ is not only sufficient but also necessary. For details on this topic see De La Vega [@vega] for construction, and [@le-er] for recent results. Functional central limit theorem for general class of Gaussian processes were studied in Perrin [@perrin]. More recently Kubilius and Melichov [@kub-mel] defined a modified Gladyshev’s estimator and the authors also studied the rate of convergence. Norvaiŝa [@norvaisa] have extended Gladyshev’s theorem to a more general class of Gaussian processes. Finally, we can mention a paper by Malukas [@malukas] who extended the results of Norvaiŝa to irregular partitions, and derived sufficient conditions for the mesh in order to obtain almost sure convergence. The case of fractional Brownian motion with Hurst index $H\in(0,1)$ were studied in details by Gyons and Leons [@gyo-leo] where the authors showed that appropriately scaled first order quadratic variation (that is, the one based on differences $X_{t_k} - X_{t_{k-1}}$) converges to a Gaussian limit only if $H<\frac{3}{4}$. To overcome this problem, a generalisations of quadratic variations were used in [@I-L], [@benassi-et-al], [@co], and [@cohen-et-al]. The most commonly used generalisation is second order quadratic variations based on differences $X_{t_{k+1}}-2X_{t_k}+X_{t_{k-1}}$ which was studied in details in a series of papers by Begyn [@begyn-ejp; @begyn-spa; @begyn-ber] with applications to fractional Brownian sheet and time-space deformed fractional Brownian motion. In particular, in [@begyn-ejp] the sufficient condition for almost sure convergence was studied with non-uniform partitions. The central limit theorem and its functional version were studied in [@begyn-ber] and [@begyn-spa] with respect to a standard uniform divided partitions. Furthermore, the authors in papers [@Breu-major; @taqqu] have studied more general variations assuming that the underlying Gaussian process have stationary increments. For another generalisation, the localised quadratic variations were introduced in [@benassi-et-al2] in order to estimate the Hurst function of multifractional Gaussian process. These results have been generalised in [@coeurjolly; @lacaux]. The fractional Brownian motion has received a lot of attention in modelling as a (relatively) simple generalisation of a standard Brownian motion. However, for some applications the assumption of stationary increments is an unwanted feature. For this reason there is a need for extensions of fractional Brownian motion. Recent such generalisations are sub-fractional Brownian motion depending on one parameter $H\in(0,1)$ introduced by Bojdecki et al. [@bojdecki-et-al], and bifractional Brownian motion depending on two parameters $H\in(0,1)$ and $K\in(0,1]$ (the case $K=1$ corresponding to the fractional Brownian motion) introduced by Houdré and Villa [@h-v], and later studied in more details by Russo and Tudor [@ru-tu]. Furthermore, bifractional Brownian motion was extended for values $H\in(0,1)$, $K\in(1,2)$ satisfying $HK\in(0,1)$ in [@bar-seb]. Recently there has also been interest in general Hermite variations which are partially motivated by the contributing paper by Breuer & Major [@Breu-major]. For results related to fractional Brownian motion, we refer to [@bre-nou; @nou-et-al; @nou-rev]. The integrals of fractional Brownian motion were studied in [@cor-et-al]. Moreover, fractional Brownian sheet have received attention at least in papers [@rev; @rev-et-al; @pak-rev]. Furthermore, generalisations are studied in [@pakkanen] and [@barn-et-al]. The mentioned papers studying general Hermite variations are not focused on practical importance of quadratic variations. Furthermore, the proofs are based on now well-developed Malliavin calculus. This paper aims to take a practical, instructive, and a general approach to quadratic variations. Firstly, our aim is to provide easy to check conditions for practitioners which still covers many interesting cases. Secondly, with our simplified approach we are able to provide intuitively clear explanations such as the discussion in subsection \[subsec:a-qv\] rather than get lost into technical details. Finally, to obtain the generality we study sequences of general $n$-dimensional Gaussian vectors $Y^n=\left(Y_1^{(n)}, \ldots, Y_n^{(n)}\right)$, where each component $Y^{(n)}_k$ may depend on $n$, and we study the asymptotic behaviour of the vector $Y^n$ or its quadratic variation defined as the limit $$\lim_{n\to \infty} \sum_{k=1}^n \left[Y^{(n)}_k\right]^2$$ provided it exists in some sense. As such, different cases such as first or second order quadratic variations can be obtained by choosing the vectors $Y^n$ suitably, and this fact will be illustrated in the present paper. We begin by providing sufficient and necessary conditions for the convergence in probability which, applied to some quadratic functional of a given process, can be used to construct consistent estimators of the model parameters. Furthermore, we show that in this case the convergence holds also in $L^p$ for any $p\geq 1$. We will also apply the well-known Gaussian concentration inequality for Hilbert-valued Gaussian random variables which provides a simple condition that guarantees the almost sure convergence. This condition is applied to quadratic variations of Gaussian processes with non-uniform partitions for which we obtain sufficient conditions for the convergence. More importantly, this result is shown to hold for many cases of interest and in the particular case of standard Brownian motion, this condition corresponds to the known sufficient and necessary condition. Compared to the existing literature, in many of the mentioned studies the almost sure convergence is obtained by the use of Hanson and Wright inequality [@hanson-wright] together with some technical computations. In this paper we show how these results follow easily from Gaussian concentration phenomena. We will also study central limit theorem in our general setting. We begin by providing sufficient and necessary conditions under which appropriately scaled quadratic variation converges to a Gaussian limit. To obtain this result we apply a powerful fourth moment theorem proved by Nualart and Peccati [@nu-pe] which, thanks to the recent results by Sottinen and the current author [@sot-vii], can essentially be applied always. We will also show how a version of Lindeberg’s central limit theorem for this case follows easily. Finally, we will use some well-known matrix norm relations to obtain surprisingly simple way to obtain a convergence towards a normal random variable. More remarkably, it seems that this simple condition is essentially the one used in many of the mentioned studies while the result is derived case by case. We will also provide a Berry-Esseen type bound that holds in our general setting which, to the best of our knowledge, is not present in the literature excluding some very special cases. Furthermore, our approach does not require the knowledge of Malliavin calculus and should be applicable to anyone with some background on linear algebra and Gaussian vectors. To summarise, in this paper we give sufficient and necessary conditions for the convergence of guadratic variations of general Gaussian vectors which can be used to reproduce and generalise the existing results. Furthermore, we give easily checked sufficient conditions how one can obtain the wanted convergence results. As such, with our approach we are able to generalise the existing results as well as simplify the proofs considerably by relying on different techniques. At the best, the methods and results of this paper would provide new tools to attack the problem under consideration while classically the problem is studied by relying on Hanson and Wright inequality together with Lindeberg’s central limit theorem. The rest of the paper is organised as follows. In section \[sec:conv\] we study general Gaussian vectors and provide our main results. In section \[sec:qv\] we illustrate how our results can be used to study quadratic variations. We will consider non-uniform sequences and generalise some of the existing results. The main emphasis is on first order quadratic variations which is more closely related to stochastic calculus while we also illustrate how second order quadratic variations can be studied with our approach. We end section \[sec:qv\] with a discussion on general quadratic variations which is close in spirit with the work by Istas and Lang [@I-L]. Finally, section \[sec:ex\] is devoted to examples. Convergence of Gaussian sequences {#sec:conv} ================================= Notation and first results -------------------------- Let $Y^n=\left(Y_1^{(n)},Y_2^{(n)},\ldots, Y_n^{(n)}\right)$ be a zero mean Gaussian vector, where each $Y_k^{(n)}$ possibly depends on $n$. We consider properties of sequences $Y^n$ as $n$ tends to infinity. Throughout the paper we will also use Landau notation, i.e. for a sequences $a_n$ and $b_n$ we denote; - $a_n = O(b_n)$ if $\sup_{n\geq 1}\frac{|a_n|}{|b_n|} < \infty$, - $a_n = o\left(b_n\right)$ if $\lim_{n\to\infty}\frac{|a_n|}{|b_n|} = 0 $. We also denote $a_n \sim b_n$ as $n\to \infty$ if $\lim_{n\to \infty}\frac{a_n}{b_n} \to 1$. We begin with the following definition which is a discrete analogy to the similar concepts introduces in [@ru-va2]. 1. We say that the sequence $Y=(Y^n)_{n=1}^\infty$ has a quadratic variation ${\langle}Y{\rangle}$ if the random variable ${\langle}Y{\rangle}:=\lim_{n\to \infty} \sum_{k=1}^n \left(Y_k^{(n)}\right)^2$ exists as a limit in probability. 2. The energy of a sequence $Y=(Y^n)_{n=1}^\infty$ is defined as the limit $$\varepsilon(Y):=\lim_{n\to \infty} \sum_{k=1}^n {\mathbb{E}}\left(Y_k^{(n)}\right)^2$$ provided the limit exists. 3. We say that $Y$ has $2$-planar variation defined as the limit $$\Upsilon(Y):=\lim_{n,m\to \infty} \sum_{k=1}^n \sum_{j=1}^m\left[{\mathbb{E}}\left[Y_k^{(n)}Y_j^{(m)}\right]\right]^2$$ provided the limit exists. We will also denote $$\label{eq:V_n} V_n = \sum_{k=1}^n \left[\left(Y_k^{(n)}\right)^2 - {\mathbb{E}}\left(Y_k^{(n)}\right)^2\right]$$ for the centered quadratic variation. Let $X=(X_t)_{t\in[0,T]}$ be a centred Gaussian process, $t_k^n = \frac{k}{n}$ and let $\Delta_k X$ denote some difference of the Gaussian process $X$. By setting $Y^{(n)}_k = \frac{\Delta_k X}{n{\mathbb{E}}(\Delta_k X)^2}$ we have ${\mathbb{E}}\left(Y_k^{(n)}\right)^2 = n^{-1}$ and $$V_n = \frac{1}{n} \sum_{k=1}^n \frac{\Delta_k X}{{\mathbb{E}}(\Delta_k X)^2} - 1.$$ Hence by setting $\Delta_k X = X_{t_k^n} - X_{t^n_{k-1}}$ we obtain the first order quadratic variation with respect to uniform partition. Similarly, by setting $\Delta_k X = X_{t^n_{k+1}} - 2X_{t^n_k} + X_{t^n_{k-1}}$ we obtain the second order quadratic variation with respect to uniform partition. The Euclidean distance of the vector $Y^n=(Y_1^{(n)},Y_2^{(n)},\ldots, Y_n^{(n)})$ is given by $$\Vert Y^n\Vert_2 = \sqrt{\sum_{k=1}^n \left(Y_k^{(n)}\right)^{2}}.$$ Note that the norm $\Vert \cdot \Vert_2$ also depends on the dimension $n$ which we will omit on the notation. We will denote by $\Gamma^{(n)}$ the covariance matrix of the vector $Y^n$, i.e. the $n\times n$-matrix with elements $$\Gamma_{jk}^{(n)} = {\mathbb{E}}\left(Y_j^{(n)}Y_k^{(n)}\right).$$ Note now that with this notation the energy of a process $Y$ is simply the limit of the trace of the matrix $\Gamma$, i.e. $\varepsilon(Y) = \lim_{n\to\infty}\text{trace}(\Gamma^n)$. Similarly, $Y^n$ has a quadratic variation if the limit $\Vert Y^n\Vert_2^2$ converges as $n$ tends to infinity. Recall next that the Frobenius norm of a matrix $\Gamma^{(n)} = (\Gamma^{(n)}_{ij})_{i,j=1,\ldots,n}$ is given by $$\Vert \Gamma^{(n)} \Vert_F = \sqrt{\sum_{i,j=1}^n \left(\Gamma_{ij}^{(n)}\right)^2}.$$ Hence we have $$\label{eq:frob_limit} \lim_{n\to\infty} \Vert \Gamma^{(n)} \Vert_F^2 = \lim_{n\to\infty} \sum_{k=1}^n \sum_{j=1}^n \left[{\mathbb{E}}\left[Y_k^{(n)}Y_j^{(n)}\right]\right]^2.$$ We will later show that in interesting cases we also have $$\lim_{n,m\to\infty} \sum_{k=1}^n \sum_{j=1}^m\left[{\mathbb{E}}\left[Y_k^{(n)}Y_j^{(m)}\right]\right]^2 = \lim_{n\to\infty} \sum_{k=1}^n \sum_{j=1}^n \left[{\mathbb{E}}\left[Y_k^{(n)}Y_j^{(n)}\right]\right]^2$$ which, in view of , shows that the $2$-planar variation $\Upsilon(Y)$ is given by $$\Upsilon(Y) = \lim_{n\to \infty} \Vert \Gamma^{(n)} \Vert_F^2.$$ The following first result concerns convergence in probability. The proof follows essentially the arguments presented in [@ru-va2] and is based on cumulant formulas for Gaussian random variables. The main difference is that we also prove the convergence in $L^p$ for any $p\geq 1$ which have some important consequences in stochastic analysis (to be detailed in a forthcoming paper) while in [@ru-va2] the authors considered only $L^2$-convergence. The proof follows the ideas presented in [@ru-va2] but we will present the key points for the sake of completeness. \[thm:QV-ex\] Let $Y=(Y^n)_{n=1}^\infty$ be a sequence of Gaussian vectors with finite energy. Then quadratic variation exists as a limit in probability for every $t\geq 0$ if and only if the sequence $(Y^n)_{n=1}^\infty$ has 2-planar variation. In this case, the convergence holds also in $L^p$ for any number $p\geq 1$. Denote $$Z_n = \Vert Y^n \Vert_{2}^{2}=\sum_{k=1}^n \left(Y_k^{(n)}\right)^2.$$ We start by showing that $$Z^p_n = \left[\sum_{k=1}^n \left(Y^{(n)}_k\right)^2\right]^p$$ is uniformly integrable for any $p\geq 1$. By Minkowski’s inequality for measures we get $$\begin{split} &\left[{\mathbb{E}}\left(\sum_{k=1}^n\left(Y_k^{(n)}\right)^2\right)^p\right]^{\frac{1}{p}}\\ &\leq \sum_{k=1}^n \left[{\mathbb{E}}\left(Y_k^{(n)}\right)^{2p}\right]^{\frac{1}{p}}\\ &= C_p \sum_{\pi_n} {\mathbb{E}}\left(Y_k^{(n)}\right)^2 \end{split}$$ by the fact that $Y_k$ is Gaussian. Now this upper bound converges to $\varepsilon(Y)$, and consequently the quantity $\left(\sum_{k=1}^n \left(Y^{(n)}_k\right)^2\right)^p$ is uniformly integrable for any $p\geq 1$. Now we have $${\mathbb{E}}(Z_n-Z_m)^2 = {\mathbb{E}}Z_n^2 + {\mathbb{E}}Z_m^2 - 2 {\mathbb{E}}(Z_nZ_m),$$ where $${\mathbb{E}}(Z_nZ_m)=\sum_{k=1}^n\sum_{j=1}^m {\mathbb{E}}\left[\left(Y^{(n)}_k\right)^2\left(Y^{(m)}_j\right)^2\right]$$ Here we have $$\label{eq:basic_cumulant} {\mathbb{E}}\left[\left(Y^{(n)}_k\right)^2\left(Y^{(m)}_j\right)^2\right] = 2\left[{\mathbb{E}}\left(Y^{(n)}_kY^{(m)}_j\right)\right]^2 + {\mathbb{E}}\left(Y^{(n)}_k\right)^2{\mathbb{E}}\left(Y^{(m)}_j\right)^2.$$ Hence we have $$\sum_{k=1}^n\sum_{j=1}^m {\mathbb{E}}\left[\left(Y^{(n)}_k\right)^2\left(Y^{(m)}_j\right)^2\right] = 2\sum_{k=1}^n\sum_{j=1}^m \left[{\mathbb{E}}\left(Y^{(n)}_kY^{(m)}_j\right)\right]^2 + \sum_{k=1}^n {\mathbb{E}}\left(Y^{(n)}_k\right)^2 \times \sum_{k=1}^n {\mathbb{E}}\left(Y^{(n)}_k\right)^2.$$ Consequently, $$\label{eq:key_relation} {\mathbb{E}}(Z_nZ_m) = 2\sum_{k=1}^n\sum_{j=1}^m \left[{\mathbb{E}}\left(Y^{(n)}_kY^{(m)}_j\right)\right]^2 + \sum_{k=1}^n {\mathbb{E}}(Y^{(n)}_k)^2 \times \sum_{k=1}^n {\mathbb{E}}(Y^{(n)}_k)^2,$$ where $$\sum_{k=1}^n {\mathbb{E}}(Y^{(n)}_k)^2 \times \sum_{k=1}^n {\mathbb{E}}(Y^{(n)}_k)^2 \rightarrow \varepsilon(Y)^2$$ by the fact that $Y$ has finite energy. Recall also that 2-planar variation is given by $$\Upsilon(Y) =\lim_{n,m\to\infty} \sum_{k=1}^n\sum_{j=1}^m \left[{\mathbb{E}}\left(Y^{(n)}_kY^{(m)}_j\right)\right]^2.$$ Hence relation implies the result. Indeed, assuming that $Z_n$ converges in probability, then uniform integrability implies that ${\langle}Y{\rangle}\in L^p$ and the convergence holds also in $L^p$. Hence ${\mathbb{E}}(Z_nZ_m)$ converges, and implies that 2-planar variation exists. Conversely, if 2-planar variation exists, then implies that ${\mathbb{E}}(Z_nZ_m)$ converges to the same limit as ${\mathbb{E}}(Z_n^2)$ which concludes the proof. [It is straightforward to check that the $L^p$-convergence takes place also in a continuous setting of Russo and Vallois [@ru-va2]. ]{} The following theorem gives condition when the quadratic variation is deterministic. It seems that this is indeed true in many cases of interest. \[pro:deterministic\_QV\] Let $Y=(Y^n)_{n=1}^\infty$ be a sequence of centred Gaussian vectors such that $Y$ has finite energy. Then quadratic variation exists as a limit in probability and is deterministic if and only if the 2-planar variation is zero. In this case ${\langle}Y{\rangle}$ equals to the energy of the process and the convergence holds also in $L^p$ for any $p\geq 1$. The result follows directly from $$\lim_{n\to\infty}{\mathbb{E}}(Z_n - \varepsilon(Y))^2 = {\mathbb{E}}(Z_n^2) - \varepsilon(Y)^2$$ and the relation with $m=n$. Finally, the convergence in $L^p$ follows directly from Theorem \[thm:QV-ex\]. [ \[rem:alpha-var\] A generalisation of quadratic variation is $\alpha$-variation which is defined as a limit of $ \sum_{k=1}^n \left|Y_k^{(n)}\right|^\alpha $ provided the limit exists. Similarly, we say that $Y$ has finite $\alpha$-energy if the limit $$\varepsilon_\alpha(Y) := \lim_{n\to\infty}\sum_{k=1}^n {\mathbb{E}}\left|Y_k^{(n)}\right|^\alpha$$ exists. It is straightforward to show that if the sequence $Y=(Y_n)_{n=1}^\infty$ has finite $\alpha$-energy and $$\sum_{k=1}^n \left|Y_k^{(n)}\right|^\alpha$$ converges in probability, then the convergence holds also in $L^p$ for any $p\geq 1$. In this case however, the concept of 2-planar variation becomes much more complicated. ]{} Almost sure convergence ----------------------- In this subsection we address the question when the convergence takes place almost surely. The key ingredient for our results is the concentration inequality for Gaussian measures and by use of this inequality we show that, rather surprisingly, the quadratic variation always converges to the energy of the process whether or not the energy is finite provided that 2-planar variation vanishes. Before stating our results we recall the following concentration inequality taken from [@b-h] for Gaussian processes. We present the result using our notation. \[lma:concentration\] Let $Y^n$ be a Gaussian vector with covariance matrix $\Gamma^{(n)}$, and denote $T = \sqrt{\text{trace}\left(\Gamma^{(n)}\right)}$. Then for any $h>0$ we have $$\label{eq:concentration} {\mathbb{P}}\left(\left|\Vert Y\Vert_2 - T\right| \geq h\right) \leq exp\left(-\frac{h^{2}}{4\Vert\Gamma^{(n)}\Vert_2}\right).$$ [ The result holds for any Hilbert-valued Gaussian random variables, not only finite dimensional. Similarly, for general $\alpha$-variations one can use the concentration inequality $${\mathbb{P}}\left(|\Vert X\Vert_{\mathcal{B}} - {\mathbb{E}}\Vert X\Vert_{\mathcal{B}}|\geq h\right) \leq \exp\left(-\frac{h^2}{2\sigma^2}\right),$$ where $(\mathcal{B},\Vert\cdot\Vert_{\mathcal{B}})$ is a Banach space, $X$ is a Banach-valued Gaussian random variable and $\sigma^2 = \sup_{L\in \mathcal{B}':\Vert L\Vert \leq 1} {\mathbb{E}}L(X)^2$. Applying this to ${\mathbb{R}}^n$ equipped with the norm $\Vert\cdot\Vert_{\alpha}$ together with Riesz representation theorem one has $$\label{eq:banach-concentration} {\mathbb{P}}\left(|\Vert Y\Vert_{\alpha} - {\mathbb{E}}\Vert Y\Vert_{\alpha}|\geq h\right) \leq \exp\left(-\frac{h^2}{2\sigma^2}\right),$$ where $\sigma^2 = \sup_{\Vert a\Vert_{q}\leq 1} \sum_{k=1}^n\sum_{j=1}^n a_ka_j{\mathbb{E}}[Y_kY_j]$ with $\frac{1}{\alpha}+\frac{1}{q}=1$. ]{} We now turn to address the question when one obtains almost sure convergence. Clearly, the idea is to find an upper bound on $\Vert \Gamma^{(n)}\Vert_2$, say, $\Vert \Gamma^{(n)}\Vert_2 \leq\phi(n)$. Then one can hope that the bound is good enough such that $$\label{as-convergence} \sum_{n=1}^\infty exp\left(-\frac{\epsilon^{2}}{4\phi(n)}\right) < \infty$$ from which we obtain the almost sure convergence immediately by Borel-Cantelli Lemma. \[thm:QV-conv-general\] Let $Y=(Y^n)_{n=1}^\infty$ be a sequence of Gaussian vector such that $\Vert \Gamma^{(n)} \Vert_2 \rightarrow 0$ and $$\label{eq:UI_general} \sup_{n\geq 1}\sum_{k=1}^n {\mathbb{E}}\left(Y^{(n)}_k\right)^2 < \infty.$$ Then, as $n\to \infty$, $$\label{eq:gen_conv} \left|\sum_{k=1}^n \left(Y^{(n)}_k\right)^2 - \sum_{k=1}^n {\mathbb{E}}\left(Y^{(n)}_k\right)^2\right| \rightarrow 0$$ in probability and in $L^p$ for any $p\geq 1$. Furthermore, the convergence holds almost surely for any partition satisfying $\Vert \Gamma^{(n)} \Vert_2 =o\left(\frac{1}{\log n}\right)$. The convergence From Lemma \[lma:concentration\] we get that $$\left|\sqrt{\sum_{k=1}^n \left(Y^{(n)}_k\right)^2} - \sqrt{\sum_{k=1}^n {\mathbb{E}}\left(Y^{(n)}_k\right)^2}\right| \rightarrow 0$$ in probability follows immediately from Lemma \[lma:concentration\], and the almost sure convergence follows by applying Borel-Cantelli Lemma. Now the convergence follows from decomposition $|a-b|=|\sqrt{a}-\sqrt{b}|(\sqrt{a}+\sqrt{b})$ together with the uniform integrability condition which also implies convergence in $L^p$. [ Note that if $\Vert \Gamma^{(n)}\Vert_2 \to 0$, we have $$\left|\sqrt{\sum_{k=1}^n \left(Y^{(n)}_k\right)^2} - \sqrt{\sum_{k=1}^n {\mathbb{E}}\left(Y^{(n)}_k\right)^2}\right| \rightarrow 0$$ in probability even if $\sup_{n\geq 1}\sum_{k=1}^n {\mathbb{E}}\left(Y^{(n)}_k\right)^2 =\infty$, i.e. we have convergence in probability (or almost surely) while the energy might be infinite. For example, it can be shown that this is the case for fractional Brownian motion $B^H$ with Hurst index $H\in\left(0,\frac12\right)$ and its non-scaled quadratic variation, i.e. the one corresponding to a vector $Y^{(n)}_k = B^H_{t_k} - B^H_{t_{k-1}}$. ]{} [ Note that for finite energy processes such that 2-planar variation vanishes this result shows that $$\lim_{n,m\to\infty} \sum_{k=1}^n \sum_{j=1}^m\left[{\mathbb{E}}\left[Y_k^{(n)}Y_j^{(m)}\right]\right]^2 = \lim_{n\to\infty} \sum_{k=1}^n \sum_{j=1}^n \left[{\mathbb{E}}\left[Y_k^{(n)}Y_j^{(n)}\right]\right]^2$$ or more compactly, $\Upsilon(Y) = \lim_{n\to\infty} \Vert \Gamma^{(n)}\Vert_F$. Indeed, from the well-known relation $\Vert A\Vert_2 \leq \Vert A \Vert_F$ we obtain that if $\lim_{n\to\infty}\Vert \Gamma^{(n)}\Vert_F \rightarrow 0$, then $\Vert \Gamma^{(n)}\Vert_2 \rightarrow 0$. Consequently, Theorem \[thm:QV-ex\] implies that 2-planar variation vanishes. This answers to question raised in [@ru-va2 Remark 3.12] in our discrete setting. Similarly, one can use general concentration inequality to give analogous answer in a continuous case. ]{} To compute the spectral norm $\Vert \Gamma^{(n)}\Vert_2$, or equivalently the largest eigenvalue, can be a challenging task. One way to overcome this challenge is to use Frobenius norm $\Vert\cdot \Vert_F$ which provides an upper bound and is easier to analyse. Unfortunately however, it provides quite rough estimates even in a simple case of standard Brownian motion as will be shown in subsection \[subsec:bm\]. A way to obtain general conditions is to use matrix norm $\Vert \cdot \Vert_1$ which is also the main approach applied in the literature. This is the topic of the next general theorem. The proof is based on some well-known relations for matrix norm and we do not claim originality here. Let $Y=(Y^n)_{n=1}^\infty$ be a sequence of Gaussian vectors such that holds. Furthermore, assume there exists a function $\phi(n)$ such that $$\max_j\sum_{k=1}^n \left|{\mathbb{E}}\left[Y_k^{(n)}Y_j^{(n)}\right]\right| \leq \phi(n).$$ If $\phi(n) \to 0$, then the convergence $$\left|\sum_{k=1}^n \left[Y_k^{(n)}\right]^2 - \sum_{k=1}^n {\mathbb{E}}\left[Y_k^{(n)}\right]^2\right| \rightarrow 0$$ holds in probability and in $L^p$ for any $p\geq 1$. Furthermore, if $$\phi(n) = o\left(\frac{1}{\log n}\right),$$ Then the convergence holds almost surely. Recall the well-known bound for matrix norm $\Vert A\Vert_2$ given by $ \Vert A\Vert_2 \leq \sqrt{\Vert A\Vert_1 \Vert A\Vert_{\infty}}, $ where $ \Vert A\Vert_1 = \max_{1\leq j\leq n} \sum_{k=1}^n |a_{kj}| $ and $ \Vert A\Vert_{\infty} = \max_{1\leq j\leq n} \sum_{k=1}^n |a_{jk}|. $ Hence symmetry of $\Gamma^{(n)}$ gives $ \Vert \Gamma^{(n)} \Vert_2 \leq \Vert \Gamma^{(n)} \Vert_1$, where $$\Vert \Gamma^{(n)} \Vert_1 = \max_j\sum_{k=1}^n \left|{\mathbb{E}}\left[Y_k^{(n)}Y_j^{(n)}\right]\right|.$$ This proves the claim together with Theorem \[thm:QV-conv-general\]. The following final result of this section can be useful for stationary sequences. Let $Y=(Y^n)_{n=1}^\infty$ be a sequence of Gaussian vectors such that holds. Moreover, assume there exists a positive symmetric function $r$ such that $$\left|{\mathbb{E}}\left[Y_k^{(n)}Y_j^{(n)}\right]\right| \leq r(k-j),\quad k,j=1,\ldots,n$$ and assume that there exists a function $\phi(n)$ such that $$\sum_{k=1}^n r(k) \leq \phi(n).$$ If $\phi(n)\to 0$ as $n$ tends to infinity, then the convergence $$\left|\sum_{k=1}^n \left[Y_k^{(n)}\right]^2 - \sum_{k=1}^n {\mathbb{E}}\left[Y_k^{(n)}\right]^2\right| \rightarrow 0$$ holds in probability and in $L^p$. Furthermore, the convergence holds almost surely provided that $ \phi(n) = o\left(\frac{1}{\log n}\right). $ Now for any $j\geq 1$ we have $$\sum_{k=1}^n \left|{\mathbb{E}}\left[Y_k^{(n)}Y_j^{(n)}\right]\right| \leq \sum_{k=1}^n r(k-j) \leq \sum_{k=0}^{j-1}r(k) + \sum_{k=1}^{n-j}r(k) \leq 2\sum_{k=0}^n r(k)$$ from which the result follows. Central limit theorem --------------------- In this section we provide sufficient and necessary condition for the central limit theorem (CLT) to hold. More importantly, we obtain two simple corollaries where the first one gives a version of Lindeberg’s CLT for quadratic variations, and the second one gives simple to check condition which actually holds in most of the studies cited in the introduction. In particular, the second corollary can be used to simplify the used techniques (namely, the Lindeberg’s CLT) considerably. Our necessary and sufficient condition is based on the fourth moment theorem by Nualart and Peccati [@nu-pe]. Hence we begin by recalling some basic facts on Wiener chaos. For more details we refer to monographs [@nualart; @pe-t; @n-p]. Let now $X$ be a separable Gaussian centered process and let $q\geq 1$ be fixed. The symbol $\mathcal{H}_{q}$ denotes the $q$th Wiener chaos of $X$, defined as the closed linear subspace of $L^2(\Omega)$ generated by the family $\{H_{q}(X(h)) : h\in {\mathcal{H}},\left\| h\right\| _{ {\mathcal{H}}}=1\}$, where $H_{q}$ is the $q$th Hermite polynomial. The mapping $I^{X}_{q}(h^{\otimes q})=H_{q}(X(h))$ can be extended to a linear isometry between the symmetric tensor product $ {\mathcal{H}}^{\odot q}$ and the $q$th Wiener chaos $\mathcal{H}_{q}$, and for any $h \in {\mathcal{H}}^{\odot q}$ the random variable $I^X_q(h)$ is called a multiple Wiener integral of order $q$. [ If $X=W$ is a standard Brownian motion, then ${\mathcal{H}}$ is simply the space $L^2([0,T],{\mathrm{d}}t)$. In this case the random variable $I^X_q(h)$ coincides with the $q$-fold multiple Wiener-Itô integral of $h$ (see [@nualart]). ]{} [ Let $X$ be a separable Gaussian process on $[0,T]$. It was proved in [@sot-vii] that $X$ admits a Fredholm integral representation $$X_t = \int_0^T K(t,s){\mathrm{d}}W_s,$$ where $K\in L^2([0,T]^2)$ and $W$ is a Brownian motion, if and only if $\int_0^T {\mathbb{E}}X_u^2 {\mathrm{d}}u < \infty$. Furthermore, it was shown that this representation can be extended to a transfer principle which can be used to develop stochastic calculus with respect to $X$. In particular, the transfer principle can be used to define multiple Wiener integrals as multiple Wiener integrals with respect to a standard Brownian motion. This definition coincides with the one defined via Hermite polynomials. ]{} Finally, we are ready to recall the following characterisation of convergence towards a Gaussian limit. [@nu-pe]\[thm:4th\] Let $\{F_{n}\}_{n\geq 1}$ be a sequence of random variables in the $q$th Wiener chaos, $q\geq 2$, such that $\lim_{n\rightarrow \infty } {\mathbb{E}}(F_{n}^{2})=\sigma ^{2}$. Then, as $n \to \infty$, the following asymptotic statements are equivalent: - $F_n$ converges in law to $\mathcal{N}(0,\sigma^2)$. - ${\mathbb{E}}F_n^4$ converges to $3 \sigma ^{4}$. [ In this paper we are studying quadratic variations of Gaussian sequences. Hence, thanks to Fredholm representation from [@sot-vii], such objects can be viewed as a sequences in the second chaos. ]{} [ The case of the second chaos was studied in details in [@no-po; @no-po1] where the authors characterised all possible limiting laws. More precisely, authors in [@no-po] proved that if a sequence in the second chaos converges in law to some random variable $F$, then $F$ can be viewed as a sum of normal random variable and an independent random variable living in the second chaos. ]{} [ It was proved in [@n-o] that instead of fourth moment one can also study the convergence of $ \|DF_n\|^2_{{\mathcal{H}}} $ in $L^2$, where $D$ stands for Malliavin derivative operator, and this approach have turned out to be very useful in some cases. For our purposes however it seems that working with the fourth moment is more convenient. ]{} Finally, we recall the following Berry-Esseen type estimate taken from [@pe-t Theorem 11.4.3]. \[thm:B-E-bound\] Let $\{F_n\}_{n \ge 1}$ be a sequence of elements in the $q$th Wiener chaos such that ${\mathbb{E}}(F_n^2) =1$ and let $Z$ denote a standard normal random variable. Then there exists a constant $C_q$ depending only on $q$ such that $$\sup_{x\in{\mathbb{R}}}\Big\vert {\mathbb{P}}(F_n < x) - {\mathbb{P}}(Z<x)\Big \vert \leq C_q\sqrt{{\mathbb{E}}F_n^4 - 3}.$$ We are now ready to prove our results. We begin with the following auxiliary technical lemma whose proof is postponed to the appendix. \[lma:2nd\_4th\] For $V_n$ given by we have $${\mathbb{E}}V_n^2 = 2\sum_{k,j=1}^n \left({\mathbb{E}}\left[Y^{(n)}_kY^{(n)}_j\right]\right)^2$$ and $$\begin{split} {\mathbb{E}}V_n^4 &= 12\left[\sum_{k,j=1}^n \left({\mathbb{E}}\left[Y^{(n)}_kY^{(n)}_j\right]\right)^2\right]^2\\ &+ 24\sum_{i,j,k,l=1}^n {\mathbb{E}}\left[Y^{(n)}_kY^{(n)}_j\right]{\mathbb{E}}\left[Y^{(n)}_jY^{(n)}_i\right]{\mathbb{E}}\left[Y^{(n)}_iY^{(n)}_l\right]{\mathbb{E}}\left[Y^{(n)}_lY^{(n)}_k\right]. \end{split}$$ With the help of this lemma we are ready to proof our main theorem. \[thm:CLT\_iff\] Let $Y=(Y^n)_{n=1}^\infty$ be a sequence of Gaussian vectors such that for every $n\geq 1$ the elements $Y_k^{(n)},\quad k=1,\ldots, n$ belong to the first Wiener chaos, and let $\Gamma^{(n)}$ denote the covariance matrix of $Y^n$ with eigenvalues $\lambda_1^n, \lambda_2^n, \ldots, \lambda_n^n$. Then, as $n$ tends to infinity, the following are equivalent. - $\frac{V_n}{\sqrt{Var(V_n)}} \rightarrow \mathcal{N}(0,1)$, - $$\sum_{i,j,k,l=1}^n {\mathbb{E}}\left[Y^{(n)}_kY^{(n)}_j\right]{\mathbb{E}}\left[Y^{(n)}_jY^{(n)}_i\right]{\mathbb{E}}\left[Y^{(n)}_iY^{(n)}_l\right]{\mathbb{E}}\left[Y^{(n)}_lY^{(n)}_k\right] = o\left(Var(V_n)^2\right),$$ - $$\sum_{k=1}^n \left(\lambda_k^n\right)^4 =o\left(\left[\sum_{k=1}^n (\lambda_k^n)^2\right]^2\right).$$ By assumption we are able to use fourth moment theorem \[thm:4th\] from which the equivalence of items (1) and (2) follows with the help of Lemma \[lma:2nd\_4th\]. To obtain equivalence of (1) and (3), it is well-known that $\sum_{k=1}^n \left[Y^{(n)}_k\right]^2$ can be decomposed as $$\sum_{k=1}^n \left[Y^{(n)}_k\right]^2 {\stackrel{\text{law}}{=}}\sum_{k=1}^n \lambda_{k}^n \left[\xi_k^{(n)}\right]^2,$$ where $\lambda_k^n$ are the eigenvalues of the covariance matrix $\Gamma^{(n)}$ and $\xi^{(n)}_k$ are independent standard normal random variables. Denoting $$\tilde{V}_n = \sum_{k=1}^n \left[\lambda_{k}^n \left[\xi_k^{(n)}\right]^2- \lambda_{k}^n\right]$$ and using Lemma \[lma:2nd\_4th\] again we obtain $${\mathbb{E}}\tilde{V}_n^2 = 2\sum_{k=1}^n [\lambda_k^n]^2$$ and $${\mathbb{E}}\tilde{V}_n^4 = 12\left[\sum_{k=1}^n [\lambda_k^n]^2\right]^2 + 6\sum_{k=1}^n [\lambda_k^n]^4$$ which concludes the proof. As a simple corollary we obtain the following result which corresponds to Lindeberg’s CLT and is mainly used in the references given in the introduction. \[cor:lindeberg\] Assume that assumptions of Theorem \[thm:CLT\_iff\] prevail. If $$\lambda^*(n):=\max_{k=1,\ldots,n}|\lambda_k^n| = o\left(\sqrt{Var(V_n)}\right), \quad n\to \infty,$$ then $$\frac{V_n}{\sqrt{Var(V_n)}} \rightarrow \mathcal{N}(0,1).$$ We have $$\sum_{k=1}^n (\lambda_k^n)^4 \leq \max_{k=1,\ldots,n}|\lambda_k^n|^2 \sum_{k=1}^n (\lambda_k^n)^2$$ and since $Var(V_n) = \sum_{k=1}^n (\lambda_k^n)^2$, the result follows at once. [ Note that since Lindeberg’s CLT can be proved without the theory of Wiener chaos, the above result is valid for arbitrary sequences $Y^n$. ]{} Finally, the following theorem justifies that in many cases it is sufficient to find an upper bound for $\lambda^*(n)$, or even for $\max_{1\leq j\leq n}\sum_{k=1}^n \left|{\mathbb{E}}\left[Y^{(n)}_kY^{(n)}_j\right]\right|$. While the proof follows from simple relations for matrix norm, the result turn out to be very useful for many practical applications. In particular, the following result covers many of the cases studied in the literature. Furthermore, in this case it easy to give a Berry-Esseen bound. \[thm:CLT\_main\] Let the assumptions of Theorem \[thm:CLT\_iff\] prevail, and assume that $Y^n$ is a Gaussian vector with finite non-zero energy. Then there exists a constant $C>0$ such that $$\sup_{x\in{\mathbb{R}}}\left| {\mathbb{P}}\left(\frac{V_n}{\sqrt{Var(V_n)}} < x\right) - {\mathbb{P}}(Z<x)\right| \leq C\sqrt{n}\lambda^*(n).$$ where $Z$ is a standard normal random variable. Hence if $$\lambda^*(n) =o\left(\frac{1}{\sqrt{n}}\right),$$ then $\frac{V_n}{\sqrt{Var(V_n)}} \rightarrow \mathcal{N}(0,1)$. In particular, if $$\max_{1\leq j\leq n}\sum_{k=1}^n \left|{\mathbb{E}}\left[Y^{(n)}_kY^{(n)}_j\right]\right| = o\left(\frac{1}{\sqrt{n}}\right),$$ then $\frac{V_n}{\sqrt{Var(V_n)}} \rightarrow \mathcal{N}(0,1)$. Recall the trace norm is given by $\Vert \Gamma^{(n)} \Vert_* = \sum_{k=1}^n \lambda_k^n$. By Cauchy-Schwartz inequality we get the well-known matrix norm inequality $$\Vert \Gamma^{(n)} \Vert_* \leq \sqrt{n}\sqrt{\sum_{k=1}^n [\lambda_k^n]^2} = \sqrt{n}\Vert \Gamma^{(n)} \Vert_F.$$ By assumption, $Y^n$ has finite non-zero energy. Hence by observing that $\lim_{n\to\infty} \Vert \Gamma^{(n)}\Vert_* = \epsilon(Y) > 0$, we obtain that for large enough $n$ we have $$0<c\leq \sqrt{n}\Vert \Gamma^{(n)}\Vert_F.$$ Next we observe that $\sqrt{Var(V_n)} = \Vert \Gamma^{(n)}\Vert_F$. Now the fourth moment of $\frac{V_n}{\sqrt{Var(V_n)}}$ is given by $$\frac{{\mathbb{E}}V_n^4}{\left({\mathbb{E}}V_n^2\right)^2} = 3+\frac{3\sum_{k=1}^n [\lambda_k^n]^4}{2\left(\sum_{k=1}^n [\lambda_k^n]^2\right)^2}$$ so that $$\frac{{\mathbb{E}}V_n^4}{\left({\mathbb{E}}V_n^2\right)^2} - 3 \leq \frac{3[\lambda^*(n)]^2}{2{\mathbb{E}}V_n^2} \leq \frac{3}{2}n[\lambda^*(n)]^2.$$ Hence the Berry-Esseen bound follows from Theorem \[thm:B-E-bound\]. [Note that the convergence towards normal random variable follow also from Corollary \[cor:lindeberg\] which does not rely on the theory of Wiener chaos. However, for a sequence living in the second chaos we also obtain a Berry-Esseen bound. ]{} Application to quadratic variations {#sec:qv} =================================== In this section we apply the results to quadratic variations of Gaussian processes. We begin by giving a general results for generalised variations which will be illustrated in the particular cases of first and second order quadratic variations. We consider arbitrary sequences of partitions $\pi_n = \{0=t^n_0 < t^n_1<\ldots < t^n_{N(\pi_n)-1}=T\}$, where $N(\pi_n)$ denotes the number of points in the partition. For notational simplicity, we drop the super-index $n$ and simply use $t_k$ instead of $t_k^n$. The mesh of the partition is denoted by $|\pi_n|=\max\{t_k-t_{k-1}, t_k \in \pi_n \slash \{0\}\}$. We also use $m(\pi_n) = \min\{t_k-t_{k-1}, t_k \in \pi_n \slash \{0\}\}$. Throughout this section we assume that $$\label{eq:partition_assumption} \frac{|\pi_n|}{m(\pi_n)} \leq k(|\pi_n|), \quad n\geq 1$$ for some function $k$. Obviously, usually the partition is chosen such that $k(|\pi_n|)\leq C < \infty$. First order quadratic variations -------------------------------- In this subsection we apply the results of previous section to study first order quadratic variations of Gaussian processes which is our main interest due to its connection to stochastic analysis. Throughout this subsection we also use the metric defined by the incremental variance of $X$, i.e. $$d_X(t,s) = {\mathbb{E}}(X_t - X_s)^2.$$ Let $X=(X_t)_{t\in[0,T]}$ be a centred Gaussian process. We say that $X$ has first order $\phi$-quadratic variation along $\pi_n$ if $$V_1(\pi_n,\phi):=\sum_{t_k^n\pi_n} \frac{(X_{t_k}-X_{t_{k-1}})^2}{\phi(t_k-t_{k-1})}$$ converges in probability as $|\pi_n|$ tends to zero. [ A natural choice for the function $\phi$ is such that $$\label{eq:fin_en_assumption} \lim_{|\pi_n|\to 0}\sum_{k=1}^{N(\pi_n)-1} \frac{{\mathbb{E}}(X_{t_k}-X_{t_{k-1}})^2}{\phi(t_k-t_{k-1})} = K < \infty.$$ In particular, in many interesting cases one has $d_X(t,s) \sim r(t-s)$ as $|t-s| \to 0$ for some function $r$. In this case a natural choice is $\phi(x)=\frac{r(x)}{x}$. ]{} To simplify the notation we denote $\tilde{V}_1(\pi_n,\phi) = V_1(\pi_n,\phi) - {\mathbb{E}}V_1(\pi_n,\phi)$. We also use $\Delta t_j = t_j - t_{j-1}$. We will begin by giving the following general theorem which generalises main results of [@malukas] by allowing us to drop some technical assumptions. The result follows directly by uniting and rewriting Theorems \[thm:QV-conv-general\] and \[thm:CLT\_main\] for sequence $Y^{(n)}_k = \frac{X_{t_k}-X_{t_{k-1}}}{\sqrt{\phi(t_k-t_{k-1})}}$. \[thm:1st\_order\_main\] Let $X$ be a Gaussian process. Assume that $$\label{eq:general_condition} \max_{1\leq j\leq N(\pi_n)-1}\sum_{k=1}^{N(\pi_n)-1} \frac{1}{\sqrt{\phi(\Delta t_k)\phi(\Delta t_j)}}|{\mathbb{E}}[(X_{t_k}-X_{t_{k-1}})(X_{t_j}-X_{t_{j-1}})]| \leq H(|\pi_n|)$$ for some function $H(|\pi_n|)$. 1. If $H(|\pi_n|) \to 0$ as $|\pi_n|$ tends to zero, then convergence $$\label{eq:convergence} \left|\sum_{k=1}^{N(\pi_n)-1} \frac{(X_{t_k}-X_{t_{k-1}})^2}{\phi(t_k-t_{k-1})} - \sum_{k=1}^{N(\pi_n)-1} \frac{{\mathbb{E}}(X_{t_k}-X_{t_{k-1}})^2}{\phi(t_k-t_{k-1})}\right| \rightarrow 0$$ holds in probability. If $H(|\pi_n|) =o\left(\frac{1}{\log n}\right)$, then convergence holds almost surely. In these cases the convergence holds also in $L^p$ for any $p\geq 1$ provided that holds. 2. Furthermore, assume that $$\label{eq:nontriv_energy} \lim_{|\pi_n|\to 0}\sum_{k=1}^{N(\pi_n)-1} \frac{{\mathbb{E}}(X_{t_k}-X_{t_{k-1}})^2}{\phi(t_k-t_{k-1})} = K > 0.$$ Then there exists a constant $C>0$ such that $$\sup_{x\in{\mathbb{R}}}\left| {\mathbb{P}}\left(\frac{\tilde{V}_1(\pi_n)}{\sqrt{Var(\tilde{V}_1(\pi_n))}} < x\right) - {\mathbb{P}}(Z<x)\right| \leq C\sqrt{N(\pi_n)}H(|\pi_n|),$$ where $Z$ is a standard normal random variable. In particular, if $H(|\pi_n|) = o\left(N(\pi_n)^{-\frac12}\right)$, then $$\frac{\tilde{V}_1(\pi_n)}{\sqrt{Var(\tilde{V}_1(\pi_n))}} \to \mathcal{N}(0,1).$$ [ In [@malukas] the author studied a particular class of Gaussian processes while here we consider arbitrary Gaussian process. Similarly, in [@malukas] the main result was derived by using some technical computations under assumption together with several additional technical assumptions. Here we have shown that is the only needed condition which generalises the class of processes considerably. Similarly, we have been able to simplify the proof since we have shown that such results follows essentially from Gaussian concentration together with some matrix algebra. ]{} Next we will provide some sufficient conditions which are easy to check. A particularly interesting case for us is Gaussian processes such that the function $$d(s,t)= {\mathbb{E}}(X_t-X_s)^2$$ is $C^{1,1}$ outside diagonal. Note that a sufficient condition for this is that the variance ${\mathbb{E}}X_t^2$ is $C^1$ and the covariance $R$ of $X$ is $C^{1,1}$ outside diagonal. Furthermore, note that this assumption is satisfied for many cases of interest. The first theorem gives a general result in a case of bounded derivative. \[thm:1st\_order\_bounded\] Let $X$ be a continuous Gaussian such that the function $d(s,t)={\mathbb{E}}(X_t-X_s)^2$ is $C^{1,1}$ outside diagonal. Furthermore, assume that there exists a positive function $f(s,t)$ such that $$|\partial_{st}d(s,t)| \leq f(s,t)$$ and $$\sup_{s,v\in[0,T]}\int_0^v f(s,t){\mathrm{d}}t < \infty.$$ Assume there exists a function $H(|\pi_n|)$ such that $$\max_{1\leq j\leq N(\pi_n)-1}\frac{d(t_j,t_{j-1})+\Delta t_j\sup_{1\leq k,j\leq N(\pi_n)-1}\frac{\sqrt{\phi(\Delta t_j)}}{\sqrt{\phi(\Delta t_k)}}}{\phi(\Delta t_j)} \leq H(|\pi_n|).$$ Then the result of Theorem \[thm:1st\_order\_main\] holds with function $H(|\pi_n|)$. Now for $j\neq k$ we have $${\mathbb{E}}[(X_{t_k}-X_{t_{k-1}})(X_{t_j}-X_{t_{j-1}})] = \int_{t_{k-1}}^{t_k}\int_{t_{j-1}}^{t_j}\partial_{st}d(s,t){\mathrm{d}}s{\mathrm{d}}t.$$ Hence we have $$\begin{split} &\sum_{k=1}^{N(\pi_n)-1} \left|{\mathbb{E}}\left[Y^{(n)}_k Y^{(n)}_j\right]\right|\\ &\leq\frac{d(t_j,t_{j-1})}{\phi(\Delta t_j)} + \frac{1}{\sqrt{\phi(\Delta t_j)}}\sum_{k=1, k\neq j}^{N(\pi_n)-1}\frac{1}{\sqrt{\phi(\Delta t_k)}}\int_{t_{k-1}}^{t_k}\int_{t_{j-1}}^{t_j}\left|\partial_{st}d(s,t)\right|{\mathrm{d}}s{\mathrm{d}}t\\ &\leq \frac{d(t_j,t_{j-1})}{\phi(\Delta t_j)} +\frac{1}{\phi(\Delta t_j)}\sum_{k=1, k\neq j}^{N(\pi_n)-1}\frac{\sqrt{\phi(\Delta t_j)}}{\sqrt{\phi(\Delta t_k)}}\int_{t_{k-1}}^{t_k}\int_{t_{j-1}}^{t_j}\left|\partial_{st}d(s,t)\right|{\mathrm{d}}s{\mathrm{d}}t\\ &\leq \frac{d(t_j,t_{j-1})}{\phi(\Delta t_j)} +\frac{\sup_{1\leq k,j\leq N(\pi_n)-1}\frac{\sqrt{\phi(\Delta t_j)}}{\sqrt{\phi(\Delta t_k)}}}{\phi(\Delta t_j)}\sum_{k=1, k\neq j}^{N(\pi_n)-1}\int_{t_{k-1}}^{t_k}\int_{t_{j-1}}^{t_j}\left|\partial_{st}d(s,t)\right|{\mathrm{d}}s{\mathrm{d}}t. \end{split}$$ Now here $$\begin{split} &\sum_{k=1,k\neq j}^{N(\pi_n)-1}\int_{t_{k-1}}^{t_k}\int_{t_{j-1}}^{t_j}\left|\partial_{st}d(s,t)\right|{\mathrm{d}}s{\mathrm{d}}t\\ &= \int_{0}^{t_{j-1}}\int_{t_{j-1}}^{t_j}\left|\partial_{st}d(s,t)\right|{\mathrm{d}}s{\mathrm{d}}t + \int_{t_{j}}^{T}\int_{t_{j-1}}^{t_j}\left|\partial_{st}d(s,t)\right|{\mathrm{d}}s{\mathrm{d}}t\\ &= O\left( \Delta t_j\right) \end{split}$$ by Tonelli’s theorem and assumptions. The claim follows at once. [ The convergence depends now on $\sup_{1\leq k,j\leq N(\pi_n)-1}\frac{\sqrt{\phi(\Delta t_j)}}{\sqrt{\phi(\Delta t_k)}}$. However, in any natural chosen sequence partition we have $\sup_{n\geq 1} k(\pi_n)<\infty$ which guarantees $\sup_{1\leq k,j\leq N(\pi_n)-1}\frac{\sqrt{\phi(\Delta t_j)}}{\sqrt{\phi(\Delta t_k)}} < \infty$ for many functions $\phi$. For example, this is obviously true for power functions $\phi(x) = x^\gamma$ which is a natural choice in many cases. ]{} As an immediate corollary we obtain the following which again seems to generalise the existing results in the literature. Most importantly, the following result shows how the lower bound for the variance plays a fundamental role. Furthermore, a standard assumption in the literature is that $d(t,s) \sim |t-s|^\gamma$ for some number $\gamma\in(0,2)$ which in particular covers the fractional Brownian motion and related processes. In the following the structure of the variance can be more complicated. For simplicity we will only present the result in the case of bounded function $k(\pi_n)$ while the general case follows similarly. \[cor:1st\_order\_bounded\] Let the notation and assumptions above prevail. Furthermore, assume that there exists a function $r$ such that $d(t,s) \sim r(t-s)$ as $|t-s| \to 0$. Let $\sup_{n\geq 1}\sup_{1\leq k,j\leq N(\pi_n)-1}\frac{\sqrt{r(\Delta t_j)}}{\sqrt{r(\Delta t_k)}} < \infty$, $\sup_{n\geq 1} k(\pi_n)<\infty$, and put $\phi(\Delta t_j) = \frac{r(\Delta t_j)}{\Delta t_j}$. 1. If $(|\pi_n|)^2 =o\left(r(|\pi_n|)\right)$, then $$|\tilde{V}_1(\pi_n,\phi)| \to 0$$ in probability and in $L^p$ for any $p\geq 1$. The convergence holds almost surely for any sequence $\frac{(|\pi_n|)^2}{r(|\pi_n|)} = o\left(\frac{1}{\log n}\right)$. 2. There is a constant $C>0$ such that $$\sup_{x\in{\mathbb{R}}}\left| {\mathbb{P}}\left(\frac{\tilde{V}_1(\pi_n)}{\sqrt{Var(\tilde{V}_1(\pi_n))}} < x\right) - {\mathbb{P}}(Z<x)\right| \leq C\frac{|\pi_n|^{\frac32}}{r(|\pi_n|)},$$ where $Z$ is a standard normal random variable. In particular, if $(|\pi_n|)^{\frac32} =o\left(r(|\pi_n|)\right)$, then $$\frac{\tilde{V}_1(\pi_n)}{\sqrt{Var(\tilde{V}_1(\pi_n))}} \to \mathcal{N}(0,1).$$ The claim follows immediately from Theorem \[thm:1st\_order\_bounded\] by noting that our choice of function $\phi$ guarantees condition . We end this section with a following result that recovers the case of fractional Brownian motion and related processes. Note again that our technical assumptions are quite minimal. \[thm:1st\_order\_fbm\] Let $X$ be a continuous Gaussian such that the function $d(s,t)={\mathbb{E}}(X_t-X_s)^2$ is in $C^{1,1}$ outside diagonal. Furthermore, assume that $$|\partial_{st}d(s,t)| =O\left(|t-s|^{2H-2}\right)$$ for some $H\in(0,1)$, $H\neq \frac12$ and assume there exists a function $H(|\pi_n|)$ such that $$\max_{1\leq j\leq N(\pi_n)-1}\frac{d(t_j,t_{j-1})+[\Delta t_j]^{\min(1,2H)}\sup_{1\leq k,j\leq N(\pi_n)-1}\frac{\sqrt{\phi(\Delta t_j)}}{\sqrt{\phi(\Delta t_k)}}}{\phi(\Delta t_j)} \leq H(|\pi_n|).$$ Then the result of Theorem \[thm:1st\_order\_main\] holds with function $H(|\pi_n|)$. Note that the case $H>\frac12$ follows directly from Theorem \[thm:1st\_order\_bounded\]. Let now $H<\frac{1}{2}$. Using Fubini’s Theorem we have $$\begin{split} &\int_{t_j}^{T}\int_{t_{j-1}}^{t_j}\left|\partial_{st}d(s,t)\right|{\mathrm{d}}s{\mathrm{d}}t\\ &\leq C\int_{t_{j-1}}^{t_j} (T-s)^{2H-1}{\mathrm{d}}s + C\int_{t_{j-1}}^{t_j} (t_j-s)^{2H-1}{\mathrm{d}}s\\ &\leq C [\Delta t_j]^{2H} \end{split}$$ for some unimportant constants $C$ which vary from line to line. Here we have used the fact that for positive numbers $a,b$ and $\gamma \in (0,1)$ we have $|a^\gamma - b^\gamma| \leq |a-b|^\gamma$. Treating integral $\int_{0}^{t_{j-1}}\int_{t_{j-1}}^{t_j}\left|\partial_{st}d(s,t)\right|{\mathrm{d}}s{\mathrm{d}}t$ similarly the result follows. [ \[rem:case\_half\] We remark that the case $H=\frac12$ can be treated similarly. In this case one obtains a condition $$\max_{1\leq j\leq N(\pi_n)-1}\frac{d(t_j,t_{j-1})+\Delta t_j |\log \Delta t_j|\sup_{1\leq k,j\leq N(\pi_n)-1}\frac{\sqrt{\phi(\Delta t_j)}}{\sqrt{\phi(\Delta t_k)}}}{\phi(\Delta t_j)} \leq H(|\pi_n|).$$ ]{} [ It is straightforward to give a version of Corollary \[cor:1st\_order\_bounded\] in this case also. Indeed, the only difference is slightly different exponents in the case $H<\frac12$. ]{} Second order quadratic variations --------------------------------- In this subsection we briefly study second order quadratic variations. In particular, we reproduce and generalise the results presented in papers [@begyn-ejp] and [@begyn-ber]. We will present our results in slightly different form. However, comparison is provided in remark \[rem:comparison\]. Usually second order quadratic variation on $[0,1]$ is defined as the limit of $\sum_{k=1}^n \left(X_{\frac{k+1}{n}} - 2X_{\frac{k}{n}} +X_{\frac{k-1}{n}}\right)$. To generalise for irregular subdivisions Begyn introduced and motivated [@begyn-ejp] a second order differences along a sequence $\pi_n$ as $$\Delta X_k = \Delta t_{k+1} X_{t_{k-1}} + \Delta t_{k} X_{t_{k+1}} - \left(\Delta t_{k+1} + \Delta t_k\right)X_{t_k},$$ and we study the second order quadratic variation defined as the limit $$V_2(\pi_n) = \sum_{k=1}^{N(\pi_n) -1} \frac{\Delta t_{k+1} \left(\Delta X_k\right)^2}{{\mathbb{E}}\left(\Delta X_k\right)^2}.$$ Again we use short notation $$\tilde{V}_2(\pi_n) = V_2(\pi_n) - {\mathbb{E}}V_2(\pi_n).$$ We also assume that the derivative $\frac{\partial^4}{\partial^2 s\partial^2 t}R(s,t)$ of the covariance function $R$ of $X$ exists outside diagonal and satisfies $$\label{eq:2nd_order_partial} \left|\frac{\partial^4}{\partial^2 s\partial^2 t}R(s,t)\right| \leq \frac{C}{|t-s|^{2+\gamma}}$$ for some number $\gamma \in(0,2)$. Finally, we make the simplifying assumption $\sup_n k(\pi_n) < \infty$ on the function $k$. Hence it is also natural to assume $$\label{eq:2nd_order_var_bound} \sup_{j,k}\frac{{\mathbb{E}}\left(\Delta X_k\right)^2}{{\mathbb{E}}\left(\Delta X_k\right)^2} < \infty.$$ In particular, the assumptions made in [@begyn-ejp] implied $$\label{eq:2nd_order_var} {\mathbb{E}}\left(\Delta X_k\right)^2 \sim \left(\Delta t_{k+1}\right)^{\frac{3-\gamma}{2}}\left(\Delta t_{k}\right)^{\frac{3-\gamma}{2}}\left(\Delta t_{k+1}+\Delta t_{k}\right)$$ in which case is clearly satisfied. Let all the notation and assumptions above prevail. 1. If $H(\pi_n):=\max_{1\leq j\leq N(\pi_n)-1}\frac{|\pi_n|^{5-\gamma}}{{\mathbb{E}}\left(\Delta X_k\right)^2}$ converges to zero, then $$\left|\tilde{V}_2(\pi_n)\right| \to 0$$ in probability and in $L^p$ for any $p\geq 1$. Furthermore, the convergence holds almost surely provided that $H(\pi_n) =o\left(\frac{1}{\log n}\right)$. 2. There is a constant $C>0$ such that $$\sup_{x\in{\mathbb{R}}}\left| {\mathbb{P}}\left(\frac{\tilde{V}_2(\pi_n)}{\sqrt{Var(\tilde{V}_2(\pi_n))}} < x\right) - {\mathbb{P}}(Z<x)\right| \leq C\sqrt{N(\pi_n)}H(|\pi_n|),$$ where $Z$ is a standard normal random variable. In particular, if $H(\pi_n) =o\left(\left[N(\pi_n)\right]^{-\frac12}\right)$, then $$\frac{\tilde{V}_2(\pi_n)}{\sqrt{Var\left(\tilde{V}_2(\pi_n)\right)}} \to \mathcal{N}(0,1).$$ Denote $$Y^{(n)}_k = \frac{\sqrt{\Delta t_k}\Delta X_k}{\sqrt{{\mathbb{E}}\left(\Delta X_k\right)^2}}.$$ For some constant $C$ we have $$\sum_{k=1}^{N(\pi_n)-1} \left|{\mathbb{E}}\left[Y^{(n)}_kY^{(n)}_j\right]\right| \leq C \frac{|\pi_n|}{{\mathbb{E}}\left(\Delta X_j\right)^2} \sum_{k=1}^{N(\pi_n)-1} \left|{\mathbb{E}}\left[\Delta X_k \Delta X_j\right]\right|$$ by and boundedness of $k(\pi_n)$. Furthermore, it was proved in [@begyn-ejp] that boundedness of $k(\pi_n)$ together with yields $$\sum_{k=1}^{N(\pi_n)-1} \left|{\mathbb{E}}\left[\Delta X_k \Delta X_j\right]\right| \leq C|\pi_n|^{4-\gamma}$$ for a constant $C$. This gives a bound for $$\sum_{k=1}^{N(\pi_n)-1} \left|{\mathbb{E}}\left[Y^{(n)}_kY^{(n)}_j\right]\right|$$ from which the result follows immediately by combining Theorems \[thm:QV-conv-general\] and \[thm:CLT\_main\]. [ \[rem:comparison\] To compare our result with the ones provided in papers [@begyn-ejp] and [@begyn-ber], first note that we were able to reproduce and generalise the main theorem of [@begyn-ejp] although we gave our result in a slightly different form. Indeed, in [@begyn-ejp] several additional technical conditions were assumed to ensure the asymptotic relation while here we have worked with general variance. This is helpful since the message of our result is that basically one has to only check the asymptotic behaviour of the variance, and quarantees the upper bound for the corresponding matrix norm. Furthermore, the central limit theorem in [@begyn-ber] was proved only for uniformly divided partition, and the central limit theorem was proved under more restrictive technical conditions, similar to those in [@begyn-ejp], and by finding a lower bound for the variance in order to apply Lindeberg’s CLT. Here we have proved that such result holds also non-uniform partitions and the result follows easily from the computations presented in [@begyn-ejp] together with Theorem \[thm:CLT\_main\]. Finally, here we also obtained Berry-Esseen bound. In particular, under we obtain bound proportional to $\sqrt{|\pi_n|}$. ]{} Remarks on generalised quadratic variations {#subsec:a-qv} ------------------------------------------- We end this section by giving some remarks on generalised quadratic variations introduced by Istas and Lang [@I-L]. Let now $a=(a_0,a_1,\ldots,a_p)$ be a vector such that $\sum_{k=0}^p a_k = 0$, where $p$ is a fixed integer. Let also $\Delta$ be a fixed small number and consider time points $t_k = k\Delta,\quad k=1,\ldots, n$. The $a$-differences of $X$ is given by $$\Delta_a X_j = \sum_{k=0}^p a_k X_{t_{j+k}}, \quad j=0,1,\ldots,n - p.$$ In [@I-L] the authors considered stationary or stationary increment Gaussian processes and studied generalised $a$-variations defined as a limit of $$V(a,n) = \frac{1}{n}\sum_{k=1}^{n-p} \frac{(\Delta_a X_k)^2}{{\mathbb{E}}[\Delta_a X_k]^2}.$$ Now since $X$ is either stationary or has stationary increments, the function $d(t,s)$ depends only on the difference $|t-s|$. The assumption in [@I-L] was that the function $v(t) = d(0,t)$ is $2D$ times differentiable ($D$ is the greatest such integer, possibly 0), and for some number $\gamma\in(0,2)$ and some constant $C>0$ we have $$v^{(2D)}(t) = v^{(2D)}(0) + Ct^{\gamma} + r(t),$$ where the remainder $r$ satisfies $r(t) = o\left(t^{\gamma}\right)$. The main results in [@I-L] was that under different set of assumptions one can obtain Gaussian limit with some rate with a suitable choice of the vector $a$, although in some cases one has to assume the observation window $n\Delta$ to increase to infinity. Obviously, by using the result of this paper we could reproduce and generalise, at least to cover more general variances as in here for first and second kind quadratic variations, these results together with a much simplified proofs. Instead of getting lost into technical details we wish to give some remarks and explanations. In [@I-L] the main message was roughly that larger the value of $D$ and $s$, then larger one has to choose the value $p$, i.e. taking account more refined discretisation. However, the reason for this can, and at least was for the present author, be lost in the technicalities. In a nutshell, the idea is to try to find a discretisation vector $a$ so that $$\max_{1\leq j\leq n}\frac{1}{n^2\sqrt{{\mathbb{E}}[\Delta_a X_j]^2}}\sum_{k=1}^{n-p}\frac{1}{\sqrt{{\mathbb{E}}[\Delta_a X_k]^2}}\left|{\mathbb{E}}\left[(\Delta_a X_k)(\Delta_a X_j)\right]\right| = o\left(n^{-\frac12}\right)$$ from which we obtain almost sure convergence and central limit theorem (with $\sqrt{Var\left(V(a,n)-{\mathbb{E}}V(a,n)\right)}$ as normalisation so one is left to analyse the asymptotic of this variance). Hence it remains to answer how one should choose the vector $a$. The idea for this comes little bit clearer from the first order variation and Corollary \[cor:1st\_order\_bounded\]. Indeed, as also pointed out in [@I-L], the number $D$ is the order of differentiability of $X$ in the $L^2$-sense. Hence if $D\geq 1$, the variance must be at least of order $(\Delta t_j)^2$ so there is no hope to obtain even convergence in probability. Hence larger the $D$, larger the value of $p$ should also be. Similarly, as $\gamma$ becomes close to number $2$ it roughly means that $D$ comes closer to $1$ so once again one needs to refine the discretisation to obtain a Gaussian limit. More precisely, as $\gamma$ comes closer to $2$ we see immediately that the variance is no longer enough to compensate $|\pi_n|^{\frac32}$ in order to obtain central limit theorem. Hence one has to consider second order quadratic variations in order to stay in a Gaussian paradise. Examples {#sec:ex} ======== This section is devoted to examples. We focus to reproduce some interesting and already studied examples to illustrate the power of our method rather than finding a complicated new examples, while now already it should be obvious how our approach can be used to study more complicated cases. More precisely, we study Brownian motion, fractional Brownian motion and related processes together with extensions sub-fractional Brownian motion and bifractional Brownian motion. Furthermore, our main focus is on first order quadratic variation for which we find sufficient conditions for the mesh to obtain almost sure convergence. In this context particularly interesting case for us is bifractional Brownian motion for which we are able to improve the sufficient condition proved in [@malukas]. For simplicity we assume that the function $k(\pi_n)$ is bounded. Standard Brownian motion {#subsec:bm} ------------------------ Let $X=W$ be a standard Brownian motion. Then it is known that the almost sure convergence holds provided that $|\pi_n| = o\left(\frac{1}{\log n}\right)$ (for recent results on the topic see [@le-er]). Furthermore, this is a sharp in a sense that one can construct a sequence with $|\pi_n| = O\left(\frac{1}{\log n}\right)$ such that almost sure convergence does not hold. Now the sufficiency of $|\pi_n| = o\left(\frac{1}{\log n}\right)$ follows easily from concentration inequality applied to the increments of Brownian motion. Indeed, in the case of standard Brownian motion the covariance matrix $\Gamma^{(n)}$ of the increments is diagonal, and we have $$\Vert \Gamma^{(n)}\Vert_1 = \Vert \Gamma^{(n)} \Vert_2 = |\pi_n|.$$ Note also that if one uses Frobenius norm $\Vert\Gamma^{(n)}\Vert_F$ to obtain the upper bound, we have $\Vert \Gamma^{(n)}\Vert_F = \sqrt{|\pi_n|}$ provided that $\frac{|\pi_n|}{m(\pi_n)} \leq C$. Hence by using Frobenius norm we can only obtain half of the best possible rate even in the case of standard Brownian motion. Finally, it is straightforward to obtain central limit theorem which, of course, is well-known already. Fractional Brownian motion -------------------------- Recall that a fractional Brownian motion $B^H$ with Hurst index $H\in(0,1)$ is a continuous centred Gaussian process with covariance function $$R(t,s) = \frac{1}{2}\left(t^{2H}+s^{2H}-|t-s|^{2H}\right).$$ The case $H=\frac12$ reduces to a standard Brownian motion. To obtain $L^p$-convergence of general $\alpha$-variations is straightforward by using Remark \[rem:alpha-var\]. Let $B^H$ be a fractional Brownian motion with $H\in(0,1)$. Then there exists a constant $C_H$ such that for $\alpha=\frac{1}{H}$ we have $$\sum_{t^n_k\in \pi_n} |B_{t_k}^H - B_{t_{k-1}}^H|^\alpha \rightarrow C_H T$$ in $L^p$ for any $p\geq 1$. [ The exact value of the constant $C_H$ is given by $C_H = {\mathbb{E}}|N|^\frac{1}{H}$, where $N$ is a standard normal random variable. ]{} We now turn to the convergence of quadratic variation which is more interesting for us. Now it is natural to take $\phi(x) = x^{2H-1}$, since for any partition of $[0,T]$ we have $$\sum_{t^n_k\in \pi_n} \frac{{\mathbb{E}}(B_{t_k}^H - B_{t_{k-1}}^H)^2}{[\Delta t_k]^{2H-1}} = T.$$ The following result is a direct consequence of Theorem \[thm:1st\_order\_fbm\]. \[pro:fbm\] Let $B^H$ be a fractional Brownian motion with $H\in(0,1)$. Then 1. $$V_n^B:=\sum_{k=1}^{N(\pi_n)-1} \frac{(B_{t_k}^H - B_{t_{k-1}}^H)^2}{[\Delta t_k]^{2H-1}} \to T$$ in probability and in $L^p$ for any $p\geq 1$. Furthermore, the convergence holds almost surely for any sequence of partitions satisfying $|\pi_n| =o\left(\frac{1}{(\log n)^\gamma}\right)$, where $\gamma = \max\left(\frac{1}{2-2H},1\right)$. 2. There exists a constant $C>0$ such that $$\sup_{x\in{\mathbb{R}}}\left| {\mathbb{P}}\left(\frac{V_n^B - T}{\sqrt{Var(V_n^B - T)}} < x\right) - {\mathbb{P}}(Z<x)\right| \leq C|\pi_n|^{\min\left(\frac12,\frac32-2H\right)},$$ where $Z$ is a standard normal random variable. In particular, central limit theorem holds for all values $H<\frac{3}{4}$. [ Note that in the case $H<\frac12$ we obtain similar sufficient condition as for standard Brownian motion. Indeed, the only difference is that since here the increments are not independent, we have to pose an additional assumption $\sup_{n\geq 1} k(\pi_n)<\infty$ to obtain the optimal condition $o\left(\frac{1}{\log n}\right)$. ]{} [ By considering uniform partitions it can be shown that via concentration inequalities one cannot obtain any better result. It would be interesting to know whether the given conditions are optimal similarly as in the case of Brownian motion. However, for Brownian motion the counter-examples are constructed relying on independence of increments and, to the best of our knowledge, there exists no method to attack the problem for general Gaussian process. ]{} [ The limit theorems for quadratic variations of fractional Brownian motion are extensively studied in the literature. However, most of the related studies rely on uniform partitions and focus on generalisations, e.g. to study Hermite variations or weighted variations rather than generalising the sequence of partitions. Furthermore, to recover the central limit theorem in the case $H<\frac34$ our approach is based only on simple linear algebra. For this reason our approach may be more applicable to generalise the results for arbitrary Gaussian processes while the obvious drawback is that it cannot provide a full answer to the problem. Finally, to the best of our knowledge the non-uniform partition are not widely studied in the literature. ]{} [ It is known that in the critical case $H=\frac34$ we also have convergence towards a normal random variable (see, e.g. [@nou-rev] and references therein) with the only difference in the rate, i.e. the variance is of order $\frac{\log n}{n}$ instead of $\frac{1}{n}$. Given a priori knowledge that the variance is of order $\frac{\log n}{n}$ it is straightforward to recover also this critical case by Corollary \[cor:lindeberg\]. Hence again, it is sufficient to study the asymptotic behaviour of the variance. ]{} sub-fractional Brownian motion ------------------------------ The sub-fractional Brownian motion $G^H$ with parameter $H\in(0,1)$ is a centred Gaussian process with covariance function $$R(s,t) = s^{2H} + t^{2H}-\frac{1}{2}\left[s^{2H}+t^{2H}-|t-s|^{2H}\right].$$ Note that as for fractional Brownian motion, value $H=\frac12$ corresponds to a standard Brownian motion. Let $G^H$ be a sub-fractional Brownian motion with $H\in(0,1)$. Then 1. $$V_n^G:=\sum_{t^n_k\in \pi_n} \frac{\left(G_{t_k}^H - G_{t_{k-1}}^H\right)^2}{[\Delta t_k]^{2H-1}}\to T$$ in probability and in $L^p$ for any $p\geq 1$. Furthermore, the convergence holds almost surely for any sequence satisfying $|\pi_n| =o\left(\frac{1}{(\log n)^\gamma}\right)$, where $\gamma = \max\left(\frac{1}{2-2H},1\right)$. 2. There exists a constant $C>0$ such that $$\sup_{x\in{\mathbb{R}}}\left| {\mathbb{P}}\left(\frac{V_n^G - T}{\sqrt{Var(V_n^G - T)}} < x\right) - {\mathbb{P}}(Z<x)\right| \leq C|\pi_n|^{\min\left(\frac12,\frac32-2H\right)},$$ where $Z$ is a standard normal random variable. In particular, central limit theorem holds for all values $H<\frac{3}{4}$. Note that the case $H=\frac12$ is already covered since it reduces back to the standard Brownian motion. Hence assume $H\neq \frac12$ and let $t>s$. We have $$\partial_{st}d_G(t,s) \leq C (t+s)^{2H-2} +C|t-s|^{2H-2}$$ for some constant $C$ and $d_G(s,t) \leq C|t-s|^{2H}$. Now $ (t+s)^{2H-2} \leq |t-s|^{2H-2} $ from which the result follows immediately as in the case of fractional Brownian motion. [ We remark that the above result was already given in [@malukas] with the same rates although there the condition for the case $H=\frac12$ was $|\pi_n||\log|\pi_n|| = o\left(\frac{1}{\log n}\right)$ which would follow from Remark \[rem:case\_half\]. Obviously however, in this case we have standard Brownian motion so that $|\pi_n| = o\left(\frac{1}{\log n}\right)$ is sufficient. Note also that in this case one cannot obtain better via concentration inequalities. Indeed, this comes from the ”fractional Brownian part” $|t-s|^{2H}$. ]{} Bifractional Brownian motion ---------------------------- Particularly interesting case for us is Bifractional Brownian motion which was also studied in [@malukas]. However, with our method we are able to improve the results of [@malukas]. The bifractional Brownian motion is an extension of fractional Brownian motion first introduced by [@h-v] and later analysed e.g. by [@ru-tu]. Finally, bifractional Brownian motion was extended for values $K\in(1,2)$ such that $HK\in(0,1)$ by [@bar-seb]. The bifractional Brownian motion is a centred Gaussian process $B^{H,K}$ with $B^{H,K}_0=0$ and covariance function $$R(t,s) = \frac{1}{2^K}\left((t^{2H}+s^{2H})^K-|t-s|^{2HK}\right)$$ with $H\in(0,1)$ and $K\in(0,2)$ such that $HK\in(0,1)$. [ Note that $K=1$ corresponds to ordinary fractional Brownian motion. It is straightforward to check that $B^{H,K}$ is $HK$-self-similar and Hölder continuous of any order $\gamma<HK$. For more details on the properties of bifractional Brownian motion we refer to [@ru-tu] and references therein. ]{} While the main emphasis in [@ru-tu] was integration via regularisation it was pointed out that one can prove that $\alpha$-variations exists as a limit in $L^1$ in our sense. Hence the following result is obvious from remark \[rem:alpha-var\]. Let $B^{H,K}$ be a bifractional Brownian motion with $H\in(0,1)$ and $K\in(0,2)$ such that $HK\in(0,1)$. Then there exists a constant $C_{H,K}$ such that for $\alpha=\frac{1}{HK}$ we have $$\sum_{t^n_k\in \pi_n} \left|B_{t_k}^{H,K} - B_{t_{k-1}}^{H,K}\right|^\alpha \rightarrow C_{H,K} T$$ in $L^p$ for any $p\geq 1$. [ In [@ru-tu] the authors considered only the case $K\in(0,1]$. However, it is straightforward to obtain the claim for the case $K>1$ by repeating the arguments. ]{} The next theorem studies the quadratic variation of bifractional Brownian motion. Let $B^{H,K}$ be a bifractional Brownian motion with $H\in(0,1), K\in(0,2)$ and $HK\in(0,1)$. Then 1. $$V_n^{H,K}:= \sum_{t^n_k\in \pi_n} \frac{\left(B_{t_k}^{H,K} - B_{t_{k-1}}^{H,K}\right)^2}{[\Delta t_k]^{2HK-1}} \to 2^{1-K}T$$ in probability and in $L^p$ for any $p\geq 1$. Furthermore, the convergence holds almost surely for any sequence of partitions satisfying $|\pi_n| =o\left(\frac{1}{(\log n)^\gamma}\right)$, where; - $\gamma = \max\left(\frac{1}{2-2HK},1\right)$ for $K\in(0,1]$, - $\gamma = \frac{1}{\min(1,2H)+1-2HK}$ for $K\in(1,2)$. 2. In the case $K\in(0,1]$ we have $$\sup_{x\in{\mathbb{R}}}\left| {\mathbb{P}}\left(\frac{V_n^{H,K} - 2^{1-K}T}{\sqrt{Var(V_n^{H,K} - 2^{1-K}T)}} < x\right) - {\mathbb{P}}(Z<x)\right| \leq C|\pi_n|^{\min\left(\frac12,\frac32-2HK\right)}$$ for some constant $C$ and a standard normal random variable $Z$. In particular, central limit theorem holds for all values $HK<\frac{3}{4}$. 3. In the case $K\in(1,2)$ we have $$\sup_{x\in{\mathbb{R}}}\left|{\mathbb{P}}\left(\frac{V_n^{H,K} - 2^{1-K}T}{\sqrt{Var(V_n^{H,K} - 2^{1-K}T)}} < x\right) - {\mathbb{P}}(Z<x)\right|\leq C|\pi_n|^{\min\left(1,2H\right)-2HK+\frac12}.$$ We assume $K\neq 1$ since the case $K=1$ reduces to ordinary fractional Brownian motion treated in Proposition \[pro:fbm\]. The function $d(s,t) = {\mathbb{E}}[B_t - B_s]^2$ is differentiable outside diagonal and we have $$\partial_{st}d(s,t) = C_1 |t-s|^{2HK-2} + C_2 \frac{(ts)^{2H-1}}{(t^{2H}+s^{2H})^{2-K}}$$ for some unimportant constants $C_1$ and $C_2$. Furthermore, we have $$d(s,t) \sim |t-s|^{2HK}$$ as $|t-s|\to 0$ which also corresponds to fractional Brownian motion. Now the term $|t-s|^{2HK-2}$ can be treated as in Theorem \[thm:1st\_order\_fbm\], and for $HK=\frac12$ this term vanishes. Consider next the term $$\frac{(ts)^{2H-1}}{(t^{2H}+s^{2H})^{2-K}}.$$ We are left to bound integrals $$I_1 = \int_0^{t_{j-1}}\int_{t_{j-1}}^{t_j} \frac{(ts)^{2H-1}}{(t^{2H}+s^{2H})^{2-K}}{\mathrm{d}}s {\mathrm{d}}t$$ and $$I_2 = \int_{t_{j}}^{T}\int_{t_{j-1}}^{t_j} \frac{(ts)^{2H-1}}{(t^{2H}+s^{2H})^{2-K}}{\mathrm{d}}s {\mathrm{d}}t.$$ We consider only $I_1$ since $I_2$ can be treated similarly, and we denote by $C$ any unimportant constant which may vary from line to line. By change of variables $x = t^{2H}$, $y=s^{2H}$ and Tonelli’s theorem we have $$\begin{split} I_1 &= C\int_0^{t_{j-1}^{2H}}\int_{t_{j-1}^{2H}}^{t_j^{2H}} \frac{(xy)^{\frac{1}{2H}(2H-1)}}{(x+y)^{2-K}}x^{\frac{1}{2H}-1}y^{\frac{1}{2H}-1}{\mathrm{d}}x {\mathrm{d}}y\\ &=C\int_0^{t_{j-1}^{2H}}\int_{t_{j-1}^{2H}}^{t_j^{2H}}(x+y)^{K-2}{\mathrm{d}}x {\mathrm{d}}y\\ &=C\int_{t_{j-1}^{2H}}^{t_j^{2H}} \int_0^{t_{j-1}^{2H}}(x+y)^{K-2}{\mathrm{d}}x {\mathrm{d}}y\\ &=C\int_{t_{j-1}^{2H}}^{t_j^{2H}} \int_0^{t_{j-1}^{2H}}(t_{j-1}^{2H}+y)^{K-1}{\mathrm{d}}x {\mathrm{d}}y\\ &-C\int_{t_{j-1}^{2H}}^{t_j^{2H}} y^{K-1} {\mathrm{d}}y \end{split}$$ Now for $K>1$ we have $y^{K-1} \leq C$ which leads to $$\int_{t_{j-1}^{2H}}^{t_j^{2H}} y^{K-1} {\mathrm{d}}y \leq C(t_j^{2H} - t_{j-1}^{2H}) \leq C|\pi_n|^{\min(1,2H)}$$ by the fact that for $T\geq a>b\geq 0$ and $\gamma\in(0,1)$ we have $a^{\gamma}-b^{\gamma} \leq (a-b)^{\gamma}$ and for $\gamma\geq 1$ we have $a^{\gamma}-b^{\gamma} \leq C(a-b)$ by the mean value theorem. Similarly, for $K<1$ we have $$\int_{t_{j-1}^{2H}}^{t_j^{2H}} y^{K-1} {\mathrm{d}}y \leq t_{j}^{2HK} - t_{j-1}^{2HK} \leq |\pi_n|^{\min(1,2HK)}.$$ Treating other integrals similarly the result follows by Theorem \[thm:1st\_order\_fbm\] with $\phi(x) =x^{2HK-1}$. [ It may look like that for case $K>1$ one gets better (i.e. larger exponent) by computing $$\int_{t_{j-1}^{2H}}^{t_j^{2H}} y^{K-1} {\mathrm{d}}y = C(t_j^{2HK} - t_{j-1}^{2HK}) \leq C|\pi_n|^{\min(1,2HK)}.$$ However, this analysis cannot be used to cover, e.g. integral $\int_{t_{j-1}^{2H}}^{t_j^{2H}} (T+y)^{K-1} {\mathrm{d}}y$. ]{} [ To compare our results with the existing literature, in [@malukas] it was proved that almost sure convergence holds, in our notation, for value $\gamma=\frac{1}{1-HK}$ for the whole range $H\in(0,1)$ and $K\in(0,1]$. Note that by putting $K=1$ and $H=\frac12$ we have a standard Brownian motion, and this produces half of the best possible rate. Now in our result we have value $\gamma =\frac{1}{2-2HK}$ which is twice better compared to $\frac{1}{1-HK}$. Furthermore, we obtained even better for the range $2HK>1$. Note also that, to the best of our knowledge, the case $K>1$ is not studied in the literature before the present paper. ]{} [ Particularly interesting case is $HK=\frac12$. In this case the quadratic variation exists in the ordinary sense which allows one to develop stochastic calculus with respect to $B^{H,K}$ although $B^{H,K}$ is not a semimartingale [@ru-tu] if $K\in(0,1)$. Now for this process we obtain similar condition to the one of standard Brownian motion. On the other hand, if $K\in(1,2)$ and $HK=\frac12$, then the process $B^{H,K}$ is a semimartingale. However, in this case we only obtain condition $|\pi_n|^{2H}=o\left(\frac{1}{\log n}\right)$ since $K>1$ and $HK=\frac12$ implies $H<\frac12$. ]{} [ In the case $K\in(0,1]$ we obtain a sufficient condition $HK<\frac34$ for the central limit theorem to hold which of course is not at all surprising. Similarly, in the case $K\in(1,2)$ we obtain sufficient condition $HK<\frac34$ provided that $H>\frac12$. However, in the case $H<\frac12$ something odd seems to happen. Indeed, if $HK\geq\frac34$, then $2H+\frac12 - 2HK \leq 0$ so that the given Berry-Esseen bound does not converge to zero. On the other hand, even if $HK<\frac34$ it is not necessarily true that $2H+\frac12 - 2HK>0$ so that condition $HK<\frac34$ is no longer sufficient. Indeed, even in the semimartingale case $2HK=1$ we have $2H+\frac12 -2HK\leq 0$ for values $H\in\left(0,\frac14\right]$. ]{} **Acknowledgements** Lauri Viitasaari was partially funded by Emil Aaltonen Foundation. Proof of Lemma \[lma:2nd\_4th\] =============================== A simple application of yields $${\mathbb{E}}V_n^2 = \sum_{k,j=1}^n \left({\mathbb{E}}\left[Y^{(n)}_kY^{(n)}_j\right]\right)^2.$$ Next we compute ${\mathbb{E}}V_n^4$. We have $$\label{eq:product} V_n^4 = \sum_{i,j,k,l=1}^n \prod_{p\in\{i,j,k,l\}}\left[\left(Y^{(n)}_p\right)^2 - {\mathbb{E}}\left(Y^{(n)}_p\right)^2\right].$$ Recall next that all information of a Gaussian vector is encoded to the covariance matrix $\Gamma^{(n)}$ so that $k$-moments of a centred Gaussian vector $(Y_1,Y_2,\ldots,Y_n)$ can be computed via formula $${\mathbb{E}}[Y_1^{k_1}Y_2^{k_2}\ldots Y_n^{k_n}] = \sum_{\sigma}{\mathbb{E}}[Y_{\sigma(1)}Y_{\sigma(2)}]\ldots {\mathbb{E}}[Y_{\sigma(n-1)}Y_{\sigma(n)}],$$ where the summation is over all permutations $\sigma$ of numbers $\{1,2,\ldots,n\}$, hence producing $n!$ terms. Applying this to vector 8-dimensional vector $\left(Y_k^{(n)},Y_k^{(n)},Y_j^{(n)},\ldots, Y_l^{(n)}\right)$ and taking expectation on $$\left[\left(Y^{(n)}_k\right)^2\left(Y^{(n)}_i\right)^2\left(Y^{(n)}_j\right)^2\left(Y^{(n)}_l\right)^2\right]$$ we obtain terms of form $$A_1(\sigma) = \left[{\mathbb{E}}\left[Y^{(n)}_{\sigma(1)}Y^{(n)}_{\sigma(2)}\right]\right]^2\left[{\mathbb{E}}\left[Y^{(n)}_{\sigma(3)}Y^{(n)}_{\sigma(4)}\right]\right]^2,$$ $$A_2(\sigma) = {\mathbb{E}}\left[Y^{(n)}_{\sigma(1)}\right]^2{\mathbb{E}}\left[Y^{(n)}_{\sigma(2)}\right]^2\left[{\mathbb{E}}\left[Y^{(n)}_{\sigma(3)}Y^{(n)}_{\sigma(4)}\right]\right]^2,$$ $$A_3(\sigma) = {\mathbb{E}}\left[Y^{(n)}_{\sigma(1)}\right]^2{\mathbb{E}}\left[Y^{(n)}_{\sigma(2)}\right]^2{\mathbb{E}}\left[Y^{(n)}_{\sigma(3)}\right]^2{\mathbb{E}}\left[Y^{(n)}_{\sigma(4)}\right]^2,$$ $$A_4(\sigma) = {\mathbb{E}}\left[Y^{(n)}_{\sigma(1)}Y^{(n)}_{\sigma(2)}\right]{\mathbb{E}}\left[Y^{(n)}_{\sigma(2)}Y^{(n)}_{\sigma(3)}\right]{\mathbb{E}}\left[Y^{(n)}_{\sigma(3)}Y^{(n)}_{\sigma(1)}\right]{\mathbb{E}}\left[Y^{(n)}_{\sigma(4)}\right]^2,$$ and $$A_5(\sigma) = {\mathbb{E}}\left[Y^{(n)}_{\sigma(1)}Y^{(n)}_{\sigma(2)}\right]{\mathbb{E}}\left[Y^{(n)}_{\sigma(2)}Y^{(n)}_{\sigma(3)}\right]{\mathbb{E}}\left[Y^{(n)}_{\sigma(3)}Y^{(n)}_{\sigma(4)}\right]{\mathbb{E}}\left[Y^{(n)}_{\sigma(4)}Y^{(n)}_{\sigma(1)}\right],$$ where $\sigma=(\sigma(1),\sigma(2),\sigma(3),\sigma(4))$ can be any permutation of indices $\{i,j,k,l\}$. Next we note by symmetry of covariance and summing over symmetric set $\{1\leq i,j,k,l\leq n\}$ we obtain that for each $p=1,\ldots, 5$ and any permutation $\sigma$ we have $ \sum_{i,j,k,l=1}^n A_p(\sigma) = \sum_{i,j,k,l=1}^n A_p(\sigma_0), $ where $\sigma_0$ is any fixed permutation. For example, we obviously have $$\sum_{i,j,k,l=1}^n \left[{\mathbb{E}}\left[Y^{(n)}_{i}Y^{(n)}_{j}\right]\right]^2\left[{\mathbb{E}}\left[Y^{(n)}_{k}Y^{(n)}_{l}\right]\right]^2 = \sum_{i,j,k,l=1}^n \left[{\mathbb{E}}\left[Y^{(n)}_{i}Y^{(n)}_{k}\right]\right]^2\left[{\mathbb{E}}\left[Y^{(n)}_{j}Y^{(n)}_{l}\right]\right]^2.$$ Consequently we obtain $$\sum_{i,j,k,l=1}^n {\mathbb{E}}\left[\left(Y^{(n)}_k\right)^2\left(Y^{(n)}_i\right)^2\left(Y^{(n)}_j\right)^2\left(Y^{(n)}_l\right)^2\right]\\ = \sum_{i,j,k,l=1}^n \sum_{p=1}^5 a_p A_p(\sigma_0)$$ for arbitrary reference permutation $\sigma_0$ and some weights $a = (a_1,\ldots,a_5)$. Note also that the weights $a_p,p=1,\ldots,5$ are independent of indices $i,j,k,l$ and the underlying Gaussian process. Now treating rest of the terms in $\prod_{p\in\{i,j,k,l\}}\left[\left(Y^{(n)}_p\right)^2 - {\mathbb{E}}\left(Y^{(n)}_p\right)^2\right]$ similarly we conclude that $${\mathbb{E}}V_n^4 = \sum_{i,j,k,l=1}^n \sum_{p=1}^5 b_p A_p(\sigma_0)$$ with some weights $b=(b_1,\ldots,b_5)$ independent of $i,j,k,l$ and the underlying Gaussian process. Next we claim that $b=(12,0,0,0,24)$. Of course the weight vector $b$ could be computed exactly via combinatorial arguments but we wish to use a more subtle argument by relying on the classical central limit theorem for a sequence of independent standard normal random variables. We begin by computing the values $b_4$ and $b_5$ which are relatively easy to compute directly. First note that terms $A_5$ are produced only by the term ${\mathbb{E}}\left[\left(Y^{(n)}_k\right)^2\left(Y^{(n)}_i\right)^2\left(Y^{(n)}_j\right)^2\left(Y^{(n)}_l\right)^2\right]$. Furthermore, terms of form $A_5$ are produced by permutations of indices $\{i,j,k,l\}$ which gives $b_5=4! = 24$. Consider next $b_4$. Terms of form $A_4$ are produced from ${\mathbb{E}}\left[\left(Y^{(n)}_k\right)^2\left(Y^{(n)}_i\right)^2\left(Y^{(n)}_j\right)^2\left(Y^{(n)}_l\right)^2\right]$ by first picking one variable, $Y^{(n)}_k$ say, to get ${\mathbb{E}}\left[Y^{(n)}_k\right]^2$ and then organising the remaining three into $3!=6$ ways which produces $4!=24$ (the first one can be picked in 4 ways). On the other hand, computing the product we obtain 4 terms of form ${\mathbb{E}}\left[\left(Y^{(n)}_k\right)^2\right]\left(Y^{(n)}_i\right)^2\left(Y^{(n)}_j\right)^2\left(Y^{(n)}_l\right)^2$ and with similar analysis we obtain that each term produces $A_4$ exactly $3!=6$ times. Due to the minus sign in terms $-{\mathbb{E}}\left[\left(Y^{(n)}_k\right)^2\right]$ and the fact $24-4\times 6 = 0$ we obtain $b_4=0$. It remains to prove that $b_1 = 12$ and $b_2=b_3 = 0$. For this purpose let $Y^{(n)}_k$ be a sequence of independent standard normal random variables $Y_k$. Then by the classical central limit theorem we have $$S_n := \frac{1}{\sqrt{2n}}\sum_{k=1}^n [Y_k^2 - {\mathbb{E}}Y_k^2] \to \mathcal{N}(0,1)$$ in distribution and consequently, ${\mathbb{E}}S_n^4 \to 3$. In this case we have $$b_1 \sum_{i,j,k,l}^n A_1(\sigma_0) = b_1n^2$$ $$b_2 \sum_{i,j,k,l=1}^n A_3(\sigma_0) = b_2n^3$$ and $$b_4 \sum_{i,j,k,l=1}^n A_2(\sigma_0) = b_3n^4$$ so that $${\mathbb{E}}S_n^4 = \frac{1}{4n^2}\left[b_1 n^2 + b_2n^3 + b_3n^4\right].$$ Now since $b_k,k=1,2,3$ is independent of $n$ and the underlying Gaussian process, the convergence ${\mathbb{E}}S_n^4 \to 3$ implies $b_2=b_3=0$ and $b_1 = 12$. This completes the proof.

ArXiv

--- abstract: 'We perform an empirical evaluation of several methods of low-rank approximation in the problem of obtaining PMI-based word embeddings. All word vectors were trained on parts of a large corpus extracted from English Wikipedia (enwik9) which was divided into two equal-sized datasets, from which PMI matrices were obtained. A repeated measures design was used in assigning a method of low-rank approximation (SVD, NMF, QR) and a dimensionality of the vectors (250, 500) to each of the PMI matrix replicates. Our experiments show that word vectors obtained from the truncated SVD achieve the best performance on two downstream tasks, similarity and analogy, compare to the other two low-rank approximation methods.' author: - Alena Sorokina - Aidana Karipbayeva - Zhenisbek Assylbekov bibliography: - 'ref.bib' title: ' Low-Rank Approximation of Matrices for PMI-based Word Embeddings' --- Introduction ============ Today word embeddings play an important role in many natural language processing tasks, from predictive language models and machine translation to image annotation and question answering, where they are usually ‘plugged in’ to a larger model. An understanding of their properties is of interest as it may allow the development of better performing embeddings and improved interpretability of models using them. One of the widely-used word embedding models is the Skip-gram with negative sampling (SGNS) of [@mikolov2013distributed ([-@mikolov2013distributed])]{}. [@levy2014neural ([-@levy2014neural])]{} showed that the SGNS is implicitly factorizing a pointwise mutual information (PMI) matrix shifted by a global constant. They also showed that a low-rank approximation of the PMI matrix by truncated singular-value decomposition (SVD) can produce word vectors that are comparable to those of SGNS. However, truncated SVD is not the only way of finding a low-rank approximation of a matrix. It is optimal in the sense that it minimizes the approximation error in the Frobenius and the 2- norms, but this does not mean that it produces optimal word embeddings, which are usually evaluated in downstream NLP tasks. The question is: Is there any other method of low-rank matrix approximation that produces word embeddings better than the truncated SVD factorization? Our experiments show that the truncated SVD is actually a strong baseline which we failed to beat by another two widely-used low-rank approximation methods. Low-Rank Approximations of the PMI-matrix ========================================== The simplest version of a PMI matrix is a symmetric matrix with each row and column indexed by words[^1], and with elements defined as $$\mathrm{PMI}(i,j)=\log\frac{p(i,j)}{p(i)p(j)},\label{pmi}$$ where $p(i, j)$ is the probability that the words $i$, $j$ appear within a window of a certain size in a large corpus, and $p(i)$ is the unigram probability for the word $i$. For computational purposes, [@levy2014neural ([-@levy2014neural])]{} suggest using a positive PMI (PPMI), defined as $$\mathrm{PPMI}(i,j)=\max(\mathrm{PMI}(i,j),0).\label{ppmi}$$ They also show empirically that the low-rank SVD of the PPMI produces word vectors which are comparable in quality to those of the SGNS. The low-rank matrix approximation is approximating a matrix by one whose rank is less than that of the original matrix. The goal of this is to obtain a more compact representation of the data with a limited loss of information. In what follows we give a brief overview of the low-rank approximation methods used in our work. Since both PMI and PPMI are square matrices, we will consider approximation of square matrices. For a thorough and up-to-date review of low-rank approximation methods see the paper by [@kishore2017literature ([-@kishore2017literature])]{}. ### Singular Value Decomposition (SVD) factorizes $\mathbf{A}\in\mathbb{R}^{n\times n}$, into the matrices $\mathbf{U}\in\mathbb{R}^{n\times n}$, $\mathbf{S}\in\mathbb{R}^{n\times n}$ and $\mathbf{V}^\top\in\mathbb{R}^{n\times n}$: $$\mathbf{A}=\mathbf{USV}^\top,$$ where $\mathbf{U}$ and $\mathbf{V}$ are orthogonal matrices, and $\mathbf{S}$ is a rectangular diagonal matrix whose entries are in descending order, $\sigma_1\ge\sigma_2\ge\cdots\ge\sigma_n\ge0$, along the main diagonal, and are known as the [singular values]{} of $\mathbf{A}$. The rank $d$ approximation (also called *truncated* or *partial SVD*) of $\mathbf{A}$, $\mathbf{A}_d$ where $d<\operatorname{rank}\mathbf{A}$, is given by zeroing out the $n-d$ trailing singular values of $\mathbf{A}$, that is[^2] $$\mathbf{A}_d=\mathbf{U}_{1:n,1:d}\mathbf{S}_{1:d,1:d}\mathbf{V}^\top_{1:d,1:n}.$$ By the Eckart-Young theorem [(@eckart1936approximation, [-@eckart1936approximation])]{}, $A_d$ is the closest rank-$d$ matrix to $A$ in Frobenius norm, i.e. $\|\mathbf{A}-\mathbf{A}_d\|_F\le\|\mathbf{A}-\mathbf{B}\|_F$, $\forall\mathbf{B}\in\mathbb{R}^{n\times n}:\,\operatorname{rank}(\mathbf{B})=d$. [@levy2014neural ([-@levy2014neural])]{} suggest factorizing the PPMI matrix with truncated SVD, and then taking the rows of $\mathbf{U}_{1:n,1:d}\mathbf{S}_{1:d,1:d}^{1/2}$ as word vectors, and we follow their approach. ### QR decomposition with column pivoting of $\mathbf{A}\in\mathbb{R}^{n\times n}$ has the form $\mathbf{A P}=\mathbf{Q R}$, where $\mathbf{Q}\in\mathbb{R}^{n\times n}$ is orthogonal, $\mathbf{R}\in\mathbb{R}^{n\times n}$ is upper triangular and $\mathbf{P}\in\mathbb{R}^{n\times n}$ is a permutation matrix. The rank $d$ approximation to $\mathbf{A}$ is then $$\mathbf{A}_d=\mathbf{Q}_{1:n,1:d}[\mathbf{RP}^\top]_{1:d,1:n}$$ which is called *truncated QR decomposition* of $\mathbf{A}$. After factorizing the PPMI matrix with this method we suggest taking the rows of $\mathbf{Q}_{1:n,1:d}$ as word vectors. However, we suspect that a valuable information could be left in the $\mathbf{R}$ matrix. A promising alternative to SVD is a Rank Reveling QR decomposition (RRQR). Assume the QR factorization of the matrix $\mathbf{A}$: $$\mathbf{AP}= \mathbf{Q}\left[ {\begin{array}{cc} \mathbf{R}_{11} & \mathbf{R}_{12} \\ 0 & \mathbf{R}_{22} \\ \end{array} } \right]$$ where $\mathbf{R}_{11}\in\mathbb{R}^{d\times d}$, $\mathbf{R}_{12}\in\mathbb{R}^{d\times (n-d)}$, $\mathbf{R}_{22}\in\mathbb{R}^{(n-d)\times (n-d)}$. For RRQR factorization, the following condition should be satisfied: $$\begin{aligned} \sigma_{\min}(\mathbf{R}_{11}) &= \mathrm{\Theta}(\sigma_{d}(\mathbf{A}))\\ \sigma_{\max}(\mathbf{R}_{22}) &= \mathrm{\Theta}(\sigma_{d+1}(\mathbf{A}))\end{aligned}$$ which suggests that the most significant entries are in $\mathbf{R}_{11}$, and the least important are in $\mathbf{R}_{22}$. Thus, we also suggest taking the columns of $[\mathbf{RP}^\top]_{1:d,1:n}$ as word vectors. ### Non negative matrix factorization (NMF). Given a non negative matrix $\mathbf{A}\in\mathbb{R}^{n\times n}$ and a positive integer $d<n$, NMF finds non negative matrices $\mathbf{W}\in\mathbb{R}^{n\times d}$ and $\mathbf{H}\in\mathbb{R}^{d\times n}$ which minimize (locally) the functional $f(\mathbf{W},\mathbf{H})=\|\mathbf{A}-\mathbf{WH}\|^2_F$. The rank $d$ approximation of $\mathbf{A}$ is simply $$\mathbf{A}_d=\mathbf{WH}.$$ When factorizing the PPMI matrix with NMF, we suggest taking the rows of $\mathbf{W}$ as word vectors. Experimental Setup ================== Corpus ------ All word vectors were trained on the `enwik9` dataset[^3] which was divided into two equal-sized splits. The PMI matrices on these splits were obtained using the `hypewords` tool of [@levy2015improving ([-@levy2015improving])]{}. All corpora were pre-processed by removing non-textual elements, sentence splitting, and tokenization. PMI matrices were derived using a window of two tokens to each side of the focus word, ignoring words that appeared less than 300 times in the corpus, resulting in vocabulary sizes of roughly 13000 for both words and contexts. A repeated measures design was used for assigning the method of factorization (SVD, QR, NMF) and dimensionality of the vectors (250, 500) to each PMI matrix replicate. We used two replicates per each level combination. Training -------- Low-rank approximations were performed using the following open-source implementations: - Sparse SVD from SciPy [(@jones2014scipy, [-@jones2014scipy])]{}, - Sparse RRQR from SuiteSparse [(@davis2011university, [-@davis2011university])]{}, and - NMF from scikit-learn [(@pedregosa2011scikit, [-@pedregosa2011scikit])]{}. For NMF we used the nonnegative double SVD initialization. We trained 250 and 500 dimensional word vectors with each method. Evaluation ---------- We evaluate word vectors on two tasks: similarity and analogy. A similarity is tested using the WordSim353 dataset of [@finkelstein2002placing ([-@finkelstein2002placing])]{}, containing word pairs with human-assigned similarity scores. Each word pair is ranked by cosine similarity and the evaluation is the Spearman correlation between those rankings and human ratings. Analogies are tested using Mixed dataset of 19544 questions such as “$a$ is to $b$ as $c$ is to $d$”, where $d$ is hidden and must be guessed from the entire vocabulary. We filter questions with out of vocabulary words, as standard. Accuracy is computed by comparing $\operatorname*{arg\,min}_{d}\|\mathbf{b} - \mathbf{a} + \mathbf{c} - \mathbf{d}\|$ to the labelled answer. Results ======= The results of evaluation are provided in Table \[tab1\], which we analyze using the two-factor ANOVA with factors being (1) low-rank approximation method and (2) dimensinality of word vectors, and response being the performance in similarity or analogy task. We analyze the tasks separately. -------------------- ---------------- ----------- ------------ --------- Method of low- Dimensionality Replicate Similarity Analogy rank approximation of vectors \# task task SVD 250 1 0.7010 0.3778 SVD 250 2 0.6969 0.3817 SVD 500 1 0.6989 0.3568 SVD 500 2 0.6914 0.3458 NMF 250 1 0.5265 0.0660 NMF 250 2 0.4780 0.0563 NMF 500 1 0.4499 0.0486 NMF 500 2 0.3769 0.0487 QR (R) 250 1 0.4077 0.1644 QR (R) 250 2 0.3822 0.1533 QR (R) 500 1 0.4717 0.2284 QR (R) 500 2 0.3719 0.1925 QR (Q) 250 1 0.2870 0.0034 QR (Q) 250 2 0.2009 0.0059 QR (Q) 500 1 0.3573 0.0165 QR (Q) 500 2 0.3048 0.0186 -------------------- ---------------- ----------- ------------ --------- : Results[]{data-label="tab1"} ![Test scores for different factorization methods on Similarity and Analogy tasks.[]{data-label="residuals1"}](research/sim_graph_cicling.pdf "fig:"){width=".49\textwidth"} ![Test scores for different factorization methods on Similarity and Analogy tasks.[]{data-label="residuals1"}](research/analog_graph_cicling.pdf "fig:"){width=".49\textwidth"} Similarity task --------------- ![ANOVA residuals for the results on Similarity task.[]{data-label="residuals1"}](research/similar_with_Q.png){height=".45\textheight"} The standard residual analysis is used to check whether the ANOVA assumptions are satisfied. From Figure \[residuals1\] we see that the residuals have constant variability around zero, are independent and normally distributed. The normality is confirmed using Shapiro-Wilk test, $p\text{-value} = 0.7923$. Source DF Sum of Squares Mean Square F Value Pr ${>}$ F ----------------- ---------- ---------------- ------------- ------------ ------------- Model 7 0.37055017 0.05293574 29.68 ${<}$ .0001 Error 8 0.01426728 0.00178341 Corrected total 15 0.38481745 R-Square Coeff Var Root MSE Score Mean 0.962925 9.126819 0.042230 0.462707 : ANOVA table for the similarity task results[]{data-label="tab2"} Source DF Sum of Squares Mean Square F Value Pr ${>}$ F --------------- ---- ---------------- ------------- --------- ------------- Factorization 3 0.35433596 0.11811199 66.23 ${<}$ .0001 Dimension 1 0.00011159 0.00011159 0.06 0.8088 Interaction 3 0.01610263 0.00536754 3.01 0.0945 : Main and Interaction Effects in the Similarity task[]{data-label="tab3"} ![ANOVA Residuals for the Analogy task results[]{data-label="fig2"}](research/analog_with_Q.png){height=".45\textheight"} Source DF Sum of Squares Mean Square F Value Pr ${>}$ F ----------------- ---------- ---------------- ------------- ------------ ------------- Model 7 0.30745304 0.04392186 424.61 ${<}$ .0001 Error 8 0.00082753 0.00010344 Corrected total 15 0.30828057 R-Square Coeff Var Root MSE Score Mean 0.997316 6.602449 0.010171 0.154043 : ANOVA Table for the Analogy task results[]{data-label="tab4"} Source DF Sum of Squares Mean Square F Value Pr ${>}$ F --------------- ---- ---------------- ------------- --------- ------------- Factorization 3 0.30365768 0.10121923 978.52 ${<}$ .0001 Dimension 1 0.00013820 0.00013820 1.34 0.2811 Interaction 3 0.00365715 0.00121905 11.78 0.0026 : Main and Interaction Effects in the Analogy task[]{data-label="tab5"} The SAS package was used to obtain ANOVA table (Table \[tab2\]), which shows the effects of the factors on the similarity score. F-test for equality of the factor level means was conducted, $F = 29.68$ and $\text{p-value}<0.0001$. Hence, it can be concluded that at least one factor level mean is different from the others. ${R}^2 = 0.962925$ shows that more than 96% of variation in the similarity score is explained by the factors considered. Proceeding with analysis of main and interaction effects, one can conduct F-test for each of the factors and the interaction between them. From Table \[tab3\], we see that the method of low-rank approximation affects the performance of words vectors in the similarity task, $F = 66.23$, p-value $< 0.0001$. The dimensionality of word vectors has no effect on the performance in the similarity task, $F = 0.06$ with $\text{p-value} > 0.8$. Also, there is no interaction between the method of factorization and the dimensionality of word vectors, $F = 3.01$ with p-value $0.0945$. Thus, SVD significantly outperforms the other factorization methods. Analogy task ------------ Again, we first need to check whether the ANOVA assumptions are satisfied. From Figure \[fig2\] we see that the residuals have constant variability around zero, are independent and normally distributed. The normality is confirmed using Shapiro-Wilk test, $\text{p-value} = 0.112$. The ANOVA Table (Table  \[tab4\]) shows that at least one level mean is different from the others. ${R}^2$ is 0.997316, thus, 99% of variation in the analogy score is explained by the considered factors. We proceed to the analysis of main and interaction effects. The method of low-rank approximation affects the performance of word vectors in the analogy task, $F = 978.52$ with $\text{p-value} < 0.0001$. The dimensionality of word vectors has no effect on the performance in the analogy task, $F = 1.34$ with $\text{p-value} > 0.2$. Unlike the similarity task, there is an interaction effect between the two factors, $F = 11.78$ with $\text{p-value} = 0.0026$. Discussion ========== **Why dimensionality is critical in similarity task for NMF?** We obtained the highest results in the similarity task using the SVD-based low-rank approximation, for which the dimensionality of word vectors did not influence the performance much. On the contrary, the performance in similarity task using the NMF method of factorization is significantly affected by the dimension of the word vector: 250-dimensional word vectors give significantly better results than 500-dimensional ones. This can be explained by the specific characteristics of the NMF method of factorization. When we look at the word vectors produced by NMF, we can see that they contain many zeros. Hence, an increase in the dimensionality makes them even sparser. Similarity task is based on finding the cosine of the angle between two word vectors. Therefore, when the vectors become sparser, the result of element-wise multiplication, which is necessary for obtaining cosine, becomes smaller. Thus, there is a much higher possibility that the cosine similarity score between two vectors, containing many zeros, will give a number closer to zero than to 1. This, as a result, leads to the worse performance in the similarity task. Our suggestion is to decrease the dimensionality of the NMF method to 100. We expect that this may give better results. **Why NMF performs poorly in the analogy task?** We provide a theoretical analysis of the poor performance of the NMF in the analogy task. We model word vectors produced by the NMF as independent and identically distributed random vectors from an isotropic multivariate Gaussian distribution $\mathcal{N}(\mathbf{4.5},\mathbf{I})$[^4], since for a 500-dimensional $\mathbf{v}\sim\mathcal{N}(\mathbf{4.5},\mathbf{I})$ there is a big chance that it is nonnegative: $$\Pr(\mathbf{v}\in[0,+\infty)^{500})=[\Pr(4.5+Z>0)]^{500}\approx0.9983,$$ where $Z\sim\mathcal{N}(0,1)$ is a standard normal random variable. For a triplet of word vectors $\mathbf{a}$, $\mathbf{b}$ and $\mathbf{c}$ we have $\mathbf{b}-\mathbf{a}+\mathbf{c}\sim\mathcal{N}(\mathbf{4.5},3\mathbf{I})$, and therefore $$\begin{aligned} &\Pr(\mathbf{b}-\mathbf{a}+\mathbf{c}\in[0,+\infty)^d)=[\Pr(3+\sqrt{3}Z\ge0)]^d\\ &=[\Pr(Z\ge-4.5/\sqrt{3})]^d<[0.9953]^d.\end{aligned}$$ When $d=500$, this probability is $\approx0.1$, i.e. there is a small chance that $\mathbf{b}-\mathbf{a}+\mathbf{c}$ is non negative, and thus we will likely not find a non-negative $\mathbf{d}$ when we minimize $\|\mathbf{b}-\mathbf{a}+\mathbf{c}-\mathbf{d}\|$. This is confirmed empirically: for *all* word triplets $(a,b,c)$ from the analogy task, the vector $\mathbf{b}-\mathbf{a}+\mathbf{c}$ has at least one negative component. **Why using $\mathbf{R}$ is better than using $\mathbf{Q}$ in the QR decomposition?** The $\mathbf{Q}$ matrix from QR factorization gives the worst results in the similarity task, and it does not depend on the dimensionality of the vector. The reason is that the necessary information is left in the $\mathbf{R}$ matrix. Truncation of $\mathbf{RP}^\top$ gives better approximation to the original matrix than the truncated $\mathbf{Q}$, because the most significant entries of $\mathbf{RP}^\top$ are in the top left quarter and remain after truncation. Conclusion ========== We analyzed the performance of the word vectors obtained from a word-word PMI matrix by different low-rank approximation methods. As it was expected, the truncated SVD provides a far better solution than the NMF and the truncated QR in both similarity and analogy tasks. While the performance of the NMF is relatively good in the similarity task, it is significantly worse in the analogy task. NMF produces only non-negative sparse vectors and we showed how this deteriorates the performance in both tasks. $\mathbf{RP}^\top$ matrix from QR factorization with column pivoting gives better word embedding than $\mathbf{Q}$ matrix in both tasks. Acknowledgement {#acknowledgement .unnumbered} =============== The work of Zhenisbek Assylbekov has been funded by the Committee of Science of the Ministry of Education and Science of the Republic of Kazakhstan, contract \# 346/018-2018/33-28, IRN AP05133700. [^1]: Assume that words have already been converted into integer indices. [^2]: $\mathbf{A}_{a:b,c:d}$ is a submatrix located at the intersection of rows $a, a+1, \ldots, b$ and columns $c, c + 1, \ldots, d$ of a matrix $\mathbf{A}$. [^3]: <http://mattmahoney.net/dc/textdata.html> [^4]: The isotropy is motivated by the work of [@arora2016latent ([-@arora2016latent])]{}; $\mathbf{4.5}$ is a vector with all elements equal to $4.5$.

ArXiv

--- abstract: 'In this paper a robust algorithm for DOA estimation of coherent sources in presence of antenna array imperfections is presented. We exploit the current advances of deep learning to overcome two of the most common problems facing the state of the art DOA algorithms (i.e. coherent sources and array imperfections). We propose a deep auto encoder (AE) that is able to correctly resolve coherent sources without the need of spatial smoothing, hence avoiding possible processing overhead and delays. Moreover, we assumed the presence of array imperfections in the received signal model such as mutual coupling, gain/ phase mismatches, and position errors. The deep AE is trained using the covariance matrix of the received signal, where it alleviates the effect of imperfections, and at the same time act as a filters for the coherent sources. The results show significant improvement compared to the methods used in the literature.' author: - bibliography: - 'conf.bib' title: 'Deep Autoencoders for DOA Estimation of Coherent Sources using Imperfect Antenna Array\' --- DOA, Coherent sources, antenna array imperfections, Deep autoencoders, MUSIC Introduction ============ Direction-of-arrival (DOA) estimation is a common problem in different fields, including wireless communication, astronomical observation, radar and sonar. There are various challenges facing DOA estimation such as accuracy and precision of estimates in non-ideal scenarios, e.g., multipaths and antennas imperfections. There are many DOA estimation techniques that can be classified to multiple categories. On one hand, there are the conventional methods which depends on the locations of peaks in the spatial spectrum to determine the DOA, i.e., delay and sum beamforming and minimum variance distortionless response (MVDR) [@mvdr]. On the other hand there are also the subspace methods e.g., MUltiple SIgnal Classification (MUSIC) [@b7] which depends on the eigenstructure of the spatial correlation matrix, offering high resolution DOA estimation. There also exists parametric-based approach like the maximum likelihood (ML) technique, which uses a statistical structure for the process of generating data [@b3]. However, the application of those methods in real life applications is very limited, as it requires the accurate knowledge of the received signal without any errors. Hence, they are limited to specific scenarios where the antenna array measurements are ideal and the sources are not correlated. For example, the MUSIC algorithm has the ability to detect and measure multiple sources at the same time with high precision measurement only under ideal array conditions in a non multipath environment. However its performance degrades significantly in the existence of array imperfections or when the sources are coherent (fully correlated) due to multi-path. In such scenario, the rank of the signal covariance matrix is less than the number of impinging signals, which significantly affects the estimation performance. To solve this problem, spatial smoothing techniques have been widely used to remove the coherence between sources or de-correlate their signals by reconstructing a full rank covariance matrix before going through the estimation algorithm [@SSMUSIC]. However, this adds processing overhead, making it difficult to use in real time applications. Moreover, spatial smoothing only solves the coherence problem, and does not solve other problems like antenna array imperfections which is a common problem in practice. As a matter of fact, those imperfections generally occur in practical applications due to the non-idealities in the antenna array such as mutual coupling, gain / phase gradual changes over time and as well as changes in the antenna locations. For instance, mutual coupling occurs due to interference from nearby antennas during transmitting [@fried], while gain and phase inconsistency can result from the aging of electrical components or thermal effects. All of those factors change the antenna array response, causing significant degradation in the performance of DOA algorithms like MUSIC [@fried]. Many approaches in the literature addressed those problems using auto calibration algorithms as in [@MCsol1; @MCsol2]. However, such algorithms require the prior knowledge of the imperfections formulations, which can be challenging in practical applications. Other approaches used the latest advances in machine learning, and deep learning to solve this problem as in [@MLpaper]. However, the authors in [@MLpaper] only focused on the imperfections problem, where they proposed a neural network followed by a classifier. However, they didn’t take into consideration potential correlation among the sources that can result from multipath.\ In this paper we propose a deep neural network framework for DOA estimation that is generic and robust against both array imperfections, along with the existence of coherent sources. The results show that our DNN can correctly estimate the directions of spatially close coherent sources without any prior spatial smoothing techniques. Furthermore, our algorithm showed significantly better results compared to the state of the art methods for coherence removal as spatial smoothing MUSIC algorithm [@SSMUSIC], and to similar algorithms for imperfections as aforementioned approach found in [@MLpaper]. System Model ============ Array Imperfections ------------------- Assume that $K$ sources are impinging on a receive array consisting of $M$ antennas, whose DOA are $\theta_1,\ldots,\theta_K$. The received signal is sampled at $N$ uniquely spaced time instants $t_1,\ldots,t_N$ to obtain multiple snapshots gathered in matrix $\mathbf{Z} = [\mathbf{z}(t_1),\ldots,\mathbf{z}(t_N)]$, with $$\boldsymbol{z}\left(t_{n}\right)=\sum_{k=1}^{K} \boldsymbol{a}\left(\theta_{k}\right) x_{k}\left(t_{n}\right)+\boldsymbol{w}\left(t_{n}\right), \quad \text { for } n=1, \ldots, N . \label{received_signal}$$ $x_k(t_n)$ is the transmit waveform of the $k$ th source, and $\boldsymbol{w}\left(t_{n}\right)$ is the zero-mean Gaussian noise. $\boldsymbol{a}\left(\theta_{k}\right)$ denotes the imperfection free steering vector and is defined as $$\boldsymbol{a}\left(\theta_{k}\right)=\frac{1}{\sqrt{M}}\left[1, e^{-j 2 \pi \frac{d}{\lambda} \sin \theta_{k}}, \cdots, e^{-j 2 \pi \frac{d}{\lambda}\left(M-1\right) \sin \theta_{k}}\right]^{T}$$ The covariance matrix of the received signal $\boldsymbol{z}$ is $$\label{covmatrix} {\boldsymbol{R}}=\frac{1}{N} \sum_{n=1}^{N} \boldsymbol{z}\left(t_{n}\right) \boldsymbol{z}^{H}\left(t_{n}\right).$$ The model in is the idealistic received signal without any imperfections in the antenna array, which is commonly used in the literature. However, it is quite impractical. Hence, we re-define as $$\label{errRx} \boldsymbol{z}\left(t_{n}\right)=\sum_{k=1}^{K} \boldsymbol{a}\left(\theta_{k}, \boldsymbol{e}\right) x_{k}\left(t_{n}\right)+\boldsymbol{w}\left(t_{n}\right), \quad n=1, \ldots, N$$ where $\boldsymbol{a}\left(\theta_{k}, \boldsymbol{e}\right)$ is the array response after adding the array imperfections. Here we consider gain and phase errors ($\mathbf{e}_{g}$ and $\mathbf{e}_{\mathrm{p}}$), antenna position error ($\mathbf{e}_{\mathrm{pos}}$), and mutual coupling error ($\mathbf{e}_{\mathrm{mc}})$. To this end, the definition of $\boldsymbol{a}\left(\theta_{k}, \boldsymbol{e}\right)$ would be as [@MLpaper] $$\begin{aligned} \boldsymbol{a}(\theta, \boldsymbol{e})&=\left(\boldsymbol{I}_{M}+\alpha_{\mathrm{mc}} \boldsymbol{E}_{\mathrm{mc}}\right) \times\left(\boldsymbol{I}_{M}+\operatorname{diag}\left(\alpha_{\mathrm{g}} \boldsymbol{e}_{\mathrm{g}}\right)\right) \\ \times & \operatorname{diag}\left(\exp \left(j \alpha_{\mathrm{e}} \boldsymbol{e}_{\mathrm{p}}\right)\right) \times \boldsymbol{a}\left(\theta, \alpha_{\mathrm{pos}} \boldsymbol{e}_{\mathrm{pos}}\right) \end{aligned}\label{steering_imperfections}$$ $$\begin{aligned} \boldsymbol{e}_{\mathrm{g}}&=[0, \underbrace{0.2, \ldots, 0.2}_{\frac{M}{2}}, \underbrace{-0.2, \ldots,-0.2}_{\frac{M}{2}-1}]^{T} \label{e} \end{aligned}$$ $$\boldsymbol{e}_{\mathrm{p}}=\left[0, \underbrace{-\frac{\pi}{6}, \ldots,-\frac{\pi}{6}}_{\frac{M}{2}}, \underbrace{\frac{\pi}{6}, \ldots, \frac{\pi}{6}}_{\frac{M}{2}-1}\right]^{T} \label{ph}$$ The position biases are $$\boldsymbol{e}_{\mathrm{pos}}=[0, \underbrace{-0.2, \ldots,-0.2}_{\frac{M}{2}}, \underbrace{0.2, \ldots, 0.2}_{\frac{M}{2}-1}]^{T} \times d \label{pos}$$ The mutual coupling coefficient vector is $$\boldsymbol{e}_{\mathrm{mc}}=\left[0, \gamma^{1}, \ldots, \gamma^{19}\right]^{T} \label{mc}$$ where $\gamma=0.3 e^{j \frac{\pi}{3}}$ is the mutual coupling coefficient between adjacent antennas, $\alpha_i \in [0,1], i \in \{g,p,\mathrm{pos},\mathrm{mc}\}$ is weighting parameter for each error. The choice of the error values in ,, and is system dependent, and can be changed accordingly. $\boldsymbol{E}_{\mathrm{mc}}$ is defined as a Toeplitz matrix with parameter vector $ \boldsymbol{e}_{\mathrm{mc}}$ [@MLpaper]. In addition to array imperfections, in any real scenario, the received signals would be highly correlated, due to the contribution of multi-paths which makes $\boldsymbol{R}$ rank deficient or singular. Hence, the next section will define the model of coherent sources. Coherent Sources ---------------- In order to generate the multi-path component of each source, we consider having $K$ sources arriving from $K$ directions. At time instant n, there are $k$ transmit signals $x_k(t_n),$ $\forall k=1, \hdots ,K$, which arrive as replica of one of them i.e. $x_1(t_n)$, but phase delayed and magnitude weighted [@SSMUSIC]. Hence, the transmit waveform $x_{k}(t_n)$ in can be redefined as $$x_{k}(t_n)=g_{k} e^{j \phi_{k}} x_1(t_n), \quad k=1, \ldots, K, \label{coherent}$$ where $g_{k}$ is the amplitude factor of source $k$ and $\phi_{k}$ is the phase change of source $k$. Such model will impose rank deficiency on the covariance matrix structure in , causing the existing DOA algorithms to fail accordingly. To solve such problem, in the next section, we propose a deep auto-encoder to remove the effects of both coherence and array imperfections. DOA ESTIMATION BASED ON DEEP LEARNING ===================================== Deep Neural Network Architecture -------------------------------- A deep neural network (DNN) is an artificial neural network (ANN) with multiple layers between the input and output layers. The DNN learns the correct mathematical manipulation to map the input into the output. It has the ability to solve complex nonlinear problems [@b4]. Here, we propose a DNN architecture which is an autoencoder (AE), where the first hidden layer performs the function of an encoder as it reduces the dimension of the input by extracting the main features of the input. Afterwards, the encoding layer is followed by four hidden layers that help in the decoding process by retrieving the information to restore back the input of the AE. Table \[table1\] shows the size of all the hidden layers. The output layer consists of six sub-regions where each region is considered as decoder by itself. Therefore, the AE has six decoders and each decoder retrieves specific information from the input, as shown in Fig. \[autoencoder\_architecture\]. The process of encoding-decoding helps decrease the impact of disturbances in the autoencoder input, through de-noising the input, retrieving only the useful information. The disturbance is our case is mainly due to array imperfections, noise, and coherent sources. =\[circle, draw=blue, minimum size = 8mm\] in [1,...,2]{} at (0, -1.5-1) \[place\] (first\_) [$r_\x$]{}; in [1,...,3]{} (0, -3.9 -0.3) circle (2pt); node at (0, -5\*1.1) \[place\] (first\_n) [$r_n$]{}; in [1]{} at (1.25, -3) \[place\] (second\_)[$\alpha_\x$]{}; in [1,...,3]{} (1.25, -3.4 -0.3) circle (2pt); node at (1.25, -5\*1) \[place\] (second\_m) [$\alpha_a$]{}; in [1,...,2]{} at (2.5, -1.5-1) \[place\] (third\_)[$\beta_\x$]{}; in [1,...,3]{} (2.5, -3.9 -0.3) circle (2pt); node at (2.5, -5\*1.1) \[place\] (third\_m) [$\beta_b$]{}; in [1,...,2]{} at (3.75, -0.9-1.3) \[place\] (forth\_)[$\omega_\x$]{}; in [1,...,3]{} (3.75, -4.1 -0.3) circle (2pt); node at (3.75, -5\*1.17) \[place\] (forth\_m) [$\omega_c$]{}; in [1,...,3]{} at (5, -0.7-1.1) \[place\] (fifth\_)[$\gamma_\x$]{}; in [1,...,3]{} (5, -4.5 -0.3) circle (2pt); node at (5, -5\*1.23) \[place\] (fifth\_m) [$\gamma_d$]{}; in [1,...,3]{} at (6.25, -1.25) \[place\] (sixth\_)[$\mu_\x$]{}; in [1,...,3]{} (6.25, -4.5 -0.3) circle (2pt); node at (6.25, -5\*1.28) \[place\] (sixth\_m) [$\mu_e$]{}; in [1]{} at (7.5, 1-1.25) \[place\] (seventh1\_)[$y^{\boldsymbol{1}}_{\x}$]{}; in [1,...,3]{} (7.5, -0.7-0.3) circle (2pt); at (7.5, -2\*1.15) \[place\] (seventh1\_m) [$y^{\boldsymbol{1}}_{n}$]{}; in [1,...,5]{} (7.5, -3.5-0.3) circle (2pt); in [1]{} at (7.5, -5-1.25) \[place\] (seventh2\_)[$y^{\boldsymbol{6}}_{\x}$]{}; in [1,...,3]{} (7.5, -6.7-0.3) circle (2pt); at (7.5, -6\*1.4) \[place\] (seventh2\_m) [$y^{\boldsymbol{6}}_{n}$]{}; (rect) at (7.5,-7.3) \[draw,thick,minimum width=1.5cm,minimum height=3.4cm,color=red\] ; (rect) at (7.5,-1.2) \[draw,thick,minimum width=1.5cm,minimum height=3.4cm,color=red\] ; iin [1,...,2]{} in [1]{} (first\_i) to (second\_); iin [1,...,2]{} (first\_i) to (second\_m); iin [1]{} (first\_n) to (second\_i); (first\_n) to (second\_m); iin [1]{} in [1,...,2]{} (second\_i) to (third\_); iin [1]{} (second\_i) to (third\_m); iin [1,...,2]{} (second\_m) to (third\_i); (second\_m) to (third\_m); iin [1,...,2]{} in [1,...,2]{} (third\_i) to (forth\_); iin [1,...,2]{} (third\_i) to (forth\_m); iin [1,...,2]{} (third\_m) to (forth\_i); (third\_m) to (forth\_m); iin [1,...,2]{} in [1,...,3]{} (forth\_i) to (fifth\_); iin [1,...,2]{} (forth\_i) to (fifth\_m); iin [1,...,3]{} (forth\_m) to (fifth\_i); (forth\_m) to (fifth\_m); iin [1,...,3]{} in [1,...,3]{} (fifth\_i) to (sixth\_); iin [1,...,3]{} (fifth\_i) to (sixth\_m); iin [1,...,3]{} (fifth\_m) to (sixth\_i); (fifth\_m) to (sixth\_m); iin [1,...,3]{} in [1]{} (sixth\_i) to (seventh1\_); iin [1,...,3]{} (sixth\_i) to (seventh1\_m); (sixth\_m) to (seventh1\_m); (sixth\_m) to (seventh1\_1); iin [1,...,3]{} in [1]{} (sixth\_i) to (seventh2\_); iin [1,...,3]{} (sixth\_i) to (seventh2\_m); (sixth\_m) to (seventh2\_m); (sixth\_m) to (seventh2\_1); Layer Size ---------------- --------------------- Input Layer $n=$ 380 Hidden Layer 1 $a=$ 190 Hidden Layer 2 $b=$ 380 Hidden Layer 3 $c=$ 570 Hidden Layer 4 $d=$ 760 Hidden Layer 5 $e=$ 950 Output Layer $6\times380 = 2280$ : Sizes of all layers[]{data-label="table1"} Learning Scheme --------------- The input of the proposed DNN is the covariance matrix of the received signal as defined by . In order to reduce the dimension of the input layer, we consider only the correlated elements in the covariance matrix. By exploiting the symmetry in the correlation matrix, only the strict upper or strict lower triangular part of the matrix can be considered [@MLpaper]. In this paper, the off-diagonal upper right matrix of the covariance matrix is considered. For example, if we have $M=3$ antennas, the covariance matrix will be $$\mathbf{R}=\left[\begin{array}{lll}{r_{11}} & {r_{12}} & {r_{13}} \\ {r_{21}} & {r_{22}} & {r_{23}} \\ {r_{31}} & {r_{32}} & {r_{33}}\end{array}\right],$$ then the following input is obtained $${\hat{\boldsymbol{r}}}=\left[\begin{array}{lll}{r_{12}} & {r_{13}} & {r_{23}} \end{array}\right]^{T}.$$ Generally, let $\mathbf{R} \in \mathbb{C}^{M \times M}$, then the input vector is ${\hat{\boldsymbol{r}}} \in \mathbb{C}^{M(M-1) / 2}$. Additionally, the input of the DNN must be real valued, hence, ${\hat{\boldsymbol{r}}}$ is converted from complex to real, by concatenating $\operatorname{Real}\left\{\hat{\boldsymbol{r}}^{T}\right\}$ with $\operatorname{Imag}\left\{\hat{\boldsymbol{r}}^{T}\right\}$ to produce $\mathbf{r} \in \mathbb{C}^{n}$, where $n= M(M-1)$. $$\boldsymbol{r}=\left[\operatorname{Real}\left\{\hat{\boldsymbol{r}}^{T}\right\}, \operatorname{Imag}\left\{\hat{\boldsymbol{r}}^{T}\right\}\right]^{T}.$$ The proposed autoencoder in Fig.\[autoencoder\_architecture\] decomposes its input into $6$ spatial subregions, each spatial subregion is a specific range of angles and all subregions are of the same size. To define these subregions, $7$ particular spatial angles were chosen such as $$\theta^{(1)}<\theta^{(1)}<\cdots<\theta^{(7)}$$ with constant gaps such that $\theta^{i+1}-\theta^{i}=\mathrm{constant}$ $\forall i, i=1, \hdots,7$, and each subregion $j$ is defined as $\left[\theta^{\mathrm{j}},\theta^{\mathrm{j+1}}\right]$ where $j = 1, \ldots, 6$. Therefore, if the input vector of the autoencoder $\boldsymbol{r}_k$ is generated using a signal impinging from source $k$ on the antenna array at angle $\Theta_{k}$ within the $j$-th subregion, then the output of the $j$-th decoder will be $\boldsymbol{r}_k$, while the output of the other decoders will be zero as there are no signals impinging from those range of angles. The AE is trained to be able to separate multiple signals transmitted from sources located in different subregions impinging onto the array simultaneously. Hence, it is able to decompose the input vector with components belonging to different subregions and extract the information that belongs to every subregion by retrieving it in the related decoders. Training Process ---------------- The data set is constructed by generating $I$ training samples of the covariance vector $\boldsymbol{r}$ corresponding to single-signal scenarios. The data was generated with random angles that spans all subregions. The output of each decoder is determined based on which subregion the generated $\boldsymbol{r}$ belong to. Alternatively, the decoder can be considered as a spatial filter. This filter extracts the covariance vector information that belongs to a particular subregion. In order to build up the training label of the entire output of the AE, the outputs of the six decoders are concatenated as follows $$\begin{aligned} \boldsymbol {y}=& [\boldsymbol {y}_{1}^{T}, \ldots, \boldsymbol {y}_{6}^{T}]^{T} \nonumber \\=&\left [{{\boldsymbol {0}}, \ldots, {\boldsymbol {0}} , \underbrace{\boldsymbol {r}^{T}(\Theta _{k})}_{j\text{ th subregion}}, {\boldsymbol {0}}, \ldots, {\boldsymbol {0}}} \right]^{T}, \label{training_label}\end{aligned}$$ in which $\boldsymbol {y}_j$ is the output of the $j$ decoder. To train the autoencoder, the squared $l_2$-norm distance between the actual output and the expected one is used as the loss function. That is, $$\boldsymbol {L }(\Theta _{k}) = \frac {1}{2} \|\tilde {\boldsymbol {y}}(\Theta _{k}) \|_{2} ^ {2}$$ where $$\tilde {\boldsymbol {y}}(\Theta _{k}) = \boldsymbol {y}(\Theta _{k}) -\hat {\boldsymbol {y}}(\Theta _{k})$$ and $\hat{\boldsymbol{y}}\left(\Theta_{k}\right)$ is the actual output of the autoencoder when $\boldsymbol{r}\left(\boldsymbol{\Theta}_{k}\right)$ is the input. The optimizer used in the training process to minimize the loss function is a RMSProp optimizer [@MLpaper]. Scanning -------- After training the network shown in Fig. \[autoencoder\_architecture\], the training label in is used to estimate the original directions of the correlated sources impinging on the array. This is done through spatial scanning of the output of each decoder (i.e. filter) $\boldsymbol {y}_j$. The scanning phase aims at calculating the gain of each filter in all directions, in which the actual source angles would have large gain values, while the other directions would have much smaller gains. Afterwards, a threshold value is used to select the angles whose gains have peaks surpassing the threshold. The gain response of each filter is obtained by $$g^{(j)}=\left|\overline{\boldsymbol{r}}^{H}\left(\Theta_{k}\right) \overline{\boldsymbol{y}}_{j}\right|, \quad j=1, \ldots, 6,% i=1, \ldots, N \label{gain}$$ where the superscript $(\bullet)^{H}$ is the conjugate transpose of the matrices and vectors, $\overline{\boldsymbol{r}}^{H}$ is the estimated value of ${\hat{\boldsymbol{r}}}$, and $\overline{\boldsymbol{y}}_{j}$ is the complex version of the output of the $j$-th decoder $\boldsymbol{y}_{j}$. $\overline{\boldsymbol{y}}_{j}$ is obtained from $\boldsymbol{y}_{j}$ by concatenating the first half that represents the real values in $\boldsymbol{y}_{j}$ with their corresponding imaginary values in the second half of $\boldsymbol{y}_{j}$, similarly $\overline{\boldsymbol{r}}^{H}\left(\Theta_{k}\right)$ is obtained from $\boldsymbol {r}^{T}(\Theta _{k})$ . Simulation Results ================== In this section, we carry out simulations to evaluate our proposed DNN. We used the python library *tensorflow* to design and process our DNN. The network is trained on $I=1200$ samples, with learning rate of $0.001$, while the batch size is $100$ and the number of epochs is $1000$ epochs. We use a uniform linear array (ULA) of size $M=20$ elements with spacing $d=\lambda/2$ to predict directions of signals impinging from sources located in the spatial range of $\left[-60^{\circ}, 60^{\circ}\right]$, which is divided equally into six subregions. The training samples are generated randomly from directions $\Theta_{i}=-60^{\circ}+0.1^{\circ}, \forall i=1, i=1,\hdots,1200$. The covariance input vector $\boldsymbol{r}\left(\Theta_{i}\right) $ is generated using $N=800$ snapshots. To evaluate the performance of our algorithm, we use forward/backward spatial smoothing along with the MUSIC algorithm (SS-MUSIC) in [@SSMUSIC2], and compared it against our DNN in multipath environment by randomly changing $g_k$ and $\phi_{k}$ in for every target $k$. Gain responses of each decoder (i.e filter) ------------------------------------------- The DNN is tested using a covariance vector obtained from two correlated sources located at $\theta_{1}=-15^{\circ}, \theta_{2}=-5^{\circ}$ respectively, which belongs to subregion 3, i.e. $\left[\mathrm{-20}^{\circ},\mathrm{0}^{\circ}\right]$, with signal to noise ratio $\mathrm{SNR}= 10$ dB. Those specific directions were chosen, because they are in the same subregion, making them spatially close, hence it would be harder to separate them compared to distant sources. The threshold value to find the peaks in the scanning process is set to $= 0.3$, which is set by experience. Fig. \[gain response\] shows the gain obtained from for all the filters. It can be depicted that the spatial gain response of the filter corresponding to the assigned subregion has higher peaks compared to the other filters. It is clear that the filter managed to differentiate between both angles despite the fact they are coherent, and the antenna array suffers from imperfection errors as stated in . Performance against SS-MUSIC ----------------------------- Fig.\[mse\] compares DOA estimation performance of our proposed DNN with SS-MUSIC in presence of correlated sources and array imperfections combined. The average root mean square error (RMSE) in degrees is used to measure the accuracy of DOA estimates for various $\mathrm{SNR}$. It can be shown from the figure that the RMSE of our DNN starts high at $\mathrm{SNR} = 0$ dB, then decreases significantly compared to SS-Music as the $\mathrm{SNR}$ increases. It can be seen from the figure that the RMSE of SS-MUSIC algorithm is heavily impacted by the presence of imperfections due to the fact that it assumes ideal steering vector model with no imperfections as in Perfomance against algorithm in [@MLpaper] ------------------------------------------- \ Fig. \[compare\] compares the detection performance of our DNN compared to the algorithm in [@MLpaper], where the authors only considered array imperfections and assumes perfectly uncorrelated sources for DOA estimation. The same training data was used for both algorithms for fair comparison. The figure shows a consistent behavior for our algorithm detecting all 8 targets, however the algorithm in [@MLpaper] shows lower probability of detection due to its failure to detect all coherent sources. Conclusion ========== In this paper we presented a framwork for DOA estimation of correlated sources in presence of array imperfections. Our approach was based on Deep AE with 5 hidden layers, one acting as encoder and 4 as decoder. From the simulations we showed that the AE acts as a denoiser, where it could successfully remove the effect of coherence and imperfections producing accurate DOA estimates compared to the commonly used SS-MUSIC. Moreover, we compared our algorithm with the approach in [@MLpaper] where the authors only deal with imperfections, and our algorithm showed better and more consistent behavior. Acknowledgment {#acknowledgment .unnumbered} ============== Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Project-ID 287022738 – TRR 196 S03

ArXiv

--- abstract: 'We show that when the pseudomagnetic fields created by long wavelength deformations are appropriately coupled with a scalar electric potential, a significant energy gap can emerge due to the formation of a Haldane state. Ramifications of this physical effect are examined through the study of various strain geometries commonly seen in experiments, such as strain superlattices and wrinkled suspended graphene. Of particular technological importance, we consider setup where this gap can be tunable through electrostatic gates, allowing for the design of electronic devices not realizable with other materials.' author: - 'T. Low$^{1,2}$' - 'F. Guinea$^3$' - 'M. I. Katsnelson$^4$' title: Gaps tunable by electrostatic gates in strained graphene --- Introduction ============ Graphene[@Netal04; @Netal05] is a material whose unique properties are a fascinating challenge in both fundamental and applied sciences. The basic properties of its electronic structure are chirality, electron-hole symmetry, and linear gapless energy spectrum, that is, charge carriers in graphene are massless Dirac fermions[@NGPNG09]. In addition, corrugations and topological defects create gauge (pseudomagnetic and even pseudo-gravitational) fields acting on electron states[@VKG10]. Recently, a novel state of matter, a quantum Hall insulator without a macroscopic magnetic field (Haldane state[@H88]), has spawned the interest in unusual topological properties of band structures, leading to the prediction of topological insulators in two and three dimensions[@KM05b; @FKM07; @QZ10; @M10; @HK10]. It was understood afterwards that such Haldane state can be realized in a graphene superlattice by a suitable combination of scalar and vector electromagnetic potentials[@S09]. A gap opens in the electronic spectrum, turning graphene into a quantum Hall insulator with protected chiral edge states. Since, long wavelength strains in graphene induce a pseudomagnetic gauge field[@noteMGV07; @GKG10], the combination of strains and a scalar potential should too open a gap in graphene[^1]. The latter, if realised, would be of general interest. This is the subject of our work. Energy gap in graphene is crucial for many applications. Its realization remains a challenging problem, as the transformation of electrons into holes i.e. Klein tunneling[@KNG06], is an significant obstacle. At present, all known ways of gap opening have a detrimental effect on the electron mobility. In both biased bilayer and chemically functionalized graphene, one arrives at a disordered semiconductor with Mott variable range hopping mobility [@TJH10; @ENM09; @NRJR10]. In graphene nanoribbons, the mobility is typically several order of magnitude smaller than in bulk graphene [@HBK10]. Our approach provides an attractive route to circumvent these limitations and might also allows for the design of new electronic devices. This paper is organized as follows. In Sec. II, we present a general argument for the gap opening due to combination of strains and a scalar potential. Sec. III considers specific realization of this effect in strain superlattices and wrinkles, Fig. \[sketch\] provide an illustrations of these two strain geometries. In particular, we consider physical setup where this gap can be tunable through electrostatic gates. Sec. IV considers related effects such as Fermi velocity renormalization, topological defects and interplay with magnetic field. \[0.6\][![ []{data-label="sketch"}](fig1.pdf "fig:")]{} \[gen\] General arguments ========================= Strain induces scalar and gauge potentials in graphene[@VKG10]. Fig. \[sketch\] illustrates the strain induced pseudo magnetic field for a strain superlattice and wrinkled graphene. In terms of the strain tensor, the scalar and vector potentials are[@SA02b; @M07; @VKG10]: $$\begin{aligned} V ( \vec{\bf r} ) &= g \left[ u_{xx} ( \vec{\bf r} ) + u_{yy} ( \vec{\bf r} ) \right] \nonumber \\ A_x ( \vec{\bf r} ) &= \frac{\beta}{a} \left[ u_{xx} ( \vec{\bf r} ) - u_{yy} ( \vec{\bf r} ) \right] \nonumber \\ A_y ( \vec{\bf r} ) &= 2 \frac{\beta}{a} u_{xy} ( \vec{\bf r} ) \label{fields}\end{aligned}$$ where $g \approx 4$eV[@CJS10], $\beta = - \partial \log ( t ) / \partial \log ( a ) \approx 2$[@HKSS88], $t \approx 3$eV is the hopping between $\pi$ orbitals in nearest neighbor carbon atoms, and $a \approx 1.42$ Å is the distance between nearest neighbor atoms. The Hamiltonian for Dirac fermions in graphene, including a scalar potential and gauge fields due to strains is given by $$\begin{aligned} {\cal H} &= {\cal H}_0 + {\cal H}_A + {\cal H}_V \nonumber \\ {\cal H}_0 &= v_F \left( i \sigma_x \tau_z \partial_x + i \sigma_y \partial_y \right) \nonumber \\ {\cal H}_A &= - v_F \left[ \sigma_x A_x ( \vec{\bf r} ) + \tau_z \sigma_y A_y ( \vec{\bf r} ) \right] \nonumber \\ {\cal H}_V &= V ( \vec{\bf r} ) \label{hamil}\end{aligned}$$ where $v_F$ is the Fermi velocity, and $\sigma_{i=x,y,z}$ and $\tau_{i=x,y,z}$ are Pauli matrices which operates on the sublattice and valley indices. Using perturbation theory in both scalar and vector potential, we obtain a self energy correction which can be written as a cross-term to the Hamiltonian (linear in $V$ and linear in ${\bf A}$): $$\begin{aligned} \nonumber \Sigma ( \omega , \vec{\bf r} - \vec{\bf r}' ) &=& {\cal H}_V \left( \omega {\cal I} - {\cal H}_0 \right)^{-1} {\cal H}_A +\\ \nonumber &&{\cal H}_A \left( \omega {\cal I} - {\cal H}_0 \right)^{-1} {\cal H}_V \\ \nonumber &=& {\cal H}_V \left( \omega {\cal I} + {\cal H}_0 \right) \left( \omega^2 {\cal I}^2 - {\cal H}_0^2 \right)^{-1} {\cal H}_A+\\ && \left( {\cal H}_V \leftrightarrow {\cal H}_A \right) \label{sigma}\end{aligned}$$ where the identity matrix ${\cal I}$ is made explicit. Since the self energy contains term proportional to $\sigma_x \sigma_y = i\sigma_z$, a gap can be opened in both valleys. Next we make explicit the necessary relations between the scalar and gauge potential for this gap opening. \[0.5\][![$\bold{(a)}$ Sketch of the diagram which describes correlations between the scalar potential and pseudomagnetic field, see Eq. \[gap\]. $\bold{(b)}$ Electronic bandstructure of graphene superlattice, where each supercell unit (as indicated in Fig. 1a) contains $40\times 40$ graphene lattice units. Out-of plane corrugations amplitude of $h_{0}=2\AA$ is used, leading to a non-homogenoeus pseudomagnetic field (see Fig. 1a) less than $100T$. Various scalar potentials as indicated are considered, where $V(\vec{r})=c_{0}B(\vec{r})$ and $V(\vec{r})=c_{0}u_{z}(\vec{r})$ leads to gap opening. $c_{0}$ is chosen such that max$[V(\vec{r})]=0.2V$.[]{data-label="diagram"}](fig2.pdf "fig:")]{} We calculate the second order diagram for the electron self energy (effective potential) as shown in Fig. \[diagram\]a. At low energies, $\omega=0$, and zero wave vector ${\bf k}=0$ (corresponding to the Dirac point), the form for the induced gap reads $$\begin{aligned} \nonumber \Delta &= - {\rm Tr} \left\{ \sigma_z \frac{2}{v_F} \int d^2 \vec{\bf k} \frac{{\rm Im} \left( V_{-\vec{\bf k}} \right) \left[ ( \vec{\bf k} \vec{\sigma} ) , ( \vec{\bf A}_{\vec{\bf k}} \vec{\sigma} ) \right]}{ | \vec{\bf k} |^2} \right\} \\ & \propto \int d^2 \vec{\bf k} \frac{{\rm Im} ( V_{-\vec{\bf k}} ) \left( k_x A^y_{\vec{\bf k}} - k_y A^x_{\vec{\bf k}} \right)}{ | \vec{\bf k} |^2} \label{gap}\end{aligned}$$ where $\left[ \cdots \right]$ is the commutator, $V_{\vec{\bf k}}$ and ${\bf A}_{\vec{\bf k}}$ are Fourier components of scalar and vector potentials, respectively. This equation shows that the gap is induced through the correlations between the scalar potential and the pseudomagnetic (synthetic magnetic) field, $B_{\vec{\bf k}} = k_x A^y_{\vec{\bf k}} - k_y A^x_{\vec{\bf k}}$. We characterize these correlations by the parameter $C$ such that $$\begin{aligned} \lim_{\vec{\bf k} \rightarrow 0} ( B V )_{\vec{k}} &= C \label{cross} \end{aligned}$$ or, alternatively, $$\begin{aligned} \lim_{| \vec{\bf r} - \vec{\bf r}' | \rightarrow \infty} \left\langle V ( \vec{\bf r} ) B ( \vec{\bf r}' ) \right\rangle &= C \delta^{(2)} ( \vec{\bf r} - \vec{\bf r}' ) \label{cross_1} \end{aligned}$$ The parameter $C$ has dimensions of energy. It is roughly given by the value of the scalar potential times the number of flux quanta due to the synthetic field over the region where the field and the scalar potential are correlated. For $C \ne 0$, the integral in Eq. (\[gap\]) diverges as $\vec{\bf k} \rightarrow 0$. Then, the lower limit of the integral should be $k_{min} \approx \Delta / v_F$, turning Eq. (\[gap\]) into a self consistent equation for $\Delta$. \[0.63\][]{} In the diffusive regime, where electrons with momentum $\vec{\bf k}$ have an elastic scattering time $\tau_{\vec{k}}$, the divergence in the integral in Eq. (\[gap\]) has to be cutoff at a momentum $k_{min}$ such that $v_F k_{min} \approx \tau_{k_{min}}^{-1}$. For a periodic superlattice, the integral in Eq. (\[gap\]) has to be replaced by a sum over reciprocal lattice vectors, $\vec{\bf G}$. For graphene in the diffusive regime, resonant scatterers[@SPG07; @WYLGK10] or substrate charges[@NM07; @AHGS07] give rise to a dependence $\tau^{-1}_{\vec{\bf k}} \propto n_i | \vec{\bf k} |^{-1}$, where $n_i$ is the concentration of scatterers, so that the lower cutoff in Eq. (\[gap\]) is $k_{min} \propto \sqrt{n_i}$. Using this cutoff, we can write $$\begin{aligned} \Delta &\approx C \, \, \, \log \left[ \frac{1}{{\rm Max} ( k_{min} , \Delta / v_F ) \, \, a} \right] \end{aligned}$$ where $a \approx 1.4\,\AA$ is the distance between nearest carbon atoms, which sets the scale of the high momentum limit in Eq. (\[gap\]). Next, we extend this general argument to specific examples. Possible physical realizations ============================== Global gap in strain superlattices ---------------------------------- We show results for the gap opening due to correlation between the synthetic magnetic field and a scalar potential in strain graphene superlattice. We assume that the graphene layer is corrugated, as depicted in Fig. \[sketch\]a. Depending on the lattice mismatch with the underlying substrate, the supercell size could vary from $\approx 10\times10$ (e.g. iridium) to $\approx 50\times50$ (e.g. boron nitride) times that of graphene unit cell. Here, we assumed supercell to be $40\times40$. The in-plane displacements were relaxed in order to minimize the elastic energy. Details of the implementation based on continuum elasticity model are given in Appendix A. \[0.5\][]{} After minimizing the elastic energy, the resultant strains lead to an underlying non-homogeneous pseudomagnetic field as depicted in Fig. \[sketch\]a. Fig. \[diagram\]b shows the calculated electronic bandstructure of the superlattice, after considering various types of scalar potentials. Strain alone, $V(\vec{\bf r})=0$, does not produce a gap. Although strains can induced a scalar potential of type $V(\vec{\bf r})\propto u_{xx}+u_{yy}$, the correlation between this potential and the synthetic magnetic field due to the same strains is also zero. On the other hand, a scalar potential proportional to the height corrugation, i.e. $V(\vec{\bf r})\propto u_{z}$, leads to the appearance of a gap, albeit a small one. Such a scalar potential could be induced by the substrate through an existence of electric field perpendicular to the graphene layer [@GL10]. Graphene superlattices induced by commensuration effects between the mismatch in the lattice constants of graphene and the substrate[@Zetal07; @Vetal08; @Metal08; @Petal08] will, in general lead to the effect considered here. The existence of a gap at the Dirac energy in strained graphene superlattice is also consistent with observations which show gaps in very clean samples which are commensurate with the substrate[@Zetal07; @LLA09; @HK10]. In essence, strains generally induce both scalar and vector potentials. However, cross correlations between the scalar potential and the synthetic magnetic field, as in Eq. (\[cross\_1\]), vanish in many cases. In particular, the gap should be zero if the system remains symmetric with respect to inversion[@noteMGV07]. A significantly large gap is obtained when the scalar potential and pseudomagnetic field are perfectly correlated, as shown in Fig. \[diagram\]b for the case when $V(\vec{\bf r})\propto B(\vec{\bf r})$. This observation is consistent with the general arguments that we presented in Sec. \[gen\]. See also Appendix A, where a general expression for $\Delta$ in strain superlattice is derived. Next, we examine a slightly different scenario, where electrostatic gates are used to engineer the scalar potential so as to realise the correlation with the underlying strain-induced pseudo magnetic field. Local gap in strain superlattices through quantum transport calculations ------------------------------------------------------------------------ Here, we examine numerically the effect a local electric scalar potential on the electronic transport properties of strained graphene superlattices. The Hamiltonian accounting for nearest neighbor interactions between $p_{z}$ orbitals is given by[@W47], $$\begin{aligned} {\cal H} = \sum_{i}V_{i}a_{i}^{\dagger}a_{i} + \sum_{ij}t_{ij}a_{i}^{\dagger}a_{j} \label{hamilTB}\end{aligned}$$ where $V_{i}$ is the on-site energy due to the scalar potential $V(\vec{\bf r})$ and $t_{ij}=t(1-\tfrac{\beta}{a}(a_{ij}-a))$ is the hopping energy. $a_{ij}$ is the new bond length after strain. To facilitate the application of various numerical techniques, the problem is partitioned into block slices as shown in Fig. \[qt\]. The retarded Green’s function in $\Omega_{0}$, the device region of interest, can then be written as (see [@DV08; @D97; @HJ10] for general theory), $$\begin{aligned} {\cal G} = \left(\epsilon_{f}{\cal I}-{\cal H}_{0}-\Sigma_{L}-\Sigma_{R}\right)^{-1}\equiv A^{-1} \label{greenrt}\end{aligned}$$ where $\epsilon_{f}$ is the Fermi energy, and $\Sigma_{L/R}$ are defined as $\Sigma_{L}=\tau^{\dagger}g_{L}\tau$ and $\Sigma_{R}=\tau g_{R}\tau^{\dagger}$ respectively. $g_{L/R}$ are the surface Green’s function, which can be obtained numerically through an iterative scheme [@SSSR85] based on the decimation technique (see e.g. [@GTFL83]). Various physical quantities of interest such as the transmission, current/charge density, local density-of-states can be obtained once ${\cal G}$ is determined. See Appendix B for a more detailed description of the numerics. We consider a finite size strain superlattice of dimension $200nm\times100nm$, as depicted in Fig. \[superlattice\]a. The corresponding pseudomagnetic field is shown in Fig. \[superlattice\]b. Transport in non-homogeneous magnetic field is dominated by bulk “magnetic” states known as snake states[@MDE08]. Snake states has been observed in high mobility 2D electron gas system, through controlled engineering of magnetic field via lithographic patterning of ferromagnetic or superconducting thin films[@ye95; @nogaret00; @carmona95]. Fig. \[superlattice\]c plots the current density due to current injection from the left contact, biased at Fermi energy of $150meV$. As depicted, current flows in regions where $B(\vec{\bf r})\approx 0$, along the direction $\pm\nabla B(\vec{\bf r})\times \hat{z}$. Unlike the case of a real magnetic field, these snake states are non-chiral, with forward and backwards going states residing in opposite valleys. In the absence of short range scatterers, these states are relatively protected. Fig. \[superlattice\]b shows three conducting snake channels which forms the backbone for the conduction. Applying the general principle described in Sec. \[gen\], we apply a scalar potential that approximately correlates with the pseudomagnetic field of the middle channel to open up a local gap. Indeed, a local gap is opened, impeding current flow along this channel, as shown in Fig. \[superlattice\]d. Such a scalar potential can be realised experimentally with electrostatic side gates. This effect could be exploited in current guiding devices [@williams11]. Opening gaps in suspended graphene via wrinkles ----------------------------------------------- Wrinkles are common feature in very clean suspended graphene samples, leading to finite strains. Partial control of these wrinkles can be achieved by adjusting the temperature, as in some cases, they are induced by the mismatch in thermal expansion coefficients between graphene and the substrate[@Betal09; @wang11]. We assume that the deformation is described by the profile proposed in [@CM03], as illustrated in Fig. \[sketch\]a. The resulting synthetic field is discussed elsewhere [@GHL09]. Fig. \[sketch\]b illustrates the accompanied pseudo-magnetic field. A sinusoidal-like scalar potential within graphene is induced by gates shown in Fig. \[wrinkle\]a, which are tailored to correlate with the synthetic magnetic field induced by the strains. The results in Fig. \[wrinkle\]b show that a gap is generated, whose magnitude is proportional to the scalar potential. We can estimate the gap induced by a combination of strains which are changed by $\delta u$ over an area of spatial scale $\ell$ and a scalar potential of value $\delta V$ on a region of the same size. The synthetic magnetic field is of order $B \sim ( \beta \delta u ) / ( a \ell )$. Then $$\begin{aligned} \Delta \sim C \approx \beta \delta u \delta V \frac{\ell}{a} \label{gap12} \end{aligned}$$ Eq. (\[gap12\]) is consistent with numerical results obtained in Fig. \[wrinkle\]b. Fig. \[wrinkle\]c and d clearly show the global nature of the gap generated. Effectively, the gate controlled gap allows the device in Fig. \[wrinkle\] to be operated as a graphene transistor. Next, we consider situation where correlation is less than perfect. From Eq. (\[gap12\]), we note that even for small variations in the strain, $\delta u \ll 1$, the gap can be of order of the potential fluctuations, $\delta V$, if the correlations between the scalar potential and the synthetic field are maintained over long distances, $\ell \gg a$. Fig. \[wrinkle\]b considers the case where there is a phase shift between the pseudomagnetic field and scalar potential. This corresponds to a decreasing $\ell$ in Eq. (\[gap12\]). Indeed the gap reduces as expected. In general, the presence of a gap is robust against reasonable degree of local disorder, since inversion symmetry is still absent in most part, see also Eq.(4) and related discussions. Discussions =========== In this section we discuss several issues related with the previous consideration and the ways of further development. Renormalization of the Fermi velocity ------------------------------------- As evident from Eq. (\[sigma\]), there are self energy corrections due to quadratic terms in the scalar and vector potentials. These terms lead to logarithmic corrections in the Fermi velocity via the renormalization of the residue of the Green function: $$\left. \frac{\partial \Sigma \left( \vec{\bf k},E\right) }{\partial E} \right|_{E=0} =-% \frac{\ln \Lambda }{2\pi v_F^2}\left( V_{\vec{\bf k}}V_{-{\vec\bf k}}+\vec{\bf A}_{\vec{\bf k}}\vec{\bf A}_{-\vec{\bf k}}\right)$$ where $\Lambda \gg \left| \vec{\bf k} \right|$ is a momentum cutoff. This renormalization has been previously found in non-linear sigma models[@F86; @L93; @NTW94; @Z98]. These logarithmic corrections also influence the kinetic equation which describes transport processes[@AK07], in these terms they describe a pseudo-Kondo effect due to interband scattering (in the Dirac point, the energies of electron and hole states coincide which provides a necessary degeneracy). Due to these corrections, the Fermi velocity decreases in the presence of scalar and gauge disorder. When this effect is studied simultaneously with the increase induced by the Coulomb interaction non trivial new phases can arise[@SGV05; @HJV08; @FA08]. Gauge field due to topological defects -------------------------------------- The sublattice and valley symmetries of graphene allow for the definition of a second gauge field, which hybridizes states from different valleys, and does not commute with the intravalley gauge field due to long wavelength strains[@NGPNG09; @VKG10]. This field can be induced by topological defects, such as heptagons and pentagons. These defects are present at dislocations and grain boundaries, and they can be ordered periodically forming superlattices. If the synthetic magnetic field associated with this field is correlated with a scalar potential, a gap inducing term is generated. The gauge field due to topological defects has the form $$\begin{aligned} {\cal H}_{\tilde{A}} &= - v_F \left[ \tau_x \tilde{A}_y ( \vec{\bf r} ) + \tau_y \sigma_z \tilde{A}_y ( \vec{\bf r} ) \right]\end{aligned}$$ By using perturbation theory, as in Eq. (\[sigma\]), we obtain a self energy which is proportional to the cross correlations between the scalar potential and the gauge field, multiplied by the operator $\tau_y \sigma_x$. Modulated Zeeman couplings can also lead to synthetic fields which act on the spin, allowing for the possibility of spin gaps as well. Interplay with magnetic field ----------------------------- The gap studied here is defined in the whole sample, although its value should be roughly inversely proportional to the ratio between the total area and the area where the synthetic magnetic field and scalar potential are correlated. The sign of the gap is determined by the scalar potential. Localized states will be formed at boundaries between regions where the gaps have different signs, similar to the edge states in topological insulators [@QZ10; @M10]. A periodic magnetic field, when correlated with a scalar potential leads to a gap whose signs are opposite in the two valleys[@S09]. A combination of this gap and the gap due to strains leads to gaps of different values in the two valleys, allowing for the control of the valley and sublattice degrees of freedom. For example, combined strain and synthetic magnetic field could be useful for valleytronics[@LG10]. The realization of other synthetic fields might open new functionalities for graphene that cannot be achieved with other materials. Conclusion ========== We have discussed a novel way in which a combination of long wavelength strains and a long wavelength correlated potential can lead to a gap in the electronic spectrum of graphene. Such situation can occur naturally, because of correlations between a periodic substrate and graphene, or it can be engineered in a controlled way using electrostatic gates. Since the effect is induced by long wavelength, smooth perturbations, a gap can be induced without increasing the amount of scattering in the system. Finally, valley polarized edge states will be generated, as the band structure of the modified system resembles the spectrum of a quantum Hall insulator. Acknowledgements ================ TL acknowledges funding from INDEX/NSF (US). FG acknowledges financial support from MICINN (Spain) through grants FIS2008-00124 and CONSOLIDER CSD2007-00010, and from the Comunidad de Madrid, through NANOBIOMAG. The work of MIK is part of the research program of the Stichting voor Fundamenteel Onderzoek der Materie (FOM), which is financially supported by the Nederlandse Organisatie voor Wetenschappelijk Onderzoek (NWO). Computational resources is provided by Network for Computational Nanotechnology (NCN) at Purdue University. \[gapexpress\] Expression for energy gap in superlattice ======================================================== Strains induce scalar and gauge potentials[@VKG10]. We study the correlations between these potentials when the strains are induced by modulations in the vertical displacement of the layer, $h ( \vec{\bf r} )$. We assume that the in plane displacements relax in order to minimize the elastic energy. The strains are[@GHL08] $$\begin{aligned} u_{ij} ( \vec{\bf k} ) &= \frac{\lambda + \mu}{\lambda + 2 \mu} \frac{k_i k_j \left[ k_x^2 h_{yy} ( \vec{\bf k} ) + k_y^2 h_{xx} ( \vec{\bf k} ) - 2 k_x k_y h_{xy} ( \vec{\bf k} ) \right]}{| \vec{\bf k} |^4} \label{strains}\end{aligned}$$ where $h_{ij} ( \vec{\bf k} )$ are the Fourier transforms of the tensor $$\begin{aligned} h_{ij} ( \vec{\bf x} ) &= \partial_i h \partial_j h\end{aligned}$$ In terms of the strain tensor, the scalar and vector potentials are[@SA02b; @M07; @VKG10]: $$\begin{aligned} V ( \vec{\bf r} ) &= g \left[ u_{xx} ( \vec{\bf r} ) + u_{yy} ( \vec{\bf r} ) \right] \nonumber \\ A_x ( \vec{\bf r} ) &= \frac{\beta}{a} \left[ u_{xx} ( \vec{\bf r} ) - u_{yy} ( \vec{\bf r} ) \right] \nonumber \\ A_y ( \vec{\bf r} ) &= 2 \frac{\beta}{a} u_{xy} ( \vec{\bf r} ) \label{fields}\end{aligned}$$ Using Eq. (\[gap\]) of the main text, we obtain $$\begin{aligned} \Delta &= \frac{g \beta}{a} \frac{(\lambda + \mu )^2}{(\lambda + 2 \mu )^2} \times \nonumber \\ &\times \int d^2 \vec{\bf k} \frac{\left| k_x^2 h_{yy} ( \vec{\bf k} ) + k_y^2 h_{xx} ( \vec{\bf k} ) - 2 k_x k_y h_y ( \vec{\bf k} ) \right|^2 \cos ( 3 \theta_{\vec{\bf k}} )}{\left| \vec{\bf k} \right|^4}\end{aligned}$$ This expression is zero, as $\theta_{- \vec{\bf k}} = \theta_{\vec{\bf k}} + \pi$. While the scalar and gauge potentials are correlated, their correlation does not contribute to the formation of a global gap. Quantum transport methods. ========================== \[0.45\][![**Numerical approach: The graphene ribbon is partitioned into block slices along the $x$-direction (transport) as indicated. Lattice interactions within each block is described by $\alpha$. Nearest neighbor blocks interactions are represented by $\tau$. Device domain $\Omega_{0}$ will include the strains and scalar potential $V(\vec{r})$. Left/right leads regions ($\Omega_{L/R}$) are assumed unstrained and electrically doped, due to charge transfer from contacts.** []{data-label="qt"}](qt.pdf "fig:")]{} The Hamiltonian accounting for nearest neighbor interactions between $p_{z}$ orbitals is given by[@W47], $$\begin{aligned} {\cal H} = \sum_{i}V_{i}a_{i}^{\dagger}a_{i} + \sum_{ij}t_{ij}a_{i}^{\dagger}a_{j} \label{hamilTB}\end{aligned}$$ To facilitate the application of various numerical techniques, the problem is partitioned into block slices as shown in Fig. \[qt\]. The retarded Green’s function in $\Omega_{0}$, the device region of interest, can then be written as (see [@DV08; @D97; @HJ10] for general theory), $$\begin{aligned} {\cal G} = \left(\epsilon_{f}{\cal I}-{\cal H}_{0}-\Sigma_{L}-\Sigma_{R}\right)^{-1}\equiv A^{-1} \label{greenrt}\end{aligned}$$ where $\epsilon_{f}$ is the Fermi energy, and $\Sigma_{L/R}$ are defined as $\Sigma_{L}=\tau^{\dagger}g_{L}\tau$ and $\Sigma_{R}=\tau g_{R}\tau^{\dagger}$ respectively. $g_{L/R}$ are the surface Green’s function, which can be obtained numerically through an iterative scheme [@SSSR85] based on the decimation technique (see e.g. [@GTFL83]). It is also useful to define the quantity, broadening function, $\Gamma_{L/R}\equiv i(\Sigma_{L/R}-\Sigma_{L/R}^{\dagger})$. Physical quantities of interest such as the transmission ${\cal T}$ is given by, $$\begin{aligned} {\cal T}=\mbox{Tr}\left([\Gamma_{L}]^{1}_{1}[{\cal G}]^{1}_{n}[\Gamma_{R}]^{n}_{n}[{\cal G}^{\dagger}]^{n}_{1}\right) \label{trandef}\end{aligned}$$ Energy gaps, as seen in Fig. 4b of main manuscript, is estimated by the onset of increase in ${\cal T}$. The electron density $n(\vec{r})$ at slice $j$ is obtained from the diagonals elements of ${\cal G}^{n}$, given by, $$\begin{aligned} [{\cal G}^{n}]^{j}_{j}=f_{L}[{\cal G}]^{j}_{1}[\Gamma_{L}]^{1}_{1}[{\cal G}^{\dagger}]^{1}_{j}+f_{R}[{\cal G}]^{j}_{n}[\Gamma_{R}]^{n}_{n}[{\cal G}^{\dagger}]^{n}_{j} \label{gndef}\end{aligned}$$ Local density-of-states (as seen in Fig. 4c-d of main manuscript) is obtained from Eq. \[gndef\] by simply setting $f_{L}=f_{R}=1$. Current density $j(\vec{r})$ (as seen in Fig. 3c-d of main manuscript), flowing from slice $j$ to $j+1$ is given by the diagonal of $J$, given by, $$\begin{aligned} [J]^{j}_{j+1}=\tfrac{2q}{h}\left([A]^{j}_{j+1}[{\cal G}^{n}]^{j+1}_{j}-[A]^{j+1}_{j}[{\cal G}^{n}]^{j}_{j+1}\right) \label{jcdef}\end{aligned}$$ where, $$\begin{aligned} \nonumber [{\cal G}^{n}]^{j+1}_{j}=f_{L}[{\cal G}]^{j+1}_{1}[\Gamma_{L}]^{1}_{1}[{\cal G}^{\dagger}]^{1}_{j}+f_{R}[{\cal G}]^{j+1}_{n}[\Gamma_{R}]^{n}_{n}[{\cal G}^{\dagger}]^{n}_{j}\\ [{\cal G}^{n}]^{j}_{j+1}=f_{L}[{\cal G}]^{j}_{1}[\Gamma_{L}]^{1}_{1}[{\cal G}^{\dagger}]^{1}_{j+1}+f_{R}[{\cal G}]^{j}_{n}[\Gamma_{R}]^{n}_{n}[{\cal G}^{\dagger}]^{n}_{j+1} \label{gndef2}\end{aligned}$$ As apparent from Eq. (\[trandef\])-(\[gndef2\]), it is not neccessary to obtain the full matrix ${\cal G}$. Through commonly used recursive formula of the Green’s function derived from the Dyson equation and the decimation technique, one could obtain these block elements of the Green’s function, $[{\cal G}]^{i}_{j}$, in a computationally/memory efficient manner. Details of this numerical recipe are described elsewhere[@LA09]. [58]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , , , , , , , , ****, (). , , , , , , , ****, (). , , , , , ****, (). , , , ****, (). , ****, (). , ****, (). , , , ****, (). , pp. (). , ****, (). , ****, (). , ****, (). , , , ****, (). , , , ****, (). , ****, (). , , , , , , , , , , , ****, (). , , , , , , , , , , , ****, (). , , , ****, (). , ****, (). , ****, (). , , , ****, (). , , , , ****, (). , , , ****, (). , , , , , ****, (). , ****, (). , , , , ****, (). , ****, (). , , , , , , , , , ****, (). , , , , , , , ****, (). , , , , , , , , , , , ****, (). , , , , , , , ****, (). , , , ****, (). , ****, (). , (). , (). , (). , , , , ****, (). , , , , ****, (). , , , ****, (). , , , , , , , ****, (). , ****, (). , , , , , , , , ****, (). , , , , ****, (). , , , , , , , ****, (). , ****, (). , ****, (). , , , ****, (). , ****, (). , ****, (). , , , ****, (). , ****, (). , ****, (). , , , ****, (). , , , ****, (). , ****, (). , ****, (). , , , ****, (). , ****, (). [^1]: In contrast with the real magnetic field, strains do not break time reversal symmetry, and the resulting insulator is not a strong topological insulator - there is a [*pair*]{} of counterpropagating edge states belonging to different valleys, however, the intravalley scattering is frequently very weak which make these states well protected, similar to the case considered in Ref.[@GKG10]. Although strong edge disorder would leads to a transport gap instead[@LG10], which might be technologically useful too.

ArXiv

Tsemo Aristide Centre de Recherche Mathematiques Universite de Montreal Case Postale 6128, Succursale Centre Ville Montreal Quebec, H3C 3J7 **Connective structures for principal gerbes.** **Introduction.** Let $N$ be a manifold, $H$ a Lie group and $P$ a $H$-principal bundle defined over $N$, a $P$-gerbe defined over $N$, is a gerbed defined over $N$ bounded by the sheaf of automorphisms of $P$. In this paper we define the fundamental notions of the differential geometry of $P$-gerbes, that is the notions of connective structure curving holonomy and characteristic classes. The interest of such definitions is to provide a geometric action in string theory. In classical physics, the variational functional of the evolution of a particle in a phase space is a function of the holonomy of a gauge connection. In string theory, the action is given by the holonomy of a Deligne connective structure, when the gauge group is the circle. It is natural to provide the definition of non abelian holonomy in order to describe the action when the gauge group is non commutative. [**Acknowledgements.**]{} The authors want to thank Pierre Deligne for helpful corrections, and Dusa McDuff and Johan Dupont for helpful discussions. [**Notations.**]{} Let $U_{i_1},...,U_{i_p}$ be open subsets of a manifolds $N$, and $C$ a presheaf defined on $N$. We will denote by $U_{i_1..i_p}$ the intersection of $U_{i_1}$,...,$U_{i_p}$. If $e_{i_1}$ is an object of $C(U_{i_1})$, ${e_{i_1}}^{i_2...i_p}$ will be the restriction of $e_{i_1}$ to $U_{i_1...i_p}$. For a map $h:e\rightarrow e'$ between two objects of $C(U_{i_1..i_p})$, we denote by $h^{i_{p+1..i_n}}$ the restriction of $h$ to a morphism between $e^{i_{p+1}...i_n}\rightarrow {e'}^{i_{p+1}...i_n}$. [**Definition.**]{} Let $N$ be a manifold, $H$ a Lie group and $P\rightarrow N$ a $H$-principal bundle defined on $N$. A [**$P$-gerbe**]{} is a gerbe bounded by the sheaf of automorphisms of $P$. More precisely it is defined as follows: To each open subset $U$ of $N$ we associate a category $C_P(U)$. The group of automorphisms of an object of $C_P(U)$ is the group of automorphisms of the restriction of $P$ to $U$. We suppose that the following conditions are satisfied: Gluing conditions for objects. Let $U$ be an open subset of $N$, $(U_i)_{i\in I}$ an open cover of $U$ and $e_i$ an object of $C_P(U_i)$. Suppose given an arrow $u_{ij}:e^i_j\rightarrow e^j_i$ between the respective restrictions of $e_j$ and $e_i$ to $U_i\cap U_j$ such that ${u_{i_1i_2}}^{i_3}{u_{i_2i_3}}^{i_1}={u_{i_1i_3}}^{i_2}$. Then there exists an object $e_U$ of $C_P(U)$ whose restriction to $U_i$ is $e_i$. Gluing conditions for arrows. Let $e$ and $e'$ be two objects of $C_P(U)$. The correspondence defined on the category of open subsets of $U$ by $V\rightarrow Hom(e_{\mid V},e'_{\mid V})$ is a sheaf of sets, where $e_{\mid V}$ and $e'_{\mid V}$ are the respective restrictions of $e$ and $e'$ to $V$. We suppose that there exists an open cover of $N$ $(U_i)_{i\in I}$ such that the category $C_P(U_i)$ is not empty, and objects of $C_P(U_i)$ are isomorphic. An example of $P$-gerbe is defined as follows: Let $G$ be a Lie group and $H$ a closed normal subgroup of $G$. The quotient $G/H$ is a Lie group. We suppose that the projection $G\rightarrow G/H$ has local sections. Consider a $G/H$-bundle $p_{G/H}:P\rightarrow N$, defined on the manifold $N$. That is a locally trivial bundle whose transition functions is defined by the trivialization $(U_i,u_{ij})$, $u_{ij}:U_i\cap U_j\rightarrow G$ defined by the coordinate changes: $$U_i\cap U_j\times G/H\longrightarrow U_i\cap U_j\times G/H$$ $$(x,y)\longrightarrow (x,yu_{ij}(x))$$ The functions $u_{ij}$ verify the following property: ${u_{i_3i_1}}^{i_3}(x){u_{i_1i_2}}^{i_3}(x){u_{i_2i_3}}^{i_1}(x)$ are in the center of $H$ to insure the $G/H$-bundle to be well-defined as the $H$-bundle $p_H$ whose transition functions are defined by: $$U_i\cap U_j\times H\longrightarrow U_i\cap U_j\times H$$ $$(x,y)\longrightarrow (x,{u_{ij}}^{-1}(x)y{u_{ij}}(x))$$ Let ${\cal H}$ be the Lie algebra of $H$ and $Ad$ the adjoint representation. We can define the locally trivial ${\cal H}$-bundle $p_{\cal H}$ over $N$ whose transition functions are defined by: $$U_i\cap U_j\times {\cal H}\longrightarrow U_i\cap U_j\times {\cal H}$$ $$(x,y)\longrightarrow (x,Ad({u_{ij}}^{-1})(x))(y))$$ [**Proposition.**]{} [*Let $U$ be an open subset of $N$, we denote by $C_H(U)$ the category of $G$-principal bundles whose quotient by $H$ is the restriction of $p_{G/H}$ to $U$. A morphism between a pair of objects $e$ and $e'$ of $C_H(U)$ is a morphism of $G$-bundles which cover the identity of the restriction of $p_{G/H}$ to $U$. The correspondence defined on the category of open subsets of $N$ by $U\rightarrow C_H(U)$ is a gerbe bounded by the sheaf of automorphisms of $p_H$.*]{} [**Proof.**]{} Gluing property for arrows: Let $U$ be an open subset of $N$, and $(U_i)_{i\in I}$ an open cover of $U$. Consider an object $e_i$ of $C_H(U_i)$, and a map $u_{ij}:e^i_j\rightarrow e^j_i$ such that ${u_{i_1i_2}}^{i_3}{u_{i_2i_3}}^{i_1}={u_{i_1i_3}}^{i_2}$ The definition of bundle implies the existence of a $G$-bundle $e$ whose restriction to $U_i$ is $e_i$. Since the quotient of $e_i$ by $H$ is the restriction of $p_{G/H}$ to $U_i$, we deduce that the quotient of $e$ by $H$ is the restriction of $p_{G/H}$ to $U$. Gluing condition of arrows. Let $e$ and $e'$ be a pair of objects of $C_H(U)$, the correspondence defined on the category of open subsets of $U$ by $V\rightarrow Hom(e_{\mid V},e'_{\mid V})$ is a sheaf of sets, since it is the sheaf of morphisms between two bundles. The bundles $e_{\mid V}$ and $e'_{\mid V}$ are the respective restrictions of $e$ and $e'$ to $V$. Consider a trivialization $(U_i,u_{ij})$ of $p_{G/H}$. The bundle $U_i\times G$ is an element of $C_H(U_i)$, thus $C_H(U_i)$ is not empty, and for each object $e$ and $e'$ of $C_H(U)$, the restrictions of $e$ and $e'$ to $U_i\cap U$ are isomorphic to the trivial bundle $U_i\cap U\times G$ by an isomorphism whose projection to $U_i\cap U\times G/H$ is the identity. Let $e$, be an object of $C_H(e)$, and $f$ an automorphism of $e$, The restriction $f_i$ of $f$ to the restriction of $e$ to $U_i\cap U$ is an automorphism of the trivial bundle $U_i\cap U\times G$ which projects to the identity on $U_i\cap U\times G/H$. We deduce that $f_i$ is defined by a map $f':U_i\cap U_j\rightarrow H$. On $U_i\cap U_j\cap U$, we have $f_j={u_{ij}}^{-1}f_i{u_{ij}}$. This implies that $f$ is a section of $p_{H}$ $\bullet$ [**The classifying cocycle of a principal gerbe.**]{} Let $(U_i)_{i\in I}$ be an open cover of $N$ such that the category $C_P(U_i)$ is not empty and the objects of $C_P(U_i)$ are isomorphic. Consider for each $i$, an object $e_i$ of $C_P(U_i)$, and arrow $u_{ij}:e^i_j\rightarrow e^j_i$. We can defined the automorphism of $e^{i_1i_2}_{i_3}$: $c_{i_1i_2i_3}={u_{i_3i_2}}^{i_1}{u_{i_2i_1}}^{i_3}{u_{i_1i_3}}^{i_2}$. [**Proposition.**]{} [*The family of maps $c_{i_1i_2i_3}$ is a non commutative Cech $2$-cocycle.*]{} [**Proof.**]{} Let ${c'_{i_1i_2i_3}}^{i_4}={u_{i_4i_3}}^{i_1i_2}{c_{i_1i_2i_3}}^{i_4}{u_{i_3i_4}}^{i_1i_2}$ On $U_{i_1i_2i_3i_4}$, we have $$c'_{i_1i_2i_3}c_{i_1i_3i_4}=c_{i_2i_3i_4}c_{i_1i_2i_4}$$ $\bullet$ [**Connective structure on $H$-gerbes.**]{} [**Definition.**]{} Consider a gerbe $C_P$ defined on a manifold $N$ whose band is $L$, the sheaf of automorphisms of the principal bundle $P\rightarrow N$. [**A connective structure**]{} on $C_P$, is a correspondence which associates to each object $e_U$ of $C_P(U)$ an affine space $Co(e_U)$, called the torsor of connections, which is a subset of the set of ${p_{{\cal H}}}_{\mid U}$-valued $1$-forms defined on $U$, where ${p_{{\cal H}}}_{\mid U}$ is the restriction of $p_{{\cal H}}$ to $U$. The following properties are supposed to be satisfied by this assignment: (i)- The correspondence $e_U\rightarrow Co(e_U)$ is functorial with respect to restrictions to smaller subsets. (ii)- For every isomorphism $h:e_U\rightarrow e'_U$ between objects of $C_P(U)$, there exists an isomorphism of torsors $h^*:Co(e_U)\rightarrow Co(e'_U)$ compatible with the composition of morphisms of $C_P(U)$, and the restrictions to smaller subsets. (iii)- For each morphism $g$ of the object $e_U$ of $C_P(U)$, and $\nabla_{e_U}$ a connection of $Co(e_U)$, $$g^*\nabla_{e_U}=Ad(g^{-1})(\nabla_{e_U})+g^{-1}dg$$ For each open subset $U$, we define $C'_P(U)$ to be the category whose objects are pair of objects $(e_U,\nabla_{e_U})$, where $\nabla_{e_U}$ is an element of $Co(e_U)$. A morphism $f:(e_U,\nabla_{e_U})\rightarrow (e'_U,\nabla_{e'_U})$ is $\nabla_{e'_U}-u^*(\nabla_{e_U})$ where $u:e_U\rightarrow e'_U$ is a morphism of $C(U)$. We suppose the correspondence $U\rightarrow C'_P(U)$ to be a gerbe $\bullet$ [**The classifying cocycle of a connective structure.**]{} Let $(U_i, h_{ij})_{i\in I}$ a trivialization of $p_{H}$ such that $C_P(U_i)$ is not empty, and $e_i$ an object of $C_P(U_i)$ and $u_{ij}$ a morphism between $e^i_j$ and $e^j_i$. Consider an element $\alpha_i$ of $Co(e_i)$. We define $c_{i_1i_2i_3}$ to be ${u_{i_3i_1}}^{i_2}{u_{i_1i_2}}^{i_3}{u_{i_2i_3}}^{i_1}$. On $U_i\cap U_j$, we can define the ${\cal H}$-valued form $\alpha_{ij}=\alpha^j_i-{u_{ij}}^*(\alpha^i_j)$. The Cech boundary of the $p_{{\cal H}}$ $1$-cocycle $\alpha_{ij}$ is: $${u_{i_1i_2}}^*(\alpha_{i_2i_3})-\alpha_{i_1i_3}+\alpha_{i_1i_2}$$ $${u_{i_1i_3}}^*(\alpha_{i_3}-{{c_{i_1i_2i_3}}}^*(\alpha_{i_3}))$$ $$=Ad({h_{i_1i_3}}^{-1})(\alpha_{i_3}-Ad({c_{i_1i_2i_3}}^{-1})(\alpha_{i_3})+{c_{i_1i_2i_3}}^{-1}d({c_{i_1i_2i_3}}))$$ We have used the fact that on the trivialization $U_i\cap U_j\times {\cal H}$, let $\alpha$ and $\alpha'$ be two elements of $Co(e_i)$, ${u_{ij}}^*(\alpha-\alpha')$ is transformed in $Ad({h_{ij}}^{-1})(\alpha-\alpha')$ by the transition functions of $p_{{\cal H}}$, since $\alpha-\alpha'$ is an element of the vector space of the affine space $Co(e_i)$. [**Example.**]{} Consider a normal subgroup $H$ of a Lie group $G$, and a $G/H$-bundle $p_{G/H}$ over the manifold $N$. We have defined a gerbe $C_H$ at page 2. We define the connective structure $Co$ on $C_H$ as follows: for each open subset $U$ of $N$, and an object $e_U$ of $C_H(U)$, $Co(e_U)$ is the set of $1$-forms $\theta':U\rightarrow C({\cal H})$ where $C({\cal H})$ is the center of ${\cal H}$. This definition is natural since if the center of $H$ is trivial, then the gerbe $C_H$ is trivial. The characteristic classes defined below are also trivial. [**Definition.**]{} [**A curving**]{} of a connective structure $Co$ is a correspondence $$D(e_U,):Co(e_U)\rightarrow D(e_U)$$ where $D(e_U)$ is an affine space whose underlying vector space is a set of $p_{\cal H}$ valued $2$-forms which satisfies the following property: (i)- For each morphism $h:e'_U\rightarrow e_U$, $(D(e_U,\nabla))=D(e'_U,h^*\nabla)$. (ii)- If $\alpha$ is a ${p_{{\cal H}}}_{\mid U}$ $1$-form on $U$ such that $\nabla+\alpha$ is an element of $Co(e_U)$, then $$D(e_U,\nabla+\alpha)=D(e_U,\nabla)+d\alpha+\alpha\wedge\alpha$$ The assignment $e_U\rightarrow D(e_U,\nabla)$ is compatible with the restrictions to smaller subsets. [**Characteristic classes.**]{} [**Definition.**]{} A polynomial function of degree $l$ $F:{\cal H}^l\rightarrow {{I\!\! R}}$ is said to be invariant if and only if for every $h\in { H}$ $F(Ad(h))=F$. Let $C_P$ be a gerbe bounded by $P$ endowed with the connective structure $C'_P$, consider $[\Omega]$ the cohomology class of the classifying cocycle $\Omega$ of $C'_P$ identified with a DeRham $3$-form using the Cech-DeRham isomorphism. For every invariant polynomial $P$ of degree $l$ we can define the $3l$-form $P(\Omega)$ by $P\circ \wedge^lL$. The cohomology classes of the forms $P(\Omega)$ are the characteristic classes of the curving. [**Holonomy of non abelian gerbes.**]{} Let $C_P$ be a principal gerbe defined over a manifold $N$ endowed with a connective structure $Co$ and a curving $Cur$. Let $l:N_2\rightarrow N$ be a differentiable map whose domain is the compact surface $N_2$. We can pull-back the gerbe $C_P$, $Co$ and $Cur$ to $N_2$ by using $l$. Let $(U_i)_{i\in I}$ be an open covering of $N_2$ such that $(l^*C_P)(U_i)$ is not empty. Consider an object $e_i$ of $(l^*(C_P)(U_i))$, $\nabla_i$ an element of $l^*(Co)(U_i)$, and $L_i$ the curving of $\nabla_i$. Since $N_2$ is a surface, $L_i$ is exact. We can set $L_i=d(L'_i)$. On $U_{i_1i_2}$ we have $L_{i_2}-L_{i_1}=d(\nabla_{i_1}-u_{i_1i_2}^*\nabla_{i_2})+(\nabla_{i_1}-u_{i_1i_2}^*\nabla_{i_2}) \wedge(\nabla_{i_1}-u_{i_1i_2}^*\nabla_{i_2})$, the form $(\nabla_{i_1}-u_{i_1i_2}^*\nabla_{i_2})\wedge (\nabla_{i_1}-u_{i_1i_2}^*\nabla_{i_2})=d(L'_{i_1i_2})$ This implies that $L'_{i_2}-L'_{i_1}=\nabla_{i_1}-u_{i_1i_2}^*\nabla_{i_2}+L'_{i_1i_2}+ d(L"_{i_1i_2})$. This implies that the Cech boundary $\delta(h_{i_1i_2})$ of $h_{i_1i_2}=\nabla_{i_1}-u_{i_1i_2}^*\nabla_{i_2}+L'_{i_1i_2}$ is a $2$-chain of closed forms. We set $\delta(h_{i_1i_2})=d(C_{i_1i_2i_3})$ The chain $C_{i_1i_2i_3}+\delta(L"_{i_ii_2})$ is a $2$-chain of constant ${\cal H}$-functions. Since $N_2$ is a surface, we can find a cover such that this chain is a cocycle. It suffices to find an open cover $(U_i)_{i\in I}$ such that $U_{i_1i_2i_3i_4}$ is empty and $U_i$ is a $1$-Eilenberg-Mclane space. Thus using the Cech-DeRham isomorphism, we identify this chain to a $l^*(p_{\cal H})$ $2$-form $H$. We define $$Hol(N_2,C_P,Co)=exp(\int_{N_2}H)$$ **Reference.** J.L Brylinski, Loops spaces, Characteristic Classes and Geometric Quantization, Progr. Math. 107, Birkhauser, 1993.

ArXiv

--- author: - | Jacek Miȩkisz\ Institute of Applied Mathematics\ and Mechanics\ Warsaw University\ ul. Banacha 2\ 02-097 Warsaw, Poland\ e-mail: miekisz@mimuw.edu.pl title: Evolutionary game theory and population dynamics --- -2cm -0.2cm Short overview ============== We begin these lecture notes by a crash course in game theory. In particular, we introduce a fundamental notion of a Nash equilibrium. To address the problem of the equilibrium selection in games with multiple equilibria, we review basic properties of the deterministic replicator dynamics and stochastic dynamics of finite populations. We show the almost global asymptotic stability of an efficient equilibrium in the replicator dynamics with a migration between subpopulations. We also show that the stability of a mixed equilibrium depends on the time delay introduced in replicator equations. For large time delays, a population oscillates around its equilibrium. We analyze the long-run behaviour of stochastic dynamics in well-mixed populations and in spatial games with local interactions. We review results concerning the effect of the number of players and the noise level on the stochastic stability of Nash equilibria. In particular, we present examples of games in which when the number of players increases or the noise level decreases, a population undergoes a transition between its equilibria. We discuss similarities and differences between systems of interacting players in spatial games maximizing their individual payoffs and particles in lattice-gas models minimizing their interaction energy. In short, there are two main themes of our lecture notes: the selection of efficient equilibria (providing the highest payoffs to all players) in population dynamics and the dependence of the long-run behaviour of a population on various parameters such as the time delay, the noise level, and the size of the population. Introduction ============ Many socio-economic and biological processes can be modeled as systems of interacting individuals; see for example econophysics bulletin [@ekono] and statistical mechanics and quantitative biology archives [@quantbiol]. One may then try to derive their global behaviour from individual interactions between their basic entities such as animals in ecological and evolutionary models, genes in population genetics and people in social processes. Such approach is fundamental in statistical physics which deals with systems of interacting particles. One can therefore try to apply methods of statistical physics to investigate the population dynamics of interacting individuals. There are however profound differences between these two systems. Physical systems tend in time to states which are characterized by the minimum of some global quantity, the total energy or free energy of the system. Population dynamics lacks such general principle. Agents in social models maximize their own payoffs, animals and genes maximize their individual darwinian fitness. The long-run behavior of such populations cannot in general be characterized by the global or even local maximum of the payoff or fitness of the whole population. We will explore similarities and differences between these systems. The behaviour of systems of interacting individuals can be often described within game-theoretic models [@maynard3; @fuden1; @fuden2; @wei; @vr; @samuel; @hof2; @young2; @gintis; @cress; @ams; @nowak; @nowakscience; @nowakbook]. In such models, players have at their disposal certain strategies and their payoffs in a game depend on strategies chosen both by them and by their opponents. The central concept in game theory is that of a [**Nash equilibrium**]{}. It is an assignment of strategies to players such that no player, for fixed strategies of his opponents, has an incentive to deviate from his current strategy; no change can increase his payoff. In Chapter 3, we present a crash course in game theory. One of the fundamental problems in game theory is the equilibrium selection in games with multiple Nash equilibria. Some two-player symmetric games with two strategies, have two Nash equilibria and it may happen that one of them is payoff dominant (also called efficient) and the other one is risk-dominant. In the efficient equilibrium, players receive highest possible payoffs. The strategy is risk-dominant if it has a higher expected payoff against a player playing both strategies with equal probabilities. It is played by individuals averse to risk. One of the selection methods is to construct a dynamical system where in the long run only one equilibrium is played with a high frequency. John Maynard Smith [@maynard1; @maynard2; @maynard3] has refined the concept of the Nash equilibrium to include the stability of equilibria against mutants. He introduced the fundamental notion of an [**evolutionarily stable strategy**]{}. If everybody plays such a strategy, then the small number of mutants playing a different strategy is eliminated from the population. The dynamical interpretation of the evolutionarily stable strategy was later provided by several authors [@tayjon; @hof1; @zee]. They proposed a system of difference or differential replicator equations which describe the time-evolution of frequencies of strategies. Nash equilibria are stationary points of this dynamics. It appears that in games with a payoff dominant equilibrium and a risk-dominant one, both are asymptotically stable but the second one has a larger basin of attraction in the replicator dynamics. In Chapter 4, we introduce [**replicator dynamics**]{} and review theorems concerning asymptotic stability of Nash equilibria [@wei; @hof2; @ams]. Then in Chapter 5, we present our own model of the replicator dynamics [@migration] with a migration between two subpopulations for which an efficient equilibrium is almost globally asymptotically stable. It is very natural, and in fact important, to introduce a [**time delay**]{} in the population dynamics; a time delay between acquiring information and acting upon this knowledge or a time delay between playing games and receiving payoffs. Recently Tao and Wang [@taowang] investigated the effect of a time delay on the stability of interior stationary points of the replicator dynamics. They considered two-player games with two strategies and a unique asymptotically stable interior stationary point. They proposed a certain form of a time-delay differential replicator equation. They showed that the mixed equilibrium is asymtotically stable if a time delay is small. For sufficiently large delays it becomes unstable. In Chapter 6, we construct two models of discrete-time replicator dynamics with a time delay [@delay]. In the social-type model, players imitate opponents taking into account average payoffs of games played some units of time ago. In the biological-type model, new players are born from parents who played in the past. We consider two-player games with two strategies and a unique mixed Nash equilibrium. We show that in the first type of dynamics, it is asymptotically stable for small time delays and becomes unstable for large ones when the population oscillates around its stationary state. In the second type of dynamics, however, the Nash equilibrium is asymptotically stable for any time delay. Our proofs are elementary, they do not rely on the general theory of delay differential and difference equations. Replicator dynamics models population behaviour in the limit of the infinite number of individuals. However, real populations are finite. Stochastic effects connected with random matchings of players, mistakes of players and biological mutations can play a significant role in such systems. We will discuss various stochastic adaptation dynamics of populations with a fixed number of players interacting in discrete moments of time. In well-mixed populations, individuals are randomly matched to play a game [@kmr; @rvr; @population]. The deterministic selection part of the dynamics ensures that if the mean payoff of a given strategy is bigger than the mean payoff of the other one, then the number of individuals playing the given strategy increases. However, players may mutate hence the population may move against a selection pressure. In spatial games, individuals are located on vertices of certain graphs and they interact only with their neighbours; see for example [@nowak0; @nowak1; @nowak2; @blume1; @ellis1; @young2; @ellis2; @linnor; @doebeli1; @doebeli2; @szabo1; @szabo2; @szabo6; @hauert1; @doebeli3; @hauert2; @hauert3] and a recent review [@szabo8] and references therein. In discrete moments of times, players adapt to their opponents by choosing with a high probability the strategy which is the best response, i.e. the one which maximizes the sum of the payoffs obtained from individual games. With a small probability, representing the noise of the system, they make mistakes. The above described stochastic dynamics constitute ergodic Markov chains with states describing the number of individuals playing respective strategies or corresponding to complete profiles of strategies in the case of spatial games. Because of the presence of random mutations, our Markov chains are ergodic (irreducible and periodic) and therefore they possess unique stationary measures. To describe the long-run behavior of such stochastic dynamics, Foster and Young [@foya] introduced a concept of stochastic stability. A configuration of the system is [**stochastically stable**]{} if it has a positive probability in the stationary measure of the corresponding Markov chain in the zero-noise limit, that is the zero probability of mistakes. It means that in the long run we observe it with a positive frequency along almost any time trajectory. In Chapter 7, we introduce the concept of stochastic stability and present a useful representation of stationary measures of ergodic Markov chains [@freiwen1; @freiwen2; @shub]. In Chapter 8, we discuss populations with random matching of players in [**well-mixed populations**]{}. We review recent results concerning the dependence of the long-run behavior of such systems on the number of players and the noise level. In the case of two-player games with two symmetric Nash equilibria, an efficient one and a risk-dominant one, when the number of players increases, the population undergoes twice a transition between its equilibria. In addition, for a sufficiently large number of individuals, the population undergoes another [**equilibrium transition**]{} when the noise decreases. In Chapter 9, we discuss [**spatial games**]{}. We will see that in such models, the notion of a Nash equilibrium (called there a Nash configuration) is similar to the notion of a ground-state configuration in classical lattice-gas models of interacting particles. We discuss similarities and differences between systems of interacting players in spatial games maximizing their individual payoffs and particles in lattice-gas models minimizing their interaction energy. The concept of stochastic stability is based on the zero-noise limit for a fixed number of players. However, for any arbitrarily low but fixed noise, if the number of players is large enough, the probability of any individual configuration is practically zero. It means that for a large number of players, to observe a stochastically stable configurations we must assume that players make mistakes with extremely small probabilities. On the other hand, it may happen that in the long run, for a low but fixed noise and sufficiently large number of players, the stationary configuration is highly concentrated on an ensemble consisting of one Nash configuration and its small perturbations, i.e. configurations where most players play the same strategy. We will call such configurations [**ensemble stable.**]{} It will be shown that these two stability concepts do not necessarily coincide. We will present examples of spatial games with three strategies where concepts of stochastic stability and ensemble stability do not coincide [@statmech; @physica]. In particular, we may have the situation, where a stochastically stable strategy is played in the long run with an arbitrarily low frequency. In fact, when the noise level decreases, the population undergoes a sharp transition with the coexistence of two equilibria for some noise level. Finally, we discuss the influence of [**dominated strategies**]{} on the long-run behaviour of population dynamics. In Chapter 10, we shortly review other results concerning stochastic dynamics of finite populations. A crash course in game theory ============================= To characterize a game-theoretic model one has to specify players, strategies they have at their disposal and payoffs they receive. Let us denote by $I=\{1,...,n\}$ the set of players. Every player has at his disposal $m$ different strategies. Let $S=\{1,...,m\}$ be the set of strategies, then $\Omega=S^{I}$ is the set of strategy profiles, that is functions assigning strategies to players. The payoff of any player depends not only on his strategy but also on strategies of all other players. If $X \in \Omega$, then we write $X=(X_{i}, X_{-i})$, where $X_{i} \in S$ is a strategy of the i-th player and $X_{-i} \in S^{I-\{i\}}$ is a strategy profile of remaining players. The payoff of the i-th player is a function defined on the set of profiles, $$U_{i}: \Omega \rightarrow R, \; \; i,...,n$$ The central concept in game theory is that of a Nash equilibrium. An assignment of strategies to players is a [**Nash equilibrium**]{}, if for each player, for fixed strategies of his opponents, changing his current strategy cannot increase his payoff. The formal definition will be given later on when we enlarge the set of strategies by mixed ones. Although in many models the number of players is very large (or even infinite as we will see later on in replicator dynamics models), their strategic interactions are usually decomposed into a sum of two-player games. Only recently, there have appeared some systematic studies of truly multi-player games [@kim; @broom; @multi; @tplatk3]. Here we will discuss only two-player games with two or three strategies. We begin with games with two strategies, $A$ and $B$. Payoffs functions can be then represented by $2 \times 2$ payoff matrices. A general payoff matrix is given by A B A a b U = B c d, where $U_{kl}$, $k,l = A, B$, is a payoff of the first (row) player when he plays the strategy $k$ and the second (column) player plays the strategy $l$. We assume that both players are the same and hence payoffs of the column player are given by the matrix transposed to $U$; such games are called symmetric. In this classic set-up of static games (called matrix games or games in the normal form), players know payoff matrices, simultaneously announce (use) their strategies and receive payoffs according to their payoff matrices. We will present now three main examples of symmetric two-player games with two strategies. We begin with an anecdote, then an appropriate game-theoretic model is build and its Nash equilibria are found. [**Example 1 (Stag-hunt game)**]{} Jean-Jacques Rousseau wrote, in his Discourse on the Origin and Basis of Equality among Men, about two hunters going either after a stag or a hare [@ordershook; @fuden1]. In order to get a stag, both hunters must be loyal one to another and stay at their positions. A single hunter, deserting his companion, can get his own hare. In the game-theory language, we have two players and each of them has at his disposal two strategies: Stag (St) and Hare (H). In order to present this example as a matrix game we have to assign some values to animals. Let a stag (which is shared by two hunters) be worth 10 units and a hare 3 units. Then the payoff matrix of this symmetric game is as follows: St H St 5 0 U = H 3 3 It is easy to see that there are two Nash equilibria: $(St,St)$ and $(H,H)$. In a general payoff matrix, if $a>c$ and $d>b$, then both $(A,A)$ and $(B,B)$ are Nash equilibria. If $a+b<c+d$, then the strategy $B$ has a higher expected payoff against a player playing both strategies with the probability $1/2$. We say that $B$ risk dominates the strategy $A$ (the notion of the risk-dominance was introduced and thoroughly studied by Harsányi and Selten [@hs]). If at the same time $a>d$, then we have a selection problem of choosing between the payoff-dominant (Pareto-efficient) equilibrium $(A,A)$ and the risk-dominant $(B,B)$. [**Example 2 (Hawk-Dove game)**]{} Two animals are fighting for a certain territory of a value V. They can be either aggressive (hawk strategy - H) or peaceful (dove strategy - D). When two hawks meet, they accure the cost of fighting $C>V$ and then they split the territory. When two dove meets, they split the territory without a fight. A dove gives up the territory to a hawk. We obtain the following payoff matrix: H D H (V-C)/2 V U = D 0 V/2, The Hawk-Dove game was analyzed by John Maynard Smith [@maynard3]. It is also known as the Chicken game [@russell] or the Snowdrift game [@doebeli3]. It has two non-symmetric Nash equilibria: $(H,D)$ and $(D,H)$. [**Example 3 (Prisoner’s Dilemma)**]{} The following story was discussed by Melvin Dresher, Merill Flood, and Albert Tucker [@axelrod; @poundstone; @sigmund]. Two suspects of a bank robbery are caught and interrogated by the police. The police offers them separately the following deal. If a suspect testifies against his colleague (a strategy of defection - D), and the other does not (cooperation - C), his sentence will be reduced by five years. If both suspects testify, that is defect, they will get the reduction of only one year. However, if they both cooperate and do not testify, their sentence, because of the lack of a hard evidence, will be reduced by three years. We obtain the following payoff matrix: C D C 3 0 U = D 5 1 The strategy $C$ is a [**dominated strategy**]{} - it results in a lower payoff than the strategy $D$, regardless of a strategy used by the other player. Therefore, $(D,D)$ is the unique Nash equilibrium but both players are much better off when they play $C$ - this is the classic Prisoner’s Dilemma. A novel behaviour can appear in games with three strategies. [**Example 4 (Rock-Scissors-Paper game)**]{} In this game, each of two players simultaneously exhibits a sign of either a scissors ($S$), a rock ($R$), or a paper ($P$). The game has a cyclic behaviour: rock crashes scissors, scissors cut paper, and finally paper wraps rock. The payoffs can be given by the following matrix: R S P R 1 2 0 U = S 0 1 2 P 2 0 1 It is easy to verify that this game, because of its cyclic behavior, does not have any Nash equilibria as defined so far. However, we intuitively feel that when we repeat it many times, the only way not to be exploited is to mix randomly strategies, i.e. to choose each strategy with the probability $1/3$. This brings us to a concept of a mixed stategy, a probability mass function on the set of pure strategies $S$. Formally, a [**mixed strategy**]{} $x$ is an element of a simplex $$\Delta=\{x \in R^{m}, 0 \leq x_{k} \leq 1, \sum_{k=1}^{m} x_{k} =1\}.$$ By the support of a mixed strategy $x$ we mean the set of pure strategies with positive probabilities in $x$. Payoffs of mixed strategies are defined as appropriate expected values. In two-player games, a player who uses a mixed strategy $x$ against a player with a mixed strategy $y$ receives a payoff given by $$\sum_{k,l \in S}U_{kl}x_{k}y_{l}.$$ In general n-player games, profiles of strategies are now elements of $\Theta=\Delta^{I}.$ We are now ready to define formally a Nash equilibrium. $X \in \Theta$ is a [**Nash equilibrium**]{} if for every $i \in I$ and every $y \in \Delta$, $$U_{i}(X_{i},X_{-i}) \geq U_{i}(y,X_{-i})$$ In the mixed Nash equilibrium, expected payoffs of all strategies in its support should be equal. Otherwise a player could increase his payoff by increasing the probability of playing a strategy with the higher expected payoff. In two-player games with two strategies, we identify a mixed strategy with its first component, $x = x_{1}.$ Then the expected payoff of $A$ is given by $ax+b(1-x)$ and that of $B$ by $cx+d(1-x)$. $x^{*}= (d-b)/(d-b+a-c)$ for which the above two expected values are equal is a mixed Nash equilibrium or more formally, a profile $(x,x)$ is a Nash equilibrium. In Examples 1 and 2, in addition to Nash equilibria in pure strategies, we have mixed equilibria, $x^{*}=3/5$ and $x^{*}=V/C$ respectively. It is obvious that the Prisoner’s Dilemma game does not have any mixed Nash equilibria . On the other hand, the only Nash equilibrium of the Rock-Scissors-Paper game is a mixed one assigning the probability $1/3$ to each strategy. We end this chapter by a fundamental theorem due to John Nash [@nash1; @nash2]. Every game with a finite number of players and a finite number of strategies has at least one Nash equilibrium. In any Nash equilibrium, every player uses a strategy which is a best reply to the profile of strategies of remaining players. Therefore a Nash equilibrium can be seen as a best reply to itself - a fixed point of a certain best-reply correspondence. Then one can use the Kakutani fixed point theorem to prove the above theorem. Replicator dynamics =================== The concept of a Nash equilibrium is a static one. Here we will introduce the classical replicator dynamics and review its main properties [@wei; @hof2; @ams]. Replicator dynamics provides a dynamical way of achieving Nash equilibria in populations. We will see that Nash equilibria are stationary points of such dynamics and some of them are asymptotically stable. Imagine a finite but a very large population of individuals. Assume that they are paired randomly to play a symmetric two-player game with two strategies and the payoff matrix given in the beginning of the previous chapter. The complete information about such population is provided by its strategy profile, that is an assignment of pure strategies to players. Here we will be interested only in the proportion of individuals playing respective strategies. We assume that individuals receive average payoffs with respect to all possible opponents - they play against the average strategy. Let $r_{i}(t)$, $i=A, B,$ be the number of individuals playing the strategy $A$ and $B$ respectively at the time $t$. Then $r(t)=r_{A}(t)+r_{B}(t)$ is the total number of players and $x(t)=\frac{r_{1}(t)}{r(t)}$ is a fraction of the population playing $A$. We assume that during the small time interval $\epsilon$, only an $\epsilon$ fraction of the population takes part in pairwise competitions, that is plays games. We write $$r_{i}(t + \epsilon) = (1-\epsilon)r_{i}(t) + \epsilon r_{i}(t)U_{i}(t); \; \; i= A,B,$$ where $U_{A}(t)= ax(t)+b(1-x(t))$ and $U_{B}(t)= cx(t)+d(1-x(t))$ are average payoffs of individuals playing A and B respectively. We assume that all payoffs are not smaller than $0$ hence $r_{A}$ and $r_{B}$ are always non-negative and therefore $0\leq x \leq 1$. The equation for the total number of players reads $$r(t + \epsilon) = (1-\epsilon)r(t) + \epsilon r(t)\bar{U}(t),$$ where $\bar{U}(t)=x(t)U_{A}(t)+(1-x(t))U_{B}(t)$ is the average payoff in the population at the time $t$. When we divide (1) by (2) we obtain an equation for the frequency of the strategy $A$, $$x(t + \epsilon) - x(t) = \epsilon\frac{x(t)[U_{A}(t) - \bar{U}(t)]} {1-\epsilon + \epsilon \bar{U}(t)}.$$ Now we divide both sides of (3) by $\epsilon$, perform the limit $\epsilon \rightarrow 0$, and obtain the well known differential replicator equation: $$\frac{dx(t)}{dt}=x(t)[U_{A}(t) - \bar{U}(t)].$$ The above equation can also be written as $\frac{dx(t)}{dt}=x(t)(1-x(t))[U_{A}(t) - U_{B}(t)]$ $$= (a-c+d-b)x(t)(1-x(t))(x(t)-x^{*})$$ For games with $m$ strategies we obtain a system of $m$ differential equations for $x_{k}(t)$, fractions of the population playing the $k$-th strategy at the time $t$, $k=1,...,m$, $$\frac{dx_{k}(t)}{dt}=x_{k}(t)[\sum_{l=1}^{m}U_{kl}x_{l}(t) - \sum_{k,l=1}^{m}U_{kl}x_{k}(t)x_{l}(t)],$$ where on the right hand-size of (6) there is a difference of the average payoff of the $k$-th strategy and the average payoff of the population. The above system of differential equations or analogous difference equations, called replicator dynamics was proposed in [@tayjon; @hof1; @zee]. For any initial condition $x^{0} \in \Delta$, it has the unique global solution, $\xi(x^{0},t)$, which stays in the simplex $\Delta$. Now we review some theorems relating replicator dynamics and Nash equilibria [@wei; @hof2; @ams]. We consider symmetric two-player games. We denote the set of strategies corresponding to symmetric Nash equilibria by $$\Delta^{NE}= \{x \in \Delta: (x,x) \; is \; a \; Nash \; equilibrium\}.$$ It follows from the definition of the Nash equilibrium (see also discussion in the previous chapter concerning mixed strategies) that $$\Delta^{NE}= \{x \in \Delta: u(i,x)=\max_{z \in \Delta}u(z,x) \; for \; every \; i \; in \; the \; support \; of \; x\}.$$ It is easy to see that $$\Delta^{0}= \{x \in \Delta: u(i,x)= u(x,x) \; for \; every \; i \; in \; the \; support \; of \; x\}$$ is the set of stationary points of the replicator dynamics. It follows that symmetric Nash equilibria are stationary points of the replicator dynamics. $S \cup \Delta^{NE} \subset \Delta^{0}$ The following two theorems relate stability of stationary points to Nash equilibria [@wei; @hof2]. If $x \in \Delta$ is Lyapunov stable, then $x \in \Delta^{NE}.$ If $x^{0} \in interior(\Delta)$ and $\xi(x^{0},t) \rightarrow_{t \rightarrow \infty} x$, then $x \in \Delta^{NE}.$ Below we present the replicator dynamics in the examples of two-player games discussed in the previous chapter. We write replicator equations and show their phase diagrams. [**Stag-hunt game**]{} $$\frac{dx}{dt}=x(1-x)(5x-3)$$ $$\bullet<-----------\bullet------->\bullet$$ $$\hspace{2mm} 0 \hspace{42mm} 3/5 \hspace{27mm} 1$$ [**Hawk-Dove game**]{} $$\frac{dx}{dt}=-x(1-x)(x-C/V)$$ $$\bullet----------->\bullet<-------\bullet$$ $$\hspace{1mm} 0 \hspace{41mm} V/C \hspace{25mm} 1$$ [**Prisoner’s Dilemma**]{} $$\frac{dx}{dt}=-x(1-x)(x+1)$$ $$\bullet<-------------------- \bullet$$ $$\hspace{1mm} 0 \hspace{30mm} \hspace{45mm} 1$$ We see that in the Stag-hunt game, both pure Nash equilibria are asymptotically stable. The risk-dominant one has the larger basin of attraction which is true in general because $x^{*}=(d-b)/(d-b+a-c)>1/2$ for games with an efficient equilibrium and a risk dominant one. In the Hawk-Dove game, the unique symmetric mixed Nash equilibrium is asymptotically stable. In the Prisoner’s Dilemma, the strategy of defection is globally asymptotically stable. In the Rock-Scissors-Paper game, a more detailed analysis has to be done. One can show, by straithforward computations, that the time derivative of $lnx_{1}x_{2}x_{3}$ is equal to zero. Therefore $lnx_{1}x_{2}x_{3}=c$ is an equation of a closed orbit for any constant $c$. The stationary point $(1/3,1/3,1/3)$ of the replicator dynamics is Lyapunov stable and the population cycles on a closed trajectory (which depends on the initial condition) around its Nash equilibrium. Replicator dynamics with migration ================================== We discuss here a game-theoretic dynamics of a population of replicating who can migrate between two subpopulations or habitats [@migration]. We consider symmetric two-player games with two strategies: $A$ and $B$. We assume that $a>d>c$, $d>b$, and $a+b<c+d$ in a general payoff matrix given in the beginning of Chapter 3. Such games have two Nash equilibria: the efficient one $(A,A)$ in which the population is in a state with a maximal fitness (payoff) and the risk-dominant $(B,B)$ where players are averse to risk. We show that for a large range of parameters of our dynamics, even if the initial conditions in both habitats are in the basin of attraction of the risk-dominant equilibrium (with respect to the standard replication dynamics without migration), in the long run most individuals play the efficient strategy. We consider a large population of identical individuals who at each time step can belong to one of two different non-overlapping subpopulations or habitats which differ only by their replication rates. In both habitats, they take part in the same two-player symmetric game. Our population dynamics consists of two parts: the standard replicator one and a migration between subpopulations. Individuals are allowed to change their habitats. They move to a habitat in which the average payoff of their strategy is higher; they do not change their strategies. Migration helps the population to evolve towards an efficient equilibrium. Below we briefly describe the mechanism responsible for it. If in a subpopulation, the fraction of individuals playing the efficient strategy $A$ is above its unique mixed Nash equilibrium fraction, then the expected payoff of $A$ is bigger than that of $B$ in this subpopulation, and therefore the subpopulation evolves to the efficient equilibrium by the replicator dynamics without any migration. Let us assume therefore that such fraction is below the Nash equilibrium in both subpopulations. Without loss of generality we assume that initial conditions are such that the fraction of individuals playing $A$ is bigger in the first subpopulation than in the second one. Hence the expected payoff of $A$ is bigger in the first subpopulation than in the second one, and the expected payoff of $B$ is bigger in the second subpopulation than in the first one. This implies that a fraction of $A$-players in the second population will switch to the first one and at the same time a fraction of $B$-players from the first population will switch to the second one - migration causes the increase of the fraction of individual of the first population playing $A$. However, any $B$-player will have more offspring than any $A$-player (we are below a mixed Nash equilibrium) and this has the opposite effect on relative number of $A$-players in the first population than the migration. The asymptotic composition of the whole population depends on the competition between these two processes. We derive sufficient conditions for migration and replication rates such that the whole population will be in the long run in a state in which most individuals occupy only one habitat (the first one for the above described initial conditions) and play the efficient strategy. Let $\epsilon$ be a time step. We allow two subpopulations to replicate with different speeds. We assume that during any time-step $\epsilon$, a fraction $\epsilon$ of the first subpopulation and a fraction $\kappa\epsilon$ of the second subpopulation plays the game and receives payoffs which are interpreted as the number of their offspring. Moreover, we allow a fraction of individuals to migrate to a habitat in which their strategies have higher expected payoffs. Let $r^i_s$ denote the number of individuals which use the strategy $s \in \{A, B\}$ in the subpopulation $i \in \{1, 2\}$. By $U^i_s$ we denote the expected payoff of the strategy $s$ in the subpopulation $i$: $$U^1_A = ax + b(1-x), \quad U^1_B = cx + d(1-x),$$ $$U^2_A = ay + b(1-y), \quad U^2_B = cy + d(1-y),$$ where $$x= {r^1_A\over r_1}, \quad y = {r^2_A\over r_2}, \quad r_1 = r^1_A + r^1_B, \quad r_2 = r^2_A + r^2_B;$$ $x$ and $y$ denote fractions of $A$-players in the first and second population respectively. We denote by $\alpha = {r_1\over r}$ the fraction of the whole population in the first subpopulation, where $r = r_1 + r_2$ is the total number of individuals. The evolution of the number of individuals in each subpopulation is assumed to be a result of the replication and the migration flow. In our model, the direction and intensity of migration of individuals with a given strategy will be determined by the difference of the expected payoffs of that strategy in both habitats. Individuals will migrate to a habitat with a higher payoff. The evolution equations for the number of individuals playing the strategy $s$, $s \in \{A, B\}$, in the habitat $i$, $i \in \{1, 2\}$, have the following form: $$r^1_A(t+\epsilon) = R^1_A + \Phi_A,$$ $$r^1_B(t+\epsilon) = R^1_B + \Phi_B,$$ $$r^2_A(t+\epsilon) = R^2_A - \Phi_A,$$ $$r^2_B(t+\epsilon) = R^2_B - \Phi_B,$$ where all functions on the right-hand sides are calculated at the time $t$. Functions $R^i_s$ describe an increase of the number of the individuals playing the strategy $s$ in the subpopulation $i$ due to the replication: $$R^1_s= (1-\epsilon) r^1_s + \delta U^1_s r^1_s,$$ $$R^2_s= (1-\kappa \epsilon) r^2_s + \kappa \epsilon U^2_s r^2_s,$$ The rate of the replication of individuals playing the strategy $s$ in the first subpopulation is given by $\epsilon U^1_s$ , and in the second subpopulation by $\kappa \epsilon U^2_s$. The parameter $\kappa$ measures the difference of reproduction speeds in both habitats. Functions $\Phi_s$, $s \in \{A, B\}$, are defined by $$\Phi_s = \epsilon \gamma (U^1_s - U^2_s)[r^2_s \Theta (U^1_s - U^2_s) + r^1_s \Theta (U^2_s-U^1_s)],$$ where $\Theta$ is the Heaviside’s function, $$\Theta(x) = \cases {1, \ \ x\ge 0; \cr 0, \ \ x<0 \cr}$$ and $\gamma$ is the migration rate. Functions $\Phi_s$ describe changes of the numbers of the individuals playing strategy $s$ in the relevant habitat due to migration. $\Phi_s$ will be referred to as the migration of individuals (who play the strategy $s$) between two habitats. Thus, if for example $U^1_A > U^2_A$, then there is a migration of individuals with the strategy $A$ from the second habitat to the first one: $$\Phi_A = \delta \gamma r^2_A (U^1_A - U^2_A),$$ and since then necessarily $U^1_B < U^2_B$ \[note that $U^1_A - U^2_A = (a-b)(x-y) \hspace{3mm} and \hspace{3mm} U^1_B - U^2_B = (c-d)(x-y)$\], there is a migration flow of individuals with strategy $B$ from the first habitat to the second one: $$\Phi_B = \epsilon \gamma r^1_B (t) (U^1_B - U^2_B).$$ In this case, the migration flow $\Phi_A$ describes the increase of the number of individuals which play the strategy $A$ in the first subpopulation due to migration of the individuals playing $A$ in the second subpopulation. This increase is assumed to be proportional to the number of individuals playing $A$ in the second subpopulation and the difference of payoffs of this strategy in both subpopulations. The constant of proportionality is $\epsilon$ times the migration rate $\gamma$. The case $\gamma=0$ corresponds to two separate populations which do not communicate and evolve independently. Our model reduces then to the standard discrete-time replicator dynamics. In this case, the total number of players who use a given strategy changes only due to the increase or decrease of the strategy fitness, as described by functions defined in (11-12). In the absence of the replication, there is a conservation of the number of individuals playing each strategy in the whole population. This corresponds to our model assumption that individuals can not change their strategies but only habitats in which they live. For $U^1_A > U^2_A$ we obtain from (7-10) equations for $r_{i}(t)$ and $r(t)$: $r_{1}(t+\epsilon)=(1-\epsilon)r_{1}(t)$ $$+ \delta r_{1}(t)[\frac{r_{A}^{1}U_{A}^{1}+r_{B}^{1}U_{B}^{1}}{r_{1}}+ \gamma\frac{r^{2}_{A}(U_{A}^{1}-U_{A}^{2})+r^{1}_{B}(U_{B}^{1}-U_{B}^{2})}{r_{1}}],$$ $r_{2}(t+\epsilon)=(1-\kappa \epsilon)r_{2}(t)$ $$+ \delta r_{2}(t)[\kappa \frac{r_{A}^{2}U_{A}^{2}+r_{B}^{2}U_{B}^{2}}{r_{2}}+ \gamma\frac{r^{2}_{A}(U_{A}^{2}-U_{A}^{1})+r^{1}_{B}(U_{B}^{2}-U_{B}^{1})}{r_{2}}],$$ $r(t+\delta)= (1-\epsilon)r_{1}(t)+ (1-\kappa \delta)r_{2}(t)$ $$+ \delta r(t)[\alpha ({\frac{r^1_A}{r_1}}U_{A}^{1}+ {\frac{r^1_B}{r_1}} U_{B}^{1})+ (1-\alpha)\kappa ( {\frac{r^2_A}{r_2}} U_{A}^{2}+ {\frac{r^2_B}{r_2}} U_{B}^{2})],$$ where all functions in square brackets depend on $t$. Now, like in the derivation of the standard replicator dynamics, we consider frequencies of individuals playing the relevant strategies in both habitats. Thus, we focus on the temporal evolution of the frequencies, $x$ and $y$, and the relative size of the first subpopulation, $\alpha$. We divide (7) by (17), (9) by (18), and (17) by (19). Performing the limit $\epsilon \rightarrow 0$ we obtain the following differential equations: $\frac{dx}{dt}=x[(1-x)(U_{A}^{1}-U_{B}^{1})$ $$+ \gamma[(\frac{y(1-\alpha)}{x\alpha}-\frac{y(1-\alpha)}{\alpha}) (U_{A}^{1}-U_{A}^{2})-(1-x)(U_{B}^{1}-U_{B}^{2})]],$$ $$\frac{dy}{dt}=y[\kappa(1-y)(U_{A}^{2}-U_{B}^{2})+\gamma[(1-y)(U_{A}^{2}-U_{A}^{1})- \frac{(1-x)\alpha}{1-\alpha}(U_{B}^{2}-U_{B}^{1})]],$$ $\frac{d\alpha}{dt}=\alpha(1-\alpha)[xU_{A}^{1}+(1-x)U_{B}^{1}-(yU_{A}^{2}+(1-y)U_{B}^{2})]$ $+ \alpha \gamma[\frac{y(1-\alpha)}{\alpha}(U_{A}^{1}-U_{A}^{2})+(1-x)(U_{B}^{1}-U_{B}^{2})]$ $$+ \alpha(1-\alpha)(\kappa-1)(1-yU^{2}_{A}-(1-y)U^{2}_{B}).$$ Similar equations are derived for the case $U^1_A < U^2_A$ (since our model is symmetric with respect to the permutation of the subpopulations, it is enough to renumerate the relevant indices and redefine the parameter $\kappa$). Assume first that $U_{A}^{1}(0)>U_{A}^{2}(0)$, which is equivalent to $x(0)>y(0)$. It follows from (7-10) that a fraction of $A$-players from the subpopulation $2$ will migrate to the subpopulation $1$ and a fraction of $B$-players will migrate in the opposite direction. This will cause $x$ to increase and $y$ to decrease. However, if $x(0)<x^{*}$ and $y(0)<x^{*}$, then $U_{A}^{1}<U_{B}^{1}$ and $U_{A}^{2}<U_{B}^{2}$, therefore $B$-players will have more offspring than $A$-players. This has the opposite effect on the relative number of $A$-players in the first subpopulation than migration. If $x(0)<y(0)$, then migration takes place in the reverse directions. The outcome of the competition between migration and replication depends, for a given payoff matrix, on the relation between $x(0)-y(0)$, $\gamma$ and $\kappa$. We are interested in formulating sufficient conditions for the parameters of the model, for which most individuals of the whole population will play in the long run the efficient strategy $A$. We prove the following theorem [@migration]. If $$\gamma [ x(0)-y(0) ] > max [\frac {d-b} {d-c}, \frac {\kappa (a-c)}{a-b}],$$ then $x(t) \rightarrow_{t \rightarrow \infty}1$ and $y(t) \rightarrow_{t \rightarrow \infty} 0$. If $\kappa<(a-1)/(d-1)$, then $\alpha(t) \rightarrow_{t \rightarrow \infty} 1$. 0.2cm If $$\gamma [ y(0)-x(0) ] > max [\frac {\kappa (d-b)} {d-c}, \frac {a-c} {a-b}],$$ then $x(t) \rightarrow_{t \rightarrow \infty}0$ and $y(t) \rightarrow_{t \rightarrow \infty} 1$. If $\kappa > (d-1)/(a-1)$, then $\alpha(t) \rightarrow_{t \rightarrow \infty} 0.$ [**Proof**]{}: Assume first that $x(0)>y(0)$. From (20-21) we get the following differential inequalities: $$\frac{dx}{dt} > x(1-x)[U_{A}^{1}-U_{B}^{1})+\gamma(U_{B}^{2}-U_{B}^{1})],$$ $$\frac{dy}{dt} < y(1-y)[\kappa(U_{A}^{2}-U_{B}^{2})+\gamma(U_{A}^{2}-U_{A}^{1})],$$ Using explicit expressions for $U_{s}^{i}$ we get $$\frac{dx}{dt} > x(1-x)[(a-c+d-b)x+b-d+\gamma(d-c)(x-y)],$$ $$\frac{dy}{dt} < y(1-y)[\kappa[(a-c+d-b)y+b-d]-\gamma(a-b)(x-y)],$$ We note that if $\gamma(d-c)(x(0)-y(0)) > d - b$ then $\gamma(d-c)(x(0)-y(0))+ b - d + (a-c+d-b)x(0) > 0,$ i.e. $dx/dt(0)>0$. Analogously, if $\gamma(a-b)(x(0)-y(0))> \kappa (a-c)$, then $\gamma(a-b)(x(0)-y(0))> \kappa [(a-c+d-b) + b-d] > \kappa [(a-c+d-b)y(o) + b-d],$ therefore $dy/dt(0)<0$. Thus, combining both conditions we conclude that $x(t)-y(t)$ is an increasing function so $x(t)>y(t)$ for all $t \geq 0$, hence we may use (20-22) all the time. We get that $x(t) \rightarrow_{t \rightarrow \infty}1$ and $y(t) \rightarrow_{t \rightarrow \infty} 0$, and the first part of the thesis follows. Now from (22) it follows that if $a-d+(\kappa-1)(1-d)>0$, i.e. $\kappa< (a-1)/(d-1)$, then $\alpha(t) \rightarrow_{t \rightarrow \infty} 1$. The second part of Theorem 5, corresponding to initial conditions $y(0)>x(0),$ can be proved analogously, starting from eqs. (7-10) written for the case $U^1_A(0) < U^2_A(0)$ and their continuous counterparts. We omit details. The above conditions for $\kappa$ mean that the population consisting of just $A$-players replicates faster (exponentially in $(a-1)t$) than the one consisting of just $B$-players (exponentially in $(d-1)\kappa t$). The same results would follow if the coefficients of the payoff matrix of the game played in one habitat would differ from those in the second habitat by an additive constant. We showed that introduction of the mechanism of attraction by the habitat with a higher expected payoff in the standard replicator dynamics helps the whole population to reach the state in which in the long run most individuals play the efficient strategy. More precisely, we proved that for a given rate of migration, if the fractions of individuals playing the efficient strategy in both habitats are not too close to each other, then the habitat with a higher fraction of such players overcomes the other one in the long run. The fraction of individuals playing the efficient strategy tends to unity in this habitat and consequently in the whole population. Alternatively, we may say that the bigger the rate of migration is, larger is the basin of attraction of the efficient equilibrium. In particular, we showed that for a large range of parameters of our dynamics, even if the initial conditions in both habitats are in the basin of attraction of the risk-dominant equilibrium (with respect to the standard replication dynamics without migration), in the long run most individuals play the efficient strategy. Replicator dynamics with time delay =================================== Here we consider two-player games with two strategies, two pure non-symmetric Nash equilibria, and a unique symmetric mixed one, that is $a<c$ and $d<b$ in a general payoff matrix given in the beginning of Chapter 3. Let us recall that the Hawk-Dove game is of such type. Recently Tao and Wang [@taowang] investigated the effect of a time delay on the stability of the mixed equilibrium in the replicator dynamics. They showed that it is asymptotically stable if a time delay is small. For sufficiently large delays it becomes unstable. We construct two models of discrete-time replicator dynamics with a time delay [@delay]. In the social-type model, players imitate opponents taking into account average payoffs of games played some units of time ago. In the biological-type model, new players are born from parents who played in the past. We show that in the first type of dynamics, the unique symmetric mixed Nash equilibrium is asymptotically stable for small time delays and becomes unstable for large ones when the population oscillates around its stationary state. In the second type of dynamics, however, the Nash equilibrium is asymptotically stable for any time delay. Our proofs are elementary, they do not rely on the general theory of delay differential and difference equations. Social-type time delay ---------------------- Here we assume that individuals at time $t$ replicate due to average payoffs obtained by their strategies at time $t-\tau$ for some delay $\tau>0$ (see also a discussion after (32)). As in the standard replicator dynamics, we assume that during the small time interval $\epsilon$, only an $\epsilon$ fraction of the population takes part in pairwise competitions, that is plays games. Let $r_{i}(t)$, $i=A, B,$ be the number of individuals playing at the time $t$ the strategy $A$ and $B$ respectively, $r(t)=r_{A}(t)+r_{B}(t)$ the total number of players and $x(t)=\frac{r_{1}(t)}{r(t)}$ a fraction of the population playing $A$. We propose the following equations: $$r_{i}(t + \epsilon) = (1-\epsilon)r_{i}(t) + \epsilon r_{i}(t)U_{i}(t-\tau); \; \; i= A,B.$$ Then for the total number of players we get $$r(t + \epsilon) = (1-\epsilon)r(t) + \epsilon r(t)\bar{U}_{o}(t-\tau),$$ where $\bar{U}_{o}(t-\tau)=x(t)U_{A}(t-\tau)+(1-x(t))U_{B}(t-\tau).$ We divide (27) by (28) and obtain an equation for the frequency of the strategy $A$, $$x(t + \epsilon) - x(t) = \epsilon \frac{x(t)[U_{A}(t-\tau) - \bar{U}_{o}(t-\tau)]} {1-\epsilon + \epsilon \bar{U}_{o}(t-\tau)}$$ and after some rearrangements we get $$x(t + \epsilon) - x(t) = -\epsilon x(t)(1-x(t))[x(t-\tau)-x^{*}]\frac{\delta}{1-\epsilon + \epsilon \bar{U}_{o}(t-\tau)},$$ where $x^{*}= (d-b)/(d-b+a-c)$ is the unique mixed Nash equilibrium of the game. Now the corresponding replicator dynamics in the continuous time reads $$\frac{dx(t)}{dt}=x(t)[U_{A}(t-\tau) - \bar{U}_{o}(t-\tau)]$$ and can also be written as $\frac{dx(t)}{dt}=x(t)(1-x(t))[U_{A}(t-\tau) - U_{B}(t-\tau)]$ $$= -\delta x(t)(1-x(t))(x(t-\tau)-x^{*}).$$ The first equation in (32) can be also interpreted as follows. Assume that randomly chosen players imitate randomly chosen opponents. Then the probability that a player who played $A$ would imitate the opponent who played $B$ at time $t$ is exactly $x(t)(1-x(t)).$ The intensity of imitation depends on the delayed information about the difference of corresponding payoffs at time $t- \tau$. We will therefore say that such models have a social-type time delay. Equations (31-32) are exactly the time-delay replicator dynamics proposed and analyzed by Tao and Wang [@taowang]. They showed that if $\tau< c-a+b-d \pi /2(c-a)(b-d)$, then the mixed Nash equilibrium, $x^{*}$, is asymptotically stable. When $\tau$ increases beyond the bifurcation value $c-a+b-d \pi /2(c-a)(b-d)$, $x^{*}$ becomes unstable. We have the following theorem [@delay]. $x^{*}$ is asymptotically stable in the dynamics (30) if $\tau$ is sufficiently small and unstable for large enough $\tau$. [**Proof:**]{} We will assume that $\tau$ is a multiple of $\epsilon$, $\tau=m\epsilon$ for some natural number $m$. Observe first that if $x(t - \tau) < x^{*}$, then $x(t + \epsilon) > x(t)$, and if $x(t - \tau) > x^{*}$, then $x(t + \epsilon) < x(t)$. Let us assume first that there is $t'$ such that $x(t'), x(t'-\epsilon), x(t'-2\epsilon),..., x(t'-\tau) < x^{*}$. Then $x(t)$, $t \geq t'$ increases up to the moment $t_{1}$ for which $x(t_{1} - \tau) > x^{*}$. If such $t_{1}$ does not exist then $x(t) \rightarrow_{t \rightarrow \infty}x^{*}$ and the theorem is proved. Now we have $x^{*}<x(t_{1}-\tau) < x(t_{1}-\tau+\epsilon)< \ldots <x(t_{1})$ and $x(t_{1}+\epsilon)<x(t_{1})$ so $t_{1}$ is a turning point. Now $x(t)$ decreases up to the moment $t_{2}$ for which $x(t_{2}-\tau) < x^{*}$. Again, if such $t_{2}$ does not exist, then the theorem follows. Therefore let us assume that there is an infinite sequence, $t_{i}$, of such turning points. Let $\eta_i = |x(t_{i}) - x^{*}|$. We will show that $\eta_i \rightarrow_{i \rightarrow \infty} 0$. For $t \in \{t_{i}, t_{i} + \epsilon, \ldots, t_{i + 1}-1\}$ we have the following bound for $x(t + \epsilon) - x(t)$: $$|x(t + \epsilon) - x(t)| < \frac{1}{4} \eta_{i} \frac{\epsilon \delta} {1 -\epsilon + \epsilon \bar{U}_{o}(t - \tau)}.$$ This means that $$\eta_{i+1} < (m+1)\epsilon K \eta_{i},$$ where $K$ is the maximal possible value of $\frac{\delta}{4(1-\epsilon + \epsilon \bar{U}_{o}(t - \tau))}.$ We get that if $$\tau < \frac{1}{K}-\epsilon,$$ then $\eta_{i}\rightarrow_{i \rightarrow \infty} 0$ so $x(t)$ converges to $x^{*}$. Now if for every $t$, $|x(t + \epsilon)-x^{*}| < \max_{k\in \{0,1,...,m\}}|x(t - k\epsilon)-x^{*}|$, then $x(t)$ converges to $x^{*}$. Therefore assume that there is $t''$ such that $|x(t'' + \epsilon)-x^{*}| \geq \max_{k\in \{0,1,...,m\}}|x(t'' - k\epsilon)-x^{*}|$. If $\tau$ satisfies (35), then it follows that $x(t+\epsilon),...,x(t+\epsilon+\tau)$ are all on the same side of $x^{*}$ and the first part of the proof can be applied. We showed that $x(t)$ converges to $x^{*}$ for any initial conditions different from $0$ and $1$ hence $x^{*}$ is globally asymptotically stable. Now we will show that $x^{*}$ is unstable for any large enough $\tau$. Let $\gamma>0$ be arbitrarily small and consider a following perturbation of the stationary point $x^{*}$: $x(t)=x^{*}, t\leq 0$ and $x(\epsilon) = x^{*}+\gamma$. It folows from (30) that $x(k\epsilon)=x(\epsilon)$ for $k=1,...,m+1$. Let $K'=\min_{x \in[x^{*}-\gamma,x^{*}+\gamma]} \frac{x(1-x)\delta}{4(1-\epsilon + \epsilon \bar{U}_{o}(t - \tau))}$. If $\frac{m}{2}\epsilon K'\gamma > 2\gamma$, that is $\tau > \frac{4}{K'}$, then it follows from (30) that after $m/2$ steps (we assume without loss of generality that $m$ is even) $x((m+1+m/2)\epsilon)<x^{*}-\gamma$. In fact we have $x((2m+1)\epsilon)< \ldots < x((m+1)\epsilon)$ and at least $m/2$ of $x's$ in this sequence are smaller than $x^{*}-\gamma$. Let $\bar{t}>(2m+1)\epsilon$ be the smallest $t$ such that $x(t)>x^{*}-\gamma$. Then we have $x(\bar{t}-m \epsilon), \ldots, x(\bar{t}-\epsilon) < x^{*}-\gamma < x(\bar{t})$ hence after $m/2$ steps, $x(t)$ crosses $x^{*}+\gamma$ and the situation repeats itself. We showed that if $$\tau >\frac{4}{K'},$$ then there exists an infinite sequence, $\tilde{t}_{i}$, such that $|x(\tilde{t}_{i})-x^{*}|>\gamma$ and therefore $x^{*}$ is unstable. Moreover, $x(t)$ oscillates around $x^{*}$. Biological-type time delay -------------------------- Here we assume that individuals born at time $t-\tau$ are able to take part in contests when they become mature at time $t$ or equivalently they are born $\tau$ units of time after their parents played and received payoffs. We propose the following equations: $$r_{i}(t + \epsilon) = (1-\epsilon)r_{i}(t) + \epsilon r_{i}(t-\tau)U_{i}(t-\tau); \; \; i= A,B.$$ Then the equation for the total number of players reads $r(t + \epsilon) = (1-\epsilon)r(t)$ $$+ \epsilon r(t)[\frac{x(t)r_{A}(t-\tau)}{r_{A}(t)}U_{A}(t-\tau) +\frac{(1-x(t))r_{B}(t-\tau)}{r_{B}(t)}U_{B}(t-\tau)].$$ We divide (37) by (38) and obtain an equation for the frequency of the first strategy, $$x(t + \epsilon) - x(t) = \epsilon \frac{x(t - \tau)U_{A}(t - \tau) - x(t)\bar{U}(t - \tau)} {(1-\epsilon)\frac{r(t)}{r(t-\tau)} + \epsilon \bar{U}(t-\tau)},$$ where $\bar{U}(t-\tau)=x(t-\tau)U_{A}(t-\tau)+(1-x(t-\tau))U_{B}(t-\tau).$ We proved in [@delay] the following $x^{*}$ is asymptotically stable in the dynamics (39) for any value of the time delay $\tau$. We begin by showing our result in the following simple example. The payoff matrix is given by $U = \left(\begin{array}{cc} 0 & 1\\ 1 & 0 \end{array}\right)$ hence $x^{*} = \frac{1}{2}$ is the mixed Nash equilibrium which is asymptotically stable in the replicator dynamics without the time delay. The equation (39) now reads $$x(t+\epsilon)-x(t)=\epsilon \frac{x(t-\tau)(1 - x(t-\tau))-2 x(t)x(t-\tau)(1-x(t-\tau))} {(1-\epsilon)\frac{r(t)}{r(t-\tau)}+2\epsilon x(t-\tau)(1-x(t-\tau))}$$ After simple algebra we get $x(t+ \epsilon) - \frac{1}{2} + \frac{1}{2} - x(t)$ $$= \epsilon (1 - 2x(t)) \frac{x(t-\tau)(1-x(t - \tau))} {(1-\epsilon)\frac{r(t)}{r(t - \tau)} + 2\epsilon x(t - \tau)(1 - x (t - \tau))},$$ so $$x(t + \epsilon) - \frac{1}{2} = (x(t)- \frac{1}{2}) \frac{1} {1 + \frac{\epsilon r(t - \tau)}{(1-\epsilon)r(t)}2x(t - \tau)(1- x(t - \tau))}$$ hence $$|x(t+\epsilon)-\frac{1}{2}| < |x(t)-\frac{1}{2}|.$$ It follows that $x^{*}$ is globally asymptotically stable. Now we present the proof for the general payoff matrix with a unique symmetric mixed Nash equilibrium. [**Proof of Theorem 7:**]{} Let $c_{t} = \frac{x(t)U_{A}(t)}{\bar{U}(t)}$. Observe that if $x(t) < x^{*}$, then $c_{t} > x(t)$, if $x(t) > x^{*}$, then $c_{t} < x(t)$, and if $x(t) = x^{*}$, then $c_{t} = x^{*}$. We can write (39) as $x(t + \epsilon) - x(t)$ $$= \epsilon\frac{x(t - \tau)U_{A}(t - \tau) - c_{t - \tau} \bar{U}(t - \tau) + c_{t - \tau} \bar{U}(t - \tau) - x(t) \bar{U}(t - \tau)} {(1-\epsilon)\frac{p(t)}{p(t - \tau)} + \epsilon \bar{U}(t - \tau)}$$ and after some rearrangements we obtain $$x(t + \epsilon) - c_{t - \tau} = (x(t) - c_{t - \tau}) \frac{1} {1 + \frac{\epsilon p(t - \tau)}{(1-\epsilon) p(t)} \bar{U}(t - \tau)}.$$ We get that at time $t + \epsilon$, $x$ is closer to $c_{t -\tau}$ than at time $t$ and it is on the same side of $c_{t -\tau}$. We will show that $c$ is an increasing or a constant function of $x$. Let us calculate the derivative of $c$ with respect to $x$. $$c' = \frac{f(x)}{(xU_{A} + (1 - x)U_{B})^2},$$ where $$f(x)= (ac + bd - 2ad)x^{2}+ 2d(a-b)x+bd.$$ A simple analysis shows that $f>0$ on $(0,1)$ or $f=0$ on $(0,1)$ (in the case of $a=d=0$). Hence $c(x)$ is either an increasing or a constant function of $x$. In the latter case, $\forall_x c(x) = x^{*},$ as it happens in our example, and the theorem follows. We will now show that $$|x(t + \tau + \epsilon) - x^{*}| < \max\{|x(t) - x^{*}|, |x(t + \tau) - x^{*}|\}$$ hence $x(t)$ converges to $x^{*}$ for any initial conditions different from $0$ and $1$ so $x^{*}$ is globally asymptotically stable. If $x(t) < x^{*}$ and $x(t + \tau) < x^{*}$, then $x(t) < c_{t} \leq x^{*}$ and also $x(t+\tau) < c_{t+\tau} \leq x^{*}$. From (44) we obtain $$\left\{\begin{array}{cc} x \left( t + \tau \right) < x \left( t + \tau + \epsilon \right) < c_t \; \; if \; \; x \left( t + \tau \right) < c_t\\ x \left( t \right) < x \left( t + \tau + \epsilon \right) = c_t \; \; if \; \; x \left( t + \tau \right) = c_t\\ x \left( t \right) < c_t < x \left( t + \tau + \epsilon \right) < x \left( t + \tau \right)\; \; if \; \; x \left( t + \tau \right) > c_t \end{array}\right.$$ hence (47) holds. If $x \left( t \right) > x^{\ast}$ and $x \left( t + \tau \right) < x^{\ast}$, then $x \left( t + \tau \right) < x^{\ast} < c_t < x \left( t \right)$ and either $x \left( t + \tau \right) < x \left( t + \tau + \epsilon \right) < x^{\ast}$ or $x^{\ast} < x \left( t + \tau + \epsilon \right) < c_t$ which means that (47) holds. The cases of $x \left( t \right) > x^{*}$, $x \left( t + \tau \right) > x^{*}$ and $x \left( t \right) < x^{*}$, $x \left( t + \tau \right)<x^{*}$ can be treated analogously. We showed that (47) holds. Stochastic dynamics of finite populations ========================================= In the next two chapters we will discuss various stochastic dynamics of populations with a fixed number of players interacting in discrete moments of time. We will analyze symmetric two-player games with two or three strategies and multiple Nash equilibria. We will address the problem of equilibrium selection - which strategy will be played in the long run with a high frequency. Our populations are characterized either by numbers of individuals playing respective strategies in well-mixed populations or by a complete profile - assignment of strategies to players in spatial games. Let $\Omega$ be a state space of our system. For non-spatial games with two strategies, $\Omega=\{0,1,...,n\}$, where $n$ is the number of players or $\Omega= 2^{\Lambda}$ for spatial games with players located on the finite subset $\Lambda$ of ${\bf Z}, {\bf Z}^{2}$, or any other infinite graph, and interacting with their neighbours. In well-mixed populations, in discrete moments of times, some individuals switch to a strategy with a higher mean payoff. In spatial games, players choose strategies which are best responses, i.e. ones which maximize the sum of the payoffs obtained from individual games. The above rules define deterministic dynamics with some stochastic part corresponding to a random matching of players or a random choice of players who may revise their strategies. We call this mutation-free or noise-free dynamics. It is a Markov chain with a state space $\Omega$ and a transition matrix $P^{0}$. We are especially interested in absorbing states, i.e. rest points of our mutation-free dynamics. Now, with a small probability, $\epsilon$, players may mutate or make mistakes of not chosing the best reply. The presence of mutatation allows the system to make a transition from any state to any other state with a positive probability in some finite number of steps or to stay indefinitively at any state for an arbitrarily long time. This makes our Markov chains with a transition matrix $P^{\epsilon}$ ergodic ones. They have therefore unique stationary measures. To describe the long-run behavior of stochastic dynamics of finite populations, Foster and Young [@foya] introduced a concept of stochastic stability. A state of the system is [**stochastically stable**]{} if it has a positive probability in the stationary measure of the corresponding Markov chain in the zero-noise limit, that is the zero probability of mistakes or the zero-mutation level. It means that along almost any time trajectory the frequency of visiting this state converges to a positive value given by the stationary measure. Let $\mu^{\epsilon}$ be the stationary measure of our Markov chain. $X \in \Omega$ is [**stochastically stable**]{} if $\lim_{\epsilon \rightarrow 0}\mu^{\epsilon}(X) > 0.$ It is a fundamental problem to find stochastically stable states for any stochastic dynamics of interest. We will use the following tree representation of stationary measures of Markov chains proposed by Freidlin and Wentzell [@freiwen1; @freiwen2], see also [@shub]. Let $(\Omega,P^{\epsilon})$ be an ergodic Markov chain with a state space $\Omega$, transition probabilities given by the transition matrix $P^{\epsilon}: \Omega \times \Omega \rightarrow [0,1]$, where $P^{\epsilon}(Y,Y')$ is a conditional probability that the system will be in the state $Y' \in \Omega$ at the time $t+1$, if it was in the state $Y \in \Omega$ at the time $t$, and a unique stationary measure, $\mu^{\epsilon}$, also called a stationary state. A stationary state is an eigenvector of $P^{\epsilon}$ corresponding to the eigenvalue $1$, i.e. a solution of a system of linear equations, $$\mu^{\epsilon}P^{\epsilon}=\mu^{\epsilon},$$ where $\mu^{\epsilon}$ is a row wector $[\mu^{\epsilon}_{1},...,\mu^{\epsilon}_{|\Omega|}]$. After specific rearrangements one can arrive at an expression for the stationary state which involves only positive terms. This will be very useful in describing the asymptotic behaviour of stationary states. For $X \in \Omega$, let an X-tree be a directed graph on $\Omega$ such that from every $Y \neq X$ there is a unique path to $X$ and there are no outcoming edges out of $X$. Denote by $T(X)$ the set of all X-trees and let $$q^{\epsilon}(X)=\sum_{d \in T(X)} \prod_{(Y,Y') \in d}P^{\epsilon}(Y,Y'),$$ where the product is with respect to all edges of $d$. We have that $$\mu^{\epsilon}(X)=\frac{q^{\epsilon}(X)}{\sum_{Y \in \Omega}q^{\epsilon}(Y)}$$ for all $X \in \Omega.$ We assume that our noise-free dynamics, i.e. in the case of $\epsilon=0$, has at least one absorbing state and there are no absorbing sets (recurrent classes) consisting of more than one state. It then follows from (50) that only absorbing states can be stochastically stable. Let us begin with the case of two absorbing states, $X$ and $Y$. Consider a dynamics in which $P^{\epsilon}(Z,W)$ for all $Z, W \in \Omega$, is of order $\epsilon^{m}$, where $m$ is the number of mistakes involved to pass from $Z$ to $W$. The noise-free limit of $\mu^{\epsilon}$ in the form (50) has a $0/0$ character. Let $m_{XY}$ be a minimal number of mistakes needed to make a transition from the state $X$ to $Y$ and $m_{YX}$ the minimal number of mistakes to evolve from $Y$ to $X$. Then $q^{\epsilon}(X)$ is of the order $\epsilon^{m(YX)}$ and $q^{\epsilon}(Y)$ is of the order $\epsilon^{m(XY)}$. If for example $m_{YX} < m_{XY}$, then $ \lim_{\epsilon \rightarrow 0}\mu^{\epsilon}(X)=1$ hence $X$ is stochastically stable. In general, to study the zero-noise limit of the stationary measure, it is enough to consider paths between absorbing states. More precisely, we construct X-trees with absorbing states $X^{k}$, $k=1,...,l$ as vertices; the family of such $X$-trees is denoted by $\tilde{T}(X)$. Let $$q_{m}(X)=max_{d \in \tilde{T}(X)} \prod_{(Y,Y') \in d}\tilde{P}(Y,Y'),$$ where $\tilde{P}(Y,Y')= max \prod_{(W,W')}P(W,W')$, where the product is taken along any path joining $Y$ with $Y'$ and the maximum is taken with respect to all such paths. Now we may observe that if $lim_{\epsilon \rightarrow 0} q_{m}(X^{i})/q_{m}(X^{k})=0,$ for every $i=1,...,l$, $i \neq k$, then $X^{k}$ is stochastically stable. Therefore we have to compare trees with the biggest products in (51); such trees are called maximal. The above characterisation of the stationary measure was used to find stochastically stable states in non-spatial [@kmr; @young1; @rvr; @vr; @young2; @mar] and spatial games [@ellis1; @ellis2]. We will use it below in our examples. In many cases, there exists a state $X$ such that $\lim_{\epsilon \rightarrow 0}\mu^{\epsilon}(X)=1$ in the zero-noise limit. Then we say that $X$ was selected in the zero-noise limit of a given stochastic dynamics. However, for any low but fixed mutation level, when the number of players is very large, the frequency of visiting any single state can be arbitrarily low. It is an ensemble of states that can have a probability close to one in the stationary measure. The concept of the ensemble stability is discussed in Chapter 9. Stochastic dynamics of well-mixed populations ============================================= Here we will discuss stochastic dynamics of well-mixed populations of players interacting in discrete moments of time. We will analyze two-player games with two strategies and two pure Nash equilibria. The efficient strategy (also called payoff dominant) when played by the whole population results in its highest possible payoff (fitness). The risk-dominant one is played by individuals averse to risk. The strategy is risk dominant if it has a higher expected payoff against a player playing both strategies with equal probabilities [@hs]. We will address the problem of equilibrium selection - which strategy will be played in the long run with a high frequency. We will review two models of dynamics of a population with a fixed number of individuals. In both of them, the selection part of the dynamics ensures that if the mean payoff of a given strategy is bigger than the mean payoff of the other one, then the number of individuals playing the given strategy increases. In the first model, introduced by Kandori, Mailath, and Rob [@kmr], one assumes (as in the standard replicator dynamics) that individuals receive average payoffs with respect to all possible opponents - they play against the average strategy. In the second model, introduced by Robson and Vega-Redondo [@rvr], at any moment of time, individuals play only one or few games with randomly chosen opponents. In both models, players may mutate with a small probability, hence the population may move against a selection pressure. Kandori, Mailath, and Rob showed that in their model, the risk-dominant strategy is stochastically stable - if the mutation level is small enough we observe it in the long run with the frequency close to one [@kmr]. In the model of Robson and Vega-Redondo, the efficient strategy is stochastically stable [@rvr; @vr]. It is one of very few models in which an efficient strategy is stochastically stable in the presence of a risk-dominant one. The population evolves in the long run to a state with the maximal fitness. The main goal of this chapter is to investigate the effect of the number of players on the long-run behaviour of the Robson-Vega-Redondo model [@population]. We will discuss a sequential dynamics and the one where each individual enjoys each period a revision opportunity with the same probability. We will show that for any arbitrarily low but a fixed level of mutations, if the number of players is sufficiently large, then a risk-dominant strategy is played in the long run with a frequency closed to one - a stochastically stable efficient strategy is observed with a very low frequency. It means that when the number of players increases, the population undergoes a transition between an efficient payoff-dominant equilibrium and a risk-dominant one. We will also show that for some range of payoff parameters, stochastic stability itself depends on the number of players. If the number of players is below certain value (which may be arbitrarily large), then a risk-dominant strategy is stochastically stable. Only if the number of players is large enough, an efficient strategy becomes stochastically stable as proved by Robson and Vega-Redondo. Combining the above results we see that for a low but fixed noise level, the population undergoes twice a transition between its two equilibria as the number of individuals increases [@banach]. In addition, for a sufficiently large number of individuals, the population undergoes another equilibrium transition when the noise decreases. Let us formally introduce our models. We will consider a finite population of $n$ individuals who have at their disposal one of two strategies: $A$ and $B$. At every discrete moment of time, $t=1,2,...$ individuals are randomly paired (we assume that $n$ is even) to play a two-player symmetric game with payoffs given by the following matrix: A B A a b U = B c d, where $a>c, d>b, a>d$, and $a+b<c+d$ so $(A,A)$ is an efficient Nash equilibrium and $(B,B)$ is a risk-dominant one. At the time $t$, the state of our population is described by the number of individuals, $z_{t}$, playing $A$. Formally, by the state space we mean the set $$\Omega=\{z, 0 \leq z\leq n\}.$$ Now we will describe the dynamics of our system. It consists of two components: selection and mutation. The selection mechanism ensures that if the mean payoff of a given strategy, $\pi_{i}(z_{t}), i=A,B$, at the time $t$ is bigger than the mean payoff of the other one, then the number of individuals playing the given strategy increases in $t+1$. In their paper, Kandori, Mailath, and Rob [@kmr] write $$\pi_{A}(z_{t})=\frac{a(z_{t}-1)+b(n-z_{t})}{n-1},$$ $$\pi_{B}(z_{t})=\frac{cz_{t}+d(n-z_{t}-1)}{n-1},$$ provided $0<z_{t}<n$. It means that in every time step, players are paired infnitely many times to play the game or equivalently, each player plays with every other player and his payoff is the sum of corresponding payoffs. This model may be therefore considered as an analog of replicator dynamics for populations with a fixed numbers of players. The selection dynamics is formalized in the following way: $$z_{t+1} > z_{t} \hspace{2mm} if \hspace{2mm} \pi_{A}(z_{t}) > \pi_{B}(z_{t}),$$ $$z_{t+1} < z_{t} \hspace{2mm} if \hspace{2mm} \pi_{A}(z_{t}) < \pi_{B}(z_{t}),$$ $$z_{t+1}= z_{t} \hspace{2mm} if \hspace{2mm} \pi_{A}(z_{t}) = \pi_{B}(z_{t}),$$ $$z_{t+1}= z_{t} \hspace{2mm} if \hspace{2mm} z_{t}=0 \hspace{2mm} or \hspace{2mm} z_{t}=n.$$ Now mutations are added. Players may switch to new strategies with the probability $\epsilon$. It is easy to see that for any two states of the population, there is a positive probability of the transition between them in some finite number of time steps. We have therefore obtained an ergodic Markov chain with $n+1$ states and a unique stationary measure which we denote by $\mu^{\epsilon}_{n}.$ Kandori, Mailath, and Rob proved that the risk-dominant strategy $B$ is stochastically stable [@kmr] $\lim_{\epsilon \rightarrow 0}\mu^{\epsilon}_{n}(0)=1$ This means that in the long run, in the limit of no mutations, all players play $B$. The general set up in the Robson-Vega-Redondo model [@rvr] is the same. However, individuals are paired only once at every time step and play only one game before a selection process takes place. Let $p_{t}$ denote the random variable which describes the number of cross-pairings, i.e. the number of pairs of matched individuals playing different strategies at the time $t$. Let us notice that $p_{t}$ depends on $z_{t}$. For a given realization of $p_{t}$ and $z_{t}$, mean payoffs obtained by each strategy are as follows: $$\tilde{\pi}_{A}(z_{t},p_{t})=\frac{a(z_{t}-p_{t})+bp_{t}}{z_{t}},$$ $$\tilde{\pi}_{B}(z_{t},p_{t})=\frac{cp_{t}+d(n-z_{t}-p_{t})}{n-z_{t}},$$ provided $0<z_{t}<n$. Robson and Vega-Redondo showed that the payoff-dominant strategy is stochastically stable [@rvr]. $\lim_{\epsilon \rightarrow 0}\mu^{\epsilon}_{n}(n)=1$ We will outline their proof. First of all, one can show that there exists $k$ such that if $n$ is large enough and $z_{t} \geq k$, then there is a positive probability (a certain realization of $p_{t}$) that after a finite number of steps of the mutation-free selection dynamics, all players will play $A$. Likewise, if $z_{t} <k$ (for any $k \geq 1$), then if the number of players is large enough, then after a finite number of steps of the mutation-free selection dynamics all players will play $B$. In other words, $z=0$ and $z=n$ are the only absorbing states of the mutation-free dynamics. Moreover, if $n$ is large enough, then if $z_{t} \geq n-k$, then the mean payoff obtained by $A$ is always (for any realization of $p_{t}$) bigger than the mean payoff obtained by $B$ (in the worst case all $B$-players play with $A$-players). Therefore the size of the basin of attraction of the state $z=0$ is at most $n-k-1$ and that of $z=n$ is at least $n-k$. Observe that mutation-free dynamics is not deterministic ($p_{t}$ describes the random matching) and therefore basins of attraction may overlap. It follows that the system needs at least $k+1$ mutations to evolve from $z=n$ to $z=0$ and at most $k$ mutations to evolve from $z=0$ to $z=n$. Now using the tree representation of stationary states, Robson and Vega-Redondo finish the proof and show that the efficient strategy is stochastically stable. However, as outlined above, their proof requires the number of players to be sufficiently large. We will now show that a risk-dominant strategy is stochastically stable if the number of players is below certain value which can be arbitrarily large. If $n<\frac{2a-c-b}{a-c}$, then the risk-dominant strategy $B$ is stochastically stable in the case of random matching of players. [**Proof:**]{} If the population consists of only one $B$-player and $n-1$ $A$-players and if $c>[a(n-2)+b]/(n-1)$, that is $n< (2a-c-b)/(a-c)$, then $\tilde{\pi}_{B}> \tilde{\pi}_{A}.$ It means that one needs only one mutation to evolve from $z=n$ to $z=0.$ It is easy to see that two mutations are necessary to evolve from $z=0$ to $z=n.$ To see stochastically stable states, we need to take the limit of no mutations. We will now examine the long-run behavior of the Robson-Vega-Redondo model for a fixed level of mutations in the limit of the infinite number of players. Now we will analyze the extreme case of the selection rule (53) - a sequential dynamics where in one time unit only one player can change his strategy. Although our dynamics is discrete in time, it captures the essential features of continuous-time models in which every player has an exponentially distributed waiting time to a moment of a revision opportunity. Probability that two or more players revise their strategies at the same time is therefore equal to zero - this is an example of a birth and death process. The number of $A$-players in the population may increase by one in $t+1$, if a $B$-player is chosen in $t$ which happens with the probability $(n-z_{t})/n$. Analogously, the number of $B$-players in the population may increase by one in $t+1$, if an $A$-player is chosen in $t$ which happens with the probability $(z_{t})/n$. The player who has a revision opportunity chooses in $t+1$ with the probability $1-\epsilon$ the strategy with a higher average payoff in $t$ and the other one with the probability $\epsilon$. Let $$r(k)=P(\tilde{\pi}_{A}(z_{t},p_{t}) > \tilde{\pi}_{B}(z_{t},p_{t})) \; \; and \; \; l(k)=P(\tilde{\pi}_{A}(z_{t},p_{t}) < \tilde{\pi}_{B}(z_{t},p_{t})).$$ The sequential dynamics is described by the following transition probabilities: if $z_{t}=0 $, then $z_{t+1}=1 $ with the probability $\epsilon$ and $z_{t+1}=0 $ with the probability $1-\epsilon$, if $z_{t}=n $, then $z_{t+1}=n-1 $ with the probability $\epsilon$ and $z_{t+1}=n $ with the probability $1-\epsilon$, if $z_{t} \neq 0,n$, then $z_{t+1}= z_{t} +1$ with the probability $$r(k)\frac{n-z_{t}}{n}(1-\epsilon)+(1-r(k))\frac{n-z_{t}}{n}\epsilon$$ and $z_{t+1}= z_{t} -1$ with the probability $$l(k)\frac{z_{t}}{n}(1-\epsilon)+(1-l(k))\frac{z_{t}}{n}\epsilon.$$ In the dynamics intermediate between the parallel (where all individuals can revise their strategies at the same time) and the sequential one, each individual has a revision opportunity with the same probability $\tau <1$ during the time interval of the lenght $1$. For a fixed $\epsilon$ and an arbitrarily large but fixed $n$, we consider the limit of the continuous time, $\tau \rightarrow 0$, and show that the limiting behaviour is already obtained for a sufficiently small $\tau$, namely $\tau < \epsilon /n^{3}$. For an interesting discussion on the importance of the order of taking different limits $(\tau \rightarrow 0, n \rightarrow \infty,$ and $\epsilon \rightarrow 0)$ in evolutionary models (especially in the Aspiration and Imitation model) see Samuelson [@samuel]. In the intermediate dynamics, instead of $(n-z_{t})/n$ and $z_{t}/n$ probabilities we have more involved combinatorial factors. In order to get rid of these inconvenient factors, we will enlarge the state space of the population. The state space $\Omega^{'}$ is the set of all configurations of players, that is all possible assignments of strategies to individual players. Therefore, a state $z_{t}=k$ in $\Omega$ consists of $\left( \begin{array}{c} n \\ k \end{array} \right)$ states in $\Omega^{'}$. Observe that the sequential dynamics on $\Omega^{'}$ is not anymore a birth and death process. However, we are able to treat both dynamics in the same framework. We showed in [@population] that for any arbitrarily low but fixed level of mutation, if the number of players is large enough, then in the long run only a small fraction of the population plays the payoff-dominant strategy. Smaller the mutation level is, fewer players use the payoff-dominant strategy. The following two theorems were proven in [@population]. In the sequential dynamics, for any $\delta >0$ and $\beta >0$ there exist $\epsilon(\delta, \beta)$ and $n(\epsilon)$ such that for any $n > n(\epsilon)$ $$\mu_{n}^{\epsilon}(z \leq \beta n) > 1- \delta.$$ In the intermediate dynamics dynamics, for any $\delta >0$ and $\beta >0$ there exist $\epsilon(\delta, \beta)$ and $n(\epsilon)$ such that for any $n > n(\epsilon)$ and $\tau< \frac{\epsilon}{n^{3}}$ $$\mu_{n}^{\epsilon}(z \leq \beta n) > 1- \delta.$$ We can combine Theorems 9, 10, and 12 and obtain [@banach] In the intermediate dynamics, for any $\delta >0$ and $\beta >0$ there exists $\epsilon(\delta, \beta)$ such that, for all $\epsilon < \epsilon(\delta, \beta)$, there exist $n_{1} < n_{2} < n_{3}(\epsilon) < n_{4}(\epsilon)$ such that if $n < n_{1}=\frac{2a-c-b}{a-c}$, then $\mu_{n}^{\epsilon}(z =0) > 1- \delta,$ if $n_{2} < n < n_{3}(\epsilon)$, then $\mu_{n}^{\epsilon}(z = n) > 1- \delta,$ if $n > n_{4}(\epsilon)$ and $\tau< \epsilon/n^{3}$, then $\mu_{n}^{\epsilon}(z \leq \beta n) > 1- \delta$. Small $\tau$ means that our dynamics is close to the sequential one. We have that $n_{3}(\epsilon), n_{4}(\epsilon), n_{3}(\epsilon)-n_{2}$, and $n_{4}(\epsilon)-n_{3}(\epsilon) \rightarrow \infty$ when $\epsilon \rightarrow 0$. It follows from Theorem 13 that the population of players undergoes several [**equilibrium transitions**]{}. First of all, for a fixed noise level, when the number of players increases, the population switches from a $B$-equilibrium, where most of the individuals play the strategy $B$, to an $A$-equilibrium and then back to $B$ one. We know that if $n>n_{2}$, then $z=n$ is stochastically stable. Therefore, for any fixed number of players, $n>n_{4}(\epsilon)$, when the noise level decreases, the population undergoes a transition from a $B$-equilibrium to $A$ one. We see that in order to study the long-run behaviour of stochastic population dynamics, we should estimate the relevant parameters to be sure what limiting procedures are appropriate in specific examples. Let us note that the above theorems concern an ensemble of states, not an individual one. In the limit of the infinite number of players, that is the infinite number of states, every single state has zero probability in the stationary state. It is an ensemble of states that might be stable [@statmech; @physica]. The concept of ensemble stability will be discussed in Chapter 9. Spatial games with local interactions ===================================== Nash configurations and stochastic dynamics ------------------------------------------- In spatial games, players are located on vertices of certain graphs and they interact only with their neighbours; see for example [@nowak0; @nowak1; @nowak2; @blume1; @ellis1; @young2; @ellis2; @linnor; @doebeli1; @doebeli2; @szabo1; @szabo2; @szabo6; @hauert1; @doebeli3; @hauert2; @hauert3] and a recent review paper [@szabo8] and references therein. Let $\Lambda$ be a finite subset of the simple lattice ${\bf Z}^{d}$. Every site of $\Lambda$ is occupied by one player who has at his disposal one of $m$ different pure strategies. Let $S$ be the set of strategies, then $\Omega_{\Lambda}=S^{\Lambda}$ is the space of all possible configurations of players, that is all possible assignments of strategies to individual players. For every $i \in \Lambda$, $X_{i}$ is a strategy of the $i-$th player in the configuration $X \in \Omega_{\Lambda}$ and $X_{-i}$ denotes strategies of all remaining players; $X$ therefore can be represented as the pair $(X_{i},X_{-i})$. Every player interacts only with his nearest neighbours and his payoff is the sum of the payoffs resulting from individual plays. We assume that he has to use the same strategy for all neighbours. Let $N_{i}$ denote the neighbourhood of the $i-$th player. For the nearest-neighbour interaction we have $N_{i}=\{j; |j-i|=1\}$, where $|i-j|$ is the distance between $i$ and $j$. For $X \in \Omega_{\Lambda}$ we denote by $\nu_{i}(X)$ the payoff of the $i-$th player in the configuration $X$: $$\nu_{i}(X)=\sum_{j \in N_{i}}U(X_{i}, X_{j}),$$ where $U$ is a $m \times m$ matrix of payoffs of a two-player symmetric game with $m$ pure strategies. $X \in \Omega_{\Lambda}$ is a [**Nash configuration**]{} if for every $i \in \Lambda$ and $Y_{i} \in S$, $$\nu_{i}(X_{i},X_{-i}) \geq \nu_{i}(Y_{i},X_{-i})$$ Here we will discuss only coordination games, where there are $m$ pure symmetric Nash equilibria and therefore $m$ homogeneous Nash configurations, where all players play the same strategy. In the Stag-hunt game in Example 1, we have two homogeneous Nash configurations, $X^{St}$ and $X^{H}$, where all individuals play $St$ or $H$ respectively. We describe now the sequential deterministic dynamics of the [**best-response rule**]{}. Namely, at each discrete moment of time $t=1,2,...$, a randomly chosen player may update his strategy. He simply adopts the strategy, $X_{i}^{t+1}$, which gives him the maximal total payoff $\nu_{i}(X_{i}^{t+1}, X^{t}_{-i})$ for given $X^{t}_{-i}$, a configuration of strategies of remaining players at the time $t$. Now we allow players to make mistakes, that is they may not choose best responses. We will discuss two types of such stochastic dynamics. In the first one, the so-called [**perturbed best response**]{}, a player follows the best-response rule with probability $1-\epsilon$ (in case of more than one best-response strategy he chooses randomly one of them) and with probability $\epsilon$ he makes a mistake and chooses randomly one of the remaining strategies. The probability of mistakes (or the noise level) is state-independent here. In the so called [**log-linear dynamics**]{}, the probability of chosing by the $i-$th player the strategy $X_{i}^{t+1}$ at the time $t+1$ decreases with the loss of the payoff and is given by the following conditional probability: $$p_{i}^{\epsilon}(X_{i}^{t+1}|X_{-i}^{t})= \frac{e^{\frac{1}{\epsilon}\nu_{i}( X_{i}^{t+1},X_{-i}^{t})}}{\sum_{Y_{i} \in S} e^{\frac{1}{\epsilon}\nu_{i}(Y_{i},X_{-i}^{t})}},$$ Let us observe that if $\epsilon \rightarrow 0$, $p_{i}^{\epsilon}$ converges pointwise to the best-response rule. Both stochastic dynamics are examples of ergodic Markov chains with $|S^{\Lambda}|$ states. Therefore they have unique stationary states denoted by $\mu_{\Lambda}^{\epsilon}$. Stationary states of the log-linear dynamics can be explicitly constructed for the so-called potential games. A game is called a [**potential game**]{} if its payoff matrix can be changed to a symmetric one by adding payoffs to its columns [@mon]. As we know, such a payoff transformation does not change strategic character of the game, in particular it does not change the set of its Nash equilibria. More formally, we have the following definition. A two-player symmetric game with a payoff matrix $U$\ is a [**potential game**]{} if there exists a symmetric matrix $V$,\ called a potential of the game, such that for any three strategies $A, B, C \in S$ $$U(A,C)-U(B,C)=V(A,C)-V(B,C).$$ It is easy to see that every game with two strategies has a potential $V$ with $V(A,A)=a-c$, $V(B,B)=d-b$, and $V(A,B)=V(B,A)=0.$ It follows that an equilibrium is risk-dominant if and only if it has a bigger potential. For players on a lattice, for any $X \in \Omega_{\Lambda}$, $$V(X)=\sum_{(i,j) \subset \Lambda} V(X_{i},X_{j})$$ is then the potential of the configuration $X$. For the sequential log-linear dynamics of potential games, one can explicitely construct stationary measures [@young2]. We begin by the following general definition concerning a Markov chain with a state space $\Omega$ and a transition matrix $P$. A measure $\mu$ on $\Omega$ satisfies a [**detailed balance condition**]{} if $$\mu(X)P_{XY}=\mu(Y)P_{YX}$$ for every $X,Y \in \Omega$ [**Lemma**]{} [*If $\mu$ satisfies the detailed balance condition then it is a stationary measure*]{} [**Proof:**]{} $$\sum_{X \in \Omega}\mu(X)P_{XY}=\sum_{X \in \Omega} \mu(Y)P_{YX}= \mu(Y)$$ The following theorem is due Peyton Young [@young2]. We will present here his proof. The stationary measure of the sequential log-linear dynamics in a game with the potential $V$ is given by $$\mu^{\epsilon}_{\Lambda}(X)=\frac{e^{\frac{1}{\epsilon}V(X)}} {\sum_{Z \in \Omega_{\Lambda}}e^{\frac{1}{\epsilon}V(Z)}}.$$ [**Proof:**]{} We will show that $\mu^{\epsilon}_{\Lambda}$ in (59) satisfies the detailed balance condition. Let us notice that in the sequential dynamics, $P_{XY}=0$ unless $X=Y$ or $Y$ differs fom $X$ at one lattice site only, say $i \in \Lambda$. Let $$\lambda =\frac{1}{|\Lambda|}\frac{1}{\sum_{Z \in \Omega_{\Lambda}}e^{\frac{1}{\epsilon}V(Z)}}\frac{1}{\sum_{Z_{i} \in S}e^{\frac{1}{\epsilon}\sum_{j \in N_{i}}U(Z_{i},X_{j})}}$$ Then $$\mu^{\epsilon}_{\Lambda}(X)P_{XY}=\lambda e^{\frac{1}{\epsilon}(\sum_{(h,k) \subset \Lambda} V(X_{h},X_{k})+\sum_{j \in N_{i}}U(Y_{i},X_{j}))}$$ $$=\lambda e^{\frac{1}{\epsilon}(\sum_{(h,k) \subset \Lambda} V(X_{h},X_{k})+\sum_{j \in N_{i}}(U(X_{i},X_{j})-V(X_{i},X_{j})+V(Y_{i},X_{j})))}$$ $$=\lambda e^{\frac{1}{\epsilon}(\sum_{(h,k) \subset \Lambda} V(Y_{h},Y_{k})+\sum_{j \in N_{i}}U(X_{i},X_{j}))}=\mu^{\epsilon}_{\Lambda}(Y)P_{YX}.$$ We may now explicitly perform the limit $\epsilon \rightarrow 0$ in (59). In the Stag-hunt game, $X^{H}$ has a bigger potential than $X^{St}$ so $\lim_{\epsilon \rightarrow 0}\mu_{\Lambda}^{\epsilon}(X^{H})=1 $ hence $X^{H}$ is stochastically stable (we also say that $H$ is stochastically stable). The concept of a Nash configuration in spatial games is very similar to the concept of a ground-state configuration in lattice-gas models of interacting particles. We will discuss similarities and differences between these two systems of interacting entities in the next section. Ground states and Nash configurations ------------------------------------- We will present here one of the basic models of interacting particles. In classical lattice-gas models, particles occupy lattice sites and interact only with their neighbours. The fundamental concept is that of a ground-state configuration. It can be formulated conveniently in the limit of an infinite lattice (the infinite number of particles). Let us assume that every site of the ${\bf Z}^{d}$ lattice can be occupied by one of $m$ different particles. An infinite-lattice configuration is an assignment of particles to lattice sites, i.e. an element of $\Omega = \{1,...,m\}^{{\bf Z}^{d}}$. If $X \in \Omega$ and $i \in {\bf Z}^{d}$, then we denote by $X_{i}$ a restriction of $X$ to $i$. We will assume here that only nearest-neighbour particles interact. The energy of their interaction is given by a symmetric $m \times m$ matrix $V$. An element $V(A,B)$ is the interaction energy of two nearest-neighbour particles of the type $A$ and $B$. The total energy of a system in the configuration $X$ in a finite region $\Lambda \subset {\bf Z}^{d}$ can be then written as $$H_{\Lambda}(X)=\sum_{(i,j) \subset \Lambda} V(X_{i},X_{j}).$$ $Y$ is a [**local excitation**]{} of $X$, $Y \sim X$, $Y,X \in \Omega$ , if there exists a finite $\Lambda \subset {\bf Z}^{d}$ such that $X = Y$ outside $\Lambda.$ For $Y \sim X$, the [**relative energy**]{} is defined by $$H(Y,X)=\sum_{(i,j) \in {\bf Z}^{d}} (V(Y_{i},Y_{j})-V(X_{i},X_{j})),$$ where the summation is with respect to pairs of nearest neighbours on ${\bf Z}^{d}$. Observe that this is the finite sum; the energy difference between $Y$ and $X$ is equal to zero outside some finite $\Lambda$. $X \in \Omega$ is a [**ground-state configuration**]{} of $V$ if $$H(Y,X) \geq 0 \; \; for \; \; any \; \; Y \sim X.$$ That is, we cannot lower the energy of a ground-state configuration by changing it locally. The energy density $e(X)$ of a configuration $X$ is $$e(X)=\liminf_{\Lambda \rightarrow {\bf Z}^{2}} \frac{H_{\Lambda}(X)}{|\Lambda|},$$ where $|\Lambda|$ is the number of lattice sites in $\Lambda$. It can be shown that any ground-state configuration has the minimal energy density [@sinai]. It means that local conditions present in the definition of a ground-state configuration force the global minimization of the energy density. We see that the concept of a ground-state configuration is very similar to that of a Nash configuration. We have to identify particles with agents, types of particles with strategies and instead of minimizing interaction energies we should maximize payoffs. There are however profound differences. First of all, ground-state configurations can be defined only for symmetric matrices; an interaction energy is assigned to a pair of particles, payoffs are assigned to individual players and may be different for each of them. Ground-state configurations are stable with respect to all local changes, Nash configurations are stable only with respect to one-player changes. It means that for the same symmetric matrix $U$, there may exist a configuration which is a Nash configuration but not a ground-state configuration for the interaction matrix $-U$. The simplest example is given by the following matrix: [**Example 5**]{} A B A 2 0 U = B 0 1 $(A,A)$ and $(B,B)$ are Nash configurations for a system consisting of two players but only $(A,A)$ is a ground-state configuration for $V=-U.$ We may therefore consider the concept of a ground-state configuration as a refinement of a Nash equilibrium. For any classical lattice-gas model there exists at least one ground-state configuration. This can be seen in the following way. We start with an arbitrary configuration. If it cannot be changed locally to decrease its energy it is already a ground-state configuration. Otherwise we may change it locally and decrease the energy of the system. If our system is finite, then after a finite number of steps we arrive at a ground-state configuration; at every step we decrease the energy of the system and for every finite system its possible energies form a finite set. For an infinite system, we have to proceed ad infinitum converging to a ground-state configuration (this follows from the compactness of $\Omega$ in the product of discrete topologies). Game models are different. It may happen that a game with a nonsymmetric payoff matrix may not posess a Nash configuration. The classical example is that of the Rock-Scissors-Paper game. One may show that this game dos not have any Nash configurations on ${\bf Z}$ and ${\bf Z}^{2}$ but many Nash configurations on the triangular lattice. In short, ground-state configurations minimize the total energy of a particle system, Nash configurations do not necessarily maximize the total payoff of a population. Ground-state configuration is an equilibrium concept for systems of interacting particles at zero temperature. For positive temperatures, we must take into account fluctuations caused by thermal motions of particles. Equilibrium behaviour of the system results then from the competition between its energy $V$ and entropy $S$ (which measures the number of configurations corresponding to a macroscopic state), i.e. the minimization of its free energy $F=V-TS$, where $T$ is the temperature of the system - a measure of thermal motions. At the zero temperature, $T=0$, the minimization of the free energy reduces to the minimization of the energy. This zero-temperature limit looks very similar to the zero-noise limit present in the definition of the stochastic stability. Equilibrium behaviour of a system of interacting particles can be described by specifying probabilities of occurence for all particle configurations. More formally, it is described by a Gibbs state (see [@geo] and references therein). We construct it in the following way. Let $\Lambda$ be a finite subset of ${\bf Z}^{d}$ and $\rho^{T}_{\Lambda}$ the following probability mass function on $\Omega_{\Lambda}=(1,...,m)^{\Lambda}$: $$\rho_{\Lambda}^{T}(X)=(1/Z^{T}_{\Lambda})\exp(-H_{\Lambda}(X)/T),$$ for every $X \in \Omega_{\Lambda}$, where $$Z^{T}_{\Lambda}=\sum_{X \in \Omega_{\Lambda}}\exp(-H_{\Lambda}(X)/T)$$ is a normalizing factor. We define a [**Gibbs state**]{} $\rho^{T}$ as a limit of $\rho^{T}_{\Lambda}$ as $\Lambda \rightarrow {\bf Z}^{d}$. One can prove that a limit of a translation-invariant Gibbs state for a given interaction as $T \rightarrow 0$ is a measure supported by ground-state configurations. One of the fundamental problems of statistical mechanics is a characterization of low-temperature Gibbs states for given interactions between particles. Let us observe that the finite-volume Gibbs state in (63) is equal to stationary state $\mu^{\epsilon}_{\Lambda}$ in (59) if we identify $T$ with $\epsilon$ and $V \rightarrow -V$. Ensemble stability ------------------ The concept of stochastic stability involves individual configurations of players. In the zero-noise limit, a stationary state is usually concentrated on one or at most few configurations. However, for a low but fixed noise and for a sufficiently large number of players, the probability of any individual configuration of players is practically zero. The stationary measure, however, may be highly concentrated on an ensemble consisting of one Nash configuration and its small perturbations, i.e. configurations where most players use the same strategy. Such configurations have relatively high probability in the stationary measure. We call such configurations ensemble stable. Let $\mu^{\epsilon}_{\Lambda}$ be a stationary measure. $X \in \Omega_{\Lambda}$ is [**$\gamma$-ensemble stable**]{} if $\mu_{\Lambda}^{\epsilon}(Y \in \Omega_{\Lambda};Y_{i} \neq X_{i}) < \gamma$ for any $i \in \Lambda$ if $\Lambda \supset \Lambda(\gamma)$ for some $\Lambda(\gamma)$. $X \in \Omega_{\Lambda}$ is [**low-noise ensemble stable**]{} if for every $\gamma>0$ there exists $\epsilon(\gamma)$ such that if $\epsilon<\epsilon(\gamma)$, then $X$ is $\gamma$-ensemble stable. If $X$ is $\gamma$-ensemble stable with $\gamma$ close to zero, then the ensemble consisting of $X$ and configurations which are different from $X$ at at most few sites has the probability close to one in the stationary measure. It does not follow, however, that $X$ is necessarily low-noise ensemble or stochastically stable as it happens in examples presented below [@statmech]. [**Example 6**]{} Players are located on a finite subset $\Lambda$ of ${\bf Z}^{2}$ (with periodic boundary conditions) and interact with their four nearest neighbours. They have at their disposal three pure strategies: $A, B,$ and $C$. The payoffs are given by the following symmetric matrix: A B C A 1.5 0 1 U = B 0 2 1 C 1 1 2 Our game has three Nash equilibria: $(A,A), (B,B)$, and $(C,C)$, and the corresponding spatial game has three homogeneous Nash configurations: $X^{A}, X^{B}$, and $X^{C}$, where all individuals are assigned the same strategy. Let us notice that $X^{B}$ and $X^{C}$ have the maximal payoff in every finite volume and therefore they are ground-state configurations for $-U$ and $X^{A}$ is not. The unique stationary measure of the log-linear dynamics (56) is is given by (59) with $U=V$ which is a finite-volume Gibbs state (63) with $V$ replaced by $-U$ and $T$ by $\epsilon$. We have $$\sum_{(i,j)\subset \Lambda}U(X^{k}_{i},X^{k}_{j})-\sum_{(i,j)\in \Lambda}U(Y_{i},Y_{j})>0,$$ for every $Y \neq X^{B} \; and \; X^{C}$, $k=B,C$, and $$\sum_{(i,j)\subset \Lambda}U(X^{B}_{i},X^{B}_{j})=\sum_{(i,j)\subset \Lambda}U(X^{C}_{i},X^{C}_{j}).$$ It follows that $\lim_{\epsilon \rightarrow 0}\mu_{\Lambda}^{\epsilon}(X^{k})=1/2$, for $k=B, C $ so $X^{B}$ and $X^{C}$ are stochastically stable. Let us investigate the long-run behaviour of our system for large $\Lambda$, that is for a large number of players. Observe that $$\lim_{\Lambda \rightarrow {\bf Z}^{2}}\mu_{\Lambda}^{\epsilon}(X)=0$$ for every $X \in \Omega = S^{{\bf Z}^{2}}$. Therefore, for a large $\Lambda$ we may only observe, with reasonably positive frequencies, ensembles of configurations and not particular configurations. We will be interested in ensembles which consist of a Nash configuration and its small perturbations, that is configurations, where most players use the same strategy. We perform first the limit $\Lambda \rightarrow {\bf Z}^{2}$ and obtain an infinite-volume Gibbs state in the temperature $T=\epsilon$, $$\mu^{\epsilon} = \lim_{\Lambda \rightarrow {\bf Z}^{2}}\mu_{\Lambda}^{\epsilon}.$$ In order to investigate the stationary state of our example, we will apply a technique developed by Bricmont and Slawny [@brsl1; @brsl2]. They studied low-temperature stability of the so-called dominant ground-state configurations. It follows from their results that $$\mu^{\epsilon}(X_{i}=C)>1-\delta(\epsilon)$$ for any $i \in {\bf Z}^{2}$ and $\delta(\epsilon) \rightarrow 0$ as $\epsilon \rightarrow 0$ [@statmech]. The following theorem is a simple consequence of (66). $X^{C}$ is low-noise ensemble stable. We see that for any low but fixed $\epsilon$, if the number of players is large enough, then in the long run, almost all players use $C$ strategy. On the other hand, if for any fixed number of players, $\epsilon$ is lowered substantially, then B and C appear with frequencies close to $1/2$. Let us sketch briefly the reason of such a behavior. While it is true that both $X^{B}$ and $X^{C}$ have the same potential which is the half of the payoff of the whole system (it plays the role of the total energy of a system of interacting particles), the $X^{C}$ Nash configuration has more lowest-cost excitations. Namely, one player can change its strategy and switch to either $A$ or $B$ and the potential will decrease by $4$ units. Players in the $X^{B}$ Nash configuration have only one possibility, that is to switch to $C$; switching to $A$ decreases the potential by $8$. Now, the probability of the occurrence of any configuration in the Gibbs state (which is the stationary state of our stochastic dynamics) depends on the potential in an exponential way. One can prove that the probability of the ensemble consisting of the $X^{C}$ Nash configuration and configurations which are different from it at few sites only is much bigger than the probability of the analogous $X^{B}$-ensemble. It follows from the fact that the $X^{C}$-ensemble has many more configurations than the $X^{B}$-ensemble. On the other hand, configurations which are outside $X^{B}$ and $X^{C}$-ensembles appear with exponentially small probabilities. It means that for large enough systems (and small but not extremely small $\epsilon$) we observe in the stationary state the $X^{C}$ Nash configuration with perhaps few different strategies. The above argument was made into a rigorous proof for an infinite system of the closely related lattice-gas model (the Blume-Capel model) of interacting particles by Bricmont and Slawny in [@brsl1]. In the above example, $X^{B}$ and $X^{C}$ have the same total payoff but $X^{C}$ has more lowest-cost excitations and therefore it is low-noise ensemble stable. We will now discuss the situation, where $X^{C}$ has a smaller total payoff but nevertheless in the long run $C$ is played with a frequency close to $1$ if the noise level is low but not extremely low. We will consider a family of games with the following payoff matrix: [**Example 7**]{} A B C A 1.5 0 1 U = B 0 $2+\alpha$ 1 C 1 1 2, where $\alpha>0$ so $B$ is both payoff and pairwise risk-dominant. We are interested in the long-run behavior of our system for small positive $\alpha$ and low $\epsilon$. One may modify the proof of Theorem 15 and obtain the following theorem [@statmech]. For every $\gamma>0$, there exist $\alpha(\gamma)$ and $\epsilon(\gamma)$ such that for every $0<\alpha<\alpha(\gamma)$, there exists $\epsilon(\alpha)$ such that for $\epsilon(\alpha)<\epsilon<\epsilon(\gamma)$, $X^{C}$ is $\gamma$-ensemble stable, and for $0<\epsilon<\epsilon(\alpha)$, $X^{B}$ is $\gamma$-ensemble stable. Observe that for $\alpha=0$, both $X^{B}$ and $X^{C}$ are stochastically stable (they appear with the frequency $1/2$ in the limit of zero noise) but $X^{C}$ is low-noise ensemble stable. For small $\alpha > 0$, $X^{B}$ is both stochastically (it appears with the frequency $1$ in the limit of zero noise) and low-noise ensemble stable. However, for an intermediate noise $\epsilon(\alpha)<\epsilon<\epsilon(\gamma)$, if the number of players is large enough, then in the long run, almost all players use the strategy $C$ ($X^{C}$ is ensemble stable). If we lower $\epsilon$ below $\epsilon(\alpha)$, then almost all players start to use the strategy $B$. $\epsilon=\epsilon(\alpha)$ is the line of the first-order phase transition. In the thermodynamic limit, there exist two Gibbs states (equilibrium states) on this line. We may say that at $\epsilon=\epsilon(\alpha)$, the population of players undergoes a sharp [**equilibrium transition**]{} from $C$ to $B$-behaviour. Stochastic stability in non-potential games ------------------------------------------- Let us now consider non-potential games with three strategies and three symmetric Nash equilibria: $(A,A), (B,B)$, and $(C,C)$. Stationary measures of such games cannot be explicitly constructed. To find stochastically stable states we will use here the tree representation of stationary measures described in Chapter 7. We will discuss some interesting examples. [**Example 8**]{} Players are located on a finite subset of the one-dimensional lattice ${\bf Z}$ and interact with their nearest neighbours only. Denote by $n$ the number of players. For simplicity we will assume periodic boundary conditions, that is we will identify the $n+1$-th player with the first one. In other words, the players are located on the circle. The payoffs are given by the following matrix: A B C A $1+\alpha$ 0 1.5 U = B 0 2 0 C 0 0 3 with $0< \alpha \leq 0.5$. As before, we have three homogeneous Nash configurations: $X^{A}, X^{B}$, and $X^{C}$. The log-linear and perturbed best-response dynamics for this game were discussed in [@statphys]. Let us note that $X^{A}$, $X^{B}$, and $X^{C}$ are the only absorbing states of the noise-free dynamics. We begin with a stochastic dynamics with a state-independent noise. Let us consider first the case of $\alpha < 0.5$. If $0< \alpha< 0.5$, then $X^{C}$ is stochastically stable in the perturbed best-response dynamics. [**Proof:**]{} It is easy to see that $q_{m}(X^{C})$ is of the order $\epsilon^{2}$, $q_{m}(X^{B})$ is of the order $\epsilon^{\frac{n}{2}+1}$, and $q_{m}(X^{A})$ is of the order $\epsilon^{n+2}$. Let us now consider the log-linear rule. If $n<2+1/(0.5-\alpha)$, then $X^{B}$ is stochastically stable and if $n>2+1/(0.5-\alpha)$, then $X^{C}$ is stochastically stable in the log-linear dynamics. [**Proof**]{}: The following are maximal A-tree, B-tree, and C-tree: $$B \rightarrow C \rightarrow A, \hspace{3mm} C \rightarrow A \rightarrow B, \hspace{3mm} A \rightarrow B \rightarrow C,$$ where the probability of $A \rightarrow B$ is equal to $$\frac{1}{1+1+e^{\frac{1}{\epsilon}(2+2\alpha)}}(\frac{1}{1+e^{-\frac{2}{\epsilon}}+e^{\frac{1}{\epsilon}(-1+\alpha)}})^{n-2} \frac{1}{1+e^{-\frac{4}{\epsilon}}+e^{-\frac{4}{\epsilon}}},$$ the probability of $B \rightarrow C$ is equal to $$\frac{1}{1+1+e^{\frac{4}{\epsilon}}}(\frac{1}{1+e^{-\frac{1}{\epsilon}}+e^{-\frac{1.5}{\epsilon}}})^{n-2}\frac{1} {1+e^{-\frac{6}{\epsilon}}+e^{-\frac{3}{\epsilon}}},$$ and the probability of $C \rightarrow A$ is equal to $$\frac{1}{1+e^{-\frac{3}{\epsilon}}+e^{\frac{3}{\epsilon}}}(\frac{1}{1+e^{-\frac{1}{\epsilon}(2.5+\alpha)} +e^{\frac{1}{\epsilon}(0.5-\alpha)}})^{n-2}\frac{1}{1+e^{-\frac{2}{\epsilon}(1+\alpha)}+e^{-\frac{2}{\epsilon}(1+\alpha)}},$$ Let us observe that $$P_{B \rightarrow C \rightarrow A}= O(e^{-\frac{1}{\epsilon}(7+(0.5-\alpha)(n-2))}),$$ $$P_{C \rightarrow A \rightarrow B}= O(e^{-\frac{1}{\epsilon}(5+2\alpha+(0.5-\alpha)(n-2))}),$$ $$P_{A \rightarrow B \rightarrow C}= O(e^{-\frac{1}{\epsilon}(6+2\alpha)}),$$ where $\lim_{x \rightarrow 0}O(x)/x =1.$ Now if $n<2+1/(0.5-\alpha)$, then $$\lim_{\epsilon \rightarrow 0}\frac{q_{m}(X^{C})}{q_{m}(X^{B})}= \lim_{\epsilon \rightarrow 0}\frac{P_{A \rightarrow B \rightarrow C}} {P_{C \rightarrow A \rightarrow B}}=0$$ which finishes the proof. It follows that for a small enough $n$, $X^{B}$ is stochastically stable and for a large enough $n$, $X^{C}$ is stochastically stable. We see that adding two players to the population may change the stochastic stability of Nash configurations. Let us also notice that the strategy $C$ is globally risk dominant. Nevertheless, it is not stochastically stable in the log-linear dynamics for a sufficiently small number of players. Let us now discuss the case of $\alpha=0.5$ [@statphys]. If $\alpha=0.5$, then $X^{B}$ is stochastically stable for any $n$ in the log-linear dynamics. [**Proof**]{}: $$\lim_{\epsilon \rightarrow 0}\frac{q_{m}(X^{C})}{q_{m}(X^{B})}= \lim_{\epsilon \rightarrow 0}\frac{e^{-\frac{4}{\epsilon}}e^{-\frac{3}{\epsilon}}}{(1/2)^{n-2}e^{-\frac{3}{\epsilon}}e^{-\frac{3}{\epsilon}}}=0.$$ $X^{B}$ is stochastically stable which means that for any fixed number of players, if the noise is sufficiently small, then in the long run we observe $B$ players with an arbitrarily high frequency. However, we conjecture that for any low but fixed noise, if the number of players is big enough, the stationary measure is concentrated on the $X^{C}$-ensemble. We expect that $X^{C}$ is ensemble stable because its lowest-cost excitations occur with a probability of the order $e^{-\frac{3}{\epsilon}}$ and those from $X^{B}$ with a probability of the order $e^{-\frac{4}{\epsilon}}$. We observe this phenomenon in Monte-Carlo simulations. [**Example 9**]{} Players are located on a finite subset $\Lambda$ of ${\bf Z}$ (with periodic boundary conditions) and interact with their two nearest neighbours. They have at their disposal three pure strategies: $A, B,$ and $C$. The payoffs are given by the following matrix [@statmech]: A B C A 3 0 2 U = B 2 2 0 C 0 0 3 Our game has three Nash equilibria: $(A,A), (B,B)$, and $(C,C)$. Let us note that in pairwise comparisons, $B$ risk dominates $A$, $C$ dominates $B$ and $A$ dominates $C$. The corresponding spatial game has three homogeneous Nash configurations: $X^{A}, X^{B}$, and $X^{C}$. They are the only absorbing states of the noise-free best-response dynamics. $X^{B}$ is stochastically stable [**Proof**]{}: The following are maximal A-tree, B-tree, and C-tree: $$B \rightarrow C \rightarrow A, \hspace{3mm} C \rightarrow A \rightarrow B, \hspace{3mm} A \rightarrow B \rightarrow C.$$ Let us observe that $$P_{B \rightarrow C \rightarrow A}= O(e^{-\frac{6}{\epsilon}}),$$ $$P_{C \rightarrow A \rightarrow B}= O(e^{-\frac{4}{\epsilon}}),$$ $$P_{A \rightarrow B \rightarrow C}= O(e^{-\frac{6}{\epsilon}}).$$ The theorem follows from the tree characterization of stationary measures. $X^{B}$ is stochastically stable because it is much more probable (for low $\epsilon$) to escape from $X^{A}$ and $X^{C}$ than from $X^{B}$. The relative payoffs of Nash configurations are not relevant here (in fact $X^{B}$ has the smallest payoff). Let us recall Example 7 of a potential game, where an ensemble-stable configuration has more lowest-cost excitations. It is easier to escape from an ensemble-stable configuration than from other Nash configurations. Stochatic stability concerns single configurations in the zero-noise limit; ensemble stability concerns families of configurations in the limit of the infinite number of players. It is very important to investigate and compare these two concepts of stability in nonpotential games. Non-potential spatial games cannot be directly presented as systems of interacting particles. They constitute a large family of interacting objects not thoroughly studied so far by methods statistical physics. Some partial results concerning stochastic stability of Nash equilibria in non-potential spatial games were obtained in [@ellis1; @ellis2; @blume1; @physica; @statphys]. One may wish to say that $A$ risk dominates the other two strategies if it risk dominates them in pairwise comparisons. In Example 9, $B$ dominates $A$, $C$ dominates $B$, and finally $A$ dominates $C$. But even if we do not have such a cyclic relation of dominance, a strategy which is pairwise risk-dominant may not be stochastically stable as in the case of Example 8. A more relevant notion seems to be that of a global risk dominance [@mar]. We say that $A$ is globally risk dominant if it is a best response to a mixed strategy which assigns probability $1/2$ to $A$. It was shown in [@ellis1; @ellis2] that a global risk-dominant strategy is stochastically stable in some spatial games with local interactions. A different criterion for stochastic stability was developed by Blume [@blume1]. He showed (using techniques of statistical mechanics) that in a game with $m$ strategies $A_{i}$ and $m$ symmetric Nash equilibria $(A_{k},A_{k})$, $k=1,...,m$, $A_{1}$ is stochastically stable if $$\min_{k>1}(U(A_{1},A_{1})-U(A_{k},A_{k})) > \max_{k>1}(U(A_{k},A_{k})-U(A_{1},A_{k})).$$ We may observe that if $A_{1}$ satisfies the above condition, then it is pairwise risk dominant. Dominated strategies -------------------- We say that a pure strategy is [**strictly dominated**]{} by another (pure or mixed) strategy if it gives a player a lower payoff than the other one regardless of strategies chosen by his opponents. $k \in S$ is strictly dominated by $y \in \Delta$ if $U_{i}(k,w_{-i}) < U_{i}(y,w_{-i})$ for every $w \in \Delta^{I}$. Let us see that a strategy can be strictly dominated by a mixed strategy without being strictly dominated by any pure strategy in its support. [**Example 10**]{} A B C A 5 1 3 U = B 2 2 2 C 1 5 3 $B$ is strictly dominated by a mixed strategy assigning the probability $1/2$ both to $A$ and $C$ but is strictly dominated neither by $A$ nor by $C$. It is easy to see that strictly dominated pure strategies cannot be present in the support of any Nash equilibrium. In the replicator dynamics (16), all strictly dominated pure strategies are wiped out in the long run if all strategies are initially present [@akin; @samzhang]. If a pure strategy $k$ is strictly dominated,\ then $\xi_{k}(t,x^{0}) \rightarrow_{t \rightarrow \infty} 0$ for any $x^{0} \in interior(\Delta)$. Strictly dominated strategies should not be used by rational players and consequently we might think that their presence should not have any impact on the long-run behaviour of the population. We will show that in the best-reply dynamics, if we allow players to make mistakes, this may not be necessarily true. Let us consider the following game with a strictly dominated strategy and two symmetric Nash equilibria [@statmech]. [**Example 11**]{} A B C A 0 0.1 1 U = B 0.1 $2+\alpha$ 1.1 C 1.1 1.1 2, where $\alpha>0$. We see that strategy $A$ is strictly dominated by both $B$ and $C$, hence $X^{A}$ is not a Nash configuration. $X^{B}$ and $X^{C}$ are both Nash configurations but only $X^{B}$ is a ground-state configuration for $-U.$ In the absence of $A$, $B$ is both payoff and risk-dominant and therefore is stochastically stable and low-noise ensemble stable. Adding the strategy $A$ does not change dominance relations; $B$ is still payoff and pairwise risk dominant. However, Example 11 fulfills all the assumptions of Theorem 16 and we get that $X^{C}$ is $\gamma$-ensemble stable at intermediate noise levels. The mere presence of a strictly dominated strategy $A$ changes the long-run behaviour of the population. Similar results were discussed by Myatt and Wallace [@wallace]. In their games, at every discrete moment of time, one of the players leaves the population and is replaced by another one who plays the best response. The new player calculates his best response with respect to his own payoff matrix which is the matrix of a common average payoff modified by a realization of some random variable with the zero mean. The noise does not appear in the game as a result of players’ mistakes but is the effect of their idiosyncratic preferences. The authors then show that the presence of a strictly dominated strategy may change the stochastic stability of Nash equilibria. However, the reason for such a behavior is different in their and in our models. In our model, it is relatively easy to get out of $X^{C}$ and this makes $X^{C}$ ensemble stable. Mayatt and Wallace introduce a strictly dominated strategy in such a way that it is relatively easy to make a transition to it from a risk and payoff-dominant equilibrium and then with a high probability the population moves to a second Nash configuration which results in its stochastic stability. This is exactly a mechanism present in Examples 8 and 9. Review of other results ======================= We discussed the long-run behaviour of populations of interacting individuals playing games. We have considered deterministic replicator dynamics and stochastic dynamics of finite populations. In spatial games, individuals are located on vertices of certain graphs and they interact only with their neighbours. In this paper, we considered only simple graphs - finite subsets of the regular $\bf Z$ or ${\bf Z}^{2}$ lattice. Recently there appeared many interesting results of evolutionary dynamics on random graphs, Barabasi-Albert free-scale graphs, and small-world networks [@szabo3; @szabo4; @szabo5; @szabo7; @szabo8; @santos1; @santos2; @santos3; @antal]. Especially the Prisoner’s Dilemma was studied on such graphs and it was shown that their heterogeneity favors the cooperation in the population [@santos1; @santos2; @santos3; @szabo8]. In well-mixed populations, individuals are randomly matched to play a game. The deterministic selection part of the dynamics ensures that if the mean payoff of a given strategy is bigger than the mean payoff of the other one, then the number of individuals playing the given strategy increases. In discrete moments of time, individuals produce offspring proportional to their payoffs. The total number of individuals is then scaled back to the previous value so the population size is constant. Individuals may mutate so the population may move against a selection pressure. This is an example of a stochastic frequency-dependent Wright-Fisher process [@fisher1; @fisher2; @wright; @burger; @ewens]. There are also other stochastic dynamics of finite populations. The most important one is the Moran process [@moran; @burger; @ewens]. In this dynamics, at any time step a single individual is chosen for reproduction with the probability proportional to his payoff, and then his offspring replaces the random chosen individual. It was showed recently that in the limits of the infinite population, the Moran process results in the replicator dynamics [@claussen1; @claussen2]. The stochastic dynamics of finite populations has been extensively studied recently [@nowak3; @nowak4; @doebeli3; @nowak5; @lieberman; @ohtsuki1; @ohtsuki2; @ohtsuki3; @traulsen1; @traulsen2]. The notion of an evolutionarily stable strategy for finite populations was introduced [@nowak3; @nowak4; @neill; @wild; @dostal; @traulsen]. One of the important quantity to calculate is the fixation probability of a given strategy. It is defined as the probability that a strategy introduced into a population by a single player will take over the whole population. Recently, Nowak et. al. [@nowak3] have formulated the following weak selection 1/3 law. In two-player games with two strategies, selection favors the strategy $A$ replacing $B$ if the fraction of A-players in the population for which the average payoff for the strategy $A$ is equal to the average payoff of the strategy $B$ if is smaller than $1/3$, i.e. the mixed Nash equilibrium for this game is smaller than $1/3$. The $1/3$ law was proven to hold both in the Moran [@nowak3; @nowak4] and the Wright-Fisher process [@nowak5]. In this review we discussed only two-player games. Multi-player games were studied recently in [@kim; @broom; @multi; @physica; @tplatk1; @tplatk2; @bulletin; @tplatk3]. We have not discussed at all population genetics in the context of game theory. We refer to [@hof2; @burger] for results and references. [**Acknowledgments**]{}: These lecture notes are based on the short course given in the Stefan Banach International Mathematical Center in the framework of the CIME Summer School “From a Microscopic to a Macroscopic Description of Complex Systems" which was held in Bȩdlewo, Poland, 4-9 September 2006. I would like to thank the Banach Center for a financial support to participate in this School and the Ministry of Science and Higher Education for a financial support under the grant N201 023 31/2069. I thank Martin Nowak and Arne Traulsen for useful comments. [99]{} Akin, E.: Domination or equilibrium. Math. Biosci., **50**, 239–250 (1980) Alboszta, J., Miȩkisz, J.: Stability of evolutionarily stable strategies in discrete replicator dynamics with time delay. J. Theor. Biol., **231**, 175–179 (2004) Antal, T., Redner, S., Sood, V.: Evolutionary dynamics on degree-heterogeneous graphs. Phys. Rev. Lett., **96**, 188104 (2006) Axelrod, R.: The Evolution of Cooperation. Basic Books, New York (1984) Blume, L.E.: The statistical mechanics of strategic interaction. Games Econ. Behav., **5**, 387–424 (1993) Blume, L.E.: How noise matters. Games Econ. Behav., **44**, 251–271 (2003) Brauchli, K., Killingback, T., Doebeli, M.: Evolution of cooperation in spatially structured populations. J. Theor. Biol., **200**, 405–417 (1999) Bricmont, J., Slawny, J.: First order phase transitions and perturbation theory. In: Statistical Mechanics and Field Theory: Mathematical Aspects. Lecture Notes in Physics, **257**, Springer-Verlag (1986) Bricmont, J., Slawny, J.: Phase transitions in systems with a finite number of dominant ground states. J. Stat. Phys., **54**, 89–161 (1989) Broom, M., Cannings, C., Vickers, G.T.: Multi-player matrix games. Bull. Math. Biol. **59**, 931–952 (1997) Bukowski, M., Miȩkisz, J.: Evolutionary and asymptotic stability in symmetric multi-player games. Int. J. Game Theory, **33**, 41–54 (2004) Bürger, R.: The Mathematical Theory of Selection, Recombination, and Mutation. Wiley (2000) Condensed Matter and Quantitative Biology archives at http://xxx.lanl.gov Cressman, R.: Evolutionary Dynamics and Extensive Form Games. MIT Press, Cambridge, USA (2003) Dostálková, I., Kindlmann, P.: Evolutionarily stable strategies for stochastic processes. Theor. Pop. Biol., **65** 205–210 (2004) Econophysics bulletin at http://www.unifr.ch/econophysics Ellison, G.: Learning, local interaction, and coordination. Econometrica, **61**, 1047–1071 (1993) Ellison, G. Basins of attraction, long-run stochastic stability, and the speed of step-by-step evolution. Rev. Econ. Stud., **67**, 17–45 (2000) Ewens, W.J.: Mathematical Population Genetics. Springer (2004) Fisher, R. A.: On the dominance ratio. Proc. R. Soc. Edinburgh, **42**, 321–341 (1922) Fisher, R. A.: The Genetical Theory of Natural Selection. Clarendon Press, Oxford (1930) Foster, D., Young, P.H.: Stochastic evolutionary game dynamics. Theor. Pop. Biol., **38**, 219–232 (1990) Freidlin, M.I, Wentzell, A.D.: Random Perturbations of Dynamical Systems. Springer Verlag, New York (1984) Fudenberg, D., Tirole, J.: Game Theory. MIT Press, Cambridge, USA (1991) Fudenberg, D., Levine, D.K.: The Theory of Learning in Games. MIT Press, Cambridge, USA (1998) Georgii, H. O.: Gibbs Measures and Phase Transitions. Walter de Gruyter, Berlin (1988) Gintis, H.: Game Theory Evolving. Princeton University Press, Princeton (2000) Gulyas, L., Płatkowski, T.: On evolutionary 3-person Prisoner’s Dilemma games on 2d lattice. Lect. Notes Comp. Sci., **3305**, 831–840 (2004) Harsányi, J., Selten, R.: A General Theory of Equilibrium Selection in Games. MIT Press, Cambridge, USA (1988) Hauert, C.: Effects of space in 2x2 games. Int. J. Bifurcation Chaos, **12**, 1531–1548 (2002). Hauert, C., Doebeli, M.: Spatial structure often inhibits the evolution of cooperation in the snowdrift game. Nature, **428**, 643–646 (2004) Hauert, C.: Models of cooperation based on the Prisoner’s Dilemma and the Snowdrift game. Ecology Letters, **8**, 748–766 (2005). Hauert, C.: Spatial effects in social dilemmas. J. Theor. Biol., **240**, 627–636 (2006) Herz, A.V.M.: Collective phenomena in spatially extended evolutionary games. J. Theor. Biol., **169**, 65–87 (1994) Hofbauer, J., Schuster, P., Sigmund, K.: A note on evolutionarily stable strategies and game dynamics. J. Theor. Biol., **81**, 609–612 (1979) Hofbauer, J., Sigmund, K.: Evolutionary Games and Population Dynamics. Cambridge University Press, Cambridge (1998) Hofbauer, J., Sigmund, K.: Evolutionary game dynamics. Bull. Am. Math. Soc., **40**, 479–519 (2003) Imhof, L.A., Nowak, M.A.: Evolutionary game dynamics in a Wright-Fisher process. J. Math. Biol., **52**, 667-681 (2006) Kamiński, D., Miȩkisz, J., Zaborowski, M.: Stochastic stability in three-player games. Bull. Math. Biol., **67**, 1195–1205 (2005) Kandori, M., Mailath, G.J, Rob, R.: Learning, mutation, and long-run equilibria in games. Econometrica, **61**, 29–56 (1993) Killingback, T., Doebeli, M.: Spatial evolutionary game theory: hawks and doves revisited. Proc. R. Soc. London, **263**, 1135–1144 (1996) Kim, Y.: Equilibrium selection in n-person coordination games. Games Econ. Behav., **15**, 203–227 (1996) Lieberman, E., Hauert, C., Nowak, M.A.: Evolutionary dynamics on graphs. Nature, **433** 312–316 (2005) Lindgren, K., Nordahl, M.G.: Evolutionary dynamics of spatial games. Physica D, **75**, 292–309 (1994) Maruta, T.: On the relationship between risk-dominance and stochastic stability. Games Econ. Behav., **19**, 221–234 (1977) Maynard Smith, J., Price, G.R.: The logic of animal conflicts. Nature **246**, 15–18 (1973) Maynard Smith, J.: The theory of games and the evolution of animal conflicts. J. Theor. Biol., **47**, 209–221 (1974) Maynard Smith, J.: Evolution and the Theory of Games. Cambridge University Press, Cambridge (1982) Monderer, D., Shapley, L.S.: Potential games. Games Econ. Behav., **14**, 124–143 (1996) Moran, P.A.P.: The Statistical Processes of Evolutionary Theory. Clarendon, Oxford (1962) Miȩkisz, J.: Statistical mechanics of spatial evolutionary games. Phys. A: Math. Gen., **37**, 9891–9906 (2004) Miȩkisz, J.: Stochastic stability in spatial games. J. Stat. Phys., **117**, 99–110 (2004) Miȩkisz, J.: Stochastic stability in spatial three-player games. Physica A, **343**, 175–184 (2004) Miȩkisz, J.: Equilibrium selection in evolutionary games with random matching of players. J. Theor. Biol., **232**, 47–53 (2005) Miȩkisz, J., Płatkowski, T.: Population dynamics with a stable efficient equilibrium. J. Theor. Biol., **237**, 363–368 (2005) Miȩkisz, J.: Long-run behavior of games with many players. Markov Processes and Relat. Fields, **11**, 371–388 (2005) Miȩkisz, J.: Equilibrium transitions in finite populations of players. In: Game Theory and Mathematical Economics. Banach Center Publications, **71**, 237–142 (2006) Myatt, D.P., Wallace, C.: A multinomial probit model of stochastic evolution. J. Econ. Theory, **113**, 286–301 (2003) Nash, J.: Equilibrium points in n-person games. Proc. Natl. Acad. Sci. USA, **36**, 48–49 (1950) Nash, J.: Non-cooperative games. Ann. Math., **54**, 287–295 (1951) Neill, D.B.: Evolutionary stability for large populations. J. Theor. Biol., **227**, 397–401 (2004) Nowak, M.A., May, R.M.: Evolutionary games and spatial chaos. Nature, **359**, 826–829 (1992) Nowak, M.A., May, R.M.: The spatial dilemmas of evolution. Int. J. Bifurcation Chaos, **3**, 35–78 (1993) Nowak, M.A., Bonhoeffer, S., May, R.M.: More spatial games. Int. J. Bifurcation Chaos, **4**, 33–56 (1994) Nowak, M.A., Sasaki, A., Taylor, C., Fudenberg, D.: Emergence of cooperation and evolutionary stability in finite populations. Nature, **428**, 646–650 (2004) Nowak, M.A., Sigmund, K.: Evolutionary dynamics of biological games. Science, **303**, 793–799 (2004) Nowak, M.A.: Five rules for the evolution of cooperation. Science **314**, 1560–1563 (2006) Nowak, M.A.: Evolutionary Dynamics. Harvard University Press, Cambridge, USA (2006) Ohtsuki, H., Nowak, M.A.: Evolutionary games on cycles. Proc. R. Soc. B., **273**, 2249–2256 (2006) Ohtsuki, H., Nowak, M.A.: The replicator equation on graphs. J. Theor. Biol., **243**, 86-97 (2006) Ohtsuki, H., Hauert, C., Lieberman, E., Nowak, M.A.: A simple rule for the evolution of cooperation on graphs and social networks. Nature, **441**, 502–505 (2006) Ordershook, P.: Game Theory and Political Theory. Cambridge University Press (1986) Poundstone, W.: Prisoner’s Dilemma. Doubleday (1992) Płatkowski, T.: Evolution of populations playing mixed multi-player games. Math. Comp. Mod., **39**, 981–989 (2004) Płatkowski, T., Stachowska, J.: ESS in multiplayer mixed games. Appl. Math. Comp., **167**, 592–606 (2005). Robson, A., Vega-Redondo, F.: Efficient equilibrium selection in evolutionary games with random matching. J. Econ. Theory **70**, 65–92 (1996) Russell, B.: Common Sense and Nuclear Warfare. Allen and Unwin Ltd., London (1959) Samuelson, L., Zhang, J.: Evolutionary stability in asymmetric games. J. Econ. Theory, **57**, 363–391 (1992) Samuelson, L.: Evolutionary Games and Equilibrium Selection. MIT Press, Cambridge, USA (1997) Santos, F.C., Pacheco, J.M.: Scale-free networks provide a unifying framework for the emergence of cooperation. Phys. Rev. Lett., **95**, 098104 (2005) Santos, F.C., Pacheco, J.M., Lenaerts, T.: Evolutionary dynamics of social dilemmas in structured heterogeneous populations. Proc. Natl. Acad. Sci. USA, **103**, 3490–3494 (2006) Santos, F.C., Rodrigues, J.F., Pacheco, J.M.: Graph topology plays a determinant role in the evolution of cooperation. Proc. R. Soc. B, **273**, 51–55 (2006) Shubert, B.: A flow-graph formula for the stationary distribution of a Markov chain. IEEE Trans. Syst. Man. Cybernet., **5**, 565–566 (1975) Sigmund, K.: Games of Life - Explorations in Ecology, Evolution and Behaviour. Oxford University Press (1993) Sinai, Y.G.: Theory of Phase Transitions: Rigorous Results. Pergamon, Oxford (1982) Szabo, G., Toke, C.: Evolutionary prisoner’s dilemma game on a square lattice. Phys. Rev. E, **58**, 69–73 (1998) Szabo, G., Antal, T., Szabo, P., Droz, M.: Spatial evolutionary prisoner’s dilemma game with three strategies and external constraints. Phys. Rev. E, **62**, 1095–1103 (2000) Szabo, G., Szolnoki, A., Izsak, R.: Rock-scissors-paper game on regular small-world networks. J. Phys. A: Math. Gen., **37** 2599-2609 (2004) Szabo, G., Vukov, J., Szolnoki, A.: Phase diagrams for an evolutionary prisoner’s dilemma game on two-dimensional lattices. Phys. Rev. E, **72**, 047107 (2005) Szabo, G., Gabor, F.: Evolutionary games on graphs. arXiv: cond-mat/0607344 (2006) Szolnoki, A., Szabo, G.: Phase transitions for rock-scissors-paper game on different networks. Phys. Rev. E, **70**, 037102 (2004) Tao, Y., Wang, Z.: Effect of time delay and evolutionarily stable strategy. J. Theor. Biol., **187**, 111–116 (1997) Taylor, C., Fudenberg, D., Sasaki, A., Nowak, M.A.: Evolutionary game dynamics in finite populations. Bull. Math. Biol., **66**, 1621-1644 (2004) Taylor, P.D., Jonker, L.B.: Evolutionarily stable strategy and game dynamics. Math. Biosci., **40**, 145–156 (1978) Traulsen, A., Claussen, J.C., Hauert, C.: Coevolutionary dynamics: from finite to infinite populations. Phys. Rev. Lett., **95**, 238701 (2005) Traulsen, A., Claussen, J.C., Hauert, C.: Coevolutionary dynamics in large but finite populations. Phys. Rev. E, **74**, 011901 (2006) Traulsen, A., Pacheco, J., Imhof, L.A.: Stochasticity and evolutionary stability. Phys. Rev. E, **74**, 021905 (2006) Traulsen, A., Nowak, M.A., Pacheco, J.M.: Stochastic payoff evaluation increases the temperature of selection. J. Theor. Biol., **244**, 349–356 (2007) Traulsen, A., Pacheco, J.M., Nowak, M.A.: Pairwise comparison and selection temperature in evolutionary game dynamics. J. Theor. Biol., doi:10.1016/j.jtbi.2007.01.002 (2007) Vega-Redondo, F.: Evolution, Games, and Economic Behaviour. Oxford University Press, Oxford (1996) Vukov, J., Szabo, G.: Evolutionary prisoner’s dilemma game on hierarchical lattices. Phys. Rev. E, **71**, 71036133 (2005) Vukov, J., Szabo, Szolnoki, A.: Cooperation in the noisy case: Prisoner’s dilemma game on two types of regular random graphs. Phys. Rev. E, **74**, 067103 (2006) Weibull, J.: Evolutionary Game Theory. MIT Press, Cambridge, USA (1995) Wild, G., Taylor, P.D.: Fitness and evolutionary stability in game theoretic models of finite populations. Proc. R. Soc. B, **271**, 2345–2349 (2004) Young, P.H.: The evolution of conventions. Econometrica, **61**, 57–84 (1993) Young, H.P.: Individual Strategy and Social Structure: An Evolutionary Theory of Institutions. Princeton University Press, Princeton (1998) Wentzell, A.D., Freidlin, M.I.: On small random perturbations of dynamical systems. Russian Math. Surveys, **25**, 1–55 (1970) Wright, S.: Evolution in Mendelian populations. Genetics, **6**, 97–159 (1931) Zeeman, E.: Dynamics of the evolution of animal conflicts. J. Theor. Biol., **89**, 249–270 (1981)

ArXiv

--- abstract: 'We provide a simple example that illustrates the advantage of adaptive over non-adaptive strategies for quantum channel discrimination. In particular, we give a pair of entanglement-breaking channels that can be perfectly discriminated by means of an adaptive strategy that requires just two channel evaluations, but for which no non-adaptive strategy can give a perfect discrimination using any finite number of channel evaluations.' author: - | Aram W. Harrow$^\ast$ Avinatan Hassidim$^\dagger$ Debbie W. Leung$^\ddagger$ John Watrous$^\ddagger$\ $^\ast$ [*Department of Mathematics, University of Bristol*]{}\ [*Bristol, United Kingdom*]{}\ $^\dagger$ [*Center for Theoretical Physics, Massachusetts Institute of Technology*]{}\ [*Cambridge, Massachusetts, USA*]{}\ $^\ddagger$ [*Institute for Quantum Computing, University of Waterloo*]{}\ [*Waterloo, Ontario, Canada*]{} date: 'September 1, 2009' title: | **Adaptive versus non-adaptive strategies for\ quantum channel discrimination** --- Introduction {#sec:introduction} ============ This paper concerns the problem of [*quantum channel discrimination*]{}. In this problem, two quantum channels $\Phi_0$ and $\Phi_1$ are fixed, and access to one of the two channels is made available. It is not known which of the two channels has been made available, however, and the goal is to correctly identify which of $\Phi_0$ and $\Phi_1$ it is. Several papers, including [@Acin01; @AharonovKN98; @ChildsPR00; @ChiribellaDP08; @D'ArianoPP01; @DuanFY09; @Hayashi08; @Kitaev97; @PianiW09; @Sacchi05; @Sacchi05b; @WangY06; @Watrous08], have discovered many interesting aspects of quantum channel discrimination. There exist related topics in the study of quantum information theory, including [*quantum parameter estimation*]{} (see, for instance [@FujiwaraI03; @ImaiH08; @JiWDFY08] and the references therein), but this paper will focus just on the specific problem of channel discrimination. A [*discrimination strategy*]{} for a quantum channel discrimination problem is a step-by-step procedure consisting of channel evaluations, along with quantum state preparations, operations, and measurements, that attempts to output the identity of the given channel. Generally speaking, one is typically interested in discrimination strategies that satisfy certain natural constraints; with one well-studied example being the discrimination strategies allowing a [*single evaluation*]{} of the unknown channel. An [*optimal*]{} discrimination strategy, among those satisfying a given collection of constraints, is simply one that maximizes the probability that the unknown channel is correctly identified, assuming it is selected according to a fixed distribution that is known ahead of time. One interesting aspect of quantum channel discrimination is that the use of an [*ancillary*]{} system is generally necessary for the optimal discrimination of two quantum channels, assuming just a single evaluation of the unknown channel is made available [@Kitaev97; @AharonovKN98; @D'ArianoPP01; @KitaevSV02]. In more precise terms, the optimal strategy to discriminate two channels may require that one first prepares the input system to the unknown channel in an entangled state with an ancillary system, followed by a joint measurement of that channel’s output together with the ancillary system. Even [*entanglement-breaking*]{} channels are sometimes better discriminated through the use of an ancillary system, despite the fact that their output systems must necessarily be unentangled with the ancillary system after their evaluation [@Sacchi05b]. There are two known special classes of channels that require no ancillary system for optimal discrimination: the unitary channels [@AharonovKN98; @ChildsPR00] and the classical channels. There is a striking possibility for quantum channel discrimination problems that cannot occur in the classical setting. If a pair of classical channels cannot be perfectly distinguished with one evaluation, then they cannot be perfectly distinguished with any finite number of evaluations. (This fact is easily proved, and a simple proof may be found later in the paper.) In contrast, it is possible for a pair of quantum channels to be discriminated perfectly when multiple evaluations are available, but not in the single evaluation case. For example, this generally happens in the case of unitary channels [@Acin01]. Another interesting aspect of quantum channel discrimination is the distinction between [*adaptive*]{} and [*non-adaptive*]{} strategies when multiple uses of the unknown channel are made available. In an adaptive strategy, one may use the outputs of previous uses of the channel when preparing the input to subsequent uses, whereas a non-adaptive strategy requires that the inputs to all uses of the given channel are chosen before any of them is evaluated. It was found in [@ChiribellaDP08] that unitary channels are insensitive to this distinction; adaptive strategies do not give any advantage over non-adaptive strategies for unitary channel discrimination. In the same paper, a pair of [*memory channels*]{} was shown to require an adaptive scheme for optimal discrimination, but the question of whether or not there exist ordinary (non-memory) channels with a similar property was stated as an open question. Although an example of [*three*]{} channels that require adaptive strategies for an optimal identification was presented in [@WangY06], we were not able to find any example of a pair of (ordinary, non-memory) channels in the literature that require adaptive strategies for optimal discrimination; and so the question appears to have been unresolved prior to this work. The purpose of the present paper is to demonstrate the necessity of adaptive schemes for optimal quantum channel discrimination. We do this by presenting an example of two quantum channels that can be perfectly discriminated given two adaptive channel evaluations, but for which [*no finite number*]{} of non-adaptive channel evaluations allows for a perfect discrimination. The channels in our example are [*entanglement-breaking*]{} channels, which provides further evidence suggesting that entanglement-breaking channels share similar properties to general quantum channels with respect to channel discrimination tasks. We note that a recent paper of Duan, Feng, and Ying [@DuanFY09] has provided a criterion for the perfect discrimination of pairs of quantum channels, as well as a general method to find adaptive strategies that allow for perfect discrimination. While no explicit examples were given in that paper, the existence of pairs of channels with similar properties to those in our example is implied. Our example was, however, obtained independently from that paper, and we hope that it offers some insight into the problem of quantum channel discrimination that is complementary to [@DuanFY09]. Finally, we note that a related (but weaker) phenomenon occurs in the context of classical channel discrimination. That is, there exist classical channels that can be better discriminated by adaptive strategies than by non-adaptive strategies, and we provide three simple examples illustrating this phenomenon. While we suspect that similar examples illustrating the advantages of adaptive discrimination strategies may be known to some researchers, we did not find any in the literature. That such examples exist is also interesting when contrasted with the fact that adaptive strategies for classical channel discrimination cannot improve the asymptotic rate at which the error probability exponentially decays with the number of channel uses [@Hayashi08]. Preliminaries ============= We will begin by summarizing some of the notation and terminology that is used in the subsequent sections of the paper. We will let $\X$, $\Y$ and $\W$ denote finite-dimensional complex Hilbert spaces, which will typically correspond to the input, output, and ancillary systems to be associated with channel discrimination tasks. The notation ${{\mathrm{L}}\left(\X,\Y\right)}$ refers to the space of all linear operators from $\X$ to $\Y$, ${{\mathrm{L}}\left(\X\right)}$ is shorthand for ${{\mathrm{L}}\left(\X,\X\right)}$, and ${{\mathrm{D}}\left(\X\right)}$ refers to the set of all density operators on $\X$. A similar notation is used for other spaces in place of $\X$ and $\Y$. The identity operator on $\X$ is denoted $\I_{\X}$. For the example to be presented in the main part of the paper, we will let $\X$ and $\Y$ be the spaces associated with two qubits and one qubit, respectively. The standard bases for these spaces are therefore $\{{\left| 00 \right\rangle},{\left| 01 \right\rangle},{\left| 10 \right\rangle},{\left| 11 \right\rangle}\}$ and $\{{\left| 0 \right\rangle},{\left| 1 \right\rangle}\}$. As is common, we will also write $${\left| + \right\rangle} = \frac{1}{\sqrt{2}}{\left| 0 \right\rangle} + \frac{1}{\sqrt{2}}{\left| 1 \right\rangle} \quad\quad\text{and}\quad\quad {\left| - \right\rangle} = \frac{1}{\sqrt{2}}{\left| 0 \right\rangle} - \frac{1}{\sqrt{2}}{\left| 1 \right\rangle},$$ and we write tensor products of these states and standard basis states in a self-explanatory way (e.g., ${\left| 1+ \right\rangle} = {\left| 1 \right\rangle}{\left| + \right\rangle}$). A [*quantum channel*]{} is a linear mapping of the form $\Phi : {{\mathrm{L}}\left(\X\right)} \rightarrow {{\mathrm{L}}\left(\Y\right)}$ that is both completely positive and trace-preserving. Every such quantum channel $\Phi$ can be expressed in Kraus form as $$\Phi(X) = \sum_{j = 1}^m A_j X A_j^{\ast}$$ for some choice of linear operators $A_1,\ldots,A_m\in{{\mathrm{L}}\left(\X,\Y\right)}$ satisfying the constraint $$\sum_{j = 1}^m A_j^{\ast} A_j = \I_{\X}.$$ The identity channel mapping ${{\mathrm{L}}\left(\W\right)}$ to itself is denoted $\I_{{{\mathrm{L}}\left(\W\right)}}$. The distinguishability of two quantum channels $\Phi_0,\Phi_1:{{\mathrm{L}}\left(\X\right)}\rightarrow{{\mathrm{L}}\left(\Y\right)}$ may be quantified by the distance induced by the [*diamond norm*]{} (or [*completely bounded trace norm*]{}) $$\label{eq:diamond} {\left\lVert{\mspace{1mu}}\Phi_0 - \Phi_1 {\mspace{1mu}}\right\rVert}_{\diamond} = \max_{\rho\in{{\mathrm{D}}\left(\X\otimes\W\right)}} {\left\lVert{\mspace{1mu}}(\Phi_0 \otimes \I_{{{\mathrm{L}}\left(\W\right)}})(\rho) - (\Phi_1 \otimes \I_{{{\mathrm{L}}\left(\W\right)}})(\rho) {\mspace{1mu}}\right\rVert}_1$$ where here $\W$ is assumed to have dimension at least that of $\X$. This quantity represents the greatest possible degree of distinguishability that can result by feeding an input state into the two channels, allowing for the possibility that the input system is entangled with an ancillary system. Assuming that a bit $a\in\{0,1\}$ is uniformly chosen at random, the quantity $$\frac{1}{2} + \frac{1}{4}{\left\lVert{\mspace{1mu}}\Phi_0 - \Phi_1 {\mspace{1mu}}\right\rVert}_{\diamond}$$ represents the optimal probability to correctly determine the value of $a$ by means of a physical process involving just a single evaluation of the channel $\Phi_a$. It therefore holds that $\Phi_0$ and $\Phi_1$ are perfectly distinguishable using a single evaluation if and only if ${\left\lVert{\mspace{1mu}}\Phi_0 - \Phi_1 {\mspace{1mu}}\right\rVert}_{\diamond} = 2$. Specification of the example and a perfect discrimination protocol ================================================================== We now describe our example of two quantum channels that are better discriminated using an adaptive strategy than by any non-adaptive strategy. First, we will give an intuitive description of the channels. The two channels, $\Phi_0$ and $\Phi_1$, both map two qubits to one and operate as follows. [2.5ex]{} Channel $\Phi_0$ measures the first input qubit with respect to the standard basis. If the result is 0, it outputs the state ${\left| 0 \right\rangle}$. If the result is 1, it measures the second qubit with respect to the standard basis. If the result is 0, then it outputs 0, and if the result is 1, then it outputs the completely mixed state $\I/2$. Channel $\Phi_1$ measures the first input qubit with respect to the standard basis. If the result is 0, it outputs the state ${\left| + \right\rangle}$. If the result is 1, it measures the second qubit with respect to the $\{{\left| + \right\rangle},{\left| - \right\rangle}\}$ basis. If the result is $+$, then it outputs 1, and if the result is $-$, then it outputs the completely mixed state $\I/2$. The intuition behind these channels is as follows. If the first input qubit is set to 0, then the output is a “key” state: ${\left| 0 \right\rangle}$ for channel $\Phi_0$ and ${\left| + \right\rangle}$ for the channel $\Phi_1$. If the first input is set to 1, and the second input qubit is the channel’s “key” state, then the channel identifies itself (i.e., $\Phi_0$ outputs 0 and $\Phi_1$ outputs 1). If, however, the first input qubit is set to 1 and the second qubit’s state is orthogonal to the channel’s “key” state, then the channel outputs the completely mixed state. This effectively means that the channel provides no information about its identity in this case. It is easy to discriminate these two channels with an adaptive strategy that requires two uses of the unknown channel. The following diagram describes such a strategy: =0.25pt (1000, 400)(0,0) (first)(200,200)[$\Phi_a$]{} (second)(800,200)[$\Phi_a$]{} (in1)(0,250)[${\left| 0 \right\rangle}$]{} (in2)(0,150)[$\rho$]{} (in3)(600,250)[${\left| 1 \right\rangle}$]{} (out)(1000,200)[${\left| a \right\rangle}$]{} (in1,first) (in2,first) (in3,second) (first,10,320,second,190,320) (second,out) (in2,first) (in3,second) (first,10,320,second,190,320) (second,out) Here, the state $\rho$ input as the second qubit of the first channel evaluation is arbitrary, as it is effectively discarded by both of the channels when the first input qubit is set to ${\left| 0 \right\rangle}$. In the interest of precision, and because it will be useful for the analysis of the next section, we note the following formal specifications of these channels. It holds that $$\Phi_0(X) = \sum_{j = 1}^5 A_j X A_j^{\ast} \quad\quad\text{and}\quad\quad \Phi_1(X) = \sum_{j = 1}^5 B_j X B_j^{\ast}$$ for [5]{} A\_1 & = [| 0 ]{}[00 |]{}, & A\_2 & = [| 0 ]{}[01 |]{}, & A\_3 & = [| 0 ]{}[10 |]{}, & A\_4 & = [| 0 ]{}[11 |]{}, & A\_5 & = [| 1 ]{}[11 |]{},\ B\_1 & = [| + ]{}[00 |]{}, & B\_2 & = [| + ]{}[01 |]{}, & B\_3 & = [| 1 ]{}[1+ |]{}, & B\_4 & = [| 0 ]{}[1- |]{}, & B\_5 & = [| 1 ]{}[1- |]{}. It is clear that $\Phi_0$ and $\Phi_1$ are both entanglement-breaking channels, as all of these Kraus operators have rank one [@HorodeckiSR03]. Sub-optimality of non-adaptive strategies ========================================= We now prove that non-adaptive strategies cannot allow for a perfect discrimination of the channels $\Phi_0$ and $\Phi_1$ defined in the previous section, for any finite number $n$ of channel uses. In more precise terms, we have $${\left\lVert{\mspace{1mu}}\Phi_0^{\otimes n} - \Phi_1^{\otimes n} {\mspace{1mu}}\right\rVert}_{\diamond} < 2$$ for all choices of $n\in\natural$. We first prove a simpler mathematical fact, which is that there does not exist a two-qubit density operator $\rho$ for which $\Phi_0(\rho)$ and $\Phi_1(\rho)$ are perfectly distinguishable. As we will see, the proof is similar when taking the tensor product of the channel with itself or with an identity channel that acts on an auxiliary system. This handles the multiple-copy case with the possible use of an ancillary space, thus establishing the more general statement above. Assume toward contradiction that there exists a density operator $\rho$ such that $\Phi_0(\rho)$ and $\Phi_1(\rho)$ are perfectly distinguishable. By a simple convexity argument, we may assume that the same is true for a pure state ${|\psi\rangle\langle\psi|}$ in place of $\rho$. In other words, there exists a unit vector ${\left| \psi \right\rangle}$ satisfying $$\label{eq:trace-zero} {\operatorname{Tr}}\( \, \Phi_1({|\psi\rangle\langle\psi|}) \, \Phi_0({|\psi\rangle\langle\psi|}) \, \) = 0 \,.$$ Expanding this equation in terms of the Kraus operators of $\Phi_0$ and $\Phi_1$ yields $$\sum_{j = 1}^5 \sum_{k = 1}^5 {\left\lvert{\mspace{1mu}}{\left\langle \psi \right|} B_j^{\ast} A_k {\left| \psi \right\rangle} {\mspace{1mu}}\right\rvert}^2 = 0 \,.$$ Each of the terms in this sum is nonnegative, and must therefore be zero, i.e., ${\left\langle \psi \right|} B_j^{\ast} A_k {\left| \psi \right\rangle} = 0$ for all choices of $j,k\in\{1,\ldots,5\}$. It follows that $$\label{eq:alpha-sum} {\left\langle \psi \right|} \sum_{j = 1}^5 \sum_{k = 1}^5 \alpha_{j,k} B_j^{\ast} A_k {\left| \psi \right\rangle} = 0$$ for every choice of complex numbers $\{\alpha_{j,k}\,:\,1\leq j,k\leq 5\}$. We will now obtain a contradiction by choosing the coefficients $\{\alpha_{j,k}\,:\,1\leq j,k\leq 5\}$ in such a way that cannot hold. In particular, by letting $$\alpha_{1,1} = \alpha_{2,2} = \sqrt{2},\quad \alpha_{3,5} = \alpha_{4,3} = 1,\quad\text{and}\quad \alpha_{4,4} = -2\sqrt{2},$$ and letting $\alpha_{j,k} = 0$ for all of the remaining values of $j$ and $k$, we find that $$\sum_{j = 1}^5 \sum_{k = 1}^5 \alpha_{j,k} B_j^{\ast} A_k = P$$ for $$P = {|00\rangle\langle00|} + {|01\rangle\langle01|} + {|11\rangle\langle11|} + {|1-\rangle\langle1-|} = \begin{pmatrix} 1 & 0 & 0 & 0\\ 0 & 1 & 0 & 0\\ 0 & 0 & 1/2 & -1/2\\ 0 & 0 & -1/2 & 3/2 \end{pmatrix}.$$ The operator $P$ is positive definite and therefore $\langle \psi | P | \psi \rangle > 0$ for every nonzero vector ${\left| \psi \right\rangle}$, which is in contradiction with . Having established a contradiction, we conclude that there cannot exist a density operator $\rho$ such that $\Phi_0(\rho)$ and $\Phi_1(\rho)$ are perfectly distinguishable as claimed. Now let us consider the general setting where an arbitrary finite number $n$ of (non-adaptive) channel uses, as well as an ancillary system of arbitrary size, are permitted. We may follow a similar proof to the one above to show that there cannot exist a unit vector ${\left| \psi \right\rangle}$ such that $$\label{eq:general-trace-zero} {\operatorname{Tr}}\left[ \( \Phi_1^{\otimes n} \otimes \I_{{{\mathrm{L}}\left(\W\right)}} \) (\,{|\psi\rangle\langle\psi|}\,) \( \Phi_0^{\otimes n} \otimes \I_{{{\mathrm{L}}\left(\W\right)}} \) (\,{|\psi\rangle\langle\psi|}\,) \right] = 0 \,,$$ where $\W$ is the space (of arbitrary finite dimension) that is to be associated with the ancillary system. We may express the relevant mappings in this expression in terms of the Kraus operators of $\Phi_0$ and $\Phi_1$ as follows: $$\begin{aligned} \(\Phi_0^{\otimes n} \otimes \I_{{{\mathrm{L}}\left(\W\right)}} \) (X) & = \sum_{1\leq j_1,\ldots,j_n\leq 5} (A_{j_1}\otimes\cdots\otimes A_{j_n}\otimes \I_{\W}) X (A_{j_1}\otimes\cdots\otimes A_{j_n}\otimes \I_{\W})^{\ast},\\ \(\Phi_1^{\otimes n} \otimes \I_{{{\mathrm{L}}\left(\W\right)}} \) (X) & = \sum_{1\leq j_1,\ldots,j_n\leq 5} (B_{j_1}\otimes\cdots\otimes B_{j_n}\otimes \I_{\W}) X (B_{j_1}\otimes\cdots\otimes B_{j_n}\otimes \I_{\W})^{\ast}.\end{aligned}$$ Now, for the same coefficients $\{\alpha_{j,k}\,:\,1\leq j,k\leq 5\}$ that were defined above, we find that $$\sum_{\substack{1\leq j_1,\ldots,j_n\leq 5\\[0.2mm]1\leq k_1,\ldots,k_n\leq 5}} \alpha_{j_1,k_1} \cdots \alpha_{j_n,k_n} \; B_{j_1}^{\ast}A_{k_1} \otimes\cdots\otimes B_{k_n}^{\ast} A_{j_n} \otimes \I_{\W} \;=\; P^{\otimes n} \otimes \I_{\W},$$ which is again positive definite. Therefore, there cannot exist a nonzero vector ${\left| \psi \right\rangle}$ for which $${\left\langle \psi \right|} B_{j_1}^{\ast}A_{k_1} \otimes\cdots\otimes B_{k_n}^{\ast} A_{j_n} \otimes \I_{\W} {\left| \psi \right\rangle} = 0$$ for all $j_1,\ldots,j_n$, $k_1,\ldots,k_n$. Consequently, does not hold for any nonzero vector ${\left| \psi \right\rangle}$, which implies that $\Phi_0$ and $\Phi_1$ cannot be perfectly discriminated by means of a non-adaptive strategy. When the number of evaluations $n$ of the unknown channel is small, one can efficiently compute the value ${\left\lVert{\mspace{1mu}}\Phi_0^{\otimes n} - \Phi_1^{\otimes n} {\mspace{1mu}}\right\rVert}_{\diamond}$ because it is the optimal value of a semidefinite programming problem [@Watrous09]. For instance, it holds that $${\left\lVert{\mspace{1mu}}\Phi_0 - \Phi_1 {\mspace{1mu}}\right\rVert}_{\diamond} = 1 + \frac{1}{\sqrt{2}} \; ,$$ and therefore the channels can be discriminated with a probability $$\frac{1}{2} + \frac{1}{4}{\left\lVert{\mspace{1mu}}\Phi_0 - \Phi_1 {\mspace{1mu}}\right\rVert}_{\diamond} \approx 0.9268$$ of correctness with just a single channel evaluation. For two non-adaptive queries, we used `CVX` [@GrantB09; @GrantB08], a package for specifying and solving convex programs in Matlab, to approximate the value $$\frac{1}{2} + \frac{1}{4} {\left\lVert{\mspace{1mu}}\Phi_0\otimes\Phi_0 - \Phi_1\otimes\Phi_1 {\mspace{1mu}}\right\rVert}_{\diamond} \approx 0.9771 \,.$$ One can also obtain an upper bound on the probability of success using any feasible solution to the dual problem. In fact, even obvious choices give fairly tight upper bounds. Thus, we establish a small, but finite, advantage of an adaptive strategy over a non-adaptive one for discriminating these channels. Remarks on classical channel discrimination {#sec:2classical} =========================================== The channels in our example above are entanglement-breaking channels, yet the optimal adaptive discriminating strategy operates in a distinctively quantum way: one out of two nonorthogonal key states is extracted from the first channel evaluation and coherently input to the second. A natural question arises, which is whether adaptive strategies also help when discriminating [*classical*]{} channels. It turns out that adaptive strategies indeed are better in the classical setting, although in a more limited respect. This section discusses a few basic facts and examples that illustrate this claim. A classical channel can, of course, be succinctly represented by a stochastic matrix $M$, where the vector $M{\left| k \right\rangle}$ represents the output distribution when the input is $k$. Throughout this section, we will let $M_0$ and $M_1$ denote the two possible channels in a classical channel discrimination problem. ### Advantages of adaptive classical strategies {#advantages-of-adaptive-classical-strategies .unnumbered} We will present three examples illustrating that adaptive strategies may give advantages over non-adaptive strategies for classical channel discrimination, restricting our attention to the special case where just two channel evaluations are permitted, and where one of two channels is given with equal probability. We have the following expression for the optimal success probability using an adaptive strategy in this setting: $$\label{eq:classical-adaptive} \frac{1}{2} + \frac{1}{4}\max_{k,f} \sum_{j} {\left\lVert{\mspace{1mu}}M_0(j,k)\,M_0{\left| f(j) \right\rangle} - M_1(j,k)\,M_1{\left| f(j) \right\rangle} {\mspace{1mu}}\right\rVert}_1.$$ In this expression, $j$ and $k$ range over all outputs and inputs, respectively, of the channels $M_0$ and $M_1$ (i.e., they are row and column indices). The function $f$ ranges over all maps from outputs to inputs (or row indices to column indices). An alternate expression for the optimal success probability is $$\frac{1}{2} + \frac{1}{4}\max_{k} \sum_{j} q(j,k) \max_{l} {\left\lVert{\mspace{1mu}}p_0(j,k)\,M_0{\left| l \right\rangle} - p_1(j,k)\,M_1{\left| l \right\rangle} {\mspace{1mu}}\right\rVert}_1,$$ where $$q(j,k) = \frac{M_0(j,k) + M_1(j,k)}{2}$$ and where $$p_a(j,k) = \frac{M_a(j,k)}{M_0(j,k)+M_1(j,k)}$$ is the probability that the unknown channel is $M_a$, conditioned on $k$ being chosen as the input and $j$ being obtained as the output. This illustrates that, at least for strategies allowing just two channel evaluations, that the optimal adaptive strategy for two uses of a classical channel can be readily found, by first finding the optimal input for each prior distribution over the chosen channel (this may be the input in the second use). We then compute the success probability given every prior distribution and one use of the channel. Finally, to choose an input to the first use of the channel, we choose an input which maximizes the probability of getting each prior times the success probability given that prior. [**Example 1.**]{} This “minimal” example shows that adaptive strategies are better than nonadaptive ones in some cases. The two channels are given by: $$M_0 = \begin{pmatrix} 1/3 & 8/9 \\[2mm] 2/3 & 1/9 \end{pmatrix}, \quad\quad M_1 = \begin{pmatrix} 0 & 1/3\\[2mm] 1 & 2/3 \end{pmatrix}.$$ One can verify that the best two-evaluation non-adaptive strategy is to input $1$ to both of the channel uses, which leads to a correct identification with probability $7/9$. The best adaptive strategy is to take $k = 2$ and $f(1) = 2,\;f(2) = 1$ in the formula , which gives a correct identification with probability $65/81$. Similar examples are abundant. [**Example 2.**]{} Here, the optimal 1-shot input is never used in the optimal non-adaptive scheme. The idea is to start with two optimal $1$-shot inputs $k, k'$ such that using $k'$ becomes more informative with $2$ parallel uses. Then we perturb the $k$-th column slightly so that $k$ becomes the unique optimal $1$-shot input. In this example, the optimal $1$-shot input $k$ still serves as the first input to the optimal adaptive scheme. Let the two channels be given by: $$M_0 = \begin{pmatrix} 0.86 & 0.45 & 1 & 0.5 \\[2mm] 0.14 & 0.1 & 0 & 0.5 \\[2mm] 0 & 0.45 & 0 & 0 \end{pmatrix}, \quad\quad M_1 = \begin{pmatrix} 0.15 & 0.1 & 0.5 & 0 \\[2mm] 0.85 & 0.8 & 0.5 & 1 \\[2mm] 0 & 0.1 & 0 & 0 \end{pmatrix}.$$ The best one-shot input is $k=1$ (probability of success is 0.855) (whereas $k'= 2$). The best parallel input pairs are $(2,3)$ and $(3,2)$ (probability of success is 0.9). Allowing adaptation, and using $k=1$ as the first input, $f(1) = 3, f(2) = 4, f(3) = 1$, the probability of success is 0.9275. [**Example 3.**]{} In this final example, the optimal $1$-shot input is [*not*]{} the first input to the optimal adaptive scheme. The idea is to have two optimal $1$-shot inputs in which one is more informative than the other if given a second use. Then, we perturb the column corresponding to the less informative input to be slightly better for the $1$-shot case. Let the two channels be given by: $$M_0 = \begin{pmatrix} 1 & 0.5 & 0.828 & 0.76\\[2mm] 0 & 0.5 & 0.092 & 0.04\\[2mm] 0 & 0 & 0.08 & 0.2 \end{pmatrix}, \quad\quad M_1 = \begin{pmatrix} 0.5 & 0 & 0.092 & 0.04\\[2mm] 0.5 & 1 & 0.828 & 0.76\\[2mm] 0 & 0 & 0.08 & 0.2 \end{pmatrix},$$ The best one-shot input is $3$ (probability of success is 0.868) but the best parallel input pairs to two uses are $(3,4)$ and $(4,3)$ (probability of success is 0.9336). The optimal adaptive scheme uses $k=4$ as the first input, and $f(j) = j$ for $j=1,2,3$, resulting in a probability of success of 0.9536. ### Perfect classical strategies {#perfect-classical-strategies .unnumbered} Finally, we give a simple proof of a fact claimed in the introduction of this paper, which is that if two classical channels are not perfectly distinguishable with a single evaluation, then they cannot be perfectly distinguished by any finite number of evaluations, even using an adaptive strategy. We will prove the contrapositive of this statement. Suppose that two classical channels $M_0$ and $M_1$ are perfectly discriminated by a discrimination strategy that uses $n$ channel evaluations. Without loss of generality we may assume the strategy takes the general form suggested in Figure \[fig:classical-discrimination\]. =0.25pt (1600, 200)(-50,0) (in)(0,100)[$F_0$]{} (out)(1600,100) (K1)(400,100)[$F_1$]{} (K2)(800,100)[$F_2$]{} (K3)(1200,100)[$F_3$]{} (M1)(200,150)[$M_a$]{} (M2)(600,150)[$M_a$]{} (M3)(1000,150)[$M_a$]{} (M4)(1400,150)[$M_a$]{} (in,M1) (M1,K1) (K1,M2) (M2,K2) (K2,M3) (M3,K3) (K3,M4) (M4,out) (in,K1) (K1,K2) (K2,K3) (K3,out) (p)(1300,100)[$q_a$]{} The assumption that the strategy perfectly discriminates $M_0$ and $M_1$ means that the final output distributions for the cases $a=0$ and $a=1$ have disjoint support. Our goal is to prove that $M_0$ and $M_1$ are perfectly discriminated with a single evaluation. The proof of this statement proceeds by induction on $n$. In case $n=1$ there is nothing to prove, so assume that $n\geq 2$. Consider the two distributions $q_0$ and $q_1$ that are illustrated in the figure. Each distribution $q_a$ represents the state of the discrimination strategy immediately before the final channel evaluation takes place, assuming the unknown channel is given by $M_a$. There are two cases: $q_0$ and $q_1$ have disjoint support, or they do not. If $q_0$ and $q_1$ do have disjoint support, then terminating the discrimination strategy after $n-1$ channel evaluations allows for a perfect discrimination, so by the induction hypothesis it is possible to discriminate the channels with a single evaluation. In the other case, where $q_0$ and $q_1$ do not have disjoint supports, there must exist a classical state $x$ of the strategy at the time under consideration for which $q_0(x)$ and $q_1(x)$ are both positive. Given that the discrimination strategy is perfect, and therefore has final distributions with disjoint supports, it must hold that evaluating $M_0$ and $M_1$ on $x$ results in distributions with disjoint supports. Therefore, $M_0$ and $M_1$ can be discriminated with a single channel evaluation as required. Conclusion ========== In this paper, we presented a pair of quantum channels that can be discriminated perfectly by a strategy making two adaptive channel evaluations, but which cannot be perfectly discriminated nonadaptively with any finite number of channel evaluations. One natural question that arises is whether our example can be generalized to show a similar advantage of general adaptive strategies making $n$ channel evaluations versus strategies that make channel evaluations with depth at most $n-1$. Although our example can be generalized in a natural way, we have not proved that it has the required properties with respect to depth $n-1$ strategies. Finally, for the example we have presented, we have found that although strategies making two non-adaptive channel evaluations cannot be perfect, they can be correct with high probability (about 97.7%). What is the largest possible gap between optimal adaptive versus non-adaptive strategies making two (or any other number of) channel evaluations? The only upper-bound we have on this gap is that channels $\Phi_0$ and $\Phi_1$ that are perfectly discriminated by two (adaptive or non-adaptive) evaluations must satisfy ${\left\lVert{\mspace{1mu}}\Phi_0 - \Phi_1 {\mspace{1mu}}\right\rVert}_{\diamond} \geq 1$, and can therefore be discriminated (with a single evaluation) with probability at least $3/4$ of correctness. Acknowledgements {#acknowledgements .unnumbered} ---------------- DL thanks Chris Fuchs for helpful comments. AWH was funded by the EPSRC grant “QIP IRC” and is grateful for the hospitality of the Perimeter Institute when some of this work was carried out. AH received support from the xQIT Keck fellowship. DL was funded by CRC, CFI, ORF, CIFAR, NSERC, and QuantumWorks. JW was funded by NSERC, CIFAR, and QuantumWorks. [JWD[[$^{+}$]{}]{}08]{} A. Ac[í]{}n. Statistical distinguishability between unitary operations. , 87(17):177901, 2001. D. Aharonov, A. Kitaev, and N. Nisan. Quantum circuits with mixed states. In [*Proceedings of the 30th Annual ACM Symposium on Theory of Computing*]{}, pages 20–30, 1998. G. Chiribella, G. D’Ariano, and P. Perinotti. Memory effects in quantum channel discrimination. , 101(18):180501, 2008. A. Childs, J. Preskill, and J. Renes. Quantum information and precision measurement. , 47(2–3):155–176, 2000. R. Duan, Y. Feng, and M. Ying. The perfect distinguishability of quantum operations. Manuscript, 2009. Available as arXiv.org e-Print 0908.0119. G. D’Ariano, P. Presti, and M. Paris. Using entanglement improves the precision of quantum measurements. , 87(27):270404, 2001. A. Fujiwara and H. Imai. Quantum parameter estimation of a generalized pauli channel. , 36(29):8093–8103, 2003. M. Grant and S. Boyd. Graph implementations for nonsmooth convex programs. In V. Blondel, S. Boyd, and H. Kimura, editors, [*Recent Advances in Learning and Control (a tribute to M. Vidyasagar)*]{}, Lecture Notes in Control and Information Sciences, pages 95–110. Springer, 2008. M. Grant and S. Boyd. : [Matlab]{} software for disciplined convex programming. Available from `http://stanford.edu/~boyd/cvx`, 2009. M. Hayashi. Discrimination of two channels by adaptive methods and its application to quantum system. Available as arXiv.org e-Print 0804.0686, 2008. M. Horodecki, P. Shor, and M. Ruskai. Entanglement breaking channels. , 15(6):629–641, 2003. H. Imai and M. Hayashi. Fourier analytic approach to phase estimation. , 11, 2009. Z. Ji, G. Wang, R. Duan, Y. Feng, and M. Ying. Parameter estimation of quantum channels. , 54(11):5172–5185, 2008. A. Kitaev. Quantum computations: algorithms and error correction. , 52(6):1191–1249, 1997. A. Kitaev, A. Shen, and M. Vyalyi. , volume 47 of [*Graduate Studies in Mathematics*]{}. American Mathematical Society, 2002. M. Piani and J. Watrous. All entangled states are useful for channel discrimination. , 102(25):250501, 2009. M. Sacchi. Entanglement can enhance the distinguishability of entanglement-breaking channels. , 72:014305, 2005. M. Sacchi. Optimal discrimination of quantum operations. , 71:062340, 2005. J. Watrous. Distinguishing quantum operations having few [Kraus]{} operators. , 8(9):819–833, 2008. J. Watrous. Semidefinite programs for completely bounded norms. Available as arXiv.org e-Print 0901.4709, 2009. G. Wang and M. Ying. Unambiguous discrimination among quantum operations. , 73(4):042301, 2006.

ArXiv

--- abstract: 'We consider classical and quantum dynamics of a free particle on SU(2) group manifold. The eigenfunctions of the Hamiltonian are constructed in terms of coordinate free objects' address: | Department of Theoretical Physics, A. Razmadze Institute of Mathematics,\ 1 Aleksidze Street, Ge 380093, Tbilisi, Georgia\ e-mails: chaffinch@apexmail.com, tskipa@usa.net author: - 'George Chavchanidze, Levan Tskipuri' title: 'Free particle on $SU(2)$ group manifold' --- Lagrangian description ====================== The dynamics of a free particle on $SU(2)$ group manifold is described by the Lagrangian $${\cal L}=\langle g^{-1}\, \dot g\, g^{-1}\,\dot g \rangle$$ where $g\in SU(2)$ and $\langle ~ \rangle$ denotes the normalized trace $\langle ~\rangle=-\frac{1}{2}Tr(~)$, which defines a scalar product in $su(2)$. The Lagrangian (1.1) defines the following dynamical equations $$\frac{d}{dt}\langle g^{-1}\dot g\rangle=0$$ otherwise ,one can notice that our Lagrangian has $SU(2)$ “right” and $SU(2)$ “left” symmetry .It means that it is invariant under the following transformations $$g\longrightarrow h_1\,g~~~~~~~~~~~~~~\mbox {"left" symmetry}$$ $$~g\longrightarrow g\,h_2~~~~~~~~~~~~~~\mbox {"right" symmetry}$$ According to the Noether’s theorem the corresponding conserving quantities are $$~R=g^{-1}\dot g ~~~~~~~\frac{d}{dt}R=0~~~~~~~~~~~~~\mbox {"right" symmetry}$$ $$L= \dot g\,g^{-1}~~~~~~~ \frac{d}{dt}L=0~~~~~~~~~~~~\mbox{"left" symmetry}$$ Now let’s introduce the basis of $su(2)$ algebra $$T_1=\left( \begin{array}{cr} i&0\\0&-i \end{array}\right),~~~~ T_2=\left( \begin{array}{cr} 0&~1\\-1&~0 \end{array}\right),~~~~ T_3=\left( \begin{array}{cr} 0&i\\i&0 \end{array}\right),$$ The elements of $su(2)$ are traceless anti-hermitian matrices, and any $A\in su(2)$ can be parameterized in the following way $$A= A^nT_n~~~~~~~~~n=1,2,3~~~~~~~~~~R= R^nT_n$$ Scalar product $AB=\langle AB\rangle=-\frac{1}{2}Tr(AB)$ provides that $$A^n=\langle A\,T_n\rangle~~~~~~~~~~~(\langle T_nT_m\rangle=\delta _{nm})$$ Now we can introduce 6 functions $$R_n=\langle T_nR \rangle~~~~~~~~~n=1,2,3~~~~~~~~~R=R^nT_n$$ $$~L_n=\langle T_nL \rangle~~~~~~~~~n=1,2,3~~~~~~~~~L=L^nT_n$$ which are integrals of motion. It is easy to find general solution of Euler-Lagrange equation $$\frac{d}{dt}g^{-1}\dot g=0~~~~~~~~\Rightarrow~~~~~~~g^{-1}\dot g=const$$ $$g=e^{Rt}g(0)$$ These are well known geodesics on Lie group. Hamiltonian description ======================= Working in a first order Hamiltonian formalism we construct new Lagrangian which is equivalent to the initial one $$\tilde {\cal L}=\langle R(g^{-1} \dot g-\tilde v)\rangle +\frac{1}{2} \langle \tilde v^2\rangle$$ in sense that variation of R provides $$g^{-1} \dot g =\tilde v$$ and $ \tilde {\cal L} $ reduces to $ \cal{L} $. Variation of $\tilde v $ gives $R=\tilde v $ and therefore we can rewrite equivalent Lagrangian $\tilde {\cal L} $ in terms of R and g $$\tilde {\cal L}=\langle R\,g^{-1} \dot g \rangle -\frac{1}{2} \langle R^2 \rangle$$ Where function $\frac{1}{2} \langle R^2 \rangle $ plays role of Hamiltonian and one-form $ \langle R\,g^{-1}\,d\,g \rangle $ is a symplectic potential $\theta $. External differential of $ \theta $ is the symplectic form $ \omega $,that determines Poisson brackets and the form of Hamilton’s equation. $$\omega =d\,\theta =-\langle g^{-1}d\,g\,\wedge d\,R \rangle -\langle R\,g^{-1} d\,g\,\wedge \, g^{-1}d\,g \rangle$$ $\omega $ provides isomorphism between vector fields and one-forms $$X~~\rightarrow ~~~X\rfloor \omega$$ Let ${\cal F}(SU(2))$ denote the real-valued smooth function on $SU(2)$.For an $f\in SU(2)$ there exists a Hamiltonian vector field satisfying $$X_f:~~~~~~~ X\rfloor \omega =-df$$ Where $ X\rfloor \omega $ denotes the contraction of $X$ with $\omega $. $X_f$ is called Hamiltonian vector field associated with $f$. According to the definition Poisson bracket of two function is as follows $$\{f,g\} = {{\cal L}_X}_f g=X_f\rfloor dg =\omega(X_f,X_g)$$ Where ${{\cal L}_X}_f g$ denotes Lie derivative of $g$ with respect to $X_f$. The skew symmetry of $\omega $ provides skew symmetry of Poisson bracket . Hamiltonian vector fields that correspond to $R_n,L_m $ and $g $ functions are the following $$X_n=X_{R_n}=(~[R,T_n]~,~g\,T_n~)=(~X_{R_n}^{(R)}~,X_{R_n}^{(g)})$$ $$Y_n=X_{L_m}=(~[R,g\,T_mg^{-1}]~,~T_mg~)=(~X_{L_m}^{(R)}~,X_{L_m}^{(g)})$$ Therefore Poisson brackets are $$\{ L_n,L_m\} =-2\epsilon_{nm}^{~~~k}\,L_k ~~~\{ R_n,R_m\} =2\epsilon_{nm}^{~~~k}\,R_k$$ $$\{ R_n,L_m\} =0$$ $$\{ R_n,g\} = g\,T_n ~~~~~\{ L_m,g\} = T_mg$$ the results are natural. $R$ and $L$ that correspond respectively to the “right” and “left” symmetry commute with each other and independently form $su(2)$ algebras. It is easy to write down Hamilton’s equations $$\dot g=\{ H,\,g\}=g\,R$$ $$\dot R=\{ H,\,R\}=0$$ We consider case of $SU(2)$ , but the same constructions can be applied to the other Lie groups. Quantization ============ Let’s introduce operators $$\hat R_n =~\frac{i}{2}{{\cal L}_X}_n~~$$ $$\hat L_m =-\frac{i}{2}{{\cal L}_Y}_m$$ They act on the square integrable functions(see Appendix A) on $SU(2)$ and satisfy quantum commutation relations $$[{\hat L}_n,{\hat L}_m] =i\epsilon_{nm}^{~~~k}\,{\hat L}_k$$ $$[{\hat R}_n,{\hat R}_m] =i\epsilon_{nm}^{~~~k}\,{\hat R}_k$$ $$[ {\hat R}_n,{\hat L}_m] =0~~~~~~~~$$ The Hamiltonian is defined as $$\hat H=\hat R^2=\hat L^2$$ Therefore the complete set of observables that commute with each other is as follows $$\hat H,~~~~\hat R_a,~~~~~\hat L_b$$ Where a and b unlike n and m are fixed.Using a simple generalization of a well known algebraic construction (see Appendix B)one can check that the eigenvalues of the quantum observables $ \hat H $,$ {\hat R}_a $ and $ {\hat L}_b $ are as follows $$\hat H~\psi_{lr}^j=j(j+1)~\psi_{lr}^j$$ where $j$ takes positive integer and half integer values $$j=0,~~\frac{1}{2},~~1,~~\frac{3}{2},~~2~...$$ $$\hat R_a~\psi_{lr}^j=r~\psi_{lr}^j$$ $$\hat L_b~\psi_{lr}^j =l~\psi_{lr}^j$$ with $r$ and $l$ taking values in the following range $$-j,~~-j+1,~...~,~~j-1,~~j$$ The main aim of the article is construction of the corresponding eigenfunctions $~\psi_{lr}^j$.The first step of this construction is proposition 1\ [**proposition 1.**]{} The function $\langle \tilde T\,g \rangle $ where $\tilde T=(I+iT_a)(I+iT_b)$ is an eigenfunction of $\hat H$,${~~\hat R}_a$ and ${\hat L}_b $ with eigenvalues, respectively $\frac{3}{4},~\frac{1}{2}$ and $\frac{1}{2}$ Proof of this proposition is straightforward. Using $\langle \tilde T\,g \rangle$ we construct the complete set of eigenfunctions of $\hat H$,$~~{\hat R}_a$ and ${\hat L}_b$ operators \[ eigenfunctions\] $$~\psi_{lr}^j={\hat L}_-^{j-l}{\hat R}_-^{j-r}~{\langle \tilde T\,g \rangle }^{2j}$$ (for the definition of the ${\hat R}_- $ and ${\hat L}_- $ operators see Appendix) that are defined up to a constant multiple. Indeed , acting on (40) with $\hat H$,${~~\hat R}_a$ and ${\hat L}_b$ operators and using commutation relations (see Appendix B) one can prove that equations (35-38) hold for $~\psi_{lr}^j$ defined by (40)\ Free particle on $S^2$ as a $SU(2)/U(1)$ coset model ==================================================== Free particle on $2D$ sphere can be obtained from our model by gauging $U(1)$ symmetry. In other words let’s consider the following local gauge transformations $$g \longrightarrow h(t)\,g \label{eq:gauge}$$ Where $h(t)\in U(1) \subset SU(2)$ is an element of $U(1).$ Without loss of generality we can take $$h=e^{\,\beta (t)\,T_3}$$ Since $T_3$ is antihermitian $h(t) \in U(1)$ and since $h(t)$ depends on $t$ Lagrangian $${\cal L}=\langle g^{-1}\, \dot g\, g^{-1}\,\dot g \rangle \label{eq:lag}$$ is not invariant under (\[eq:gauge\]) local gauge transformations. To make (\[eq:lag\]) gauge invariant we should replace $\frac{d}{dt}$ with a covariant derivative $\nabla g=(\frac{d}{dt}+B)\,g$ Where $B$ can be represented as follows $$B=b\,T_3\in su(2)$$ with transformation rule $$B \longrightarrow h\,B\,h^{-1}-\dot h\,h^{-1}$$ in the other words $$b \longrightarrow b- \dot \beta$$ The new Lagrangian $${\cal L}_G=\langle g^{-1}\, \nabla g\, g^{-1}\,\nabla g \rangle \label{eq:newlag}$$ is invariant under (\[eq:gauge\]) local gauge transformations. But this Lagrangian as well as every gauge invariant Lagrangian is singular. It contains additional non-physical degrees of freedom. To eliminate them we should eliminate $B$ using Lagrange equations $$\frac{\partial {\cal L}_G}{\partial B}=0~~~~ \Longrightarrow ~~~~b=-\langle \dot g\,g^{-1}\,T_3 \rangle$$ put it back in (\[eq:newlag\]) and rewrite last obtained Lagrangian in terms of gauge invariant (physical) variables. $${\cal L}_G=\langle (g^{-1}\, \dot g\,-L_3T_3)^2 \rangle$$ It’s obvious that the following $$X=g^{-1}T_3\,g~\in ~su(2)$$ element of $su(2)$ algebra is gauge invariant. Since $X \in su(2)$ it can be parameterized as follows $$X=x^a\,T_a$$ where $x^a$ are real functions on $SU(2)$ $$x_a=\langle X\,T_a \rangle$$ So we have three gauge invariant variables $~~x^a~~~a=1,2,3$ but it’s easy to check that only two of them are independent. Indeed $$\langle X^2\rangle =\langle g^{-1}T_3\,g\,g^{-1}T_3\,g\rangle =\langle T_3^2\rangle =1$$ otherwise $$\langle X^2\rangle =\langle x^a\,T_a\,x^b\,T_b\rangle =x^a\,x_a$$ So physical variables take values on $2D$ sphere. In other words configuration space of $SU(2)/U(1)$ model is sphere. By direct calculations one can check that having been rewritten in terms of gauge invariant variables ${\cal L}_G$ takes the form $${\cal L}_G=\frac{1}{4}\langle X^{-1}\, \dot X\, X^{-1}\,\dot X \rangle$$ This Lagrangian describes free particle on the sphere. Indeed, since $X=x^a\,T_a$ it’s easy to show that $${\cal L}_G=\frac{1}{4}\langle X^{-1}\, \dot X\, X^{-1}\,\dot X \rangle = \frac{1}{4}\langle X\, \dot X\, X\,\dot X \rangle =\frac{1}{2}\dot x^a\,\dot x_a$$ So $SU(2)/U(1)$ coset model describes free particle on $S^2$ manifold. Quantization of the coset model. ================================= Working in a first order Hamiltonian formalism (see (14)-(16))we get $$\tilde {\cal L}_G=\langle R(g^{-1} \dot g-\tilde u)\rangle +\frac{1}{2} \langle (\tilde u+g^{-1}\,B\,g)^2\rangle$$ $$\tilde {\cal L}=\langle R\,g^{-1} \dot g \rangle -\frac{1}{2} \langle R^2 \rangle$$ variation of $\tilde u$ provides: $$R=\tilde u + g^{-1}\,B\,g$$ $$\tilde u=R-g^{-1}\,B\,g$$ Rewriting $\tilde {\cal L}_G $ in terms of $R$ and $g$ leads to $$\begin{aligned} \tilde {\cal L}_G=\langle R\,g^{-1} \dot g \rangle -\frac{1}{2} \langle R^2 \rangle -\langle B\,g\,R\,g^{-1}\rangle =\langle R\,g^{-1} \dot g \rangle -\frac{1}{2} \langle R^2 \rangle -b\langle g\,R\,g^{-1}\,T_3\rangle \nonumber \\ =\langle R\,g^{-1} \dot g \rangle -\frac{1}{2} \langle R^2 \rangle -b\,L_3\end{aligned}$$ Due to the gauge invariance of (\[eq:newlag\]) we obtain constrained Hamiltonian system, where $\langle R\,g^{-1}\,dg\rangle $ is symplectic potential, $\frac{1}{2}\langle R^2\rangle$ plays role of Hamiltonian and $b$ is a Lagrange multiple, variation of which leads to the first class constrain: $$\phi =\langle g\,R\,g^{-1}\,T_3\rangle =\langle L\,T_3\rangle =L_3=0 \label{eq:constr}$$ Therefore coset model is equivalent to the initial one with (\[eq:constr\]) constrain. Using technique of the constrained quantization, instead of quantization of the coset model we can submit quantum model ,that corresponds to the free particle on $SU(2)$, to the following operator constrain $${\hat L}_3|\psi \rangle =0$$ $~$Free particle on $SU(2) ~~~~~\longrightarrow ~~~~$ Quantum particle on $SU(2)$ reduction $~~\downarrow ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\downarrow ~ ~~ $reduction $~$Free particle on $S^2~~~~~~~~~~\longrightarrow $      Quantum particle on $S^2$\ \ Hilbert space of the initial sistem (that is linear span of $\psi _{rl}^j ~~~~j=0,~\frac{1}{2},~1,~\frac{3}{2},~2,~...$ wave functions) reduces to the linear span of $\psi _{r0}^j~~~~j=0,~1,~2,~3,~...$ wave functions. Indeed, ${\hat L}_3\psi _{rl}^j$ implies $l=0$, and since $l=0~~~j$ is integer. Therefore $r$ takes $-j,~-j+1,~...,~j-1,~j$ integer values only. Wave functions $\psi _{rl}^j$ rewriten in terms of gauge invariant variables up to a constant multiple should coincide with well known spherical harmonics $$\psi _{r0}^j \sim {\cal J}_{jr}$$ One can chack the following $$\psi _{r0}^j \sim {\hat L}_-^j{\hat R}_-^{j-r}~{\langle \tilde T\,g \rangle }^{2j} \sim {\hat R}_-^{j-r}~{\langle T_+\,g^{-1}\,T_3\,g \rangle }^j \sim {\hat R}_-^{j-r}\,sin^j\,\theta \,e^{ij\theta } \sim {\hat R}_-^{j-r}\,{\cal J}_{jj} \sim {\cal J}_{jr}$$ This is an example of using large initial model in quantization of coset model.\ Appendix A =========== Scalar product in Hilbert space is defined as follows $$\langle \psi _1 | \psi _2 \rangle =\int _{SU(2)} \prod _{a=1}^{3} \langle g^{-1}\,dg\,T_a \rangle \psi _1^* \,\psi _2 \label{eq:scal}$$ It’s easy to prove that if scalar product is (\[eq:scal\]) operators ${\hat R}_n $ and $ {\hat L}_m $ are hermitian. Indeed $$\langle \psi _1 |{\hat R}_n \psi _2\rangle = \int _{SU(2)} \prod _{a=1}^{3} \langle g^{-1}\,dg\,T_a \rangle \psi _1^* \,(\frac{i}{2}{{\cal L}_X}_n \psi _2) \label{scal}=\int _{SU(2)} \prod _{a=1}^{3} \langle g^{-1}\,dg\,T_a \rangle (\frac{i}{2}{{\cal L}_X}_n\psi _1)^* \,\psi _2 \label{scal}$$ Where integration by part have been used. It’s easy to check that the additional term coming from measure $$\prod _{a=1}^{3}\langle g^{-1}\,dg\,T_a \rangle$$ vanishes since $${{\cal L}_X}_n \langle g^{-1}\,dg\,T_a \rangle$$ For more transparency one can introduce the following parameterization of $SU(2)$. For any $g \in SU(2)$. $$g=e^{q^a\,T_a}$$ Then the symplectic potential takes the form $$\langle R\,g^{-1}\,dg \rangle =R_a\,dq^a$$ and scalar product $$\langle \psi _1 |\psi _2\rangle =\int _0^{2\pi }\int _0^{2\pi }\int _0^{2\pi } d^3q\, \psi _1^* \,\psi _2$$ that coincides with (\[eq:scal\]) because of $$dq_a=\langle g^{-1}\,dg \,T_a\rangle$$ Appendix B =========== Without loss of generality we can take $\hat H$,$~\hat L_3 $ and $\hat R_3$ as a complete set of observables. Assuming that there exist at least one eigenfunctions of $\hat H$,$\hat L_3 $and $\hat R_3$ operators: $$\hat H\psi =E\psi$$ $$\hat R_3 \psi =r\psi$$ $$\hat L_3\psi =l\psi$$ It is easy to show that eigenvalues of $\hat H$ are non-negative $$E \geq 0$$ and $$E-r^2\geq 0 \label{eq:restr1}$$ $$E-l^2\geq 0 \label{eq:restr2}$$ Indeed operators $\hat R$ and $\hat L $ are selfadjoint so $$\langle \psi |\hat H |\psi \rangle =\langle \psi |{\hat R}^2 |\psi \rangle =\langle \psi |{\hat R}_a{\hat R}^a |\psi \rangle =\langle \psi |{{\hat R}_a}^{\dagger }{\hat R}^a |\psi \rangle = \langle {\hat R}_a\psi |{\hat R}^a\psi \rangle =\parallel {\hat R}_a \psi \parallel \geq 0$$ To prove (\[eq:restr1\])-(\[eq:restr2\]) we shall consider ${\hat R}_1^2+{\hat R}_2^2$ and ${\hat L}_1^2+{\hat L}_2^2$ operators $$\langle \psi |{\hat R}_1^2+{\hat R}_2^2 |\psi \rangle =\parallel {\hat R}_1 \psi \parallel + \parallel {\hat R}_2 \psi \parallel \geq 0$$ and $$\langle \psi |{\hat R}_1^2+{\hat R}_2^2|\psi \rangle =\langle \psi |\hat H-{\hat R}_3^2 |\psi \rangle =(E-r^2)\langle \psi |\psi \rangle$$ Therefore $E-r^2 \geq 0$ Now let’s introduce new operators $${\hat R}_+=i{\hat R}_1+{\hat R}_2~~~~~~~~~{\hat R}_-=i{\hat R}_1-{\hat R}_2$$ $${\hat L}_+=i{\hat L}_1+{\hat L}_2~~~~~~~~~{\hat L}_-=i{\hat L}_1-{\hat L}_2$$ These operators are not selfadjoint, but ${\hat R}_-^{\dagger}={\hat R}_+$ and ${\hat L}_-^{\dagger}={\hat L}_+$ and they fulfill the following commutation relations $$[{\hat R}_{\pm},{\hat R}_3]=\pm {\hat R}_{\pm}$$ $$[{\hat L}_{\pm},{\hat L}_3]=\pm {\hat L}_{\pm}$$ $$[{\hat R}_+,{\hat R}_-]=2{\hat R}_3$$ $$[{\hat L}_+,{\hat L}_-]=2{\hat L}_3$$ $$[{\hat R}_*,{\hat L}_*]=0~~~$$ where \* takes values +,–,3 using these commutation relations it is easy to show that if $ \psi_{rl}^\lambda $ is eigenfunction of $\hat H$,${\hat L}_3$ and ${\hat R}_3$ with corresponding eigenvalues : $$\hat H\psi _{rl}^{\lambda }=\lambda \psi _{rl}^{\lambda }$$ $${\hat L}_3{\psi _{rl}^{\lambda }}=l{\psi _{rl}^{\lambda }}$$ $${\hat R}_3{\psi _{rl}^{\lambda }}=r{\psi _{rl}^{\lambda }}$$ then ${\hat R}_{\pm}\psi _{rl}^{\lambda } $ and ${\hat L}_{\pm}\psi _{rl}^{\lambda } $ are the eigenfunctions with corresponding eigenvalues $\lambda ,l\pm 1,r$ and $\lambda ,l,r\pm 1$. Consequently using ${\hat R}_{\pm },{\hat L}_{\pm }$ operators we construct a family of eigenfunctions with eigenvalues $$l,~~~l \pm 1,~~~l \pm 2,~~~l \pm 3,~~~... \label{eq:seq1}$$ $$r,~~~r \pm 1,~~~r \pm 2,~~~r \pm 3,~~~... \label{eq:seq2}$$ but conditions (\[eq:restr1\]) and (\[eq:restr2\]) give restrictions on a possible range of eigenvalues. We should have $$\lambda -r^2 \geq 0$$ $$\lambda -l^2 \geq 0$$ In other words, in order to interrupt (\[eq:seq1\])-(\[eq:seq2\]) sequences we should have $${\hat L}_+ {\psi _{rj}^{\lambda }}=0~~~~{\hat L}_-{\psi _{r,-j}^{\lambda }}=0$$ $${\hat R}_+{\psi _{kl}^{\lambda }}=0~~~~{\hat R}_-{\psi _{-kl}^{\lambda }}=0$$ and for some $j$ and $k$ therefore $l$ and $r$ take the following values $$-j,~~~-j+1,~~~...~~j-1~~~,j$$ $$-k,~~~-k+1,~~~...~~k-1~~~,k$$ The number of values is $2j+1$ and $2k+1$ respectively. Since number of values should be integer , $j$ and $k$ should take integer or half integer values $$j=0,~~~\frac{1}{2},~~~1,~~~\frac{3}{2},~~~2,~~~...$$ $$k=0,~~~\frac{1}{2},~~~1,~~~\frac{3}{2},~~~2,~~~...$$ Now using commutation relations we can rewrite $\hat H$ in terms of $ {\hat R}_{\pm},{\hat R}_3 $ operators: $$\hat H={\hat R}_+ {\hat R}_-+{\hat R}_3^2+{\hat R}_3 \label{eq:ham}$$ ( \[eq:ham\] ) provides that $\lambda =j(j+1)=k(k+1)$ so $j=k$ and $\lambda =j(j+1)$ References {#references .unnumbered} ========== [99]{} V.I. Arnold , [*Mathematical methods of classical mechanics*]{}. (Springer, Berlin, 1978). A. Bohm, [*Quantum mechanics: foundations and applications*]{}. (Springer-Verlag, 1986). G. Jorjadze, L. O’Raifeartaigh, I. Tsitsui [*Quantization of a free relativistic particle on the $SL(2,R)$ manifold based on Hamiltonian reduction*]{}. (Physics Letters B 336 (1994) 388-394). G. Jorjadze,[*Hamiltonian reduction and quantization on symplectic manifolds*]{}. (Memoirs on Differential equations and Mathematical Phisics,Volume 13,1998,1-98). N.M.J. Woodhouse,[*Geometric Quantization*]{}. (Claredon, Oxford, 1992).

ArXiv

--- abstract: 'I report the discovery of blueshifted broad absorption line (BAL) troughs in at least six transitions of the Balmer series of hydrogen (H$\beta$ to H9) and in , and excited in the quasar SDSS J125942.80+121312.6. This is only the fourth active galactic nucleus known to exhibit Balmer absorption, all four in conjunction with low-ionization BAL systems containing excited . The substantial population in the $n=2$ shell of in this quasar’s absorber likely arises from Ly$\alpha$ trapping. In an absorber sufficiently optically thick to show Balmer absorption, soft X-rays from the quasar penetrate to large $\tau_{Ly\alpha}$ and ionize . Recombination then creates Ly$\alpha$ photons that increase the $n=2$ population by a factor $\tau_{Ly\alpha}$ since they require $\simeq\tau_{Ly\alpha}$ scatterings to diffuse out of the absorber. Observing Ly$\alpha$ trapping in a quasar absorber requires a large but Compton-thin column of gas along our line of sight which includes substantial but not too much dust. Presumably the rarity of Balmer-line BAL troughs reflects the rarity of such conditions in quasar absorbers.' author: - 'Patrick B. Hall' nocite: - '[@trump06]' - '[@wea91; @sdssbal]' - '[@cs01]' - '[@hut02]' - '[@aoki06]' - '[@dr5]' - '[@aoki06]' - '[@yor00]' - '[@gun98]' - '[@gun06]' - '[@fuk96; @sdss82; @sdss105; @sdss153; @ive04]' - '[@sdssqtarget]' - '[@bwh95]' - '[@sdss1st]' - '[@sdss123]' - '[@vea06]' - '[@sdss123]' - '[@bg92]' - '[@rvs05]' - '[@sb2]' - '[@bra02]' - '[@cbbl04]' - '[@of06]' - '[@sb2]' - '[@fn79]' - '[@of06]' - '[@sb2]' - '[@kro99]' - '[@cm69]' title: 'A Balmer-Line Broad Absorption Line Quasar' --- Background {#intro} ========== Broad absorption line (BAL) active galactic nuclei (AGN) show absorption from gas at blueshifted velocities up to 0.2$c$. They are seen in about one in four of the luminous AGN known as quasars, and while most BAL AGN only show high-ionization absorption (e.g., , ), $\sim$10% of BAL AGN also exhibit low-ionization absorption (e.g., , ), including $\sim$3% which in addition show excited absorption ([Trump]{} [et al.]{} 2006, and references therein). If all quasars have BAL troughs, these percentages represent the fractions of unobscured lines of sight toward quasars along which these three types of troughs exist. If all quasars go through a BAL phase, they represent the fractional amount of time each type exists. Reality is likely somewhere in between. On the one hand, BAL and non-BAL quasar spectra are very similar, arguing that they are drawn from a single parent population ([Weymann]{} [et al.]{} 1991; [Reichard]{} [et al.]{} 2003). On the other hand, the likelihood of observing a low-ionization BAL trough increases during a transition phase between ultraluminous infrared galaxies and quasars ([Canalizo]{} & [Stockton]{} 2001), although the origin of the BAL troughs in such objects is an open question; a disk wind origin for them remains possible. The physical parameters of BAL systems need to be constrained to understand their origin(s) and contribution to AGN feedback effects on galaxy formation and evolution. However, saturation and blending effects mean that the physical parameters of typical BAL systems cannot be easily inferred. Instead, it is often the more unusual systems that can provide such constraints. Currently the rarest known absorption in BAL AGN is Balmer-line absorption. It has previously been reported only in NGC 4151 ([Hutchings]{} [et al.]{} 2002, and references therein) and SDSS J083942.11+380526.3 ([Aoki]{} [et al.]{} 2006). Weak absorption similar to that found in those objects has also recently been seen in H$\beta$ and H$\gamma$ in the quasar FBQS J2107$-$0620 through high-resolution spectroscopy (Hutsemekers et al., in preparation). Here I report a quasar with very strong Balmer-line absorption found in the Sloan Digital Sky Survey Data Release Five ([Adelman-McCarthy]{} [et al.]{} 2007). These two new discoveries confirm the observation of [Aoki]{} [et al.]{} (2006) that Balmer-line BAL troughs are found in iron low-ionization BAL quasars with relatively strong \[\] emission. Observations {#obs} ============ The Sloan Digital Sky Survey (SDSS; [York]{} [et al.]{} 2000) is using a drift-scanning imaging camera ([Gunn]{} [et al.]{} 1998) on a 2.5-m telescope ([Gunn]{} [et al.]{} 2006) to image 10$^4$deg$^2$ of sky on the SDSS $ugriz$ AB magnitude system ([Fukugita]{} [et al.]{} 1996; [Hogg]{} [et al.]{} 2001; [Smith]{} [et al.]{} 2002; [Pier]{} [et al.]{} 2003; [Ivezi[ć]{}]{} [et al.]{} 2004). Two multi-fiber, double spectrographs are being used to obtain $R\sim1900$ spectra for $\sim$10$^6$ galaxies to $r=17.8$ and $\sim$10$^5$ quasars to $i=19.1$ ($i=20.2$ for $z>3$ candidates). As discussed in [Richards]{} [et al.]{} (2002), quasar candidates are targeted for spectroscopy because their colors differ from the colors of the stellar locus or because they are unresolved objects with radio emission detected by the FIRST survey ([Becker]{}, [White]{}, & [Helfand]{} 1995). SDSS J125942.80+121312.6 (hereafter SDSS J1259+1213) was observed on Modified Julian Date 53473 with SDSS plate \#1695 and fiber \#75. It has a Galactic extinction corrected magnitude $i=18.15\pm0.02$ and was targeted both via its colors and its radio emission. Its FIRST peak flux of $2.00\pm 0.14$ mJy/beam puts it right on the border between radio-loud and radio-quiet, with $R_i=1.00$ ([Ivezi[' c]{}]{} [et al.]{} 2002). Figure 1 shows that the quasar has a blue continuum with P Cygni-like emission and absorption in , excited-state and the Balmer series (H$\beta$ through H9, and possibly H10 and H11). The redshift of the quasar, measured from \[\] and consistent with that of \[\] $\lambda$5007, is $z=0.7517\pm0.0001$, yielding $M_i=-24.88$ ($\lambda_iL_{\lambda_i}=2\times 10^{45}$ ergs s$^{-1}$). The broad and H$\beta$ emission lines peak at $z=0.748\pm0.001$, slightly blueshifted from the narrow-line redshift. The deepest Balmer absorption is blueshifted further, to $z=0.7345\pm0.0004$, which is $2960\pm 70$kms$^{-1}$ from the narrow-line redshift. The continuum drop shortward of $\sim$2600Å rest frame is due to partial covering of the quasar by overlapping absorption from multiple transitions (e.g., [Hall]{} [et al.]{} 2002). The smaller absorption depths for neutral as compared to low-ionization and are consistent with an increasing covering factor with ionization level, often seen in BAL quasars. Emission and absorption in higher-excitation transitions of , and ([V[é]{}ron-Cetty]{} [et al.]{} 2006) can plausibly explain the complex continuum between and \[\]. (Similar absorption at 2800Å$<$$\lambda$$<$3500Å is seen in the low-ionization BAL quasar SDSS J112526.13+002901.3; [Hall]{} [et al.]{} 2002.) For example, the lower term of $d^2D^1-x^2P^o$, tentatively identified in Figure 1, is at 5.91 eV above ground. There is no sign of metastable absorption, but the limits are not strong since those transitions suffer from confusion with absorption from and H8. Lastly, the \[\] emission-line profile may have a complex blueshifted component, or it may just be confused with the strong emission in that region of the spectrum.[^1] Balmer line absorption is often seen in galaxy spectra, but this object cannot be explained as an unusual galaxy. First, the emission and broad absorption in ultraviolet and transitions is consistent only with a BAL quasar. Second, when post-starburst galaxies exhibit strong Balmer absorption, it lies at the same redshift as the \[\] and \[\] emission and is accompanied by a strong Balmer break lacking in this object. Third, when low-ionization outflows are seen in star-forming galaxies, they do not occur at velocities as high as 3000 kms$^{-1}$ unless they are clearly BAL troughs associated with an AGN ([Rupke]{}, [Veilleux]{}, & [Sanders]{} 2005). Interpretation {#interp} ============== The Balmer absorption troughs in this object have widths which are uniform within the measurement uncertainties: ${\rm FWZI}=2000\pm 200$ kms$^{-1}$. To estimate the depths of the troughs, the continuum was fitted with a fourth-order polynomial in $F_{\lambda}$ vs. $\log\lambda$ to approximate a power-law with possible slight dust reddening (see Figure 1). The maximum single-pixel depth of each trough was measured from the normalized SDSS spectrum after smoothing by a 7-pixel boxcar. The depths of the troughs (Table 1) decrease slowly as the upper term of the transition increases. The decline is about a factor of two from H$\beta$ to H9. The absorption therefore must be saturated, since the transition oscillator strengths decline by a factor of [*twenty-two*]{} from H$\beta$ to H9. The percentage absorption depth in a trough $l$ is $D_l=C(1-e^{-\tau_l})$ for an absorber with optical depth $\tau_l$ and percentage covering factor $C$ in the relevant ion (e.g., Eq. 1 of [Hall]{} [et al.]{} 2003). By assuming $C$ and $\tau_{H\beta}$, $D_l$ for every Balmer trough can be calculated since their relative $\tau$ are determined by their known oscillator strengths. Acceptable matches to the maximum absorption depths of the Balmer lines were found with $\tau_{H\beta}=19.5\pm2.5$ and $C=29\mp1\%$ at the velocity of maximum depth (Table 1); note that larger $\tau$ requires smaller $C$ for an equally good fit. The additional depth in the H$\beta$ trough is assumed to come from coverage of the H$\beta$ emission line.[^2] This approximate $\tau_{H\beta}$ assumes the Balmer-line troughs are resolved and takes into account neither possible Balmer emission from the absorber itself (as in the resonance scattering model of [Branch]{} [et al.]{} 2002; see also [Casebeer]{} [et al.]{} 2004) nor possible different covering factors for the continuum source, Balmer emission lines and emission-line blends. A high-resolution spectrum is needed to properly model all those parameters and determine how greatly they affect results inferred from an SDSS-resolution spectrum. Absorption from the $n=2$ shell of hydrogen requires a substantial population in that shell. One obvious population mechanism is collisional excitation into $n=2$ due to a high density. The relevant critical density is $n_{crit} = 8.7\times 10^{16} \tau_{Ly\alpha}^{-1}\ {\rm cm}^{-3}$ at $T$=10,000 K ([Osterbrock]{} & [Ferland]{} 2006). It is possible to estimate $\tau_{Ly\alpha}$ using our estimate for $\tau_{H\beta}$ and the relationship[^3] $\tau_{Ly\alpha}=\frac{7n_1}{8n_2}\tau_{H\beta}$, where the ratio of the populations of the $n=1$ and $n=2$ shells is $\frac{n_1}{n_2}\simeq\frac{1}{4}\exp(10.2\ {\rm eV}/kT)$. For $T$=10,000 K, $\frac{n_1}{n_2}=3.50\times 10^4$, and so $\tau_{Ly\alpha}\simeq(6.6\pm0.8)\times 10^{5}$ when $\tau_{H\beta}=19.5\pm2.5$. This would imply $n_e\simeq (1.3\pm0.2) \times 10^{11}\ {\rm cm}^{-3}$ if collisional excitation was the only mechanism populating $n=2$. However, at such large $\tau_{Ly\alpha}$ there is another effect which populates the $n=2$ shell: Ly$\alpha$ trapping ([Ferland]{} & [Netzer]{} 1979). As shown in the previous paragraph, an absorber with Balmer optical depths as large as observed here must have a very large Ly$\alpha$ optical depth. In such an absorber, soft X-rays from the quasar will ionize at large $\tau_{Ly\alpha}$ and create Ly$\alpha$ photons at those depths via recombination. Every Ly$\alpha$ photon so created will be re-absorbed approximately $\tau_{Ly\alpha}$ times before it escapes, and so the $n=2$ population for a given density and temperature will be increased by a factor of $\tau_{Ly\alpha}$. This increase changes the value of $\tau_{Ly\alpha}$ inferred for a given $\tau_{H\beta}$, and thus the column density needed to explain a given $\tau_{H\beta}$. When Ly$\alpha$ trapping is occurring, $\tau_{Ly\alpha}^{trap} \simeq [\tau_{H\beta}\frac{7}{32}\exp(10.2\ {\rm eV}/kT)]^{1/2}$. For our observed $\tau_{H\beta}$ and $T=7500$K, appropriate for partially ionized gas illuminated by a quasar (Figure 14.5 of [Osterbrock]{} & [Ferland]{} 2006), $\tau_{Ly\alpha}=5520$ and $\frac{n_1}{n_2}\simeq\frac{1}{4}(\tau_{Ly\alpha}^{trap})^{-1}\exp(10.2\ {\rm eV}/kT)=323$. The H$\beta$ optical depth yields an estimate of the column density of neutral hydrogen in the $n=2$ shell (Eq. 12 of [Hall]{} [et al.]{} 2003), assuming a uniform optical depth over the entire 2000 km s$^{-1}$ trough: $N_{\rm HI}(n=2) \simeq (2.5\pm0.3) \times 10^{16}\ {\rm cm}^{-2}$. Combining the estimates for $N_{\rm HI}(n=2)$ and $\frac{n_1}{n_2}$ yields $N_{\rm HI}\simeq8.1\times10^{18}\ {\rm cm}^{-2}$ in the absorber. As is usual for BAL AGN, $N_{\rm H} = N_{\rm HI} + N_{\rm HII}$ is likely to be much higher. Ly$\alpha$ trapping eliminates the need for an absorber with a density above the critical density needed for significant collsional excitation. In fact, $n_{crit}$ becomes an [*upper*]{} limit on the density in the absorber in order to avoid [*overpopulating*]{} the $n=2$ shell through collisions as well as Ly$\alpha$ trapping. Since that would produce Balmer-line optical depths greater than are observed, the density in an absorber at $T=7500$K must be $n_e < 1.6 \times 10^{13}\ {\rm cm}^{-3} = 8.7\times 10^{16}(\tau_{Ly\alpha}^{trap})^{-1}\ {\rm cm}^{-3}$. For Ly$\alpha$ trapping to occur, the absorber must retain a relatively large neutral hydrogen column despite exposure to the quasar’s ionizing continuum. Quasar absorbers with blueshifted neutral hydrogen columns that large located relatively close to the quasar are rare along our sightlines to bright quasars, either because such columns cover a very small total solid angle or are very short-lived phenomena. However, Ly$\alpha$ trapping is expected to occur in quasar low-ionization broad emission line regions. [SDSS J1259+0931]{} may be a case where the continuum source is partially viewed through such gas. Finally, it should be mentioned that an absorber with a high temperature rather than a high $\tau_{Ly\alpha}^{trap}$ could in principle produce the observed Balmer absorption. For example, $T=85,200$ K yields equal populations in the $n=1$ and $n=2$ shells of . However, this explanation is unlikely because gas at such temperatures is unstable to rapid cooling ([Krolik]{} 1999). Conclusions {#concl} =========== The quasar [SDSS J1259+0931]{} has the strongest Balmer-line BAL troughs discovered to date. The required population of the $n=2$ shell of most likely arises from Ly$\alpha$ trapping in an absorber exposed to the quasar’s ionizing continuum but still containing a high column of ($N_{\rm HI}\simeq8.1\times10^{18}\ {\rm cm}^{-2}$). In addition, the detection of this quasar in the rest-frame optical and ultraviolet means the absorber is neither Compton-thick nor particularly dust-rich. The rarity of such Balmer-line BAL troughs (found in $\lesssim$1% of BAL quasars) indicates they have a very small global covering factor or are very transient. However, their incidence may be underestimated because for high-redshift BAL AGN the Balmer lines are redshifted beyond the typical spectral coverage of a discovery spectrum. Even when that is not the case, weak Balmer absorption can easily be missed (e.g., the absorption in FBQS J2107$-$0620 was discovered only through high-resolution spectroscopic followup). To improve our understanding of quasar winds requires observations of specific objects for which multiple physical parameters in the absorbers can be constrained. Thus, follow-up observations of [SDSS J1259+0931]{} would be quite valuable. X-ray observations can constrain the total absorbing column. Echelle spectroscopy can put a lower limit on the ionized hydrogen column by searching for absorption from metastable , which is populated by recombination from . Such spectroscopy would also reveal any variability or structure in the Balmer-line troughs and enable detailed modelling of the Balmer and excited emission and absorption to constrain the absorber’s density and ionization state and thus its distance from the ionizing source. A near-IR spectroscopic search for Paschen absorption can constrain the population in the $n=3$ shell of . Near-IR spectroscopy of H$\alpha$ can help constrain the contribution of any ‘fill-in’ Balmer emission from the absorber itself. Such emission should be $\sim$4 times stronger in H$\alpha$ than in H$\beta$ given the Balmer and inferred Lyman optical depths of the absorber ([Cox]{} & [Mathews]{} 1969). [SDSS J1259+0931]{} is detected by 2MASS with $J=17.00\pm0.16$ and $K_s=15.32\pm0.15$, placing moderate resolution near-IR spectroscopy within the reach of 8m-class telescopes. I thank N. Murray, D. Hutsemekers, and the referee. P. B. H. is supported by NSERC. Funding for the SDSS and SDSS-II has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, the U.S. Department of Energy, the National Aeronautics and Space Administration, the Japanese Monbukagakusho, the Max Planck Society, and the Higher Education Funding Council for England. The SDSS Web Site is http://www.sdss.org/. The SDSS is managed by the Astrophysical Research Consortium for the Participating Institutions. The Participating Institutions are the American Museum of Natural History, Astrophysical Institute Potsdam, University of Basel, Cambridge University, Case Western Reserve University, University of Chicago, Drexel University, Fermilab, the Institute for Advanced Study, the Japan Participation Group, Johns Hopkins University, the Joint Institute for Nuclear Astrophysics, the Kavli Institute for Particle Astrophysics and Cosmology, the Korean Scientist Group, the Chinese Academy of Sciences (LAMOST), Los Alamos National Laboratory, the Max-Planck-Institute for Astronomy (MPIA), the Max-Planck-Institute for Astrophysics (MPA), New Mexico State University, Ohio State University, University of Pittsburgh, University of Portsmouth, Princeton University, the United States Naval Observatory, and the University of Washington. This research has also made use of the Atomic Line List v2.04 at [http://www.pa.uky.edu/\$\\sim\$peter/atomic/](http://www.pa.uky.edu/$\sim$peter/atomic/). , J., [Ag[ü]{}eros]{}, M., [Allam]{}, S., [Anderson]{}, K., [Anderson]{}, S., [Annis]{}, J., [Bahcall]{}, N., [Bailer-Jones]{}, C., [et al.]{} 2007, , submitted , K., [Iwata]{}, I., [Ohta]{}, K., [Ando]{}, M., [Akiyama]{}, M., & [Tamura]{}, N. 2006, , 651, 84 , R. H., [White]{}, R. L., & [Helfand]{}, D. J. 1995, , 450, 559 , T. A. & [Green]{}, R. F. 1992, , 80, 109 , D., [Leighly]{}, K., [Thomas]{}, R., & [Baron]{}, E. 2002, , 578, L37 , G. & [Stockton]{}, A. 2001, , 555, 719 , D., [Baron]{}, E., [Branch]{}, D., & [Leighly]{}, K. 2004, in AGN Physics with the Sloan Digital Sky Survey, ed. G. T. [Richards]{} & P. B. [Hall]{}, 231 , D. P. & [Mathews]{}, W. G. 1969, , 155, 859 , G. & [Netzer]{}, H. 1979, , 229, 274 , M., [Ichikawa]{}, T., [Gunn]{}, J. E., [Doi]{}, M., [Shimasaku]{}, K., & [Schneider]{}, D. P. 1996, , 111, 1748 , J., [Siegmund]{}, W., [Mannery]{}, E., [Owen]{}, R., [Hull]{}, C., [Leger]{}, R., [Carey]{}, L., [Knapp]{}, G., [et al.]{} 2006, , 131, 2332 , J. E., [Carr]{}, M., [Rockosi]{}, C., [Sekiguchi]{}, M., [Berry]{}, K., [Elms]{}, B., [de Haas]{}, E., [Ivezi[ć]{} ]{}, [Ž]{}., [et al.]{} 1998, , 116, 3040 , P. B., [Anderson]{}, S., [Strauss]{}, M., [York]{}, D., [Richards]{}, G., [Fan]{}, X., [Knapp]{}, G., [Schneider]{}, D., [et al.]{} 2002, , 141, 267 , P. B., [Hutsem[é]{}kers]{}, D., [Anderson]{}, S. F., [Brinkmann]{}, J., [Fan]{}, X., [Schneider]{}, D. P., & [York]{}, D. G. 2003, , 593, 189 , D., [Finkbeiner]{}, D., [Schlegel]{}, D., & [Gunn]{}, J. 2001, , 122, 2129 , J. B., [Crenshaw]{}, D. M., [Kraemer]{}, S. B., [Gabel]{}, J. R., [Kaiser]{}, M. E., [Weistrop]{}, D., & [Gull]{}, T. R. 2002, , 124, 2543 , [Ž]{}., [Menou]{}, K., [Knapp]{}, G., [Strauss]{}, M., [Lupton]{}, R., [Vanden Berk]{}, D., [Richards]{}, G., [Tremonti]{}, C., [et al.]{} 2002, , 124, 2364 , [Ž]{}., [Lupton]{}, R., [Schlegel]{}, D., [Boroski]{}, B., [Adelman-McCarthy]{}, J., [Yanny]{}, B., [Kent]{}, S., [Stoughton]{}, C., [et al.]{} 2004, AN, 325, 583 , J. H. 1999, [Active Galactic Nuclei]{} (Princeton, NJ: Princeton University Press), 331 , D. E. & [Ferland]{}, G. J. 2006, [Astrophysics of Gaseous Nebulae and Active Galactic Nuclei, Second Edition]{} (Sausalito: University Science Books), 365 , J. R., [Munn]{}, J. A., [Hindsley]{}, R. B., [Hennessy]{}, G. S., [Kent]{}, S. M., [Lupton]{}, R. H., & [Ivezi[' c]{}]{}, [Ž]{}. 2003, , 125, 1559 , T., [Richards]{}, G., [Hall]{}, P., [Schneider]{}, D., [Vanden Berk]{}, D., [Fan]{}, X., [York]{}, D., [Knapp]{}, G., [et al.]{} 2003, , 126, 2594 , G., [Fan]{}, X., [Newberg]{}, H., [Strauss]{}, M., [Vanden Berk]{}, D., [Schneider]{}, D., [Yanny]{}, B., [Boucher]{}, A., [et al.]{} 2002, , 123, 2945 , D. S., [Veilleux]{}, S., & [Sanders]{}, D. B. 2005, , 632, 751 , J., [Tucker]{}, D., [Kent]{}, S., [Richmond]{}, M., [Fukugita]{}, M., [Ichikawa]{}, T., [Ichikawa]{}, S., [Jorgensen]{}, A., [et al.]{} 2002, , 123, 2121 , J., [Hall]{}, P., [Reichard]{}, T., [Richards]{}, G., [Schneider]{}, D., [Vanden Berk]{}, D., [Knapp]{}, G., [Anderson]{}, S., [et al.]{} 2006, , 165, 1 , M.-P., [Joly]{}, M., [V[é]{}ron]{}, P., [Boroson]{}, T., [Lipari]{}, S., & [Ogle]{}, P. 2006, , 451, 851 , R., [Morris]{}, S., [Foltz]{}, C., & [Hewett]{}, P. 1991, , 373, 23 , D., [Adelman]{}, J., [Anderson]{}, J., [Anderson]{}, S., [Annis]{}, J., [Bahcall]{}, N., [Bakken]{}, J., [Barkhouser]{}, R., [et al.]{} 2000, , 120, 1579 [lccccccccc]{} Observed Depth& 36% & 29% & 27% & 23% & 18% & 17% & 7% & 58% & $\sim$50%\ Model Depth& 29.0%& 29.0%& 27.6%& 23.7%& 18.9%& 14.6%& & &\ $f_{ij}$& 0.119& 0.0446& 0.0221& 0.0127& 0.00803& 0.00543& & 0.943&\ $\lambda$& 4862.683& 4341.684& 4102.892& 3971.195& 3890.151& 3836.472& 3934.777& $\lambda\lambda$2798.75& $<$2632.106\ [^1]: Note that the standard optical emission template derived from the spectrum of I Zw 1 ([Boroson]{} & [Green]{} 1992) is not a particularly good match to the emission in [SDSS J1259+0931]{}. [^2]: If the total unabsorbed flux has continuum and line components, $F=F_C+F_L$, then the flux removed by very high optical depth absorption ($\tau\rightarrow\infty$) is $A=C_CF_C+C_LF_L$, allowing for different covering factors $C$ of each component. Since the normalization only used the continuum, the depths in Table 1 are $D=A/F_C=C_C+C_LF_L/F_C$. Thus, $D>C_C$ is possible when the line flux $F_L$ is substantial, as is the case for H$\beta$ in this object. [^3]: This relationship comes from the fact that for any transition originating in shell $k$, $n_k\propto N_H(n=k)\propto\tau_{k}\Delta v/\lambda_{k}f_{k}$ (e.g., Eq. 12 of [Hall]{} [et al.]{} 2003). Thus, $n_1/n_2=\tau_1\lambda_2f_2/\tau_2\lambda_1f_1$, assuming an absorber of fixed width $\Delta v$ with all $\tau$ independent of velocity.

ArXiv

--- abstract: 'The anomalous Hall effect in metal-insulator-semiconductor structures having thin (Ga,Mn)As layers as a channel has been studied in a wide range of Mn and hole densities changed by the gate electric field. Strong and unanticipated temperature dependence, including a change of sign, of the anomalous Hall conductance $\sigma_{xy}$ has been found in samples with the highest Curie temperatures. For more disordered channels, the scaling relation between $\sigma_{xy}$ and $\sigma_{xx}$, similar to the one observed previously for thicker samples, is recovered.' author: - 'D. Chiba' - 'A. Werpachowska' - 'M. Endo' - 'Y. Nishitani' - 'F. Matsukura' - 'T. Dietl' - 'H. Ohno' title: 'Anomalous Hall effect in field-effect structures of (Ga,Mn)As' --- Along with anisotropic magnetoresistance, the anomalous Hall effect (AHE) results from an interplay between spin-orbit interactions and spin polarization of electric current specific to ferromagnets. It has been recently realized that for a certain region of conductivities, the anomalous Hall conductivity $\sigma_{xy}$ is a measure of the Berry phase of carrier trajectories in the $k$ space and, thus, provides information on the hitherto inaccessible aspects of the band structure topology in the presence of various spin-orbit interactions [@Sundaram:1999_a; @Jungwirth:2002_a; @Jungwirth:2006_a; @Nagaosa:2009_a; @Werpachowska:2009_a]. Interestingly, the effect appears to be qualitatively immune to disorder, except for the case of linear-in-$k$ Rashba-type Hamiltonians in two-dimensional electron systems, where the contribution to $\sigma_{xy}$ vanishes [@Rashba:2004_d] unless the lifetime is spin-dependent [@Inoue:2006_a]. Furthermore, a surprisingly universal empirical scaling relation between the Hall and longitudinal conductivities, $\sigma_{xy} \sim \sigma_{xx}^{\gamma}$, $\gamma \approx 1.6$ has been found to be obeyed by a number of materials on the lower side of their conductivity values [@Fukumura:2007_a], where Anderson-Mott quantum localization effects should be important. In this Letter, we report on Hall resistance studies as a function of temperature and gate electric field carried out for metal-insulator-semiconductor (MIS) structures containing a thin conducting channel of ferromagnetic (Ga,Mn)As. We find out that in the $\sigma_{xx}$ range up to $~10^2$ S/cm, $\sigma_{xy}$ obeys a scaling relation with a similar value of the exponent $\gamma$. However, for $\sigma_{xx} \gtrsim 10^2$ S/cm the scaling relation breaks down entirely. Surprisingly, in this regime and below the Curie temperature $T_{\mathrm{C}}$, $\sigma_{xy}$ tends to decrease rather abruptly with decreasing temperature, and even reverses its sign in some cases, in the region where neither resistance $R$ nor magnetization $M$ vary significantly with temperature. The effect has not been observed in thicker films and appears to have no explanation within the existing theory, pointing to the importance of yet unrevealed confinement effects. The studied thin layers of tensile-strained (Ga,Mn)As have been deposited by low-temperature molecular beam epitaxy onto a buffer layer consisting of 4-nm GaAs/ 30-nm Al$_{0.75}$Ga$_{0.25}$As/ 500-nm In$_{0.15}$Ga$_{0.85}$As/30-nm GaAs grown on a semi-insulating GaAs (001) substrate. Upon growth, Hall bars having a channel of 30 or 40-$\mu$m width and $\sim$200-$\mu$m length are patterned by photolithography and wet etching. Subsequently, samples are annealed at 180$^{\circ}$C for 5 min or introduced directly into an atomic layer deposition chamber, where a 40-nm-thick oxide insulator is deposited at a substrate temperature of 120-150$^{\circ}$C. Finally, 5-nm Cr/ 100-nm Au gate electrode is formed. Owing to the tensile strain, the easy axis is perpendicular to the plane, so that the height of the Hall voltage hystereses provides directly the magnitude of the anomalous Hall resistance $R_{yx}$. Altogether 18 MIS structures, numbered from 1 to 18, have been investigated. As tabulated in the supplementary material[@EPAPS], they differ in nominal Mn composition ($3\% \leq x \leq 17.5\%$), thickness (3.5 nm $\leq t \leq$ 5 nm), and crystallographic orientation (\[110\] vs. \[$\bar{1}$10\]) of the Ga$_{1-x}$Mn$_{x}$As channel, as well as contain three kinds of gate insulators (Al$_2$O$_3$, HfO$_2$, and ZrO$_2$). For this set of samples, and in the employed gate electric field range (5 MV/cm $\geq E_{\mathrm{G}} \geq -5$ MV/cm), $\sigma_{xx}$ spans over 3 orders of magnitude, and the corresponding $T_{\mathrm{C}}$ varies between 35 and 165 K. The lateral homogeneity of the grown wafers is proved by results displayed in Fig. S1 in EPAPS, whereas the data in Fig. S2 demonstrate that the magnitudes of $T_{\mathrm{C}}$ in the MIS structures and thick layers are virtually identical. Since $t$ is larger than the mean free path but shorter than the phase coherence length of holes in (Ga,Mn)As, whose lower limit is provided by the conductance studies in 1D systems [@Wagner:2006_a; @Vila:2007_a], the density of states preserves a 3D shape, whereas localization phenomena acquire a 2D character. As an example of experimental findings, Figs. 1(a) and 1(b) show the temperature dependence of remanent Hall resistance $R_{yx}$ and sheet resistance $R_{xx}$ at various gate electric fields for the MIS structure containing a 5-nm thick Ga$_{0.949}$Mn$_{0.051}$As channel and the Al$_2$O$_3$ gate insulator. In the inset, the results of the Hall measurements as a function of the external magnetic field $H$ at 10 K are presented. The squareness of the hysteresis loop indicates the perpendicular-to-plane orientation of the magnetization easy axis, whereas its counter-clockwise chirality, as in the case of an ordinary $M$ vs. $H$ loop, demonstrates that the anomalous Hall coefficient is positive in this case. The vertical axis of Fig. 1(a) presents $R_{yx}(T)$, as determined by hysteresis heights at zero magnetic field. As seen, $T_{\mathrm{C}}$ of the channel layer increases (decreases) by the application of negative (positive) gate electric field which accumulates (depletes) holes in the channel, as witnessed in Fig. 1(b) by a corresponding decrease (increase) of $R_{xx}$. Importantly, according to Fig. 1(a) also the low-temperature values of $R_{yx}$ increase with $R_{xx}$ changed by the gate electric field. Figure 1(c) shows $\sigma_{xy} = R_{yx}/[t(R_{xx}^2 + R_{yx}^2)]$ as a function of $\sigma_{xx} = R_{xx}/[t(R_{xx}^2 + R_{yx}^2)]$ under three different values of $E_{\mathrm{G}}$ at $T = 10$–$30$ K $\lesssim T_{\mathrm{C}}/2$, where the spontaneous spin splitting should be only weakly temperature dependent. The dotted line is a guide for the eyes, which indicates that the values of $\sigma_{xy}$ fall on a single curve, verifying an empirical scaling behavior with $\gamma \approx 1.6$ for this structure. In 12 samples of all samples, $\gamma$ is found to be within the range of $1.6\pm 0.4$, showing virtually no temperature dependence below $T_{\mathrm{C}}/2$ [@EPAPS]. ![\[Color online\] (a) Hall and (b) longitudinal sheet resistances of 5-nm thick Ga$_{0.949}$Mn$_{0.051}$As (sample 3) at various gate electric fields. Inset to (a) shows hysteresis loops of the Hall resistance. (c) Relation between Hall and longitudinal conductivities obeying a simple scaling law shown by dotted line.[]{data-label="fig:1"}](AHE_fig1.eps){width="3.3in"} Experimental results illustrating an entirely different and so-far not reported behavior of $R_{yx}(T)$ are presented in Fig. 2(a) for the MIS structure containing a 4-nm thick Ga$_{0.875}$Mn$_{0.125}$As channel and the HfO$_2$ gate insulator. While, according to Fig. 2(b) the temperature dependence of $R_{xx}$ is rather standard, $R_{yx}$ shows a nonmonotonic temperature dependence attaining a maximum at about 20 K below $T_{\mathrm{C}} \approx 120$ K, followed by a decrease of $R_{yx}$ towards zero with lowering temperature, culminating in a change of sign at $E_{\mathrm{G}} > 1.5$ MV/cm at 10 K, as shown in detail in Fig. 2(c). The data on $R_{yx}(T)$ are to be contrasted with the behavior of magnetization which, according to the results displayed in Fig. 2(d), shows a monotonic increase upon cooling. Interestingly, while the magnitude of $T_{\mathrm{C}}$ in our MIS structures attains similar values as those in thicker films, a non-Brillouin character of $M(T)$ indicates that magnetization fluctuations are rather strongly enhanced by the confinement. ![\[Color online\] (a) Hall and (b) longitudinal sheet resistances of 4-nm thick Ga$_{0.875}$Mn$_{0.125}$As (sample 16) at various gate electric fields. (c) Hysteresis loops at various gate electric fields documenting the sign change of the anomalous Hall effect. (d) Temperature dependence of remanent magnetization without gate bias. []{data-label="fig:2"}](AHE_fig2.eps){width="3.0in"} The findings for this and other samples showing similar properties, collected in Fig. 3, demonstrate clearly that the relation between the Hall resistance and magnetization or carrier polarization can be rather complex. In particular, despite that $R_{xx}$ is virtually temperature independent and the hole liquid degenerate, $R_{yx}$ and, thus, $\sigma_{xy}$ decrease abruptly or even change sign on lowering temperature, as shown in Figs. 2 and 3. In this range, the temperature derivatives of $|R_{yx}(T)|$ and $M(T)$ acquire opposite signs, indicating that Hall measurements for temperature dependent magnetometry should be applied with care. Furthermore, the revealed change of sign of $R_{yx}$ calls into question the generality of a simple scaling formula. ![\[Color online\] Hall $R_{yx}$ and longitudinal $R_{xx}$ sheet resistances at various values of the gate electric fields in MIS structures \[(a) and (b) for sample 11, (c) and (d) for sample 12, (e) and (f) for sample 14, and (g) and (h) for sample 15\] showing a strong temperature dependence of $R_{yx}$, similar to the one presented in Fig. 2.[]{data-label="fig:3"}](AHE_fig3.eps){width="3.3in"} According to the recent theory of the intrinsic AHE in (Ga,Mn)As [@Werpachowska:2009_a], an appropriately strong tensile strain or the bulk inversion asymmetry— the Dresselhaus effect[@Dresselhaus]—can result in a negative sign of the AHE for a sufficiently small magnitude of scattering broadening $\Gamma$. In fact, a negative sign of the $R_{yx}$ has been observed in films of (In,Mn)As [@Munekata:1997_a; @Matsukura:2002_c], (In,Mn)Sb [@Yanagi:2004_a; @Mihaly:2008_a], and (Ga,Mn)Sb [@Eginligil:2008_a], where the bulk asymmetry is expected to be much stronger. However, the striking behavior of $R_{yx}(T)$, including the change of sign revealed here, has not been anticipated theoretically \[4,5\]. It appears to be related to the reduced dimensionality, as it has not been observed in thicker (Ga,Mn)As films grown by us with similar conductivity $T_{\mathrm{C}}$ and strain [@Chiba:2006_a; @Pu:2008_a]. In an attempt to interpret our findings we note that thin layers are expected to exhibit structure inversion asymmetry constituting an additional source of spin-orbit coupling. Despite that the gate electric field does not affect significantly the temperature dependence of $R_{yx}$, as seen in Fig. 3, the presence of two different interfaces can account for the structure asymmetry and the corresponding lowering of the point symmetry from $D_{2d}$ to $C_{2v}$. In order to explain the data within this scenario, the structure asymmetry contribution to $R_{yx}$ should be negative and its amplitude take over the bulk terms at sufficiently high magnetization values. ![Hall conductivity vs. conductivity for 4-nm thick Ga$_{0.875}$Mn$_{0.125}$As channels oriented along \[110\] (sample 16) and \[$\bar{1}$10\] (sample 17) in-plane crystallographic orientations.[]{data-label="fig:4"}](AHE_fig4.eps){width="3.3in"} Since the in-plane anisotropy of the conductivity tensor would provide a direct proof that the symmetry is lowered to $C_{2v}$, we have examined MIS structures originating from the same wafer but having the channels oriented along either \[110\] or \[$\bar{1}$10\] crystal axis. As shown in Fig. 4, similar values of Hall and longitudinal conductance are observed in both cases. Though the negative result of this experiment does not provide a support for the structural asymmetry model, it does not disprove it, as an accidental degeneracy cannot be excluded ($e.g.$, due to a compensation of asymmetry in the values of the density of states and relaxation time). Another possibility is that dimensional quantization of the transverse motion leads to a significant reconstruction of the topology of the Fermi surface introducing, in particular, a number of additional subband crossings. This scenario may explain why the negative sign of $\sigma_{xy}$ appears on depleting the channel in one structure \[Fig. 2(a)\], whereas it shows up in the accumulation regime in another sample \[Fig. 3(a)\]. A detailed Fermi sheet behavior in this region ($e.g.$, crossing vs. anticrossing) may be sensitive to symmetry lowering and phase breaking mechanisms such as spin-orbit as well as temperature dependent inelastic and spin-disorder scattering. At the same time, the confinement-induced upward shift of the hole energies may enhance the importance of the Dresselhaus contribution determined by the admixture of the conduction band wave functions. Within this scenario, the behavior of $\sigma_{xy}$ as a function of temperature reflects a subtle and spin polarization-dependent balance between positive and negative terms originating from the intra-atom and bulk inversion asymmetry electric fields, respectively. In Fig. 5 we summarize our findings by reporting $\sigma_{xy}$ as a function of $\sigma_{xx}$ at 10 K for various gate electric fields. The results [@Matsukura:1998_a; @Chiba:2006_a; @Pu:2008_a] obtained previously for thick films are also shown for comparison. Despite the fact that the particular samples may differ in magnetization values—which may likely affect the form of the scaling [@Chun:2007_a]—we see that over a wide range of conductivities up to $10^2$ S/cm the relation $\sigma_{xy} \sim \sigma_{xx}^{\gamma}$ is obeyed and implies $\gamma\sim 1.6$, in agreement with the value $\gamma = 1.5$ found for thick films of (Ga,Mn)As in the magnetic field [@Shen:2008_a]. However, the scaling relation breaks down rather severely in the case of MIS structures, as discussed above. We also note that $\gamma$ determined at each temperature has almost no temperature dependence below $T_\mathrm{C}/2$, according to the results shown in Fig. S2 of the supplementaly material. However, on approaching $T_\mathrm{C}$ we have found the scaling to be not obeyed (Fig. S4). ![\[Color online\] Relation between Hall and longitudinal conductivities at various values of gate electric fields $ E_{\mathrm{G}}$ at 10 K (30 K for Sample 1) for the studied MIS structures of Ga$_{1-x}$Mn$_{x}$As. The legend shows the values of the Mn content $x$, channel thickness $t$, as well as the sample number. The asterisks (\*) mark the samples which were annealed before deposition of insulator layer. Half-filled symbols show the absolute value of negative $\sigma_{xy}$. The previous data for thick (Ga,Mn)As films are shown by open symbols (circles [@Matsukura:1998_a], triangles up [@Chiba:2006_a], and diamonds [@Pu:2008_a]). The dotted line shows the dependence $\sigma_{xy} \sim \sigma_{xx}^{\gamma}$, $\gamma = 1.6$.[]{data-label="fig:5"}](AHE_Fig5f.eps){width="3.3in"} We note that the empirical scaling found for thick films as well as for thin layers with low conductance cannot be explained by the intrinsic mechanisms of the AHE, as the corresponding theory predicts a decrease of $\sigma_{xy}$ with both hole density [@Jungwirth:2002_a; @Werpachowska:2009_a; @Dietl:2003_c] and scattering time [@Werpachowska:2009_a; @Jungwirth:2003_b] in a wide range of relevant hole concentrations and spin-splittings. This suggests that the physics of AHE is dominated by the proximity to the Anderson-Mott localization boundary [@Belitz:1994_a; @Dietl:2008_c]. So far the influence of quantum interference effects on the anomalous Hall conductance have been studied considering the side-jump and skew-scattering terms within the single-particle theory [@Dugaev:2001_a]. The data summarized in Fig. 5 call for the extension of the theory towards the intrinsic AHE with the effects of disorder on interference of carrier-carrier scattering amplitudes taken into account. In summary, we have found that the anomalous Hall effect of the magnetic semiconductor (Ga,Mn)As acquires qualitatively new and not anticipated features when dimensionality and disorder are reduced. The revealed striking temperature dependence of the Hall conductance indicates a significant contribution of confinement phenomena. At higher levels of disorder, the scaling relation between $\sigma_{xy}$ and $\sigma_{xx}$, similar to that observed previously for thicker samples and so-far unexplained, is recovered. The authors thank S. Murakami and N. Nagaosa for useful discussions, and L. Ye for experimental support. The work in Sendai was supported in part by Grant-in-Aids from MEXT/JSPS, the GCOE Program at Tohoku University, and the Research and Development for Next-Generation Information Technology Program from MEXT, whereas the work in Warsaw was supported by the President of the Polish Academy of Sciences for doctoral students (A. W.) and by FunDMS Advanced Grant of European Research Council within “Ideas” 7th Framework Programme of EC (T. D.). [26]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , ****, (). , , , ****, (). , , , , , ****, (). N. Nagaosa, J. Sinova, S. Onoda, A. H. MacDonald, and N. P. Ong, arXiv:0904.4154. , (). , ****, (). , , , , , , ****, (). , , , , , , ****, (). See EPAPS Document No. E-PRLTAO-xxx for a table and figures summarizing sample properties and uniformity. For more information on EPAPS, see http://www.aip.org/pubservs/epaps.html. , , , , , ****, (). L. Vila, R. Giraud, L. Thevenard, A. Lemaitre, F. Pierre, J. Dufouleur, D. Mailly, B. Barbara, and G. Faini, Phys. Rev. Lett. [**98**]{}, 027204 (2007). , , , , , , , , , ****, (). , , , , , , , , ****, (). , ****, (). , , , , ****, (). , , , , , , , , ****, (). , , , , , , ****, (). , , , , , ****, (). , , , , , ****, (). , , , , ****, (). , , , , , , , , , , , ****, (). , , , , , , ****, (). , , , , , in **, edited by (, ), p. , . , , , , , , , , , ****, (). , ****, (). , ****, (). , , , ****, ().

ArXiv

--- abstract: 'We evaluate validity of NLFFF extrapolation performed with Optimization class (OPTI) codes. While explaining inevitable for OPTI partial non-solenoidality caused by the gas pressure notable role in pressure balance at photospheric heights and by mathematical aspects related to optimization and BVP (boundary value problem), we justify elimination of the non-solenoidal component (postprocessing) from the OPTI result obtained. In essence, postprocessing converts the entire non-solenoidal part into a solenoidal force part, which possibly reflects factual deviation of magnetic field from its force-free pproximation on the photosphere and in the solar corona. Two forms of postprocessing have been analyzed in this paper. Postprocessing I eliminates the non-solenoidal component without changing transverse field at the measurement level, and Postprocessing II leaves the field normal component unchanged. Extrapolation, postprocessing, and then comparison of metric and energy characteristics are performed over AR 11158 active region for a small fragment of its evolution containing the February X-class flare. Our version of OPTI code showed that free energy decreased by $\sim 10^{32}$ erg within 1 hour, which corresponds to theoretical estimations of the flare-caused magnetic energy loss. This result differs significantly from the one in [@Sun]. Therefore, we also comment on some features of our OPTI code implementation, which may cause significant differences between our results and those obtained using the [@Wieg4] version of OPTI code in study by [@Sun].' address: 'Institute of Solar-Terrestrial Physics SB RAS, Lermontov St. 126, Irkutsk 664033, Russia' author: -   -   title: Validity of NLFFF Optimization Reconstruction --- Introduction ============ As it is commonly believed, extreme solar activity is provided by magnetic energy of active regions [@Forbes]. According to Thomson’s (or Dirichlet’s) theorem, e.g., [@Lawrence], any solenoidal field can be decomposed into the current and potential parts. The energy of the latter corresponds to the minimum possible under equal conditions for the normal component at the region boundary. For this reason, only energy of the current component (free energy) is responsible for the magnetic region ability to release energy, and can be used for predicting solar flares [@Barnes_2016]. Obtaining magnetic structures and their energy characteristics from measurement data accessible at photospheric heights is only possible if approximation of the force-free field is used. This approximation is applicable for the upper corona due to low $\beta$ – the ratio of gas-kinetic pressure (exponentially falling with height) to magnetic pressure (falling slower). At photospheric heights, parameter $\beta$ can be conventionally considered low only in regions with rather high field module $|{\bf B}|$. In regions with lower $|{\bf B}|$, parameter $\beta$ can take values of order of $1$ and greater. Because it is a strong field that generates the main structure of the coronal magnetic field, we can notionally use the force-free approximation. The key point for our work will exactly be the fact that getting a field model of a real active region from the force-free approximation, we will inevitably face disagreement with the real magnetic field containing a certain force component. In particular, free energy of the force-free approximation will also differ from free energy of a real solenoidal field with the force component. In practical extrapolation, many methods are developed from the force-free approximation and applied (see, e.g. [@Schrijver_2006; @Metcalf]). In this paper we examine the code based on the optimization method (OPTI), [@Wheatland], one of the most used in magnetography applications of active regions. The expertise is of fundamental importance in the context of significant non-solenoidality of numerical OPTI results discussed in literature [@Valory_2013; @Mastrano_2018]. Notable non-solenoidality of OPTI results is quite natural and is associated with two factors. First, the OPTI algorithm implements the search for solution in the functional space of arbitrary non-solenoidal fields, on equal terms minimizing both Lorenz’s power and divergence. If we solve a problem with boundary values (BVs) providing a force-free result, the OPTI algorithm (as known from the [@Low] analytic force-free models) provides a smallness of the non-solenoidal component at the level of numerical errors. In case there is no force-free solution satisfying the BVP, it is natural to expect that the algorithm will lead to occurrence of the finite Lorenz force and divergence in the solution. Secondly, the BVP specificity for OPTI (a full set of the field boundary components $B_x,B_y,B_z$ is specified) concedes the absence of any force-free solution. In fact, we can cite the following example. If it is a force field, then quite natural are cases when the normal field on the solar surface is zero with the nonzero component of the current normal component. In this case, if we try to define boundary conditions for a force-free field using these data, we come to a non-realistic infinite value of the force-free parameter $\alpha$. That is, for this case, it is notionally impossible to find a suitable force-free field, and we have the right to extend this statement to any real photospheric field. Thus, in OPTI solutions, we will always have the force and non-solenoidal components consistent with the results by [@Valory_2013; @Mastrano_2018]. Desirable estimates of non-solenoidality for other extrapolation techniques considered in [@Mastrano_2018], apparently, can also be explained in terms of their features and specifics of their BVP. For example, methods that belong to Grad—Rubin CFIT, XTRAPOL and FEMQ class from [@Mastrano_2018] are realized (exactly or almost exactly) in the class of solenoidal fields and this is first. Second, their BVP is defined by the pair $B_n$ and $\alpha(B_n>0)$. These two moments naturally lead to force-free solutions, as shown in [@Mastrano_2018]. So, we have two basic differences in two approaches to magnetic field extrapolation. When trying to save accurate BVs of all field components, the OPTI technique leads to an approximate result with the force and non-solenoidal components. In the second approach, we get some force-free and solenoidal result, while giving away accurate reproduction of the boundary field that we obtained in measurements. Since the field on the photosphere is different from the force-free field, the force-free parameter $\alpha $, strictly speaking, does not make sense in its exact meaning, i.e. we to some extent select boundary values of the field. In this situation, no one can say results of which of the two approaches are more realistic. It is likely that in an effort to obtain an accurate force-free field, we can move further away from the real field, which has a noticeable force component. In order to clarify the OPTI problem highlighted in studies by [@Valory_2013; @Mastrano_2018], we implement one of the two types of postprocessing. Both types of postprocessing completely remove non-solenoidality from our solution, and convert it into an additional force component. Postprocessing I leaves the transverse field unchanged at the bottom boundary of the extrapolation box using the divergence cleaner algorithm by [@Valory_2013], Formula (B.4). On the contrary, Postprocessing II leaves the normal field unchanged by adding a compensating potential component to the field of Postprocessing I. After any postprocessing, final solution includes a certain non-zero force component. This force component, if not taken as a mistake, is likely to indicate the above-mentioned deviation of the real solar magnetic field from the force-free field. Our research is conducted on a well-studied part of AR 11158 evolution that contains the X-class flare on February 15, 2011 01:44 UT. Extrapolation is carried out with our own version of OPTI technical implementation. We compare the AR 11158 energy characteristics within the selected time interval and the overall structure of field lines with the results of [@Sun]. It is important that our results before and after any postprocessing show approximately the same $\sim 10^{32}$ erg drop in free energy during one hour after the flare onset. This matches the estimates of the required reserve energy for X-class events [@Hudson; @Bleybel]. Such a great discrepancy with the result by [@Sun] seems to be caused by the fundamental difference in some aspects of our OPTI implementation that we point out here. For results before and after one of the two postprocessings, we perform a detailed analysis of the extrapolation metric parameters. Determining extrapolation box and constructing boundary conditions ================================================================== For the entire series of calculations we use the same rectangular 3D box $[214.56\times 213.12\times 183.6]$Mm. The center of its bottom face corresponds to approximate center of AR 11158 at $-20.85^o$ latitude and $34.9^o$ longitude. The $X$ axis of the bottom face is directed towards the North Pole. Coordinates $x$, $y$ correspond to the Lambert equal area projection [@Calabretta; @Thompson]. The initial $z$ -coordinate corresponds to the solar radius. For calculation, grid $(299\times297\times256)$is used, with a relevant pixel size of $720$ km. For extrapolation, we use SDO/HMI (hmi.B\_720s) vector magnetograms without azimuthal ambiguity. We remap data of the field of view magnetograms ($B_x,B_y,B_z$) to the box coordinate system on the bottom boundary of the box. Thus, we will consider BVs for the extrapolation problem at the bottom boundary as given. To specify BVs at the other boundaries of the box, we use a potential solenoidal field, whose normal component coincides at the bottom boundary with the magnetic field normal component defined above. We calculate the potential field with the Fourier decomposition algorithm similar to [@Alissandrakis] using FFT. At the same time, in order avoid the zero harmonic problem in case the initial data give non-zero total flow of the normal field, we specify the normal component on the extended region that contains 4 initial quadrants of the bottom boundary of our box. In the first lower quadrant we set boundary values of our box. In the diagonal one – the same boundary values. In the rest quadrants there are boundary values of opposite sign. Next, we calculate the potential field in 3D box only above the first quadrant. Thus, we ensure that the calculated normal component coincides with the one specified at the bottom boundary. The potential field calculated in this way will be referred to as the field of $BVP_{potI}$ Using the $BVP_{potI}$ field, we define boundary values for ($B_x,B_y,B_z$) at the other boundaries of the box. We use the same potential field to define the starting field in 3D box for OPTI extrapolation exactly according to the algorithm described in [@Wheatland] without using the weight function. To calculate the OPTI field free energy (before postprocessing), further we also use this potential field. We calculate energies of OPTI and $BVP_{potI}$ fields in the root box $[169.26\times167.766\times159.12]$ Mm with grid $(235\times233\times221)$. To arrange Postprocessing II and to calculate free energies of Postprocessing I and Postprocessing II fields, we need a potential field corresponding to the given normal field on all faces of the box ($BVP_{potII}$). To obtain this potential field, we use the code described in [@box]. Further, for convenience, we will use the following abbreviations: - $B_{all}$ - full boundary of the box. - $B_b$ -bottom boundary of the box. - $B_{st}$ - boundary including side and upper boundaries of the box. - $BV_{all}$ - boundary values $B_x,B_y,B_z$ at $B_{all}$. - $BV_b$ - boundary values $B_x,B_y,B_z$ at $B_b$. - $BV_{st}$ - boundary values $B_x,B_y,B_z$ at $B_{st}$. - $BVn_{all}$ - boundary values $B_n$ at $B_{all}$. - $BVn_b$ - boundary values $B_n$ at $B_b$. - $BVn_{st}$ - boundary values $B_n$ at $B_{st}$. - $BVt_{all}$ - boundary values of field components ($B_t$) at $B_{all}$. - $BVt_b$ - boundary values $B_t$ at $B_b$. - $BVt_{st}$ - boundary values $B_t$ at $B_{st}$. Therefore, the OPTI extrapolation problem is formulated as NLFF extrapolation with $BVP_{OPTI}$: when $BV_b$ from vector magnetograms data and\ $B_{st}$=$B_{st}(BVP_{potI})$. Features of OPTI extrapolation code =================================== The code we use for OPTI extrapolation corresponds exactly to the [@Wheatland] algorithm without using the weight function with the fixed shift of time parameter of the optimization process $dt=0.05\Delta^2$, where $\Delta$ is the pixel size of grid.The number of optimization process steps is always constant and amounts $400000$. It is with this number of steps that the required quality of extrapolation results is achieved. The algorithm is built on parallelized calculations. Calculation of one problem was performed using a supercomputer with 36 processors. One extrapolation for a box with the characteristics described above is performed within $\sim20$ h. Since our extrapolation (R1) results can be compared with those (R2) by [@Sun]), due to their significant differences, we will list common and different elements of R1 and R2: - for R1 and R2, extrapolation boxes and grids are almost the same. - in R1 and R2, the potential field is calculated in different manner (their energies are close to each other). - preprocessing is not used in R1, but it is used in R2 (not to change measurement data is our principled position). - in R1, weight function is not used in the [@Wheatland] minimizing functional, in R2 – it is used. - R1 uses a fixed $dt$ step and a fixed number of steps, R2 uses automatic adjustment of $dt$ step and the end of the optimization process based on criterion of convergence of the functional $L$ metric from [@Wheatland] . In the second case, apparently, the optimization ends too soon, bumping into the local minimum of $L$ in the functional space. In the first case, the optimization process occurs with regular reduction of $L$, without the requirement of reducing it at each step. Most likely, this last point leads to a significant difference in OPTI extrapolation. Postprocessing ============== Suppose we have a non-solenoidal OPTI magnetic field $\bf{B}$ complying with\ $BVP_{OPTI}$. Since the field flow is zeroth, the following equations are valid at $B_{all}$: $$\label{eq1} \int_S{\bf B}\cdot d{\bf s}=\int_V\nabla \cdot {\bf B}dv=0.$$ For this $BVP_{OPTI}$ field, we will consider two ways of cleaning the divergence. Postprocessing I ---------------- According to Formula (B.4) from [@Valory_2013], from $\bf{B}$ we can distinguish the following solenoidal ${\bf B}_{sI}$ component: $$\label{eq2} {\bf B}_{sI}={\bf B}+\widehat{{\bf z}}\times\int_z^{z_2}(\nabla \cdot {\bf B})dz',$$ where $z_2$ is coordinate $z$ of the box upper boundary. This equation (\[eq2\]) does not change components $B_x,B_y$ throughout the box. If field $\bf{B}$ were strictly solenoidal, the error in $B_z$ would have the order of magnitude $B_z$ at the top boundary of the box, i.e. it would be negligible (in numerical implementation of Equation (\[eq2\])). Non-solenoidality of $\bf{B}$ leads to a significant change in $B_z(z=z_0)$, therefore, the reference potential field ($BVP_{potII}$) for ${\bf B}_{sI}$ should change significantly. Postprocessing II ----------------- Postprocessing II is to find the solenoidal field of $\bf{B}_{sII}$ in the form $$\label{eq3} {\bf B}_{sII}={\bf B}_{sI}+\nabla \psi',$$ where the solenoidal potential field $\nabla \psi'$ provides equality $$\label{eq3} ({\bf B}_{sII})_z=B_z \qquad for \quad z=z_0,$$ i.e. $\nabla \psi'$ provides equal $BV_b$ of fields $\bf{B}$ and ${\bf B}_{sII}$ compensating for $BV_b$ changes in the ${\bf B}_{sI}$ field. We find field $\nabla \psi'$ by solving the problem of $BVP_{potI}$ numerically. Thus, in Postprocessing II, situation is opposite – at the bottom boundary of the box, the transverse field changes significantly while the box normal component remains unchanged. Changes in $BV_{st}$ are usually negligible for a compact magnetic region far enough from the boundaries. In addition, we don’t in urgent need of preserving $BV_{st}$ that have rather artificial nature of origin. In practice, energies of potential reference fields of $\bf{B}$ and ${\bf B}_{sII}$ are very close to each other. We should note a very important point of the presence of the common feature for procedures of Postprocessing I and II, to which we will return again a little later – both procedures accurately retain the observed $z$ component of current $\bf{J}$ on $V_b$. Analysis of OPTI results and their modifications ================================================ Configuration of field lines ---------------------------- For the same time point, for which the overall mapping of the field lines configuration for AR 11158 was given in [@Sun], we give a similar representation of field lines in Figures \[fig1\]-\[fig3\].We see that our line images are well correlated with those in [@Sun] and result in approximately the same compliance with the real structure of field lines highlighted on the AIA image. It is quite difficult to prefer one of our results or the result by [@Sun]. Note that without postprocessing, OPTI field results in many broken non-real field lines (mostly coming from the weak field regions). However, we see that the basic structure is shown quite well. Procedures of Postprocessing I and II yield “smoother” fields. The expected lines are easier to find. From value judgment, Postprocessing II results in a slightly “smoother” field in relation to field of Postprocessing I. ![ A case without postprocessing. Left: Selected field lines from the NLFFF extrapolation plotted over a frame of Brmap. The lines are color-coded by vertical current density at their footpoints (see the color bar). Dashed line outlines the bottom of the core domain of NLFFF extrapolation. Right: Same field lines over the AIA image frame.[]{data-label="fig1"}](Lines_Br.png "fig:"){width="48.00000%"} ![ A case without postprocessing. Left: Selected field lines from the NLFFF extrapolation plotted over a frame of Brmap. The lines are color-coded by vertical current density at their footpoints (see the color bar). Dashed line outlines the bottom of the core domain of NLFFF extrapolation. Right: Same field lines over the AIA image frame.[]{data-label="fig1"}](Lines_AIA.png "fig:"){width="48.00000%"} ![Same as in Figure \[fig1\] for the Postprocessing I case.[]{data-label="fig2"}](Lines_Br_cld_nochange.png "fig:"){width="48.00000%"} ![Same as in Figure \[fig1\] for the Postprocessing I case.[]{data-label="fig2"}](Lines_AIA_cld_nochange.png "fig:"){width="48.00000%"} ![Same as in Figure \[fig1\] for the Postprocessing II case.[]{data-label="fig3"}](Lines_Br_cld.png "fig:"){width="48.00000%"} ![Same as in Figure \[fig1\] for the Postprocessing II case.[]{data-label="fig3"}](Lines_AIA_cld.png "fig:"){width="48.00000%"} Metric analysis --------------- We analyze the metrics of calculation results for the same point in time as in the previous subsection 20011-02-14 20:36 UT. We use the following set of metrics for analysis. $$\label{eq5} E=\frac{1}{8\pi}\int_V{\bf B}^2dv,$$ where $E$ is the energy field. $$\label{eq6} E_{pot}=\frac{1}{8\pi}\int_V{\bf B}_{pot}^2dv,$$ where $E_{pot}$ is the energy reference potential field calculated as $BVP_{potII}$ for relevant ${\bf B}$ $$\label{eq7} E_{free}=E-E_{pot},$$ where $E_{free}$ is the value corresponding to determination of free energy for solenoidal field [**B**]{}. $$\label{eq8} E_{free}^*=\frac{1}{8\pi}\int_V\left({\bf B}-{\bf B}_{pot}\right)^2dv,$$ where $E_{free}^*$ matches $E_{free}$ from Equation (\[eq7\]) for solenoidal field [**B**]{}. $$\label{eq9} \varepsilon_1= \left\vert \frac{E_{free}-E_{free}^*}{E}\right\vert,$$ where $\varepsilon_1$ describes the relative contribution of non-solenoidality in [**B**]{}. $$\label{eq10} \varepsilon_2= \left\vert \frac{E_{free}-E_{free}^*}{E_{free}}\right\vert.$$ Equation (\[eq10\]) exactly corresponds to definition of $\varepsilon$ in Equation 15 from [@Mastrano_2018]. This value evaluates the error of the free energy related to field [**B**]{} non-solenoidality. We also use $$\label{eq11} \theta _{j}=\arcsin \left( \frac{\sum_i^N \left\vert {\bf J}\right\vert _{i}\sigma _{i}}{\sum_i^N \left\vert {\bf J}\right\vert _{i}}\right) ,\quad \sigma _{i}=\frac{\left\vert {\bf J}\times {\bf B}\right\vert _{i}}{\left\vert {\bf J}\right\vert _{i}\left\vert {\bf B}\right\vert_{i} },$$ $$\label{eq12} f=\frac{1}{N}\sum_i^N\frac{\left\vert \nabla \cdot{\bf B}\right\vert_i}{6\left\vert {\bf{B}}\right\vert_{i}}\Delta.$$ The metrics of Equations (\[eq11\]) and (\[eq12\]) are similar to those introduced by [@Wheatland]. The former reflects deviation of [**B**]{} from the force-free approximation, the latter – solenoidality. For the time investigated, metrics of Equations (\[eq5\])-(\[eq12\]) are presented in Tables (\[table1\]) and (\[table2\]). $E$ $(\times10^{32}erg)$ $E_{pot}$ $(\times10^{32}erg)$ $E_{free}$ $(\times10^{32}erg)$ $E_{free}^*$ $(\times10^{32}erg)$ -------------------- -------------------------- -------------------------------- --------------------------------- ----------------------------------- OPTI $10.45$ $8.46$ $1.99$ $3.54$ +Postprocessing I $9.53$ $6.56$ $2.96$ $3.03$ +Postprocessing II $11.41$ $8.49$ $2.92$ $3.03$ : Metrics of energy characteristics []{data-label="table1"} $\varepsilon_1$ $\varepsilon_2$ $\theta _{j}$ (Degree) $f$ -------------------- ----------------- ----------------- ------------------------ --------- OPTI $0.14$ $0.78$ $11.91$ $0.008$ +Postprocessing I $0.0067$ $0.021$ $15.55$ $0.002$ +Postprocessing II $0.0089$ $0.034$ $17.37$ $0.002$ : Metrics of relative errors[]{data-label="table2"} We see that OPTI+Postprocessing I reduces, and\ OPTI+Postprocessing II, on the contrary, increases the energy relative to OPTI by approximately the same values of $10^{32} erg$. OPTI+Postprocessing II practically does not change the potential field energy, while it decreases significantly in OPTI+Postprocessing I. Free energies after both types of postprocessing are equally (significantly) increased by the same order of magnitude $10^{32} erg$, while the OPTI+Postprocessing II energy is $~2\times10^{32} erg$ greater than the OPTI+Postprocessing I energy. The OPTI result gives a relatively small $\varepsilon_1$, i.e. formally we can consider this solution close to the solenoidal one. If this value were not small, notionally no compliance of OPTI extrapolation with the real field could be expected. At the same time, as we see from $\varepsilon_2$, we cannot estimate the real free energy with fair accuracy. Note that $\varepsilon_2$ is consistent with the result obtained for OPTI in [@Mastrano_2018]. Values of $\varepsilon_1$ and $\varepsilon_2$ for OPTI+Postprocessing I, II show that we succeed in getting rid of non-solenoidality. Besides, values of $\varepsilon_2$ are well consistent with similar ones [@Mastrano_2018] for NLFFF extrapolation methods operating with the class of solenoidal fields. From $f$ metric (Table \[table2\]), we can also see the result of eliminating the non-solenoidality to the level of numerical errors, we hope. After elimination of non-solenoidality we naturally obtain a certain increase in ${\theta _{j}}$ metric after postprocessing, indicating deviation of the solution from the force-free approximation. We cannot estimate, to which extent this value reflects the error or effect of the real deviation of the field from the force-free field. Applying Equation \[eq11\] to each $z$ layer and displaying ${\theta _{j}}$ in Figure \[fig4\] , we provide better representation of the force-free characteristic of our results. ![Altitude dependence of ${\theta _{j}}$: red - OPTI, green - OPTI+Postprocessing I, blue -OPTI+Postprocessing II.[]{data-label="fig4"}](theta_j1.png "fig:"){width="48.00000%"} ![Altitude dependence of ${\theta _{j}}$: red - OPTI, green - OPTI+Postprocessing I, blue -OPTI+Postprocessing II.[]{data-label="fig4"}](theta_j2.png "fig:"){width="48.00000%"} We see that, starting at the same level, all three characteristics decrease with height. Postprocessing II shows the highest value. Starting from the height of $\sim 110$ Mm, all three characteristics merge and grow to a large value typical of the potential field. The OPTI algorithm may not significantly change the initial potential field at these altitudes. In any case, the field at these altitudes is quite small and does not essentially contribute to the energy we are interested in. Evolution of AR 11158 energy in OPTI results ============================================ Consider the results of the NLFF extrapolation of AR 11158 in the time interval with the X-class flare that began at 2011/02/15 01:44UT. Figure \[fig5\] describes the time behavior of energies of all three of our extrapolations (solid lines) and energies corresponding to the reference potential fields (dashed lines). These dependencies reflect the X-flare event well, but they have significant shifts along the energy axis, same as in the case discussed in the previous section. The lowest energies result from OPTI+Postprocessing I, then OPTI and the highest belongs to OPTI+Postprocessing II. Energies of the reference potential fields of OPTI and OPTI+Postprocessing II, as expected, are very close to each other. Energies of the OPTI+Postprocessing I reference potential field is lower by $2.5\times10^{32}$ erg. For OPTI, energies of fields and reference potential field were calculated (as it is normally done) in the root domain of the box, using $BVP_{potI}$ for the OPTI reference potential field. For all other cases, energies were calculated for the entire box using $BVP_{potII}$ for the reference potential fields In general, we can say that energies calculated for the entire box or its root domain always have little difference between each other, which can be clearly seen from graphs for the OPTI and OPTI+Postprocessing II reference potential fields. They almost coincide with each other despite difference of their $BVn_{st}$. ![Time plot of energies for the AR 11158 X-class flare. Solid curve is the GOES flux profile. Solid colored lines -$E$, dashed colored lines - $E_pot$, red - OPTI, green - OPTI+Postprocessing I, blue -OPTI+Postprocessing II.[]{data-label="fig5"}](e_X_cld.pdf){width="70.00000%"} Figures \[fig6\] and \[fig7\] describe the time behavior of free energies. Note that all three current components of our solutions with zero $BVn_{all}$ differ from each other by only some potential (for OPTI non-solenoidal) fields. Therefore, due to the solenoidality of OPTI+Postprocessings I, II, their free energies must coincide with each other. That’s what we can observe in Figures \[fig6\] and \[fig7\]. Influence of non-solenoidality in OPTI leads to a general decrease in free energy without fundamental changes in general nature of its behavior. ![Time plot of free power for X-class flare of AR 11158. Solid curve is the GOES flux profile.Triangles – for OPTI, asterisk – OPTI+Postprocessing I.[]{data-label="fig6"}](free_X_cld_nochange.pdf){width="70.00000%"} ![Time plot of free power for X-class flare of AR 11158. Solid curve is the GOES flux profile. Triangles – for OPTI, asterisk – OPTI+Postprocessing II.[]{data-label="fig7"}](free_X_cld.pdf){width="70.00000%"} As can be seen from Figures \[fig6\] and \[fig7\], all three dependencies of free energy give the order of its dropping $-\Delta E_{free}~10^{32}$ erg during one hour after the X-flare onset. This corresponds to the estimate of released energy in X-class events [@Hudson; @Bleybel]. To explain significant difference between this result and that in [@Sun], we compare the height dependence of the mean free energy density in Figure \[fig8\] with the same dependence for the same point in time shown in Figure 4 from [@Sun]. While heights and the value maxima coincide, our mean free energy density has a much larger scale of descent. In [@Sun], this value shows the field proximity to the potential field starting already at $~20Mm$ heights. We have the same situation developing much higher – at $\sim 100$Mm altitudes (which is also consistent with the behavior of the $\theta _{j}$ force-free parameter in Figure \[fig4\]). This difference in behavior of the mean free energy density seems to be related, as mentioned in Section 3, to the earlier termination of the optimization process in [@Sun]. In our case, we go further, thus calculating free energy more accurately. ![Height profile of OPTI mean free energy density of AR 11158 for T=2011-02-15 00:00:00 UT.[]{data-label="fig8"}](free_dens_X.pdf){width="70.00000%"} Another important difference from [@Sun] is manifested in negative values we obtain (for the entire time interval in question) of the mean free energy density at zero height ( Figure \[fig8\])). This result is interesting from the physical point of view. As it was shown in [@Livshits] ), for any force-free field above the sphere the following consequence of the virial theorem is true – on the sphere ($r=R$)), a field is more radial in relation to the reference potential field: $$\label{eq13} \int_{S}\left( {\bf B}_{pot}\right) _{t}^{2}ds>\int_{S}({\bf B}_{t})^{2}ds, \qquad (r=R).$$ From Equation \[eq13\], considering equality of the normal components $({\bf B})_n$ and $({\bf B}_{pot})_n$ it automatically follows: $$\label{eq14} \int_{S}({\bf B}^2-{\bf B_{pot}} ^2)ds <0, \qquad (r=R).$$ Since the mean free energy density is proportional to the integral of Equation \[eq14\] , it should also be strictly negative for the force-free field above the sphere. In the case of rectangular box geometry, the statement above, strictly speaking, is not necessarily valid. However, the real bottom boundary of our box corresponds to the sphere, magnetic regions are mostly compact, and generally strong fields should be force-free on the photosphere. Hence, negativity of the mean free energy density at zero altitude is likely to manifest itself. That’s what we have. Interestingly, the [@Low] model has the same result for rectangular box geometry $[61\times 61\times 31]$, $[36\times36\times1.5]$Mm, with parameters $l=0.5$, $\varphi =1.4$, core ($[46,46,27]$,$[2.256\times2.256\times1.3]$Mm) and $\max(B_z)=800 Gs$ (see Figure \[fig9\]). In Figure \[fig9\], we see that this property of mean free energy density to take negative values near zero altitude is equally manifested not only in the numerical OPTI result, but also in the analytical field model. ![Height profile of the mean free energy density for the Low model.[]{data-label="fig9"}](free_dens_low.pdf){width="70.00000%"} Conclusion ========== Evaluation of the OPTI extrapolation validity has shown significant influence of partial non-solenoidality on the free energy OPTI results. This effect is displayed in a significant shift in the free energy, but does not significantly affect its time behavior character. Both with and without divergence cleaning, the OPTI extrapolation gives the same reasonable estimates of changes in the active region energy during powerful eruptive X-class events. The divergence cleaning of OPTI results, regardless of the implementation method, leads to the same values of free energy, only the energy levels being changed. Thus, we can trust behavior of changes in free energy, both with and without divergence cleaning. Converting the non-solenoidal part into the force component, divergence cleaning is beneficial to the obtained results. It significantly reduces the number of non-realistic entangled magnetic lines, without essential changes in the field structure. Divergence cleaning (Postprocessings II), which preserves the photospheric field normal component, does not change the reference potential field energy. Divergence cleaning (Postprocessings I), which preserves the photospheric field transverse components, reduces significantly the energy of its reference potential field. In terms of undesirability of changes in the photospheric field normal component and obtaining high energies of the potential field and energies of the calculated field, Postprocessing II looks more preferable. In our opinion, the main advantage of the OPTI approach to NLFF extrapolation (both with and without divergence cleaning) is preserving the unchanged photospheric $({\bf J})_z$ current component, which, in fact, allows to obtain adequate estimate of the real energy release of the active region. From this point of view, a significant difference can be expected between the reality and results of NLFF extrapolations allowing modification of the $({\bf J})_z$ component. This, for example, takes place in the approach based on the algorithms of Grad Rubin class. For the same reason, preprocessing (the procedure removes most of the net force and torque from the data, [@Wiegelmann_2006]) often used in practice (among all, to OPTI extrapolation) is undesirable, because it inevitably leads to local changes in $({\bf J})_z$. The work was performed with budgetary funding of Basic Research program II.16.The authors thank Irkutsk Supercomputer Center of SB RAS for providing the access to HPC-cluster Akademik V.M. Matrosov (Irkutsk Supercomputer Center of SB RAS, Irkutsk: ISDCT SB RAS; <http://hpc.icc.ru>, accessed 16.05.2019). The authors declare that they have no conflicts of interest. [19]{} \#1[ISBN \#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[*\#1*]{} \#1[**\#1**]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1 \#1\#2 \#1[\#1]{} \#1[\#1]{} \#1[*\#1*]{} \#1[\#1]{} \#1[**\#1**]{} \#1[\#1]{} \#1[*\#1*]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1 \#1[\#1]{} \#1[\#1]{} \#1[[](http://dx.doi.org/#1)]{} \#1[[](http://arxiv.org/abs/#1)]{} \#1[[](http://adsabs.harvard.edu/abs/#1)]{} \#1 \#1 \#1 \#1 \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} : , . , . . , , , , , , , , , , , , , , , , , : , . (), . . . , , , : , . , . . . , : , . , . . . : , . , . . . : , . , . . . : , , . . . , , , : , . , . . . , : , . , . . . , , : , . , . . . , , , , , , , , : , . , . . , : , . , . . . , , , , , , , , , : , . , . . . , , , , , , : , . , . . . : , . , . . . , , , : , . , . . . , , : , . , . . . : , . , . . . , , : , . , . . .

ArXiv

--- abstract: 'In analysis and control of large-scale nonlinear dynamical systems, a distributed approach is often an attractive option due to its computational tractability and usually low communication requirements. Success of the distributed control design relies on the separability of the network into weakly interacting subsystems such that minimal information exchange between subsystems is sufficient to achieve satisfactory control performance. While distributed analysis and control design for dynamical network have been well studied, decomposition of nonlinear networks into weakly interacting subsystems has not received as much attention. In this article we propose a vector Lyapunov functions based approach to quantify the energy-flow in a dynamical network via a model of a comparison system. Introducing a notion of power and energy flow in a dynamical network, we use sum-of-squares programming tools to partition polynomial networks into weakly interacting subsystems. Examples are provided to illustrate the proposed method of decomposition.' author: - - - bibliography: - 'IEEEabrv.bib' - 'references.bib' - 'RefKundu.bib' title: Decomposition of Nonlinear Dynamical Networks via Comparison Systems --- Introduction ============ Control of large-scale dynamical systems is challenging because of the computational complexity. In such scenarios, distributed control design is an attractive option which often offers a trade off between computational tractability and controller performance [@Bullo:2009; @Siljak:2010]. Distributed control design has become a critical challenge with the advent of interdependent large-scale infrastructure systems, smart cities, cloud architectures, and coordination and control problems for large fleets of independent agents. The combination of scale and dynamical complexity of these systems makes it computationally difficult to design global control solutions. Design of distributed control architectures can be roughly described as a three step process: 1) decomposition of the network into subsystems, 2) distributed control design to stabilize isolated subsystems, and 3) verification of stability of the closed-loop network under distributed control. The first step of this process is critical, and is the focus of this paper, since a poor system or model decomposition can affect all subsequent steps of distributed control design. In some applications, the physical layout or construction of a network dictates the subsystem structure. For example, in critical infrastructure systems the subsystem structure is traditionally defined based on distance and connectivity of buses, nodes or junctions. The increasing integration and instrumentation of heterogeneous infrastructure systems, e.g. power-gas systems [@chiang2016large; @li2008interdependency] or water-power systems [@pereira2016joint], introduces spatial and temporal overlap in time-scales across different systems, removing the insularity required previously to guarantee control performance. Moreover, there are scenarios where a suitable system decomposition may not be known [*a priori*]{}, e.g. design of large-scale or ad-hoc communication networks [@chiang2007layering; @chiang2016large] or cyber-physical systems made of agile teams of agents. These reasons motivate the development of new system decomposition algorithms that identify appropriate subsystems that facilitate and enhance distributed control design and performance. While the idea of using a decomposed network model for distributed stability analysis and control design has been proposed in the literature [@Bullo:2009; @Siljak:2010; @Kundu:2015NecSys; @Kundu:2017Multiple], decomposition of nonlinear dynamical networks for distributed analysis and control has not received as much attention. In [@Anderson:2011; @Anderson:2012], authors proposed an algorithm to decompose nonlinear networks into sub-networks to perform stability analysis using composite Lyapunov functions. However, the proposed method involved linearization of the dynamics, and hence 1) is not suitable for networks that have nonlinear interactions, and 2) provides a decomposition that is only appropriate in the close neighborhood of the operating point. In a recently concluded work [@Liu:2017], authors used data-driven deep learning techniques to decompose a nonlinear dynamical network using Koopman grammians. In this paper we propose a model-based decomposition method for nonlinear dynamical networks using comparison systems representation. Vector Lyapunov functions and comparison systems have been used for distributed stability analysis and control design for nonlinear networks [@Bailey:1966; @Araki:1978; @Michel:1983; @Siljak:1972; @Weissenberger:1973]. In [@Kundu:2015CDC; @Kundu:2015NecSys; @Kundu:2017Multiple] authors proposed sum-of-squares programming methods to compute the comparison systems for generic polynomial dynamical networks. In this paper, we use vector Lyapunov functions and comparison systems based approach to quantify the energy-flow in the network and perform a spectral clustering of the network such that the energy-flow between the sub-networks is minimized. The proposed method is dynamic since it allows to find a decomposition that best suits a given set of operating conditions. The paper is structured as follows. In Section\[S:problem\], we describe the problem. Section\[S:prelim\] presents a brief overview of the necessary background. The proposed algorithms is discussed in details in Section\[S:method\], while numerical examples are presented in Section\[S:results\]. The paper is concluded in Section\[S:concl\]. Problem Description {#S:problem} =================== Consider an $n$-dimensional nonlinear dynamical system of the form $$\begin{aligned} \label{E:f} \dot{x}(t) &= f(x(t))~ \forall t \geq 0\,,~x\in\mathbb{R}^n\,,\end{aligned}$$ with an equilibrium at the origin $(f(0)= 0)$, where $f:\mathbb{R}^n \rightarrow \mathbb{R}^n$ is locally Lipschitz. For brevity, we would drop the argument $t$ from the state variables, whenever obvious. Often complex real-life systems can be modeled in the general form . Analysis of such systems for control and decision making, however, can be non-trivial due to the nonlinearities. Especially, computational techniques for nonlinear systems analysis generally scale poorly with the size of the system. Furthermore, in real-time operations and decision making, often the information required for a global system-wide analysis is unavailable, rendering a decomposition-based distributed analysis the preferred option. The system in can be viewed as a nonlinear network of $m$ dynamical subsystems as follows: \[E:fi\] $$\begin{aligned} \mathcal{S}_i\,:\quad& \dot{x}_i=f_i(x_i)+\!\!\sum_{j\in\mathcal{N}_i\backslash{i}}\!\!g_{ij}(x_i,x_j)\,,\\ \text{where,}\,~&~x_i\in\mathbb{R}^{n_i},\,x_j\in\mathbb{R}^{n_j}~\,\forall j\in\mathcal{N}_i\backslash{i}\,,\\ &f_i(0)=0\text{ and }g_{ij}(x_i,0)=0~\,\forall x_i\,~\forall j\in\mathcal{N}_i\backslash{i}\,.\end{aligned}$$ Each subsystem $\mathcal{S}_i$ has some local dynamical state variables $x_i$ associated with it, while $\mathcal{N}_i$ (with $i\in\mathcal{N}_i$) represent the neighboring nodes that interact with the subsystem-$i$. $f_i:\mathbb{R}^{n_i}\rightarrow\mathbb{R}^{n_i}$ represents the isolated subsystem dynamics, while the terms $g_{ij}:\mathbb{R}^{n_i}\!\times\!\mathbb{R}^{n_j}\!\rightarrow\!\mathbb{R}^{n_i}$ represent the interactions from the neighbors. Note that the interaction model (total interaction as a sum of the $g_{ij}$’s) used here is chosen for simplicity, but more generic interaction models can also be considered. Also, it is assumed that the decomposition is overlapping in a way that no two subsystems share any common state variable, and $$\begin{aligned} {\sum}_{i=1}^mn_i=n\,. \end{aligned}$$ A decomposed system model of the form has been used for stability analysis studies in the literature (see [@Araki:1978; @Michel:1983; @Siljak:1972; @Weissenberger:1973; @Kundu:2017Multiple; @Kundu:2015ACC]). However, finding a suitable decomposition of the large system that facilitates distributed analysis is not trivial. In [@Anderson:2012; @Anderson:2011] authors applied spectral graph partitioning algorithms on linearized system dynamics to obtain a system decomposition that aims to minimize worst-case energy flows between subsystems. However, because of the linearization, such methods are only applicable in the close neighborhood of the operating points. For systems under large disturbance, i.e. far away from the operating point, the decomposition structure is expected to change. Furthermore, quadratic and higher order terms in the interactions do not show up in the linearized representation and are therefore completely ignored in the decomposition. In this article, we propose a decomposition algorithm to partition a nonlinear dynamical network into weakly-interacting subsystems, whereby the nonlinearity of the system is explicitly considered in the decomposition algorithm, via the use of a *comparison systems* representation, as explained in the following sections. Preliminaries {#S:prelim} ============= Let us use $\left|\,\cdot\,\right|$ to denote both the Euclidean norm of a vector and the absolute value of a scalar; and use $\mathbb{R}\left[x\right]$ to denote the ring of all polynomials in $x\in\mathbb{R}^n$. Stability Analysis ------------------ The equilibrium point at the origin of is Lyapunov stable if, for every $\varepsilon\!>\!0$ there is a $\delta\!>\!0$ such that $\left\vert x(t)\right\vert\!<\!\varepsilon~\forall t\!\geq\! 0$ whenever $\left\vert x(0)\right\vert\!<\!\delta\,.$ Moreover, it is asymptotically stable in a domain $\mathcal{D}\!\subseteq\!\mathbb{R}^n,\,0\!\in\!\mathcal{D},$ if it is Lyapunov stable and $\lim_{t\rightarrow\infty}\left\vert x(t)\right\vert \!=\!0\,$ for every $x(0)\!\in\!\mathcal{D}$. \[T:Lyap\] (Lyapunov, [@Lyapunov:1892], [@Khalil:1996]) If there exists a domain $\mathcal{D}\!\!\subseteq\!\!\mathbb{R}^n$, $0\!\in\!\!\mathcal{D}$, and a continuously differentiable positive definite function [$\tilde{V}\!\!:\!\mathcal{D}\!\rightarrow\! \mathbb{R}_{\geq 0}$]{}, i.e. the ‘Lyapunov function’ (LF), then the origin of is asymptotically stable if $\nabla{\tilde{V}}^T\!\!f(x)$ is negative definite in $\mathcal{D}$. An estimate of the region-of-attraction (ROA) can be given by [@Genesio:1985] $$\begin{aligned} \label{E:ROA} &~~\mathcal{R}:=\left\lbrace x\in\mathcal{D}\left| {V}(x)\leq 1\right.\right\rbrace\,,~\text{with}~{V}(x)= \frac{\tilde{V}(x)}{\gamma^{max}},\\ &\text{where}~\gamma^{max}:=\max\left\lbrace \gamma\, \left\vert\, \left\lbrace x\in\mathbb{R}^n\left| \tilde{V}(x)\leq\gamma\right.\right\rbrace \subseteq \mathcal{D}\right.\right\rbrace\,,\notag\end{aligned}$$ i.e. the boundary of the ROA is estimated by the unit level-set of a suitably scaled LF ${V}(x)$. Relatively recent studies have explored how sum-of-squares (SOS) based methods can be utilized to find LFs by restricting the search space to SOS polynomials [@Wloszek:2003; @Parrilo:2000; @Tan:2006; @Anghel:2013]. A (multivariate) polynomial $p\in\mathbb{R}\left[x\right],~x\in\mathbb{R}^n$, is called a *sum-of-squares* (SOS) if there exist some polynomial functions $h_i(x), i = 1\ldots s$ such that $p(x) = \sum_{i=1}^s h_i^2(x)$. We denote the ring of all SOS polynomials in $x\in\mathbb{R}^n$ by ${\Sigma}[x]$. Checking if $p\!\in\!\mathbb{R}[x]$ is an SOS is a semi-definite problem which can be solved with a MATLAB$^\text{\textregistered}$ toolbox SOSTOOLS [@sostools13; @Antonis:2005a] along with a semidefinite programming solver such as SeDuMi [@Sturm:1999]. The SOS technique can be used to search for polynomial LFs by translating the conditions in Theorem\[T:Lyap\] to equivalent SOS conditions [@sostools13; @Wloszek:2003; @Wloszek:2005; @Antonis:2005; @Antonis:2005a; @Chesi:2010a]. An important result from algebraic geometry, called Putinar’s Positivstellensatz (P-Satz) theorem [@Putinar:1993; @Lasserre:2009], helps in translating the SOS conditions into SOS feasibility problems. \[T:Putinar\] Let $\mathcal{K}\!\!=\! \left\lbrace x\in\mathbb{R}^n\left\vert\, k_1(x) \geq 0\,, \dots , k_m(x)\geq 0\!\right.\right\rbrace$ be a compact set, where $k_j\!\in\!\mathbb{R}[x]$, $\forall j\in\left\lbrace 1,\dots,m\right\rbrace$. Suppose there exists a $\mu\!\in\! \left\lbrace \sigma_0 + {\sum}_{j=1}^m\sigma_j\,k_j \left\vert\, \sigma_0,\sigma_j \in \Sigma[x]\,,\forall j \right. \right\rbrace$ such that $\left\lbrace \left. x\in\mathbb{R}^n \right\vert\, \mu(x)\geq 0 \right\rbrace$ is compact. Then, $$\begin{aligned} p(x)\!>\!0~\forall x\!\in\!\mathcal{K} \!\implies\! p \!\in\! \left\lbrace \sigma_0 \!+\! {\sum}_j\sigma_jk_j\!\left\vert\, \sigma_0,\sigma_j\!\in\!\Sigma[x],\forall j\!\right.\right\rbrace\!.\end{aligned}$$ Linear Comparison Principle --------------------------- [In]{} [@Conti:1956; @Brauer:1961] the authors proposed to view the LF as a dependent variable in a first-order auxiliary differential equation, often termed as the ‘comparison equation’ (or, ‘comparison system’). [It was shown in [@Bellman:1962; @Bailey:1966] that, under certain conditions,]{} the comparison equation can be effectively reduced to a set of linear differential equations. Noting that all the elements of the [ matrix]{} $e^{At},~ t\geq 0$, where $A=\left[a_{ij}\right]\in\mathbb{R}^{m\times m}$, are non-negative if and only if $a_{ij}\geq 0, i\neq j$, it was shown in [@Beckenbach:1961; @Bellman:1962]: \[L:comparison\] Let $A\!\in\!\mathbb{R}^{m\times m}$ have non-negative off-diagonal elements, $v:[0,\infty)\!\rightarrow\!\mathbb{R}^m$ and $r:[0,\infty)\!\rightarrow\!\mathbb{R}^m$. If $v(0)\!=\!r(0)\,$, $\dot{v}(t)\!\leq\!A\,v(t)$ and $\dot{r}(t)\!=\!A\,r(t)\,,$ then $v(t)\!\leq\! r(t)~\forall t\!\geq\! 0\,$. We henceforth refer to Lemma\[L:comparison\] as the ‘comparison principle’; the [linear time-invariant system $\dot{r}(t)\!=\!A\,r(t)$]{} as a ‘comparison system’ (CS); and the matrix $A$ as the comparison matrix (CM). Similarity Graphs and Spectral Partitioning ------------------------------------------- Consider a graph $\mathcal{G}(\mathcal{V},\mathcal{E})$ where $\mathcal{V}= \{1, 2,\dots, m\}$ represent a set of $m$ vertices (or, nodes) and $\mathcal{E}\subseteq\mathcal{V}\times\mathcal{V}$ denote a set of $l$ edges between the vertices (i.e. $\mathbf{card}({\mathcal{E}})=l$). A notion of *similarity* (or, edge weight), denoted by a scalar $w_{ij}\geq 0$, is associated with each pair of nodes $(i,j)$ in the graph, such that $w_{ij}>0$ if an edge exists between the nodes $i$ and $j$, i.e. $(i,j)\in\mathcal{E}$, and $w_{ij}=0$ if there does not exist an edge between the nodes $i$ and $j$ (i.e. $(i,j)\notin\mathcal{E}$). The edge weights quantify how similar two nodes in the graph are to each other, i.e. higher the value of the edge weights more similar the corresponding vertices are. The spectral clustering technique used in this paper for graph partitioning is applicable for graphs that are *undirected*, i.e. the edge weights (or similarity values) satisfy $w_{ij}=w_{ji}\,\forall i\neq j$ (clearly, for undirected graphs, $(i,j)\in\mathcal{E}$ implies that $(j,i)\in\mathcal{E}$). If a graph is *directed*, we can convert it into an *undirected* graph by assigning $$\begin{aligned} w_{ij}\gets\frac{1}{2}\left(w_{ij}+w_{ji}\right)\,,\text{ or }w_{ij}\gets\max\lbrace w_{ij},w_{ji}\rbrace\,.\end{aligned}$$ A symmetric *weighted adjacency matrix* $W(\mathcal{G})=[w_{ij}]\in\mathbb{R}^{m\times m}$ is then constructed, such that $w_{ii}=0~\forall i\in\mathcal{V}$, and $w_{ij}=w_{ji}>0~\forall (i,j)\in\mathcal{E}\,.$ Finally, a normalized (symmetric positive semi-definite) *graph Laplacian matrix*, $L_{sym}(\mathcal{G})\in\mathbb{R}^{m\times m}$, is defined as, $$\begin{aligned} L_{sym} = I_{m}-D^{-1/2}WD^{-1/2},\,~\,D=\mathbf{diag}\left(W\,\mathbf{1}_m\right)\,,\end{aligned}$$ where $I_m$ is an $m\times m$ identity matrix, $\mathbf{1}_m$ is a $m$-dimensional column vector with each entry equal to $1$, $D$ is a diagonal matrix with diagonal entries equal to the sum of the corresponding row in $W$; and $\left(\cdot\right)^{-1/2}$ denotes the inverse of the square root a matrix. The idea behind spectral graph partitioning is to partition the graph $\mathcal{G}$ into sub-graphs in such a way that the vertices within a sub-graph are *similar* to each other (i.e. have high edge weights), while vertices from two different sub-graphs are *dissimilar* (i.e. have low edge weights). In particular, the $K$-means spectral clustering algorithm uses the normalized graph Laplacian $L_{sym}$ to partition the vertices into $K$ clusters. The first $K$ eigenvectors of the graph Laplacian, in the order of increasing eigenvalues, are used as columns to construct a matrix $U\in\mathbb{R}^{m\times K}$. Then the normalized rows of the matrix $U$ are clustered using the $K$-means clustering algorithm. For further details on the algorithm, please refer to [@Von:2007; @Ng:2002]. Decomposition via Comparison Systems {#S:method} ==================================== In this section we describe our proposed approach of decomposing a nonlinear (polynomial) dynamical network using a comparison system. The proposed approach is composed of two steps - 1) computing a comparison system, and 2) apply spectral graph partitioning on the comparison system. Constructing Comparison System ------------------------------ Given a polynomial network of the form , vector Lyapunov functions of the subsystems can be used to construct a linear comparison system. In [@Kundu:2015CDC] authors used sum-of-squares programming to compute the vector Lyapunov functions and the comparison matrix for polynomial dynamical networks. It is assumed that, in absence of any interaction, each subsystem $\mathcal{S}_i$ is locally asymptotically stable at the origin. Specifically, we assume that for each $i\in\lbrace 1,\dots,m\rbrace$ $$\begin{aligned} \mathcal{S}_i\,\text{ (isolated) }\,:~\dot{x}_i=f_i(x_i)\end{aligned}$$ has an asymptotically stable equilibrium point at the origin, and admits a polynomial Lyapunov function $V_i(x_i)$. The expanding interior algorithm [@Wloszek:2003; @Anghel:2013] can be used to compute these LFs and a corresponding region of attraction such that $\mathcal{R}_i:= \left\lbrace x_i\in\mathbb{R}^{n_i}\left| V_i(x_i)\leq 1\right.\right\rbrace$ is an estimate of the region of attraction of the $i$-th isolated subsystem. The vector LFs thus computed are used to define the states of the *comparison system*, $$\begin{aligned} v(t)&:=\begin{bmatrix} V_1(x_1(t)) & V_2(x_2(t)) & \dots & V_m(x_m(t)) \end{bmatrix}^T\end{aligned}$$ such that the comparison system is a *positive system*, i.e. each state of the comparison system only takes non-negative values at all time $t$. Then the objective is to compute the comparison matrix (CM) $A=[a_{ij}]$ such that in some domain $$\begin{aligned} \mathcal{D}&:= \left\lbrace x\in\mathbb{R}^n\left| \,V_i(x_i)\leq \gamma_i~\forall i\right.\right\rbrace\subseteq \left\lbrace x\in\mathbb{R}^n\left| \,x_i\in\mathcal{R}_i~\forall i\right.\right\rbrace\end{aligned}$$ where $\gamma_i\leq 1$, the following comparison equations hold: $$\begin{aligned} \dot{v}\leq A\,v\,~\forall x\in\mathcal{D}\,,\text{ where }A=[a_{ij}]\,,\,a_{ij}\geq 0\,\forall i\neq j\,.\end{aligned}$$ Sum-of-squares programming can be used to compute the CM which has non-negative off-diagonal elements. In order to obtain a stable comparison system, the CM $A$ has to be Hurwitz. While solving for a generic Hurwitz CM for a large system is difficult, a sufficient condition for CM to be Hurwitz is given by the Gershgorin circle theorem [@Bell:1965] which says that a strictly diagonally dominant matrix[^1] with negative diagonal elements is Hurwitz. This condition motivates us to solve for each row of the CM, in a parallel and scalable way, by solving the following problem: $$\begin{aligned} \forall i: ~\underset{a_{ij}}{\min}~{\sum}_{j\in\mathcal{N}_i} a_{ij},\text{ s.t. }\,\dot{V}_i\leq {\sum}_{j\in\mathcal{N}_i} a_{ij}V_j~\forall x\in\mathcal{D},\end{aligned}$$ which can be formulated as an SOS optimization: $$\begin{aligned} \forall i\!:& ~\underset{a_{ij},\sigma_{ij}}{\text{minimize}}~{\sum}_{j\in\mathcal{N}_i} a_{ij}\\ \text{s.t.}&~ a_{ii}\in\mathbb{R}\,,~a_{ij}\!\geq 0~\forall i\neq j\,,~\sigma_{ij}\in\Sigma[x_{\mathcal{N}_i}]\,,\notag\\ &\!-\!\nabla V_i^T\!(f_i\!+\!\sum_{j\neq i} g_{ij}\!)\!+\!\!\sum_{j\in\mathcal{N}_i}\!\!\left(a_{ij}V_j\!-\!\sigma_{ij}(\gamma_j\!-\!V_j)\right)\!\in\!\Sigma[x_{\mathcal{N}_i}].\notag\end{aligned}$$ Here $\nabla$ is the gradient operator, and $x_{\mathcal{N}_i}$ denote the state variables that belong to the neighborhood of the subsystem $i$ (recall that $i\in\mathcal{N}_i$). The CM is guaranteed to be Hurwitz, if the sum of each row of the CM $A$ turns out to be negative. Otherwise, one can do an eigenvalue analysis to determine whether or not the CM is Hurwitz. Note that the CM is not unique, and depends on the vector LFs computed earlier, as well as the domain of definition (also referred to as the domain of interest) $\mathcal{D}$. For a given set of vector LFs, the CM could be different if the domain of interest changes. This particular feature is useful in monitoring the change in the energy-flow pattern in a nonlinear dynamical network under different operating conditions. Energy-Flow in the Network -------------------------- The comparison system (CS) $$\begin{aligned} \label{E:CS} \dot{r}(t)&=A\,r(t)\,,\quad r(0)=v(0)\,,\end{aligned}$$ where we denote the CS states as $r=\left[ r_1\,,\,\dots\,\, r_m \right]^T\!\!,$ provides an upper bound on the values of the level-set of the subsystem vector LFs, i.e. $V_i(x_i(t))\leq r_i(t)~\forall i\,\forall t\,.$ The comparison system can be represented as a *similarity graph* $\mathcal{G}(\mathcal{V},\mathcal{E})$, where $\mathcal{V}=\lbrace i\rbrace_{i=1}^m$ represent the nodes, while $\mathcal{E}=\lbrace E_k\rbrace_{k=1}^l$ denote the directional edges (where $l$ is the number of positive off-diagonal elements in the CM). We consider the case when the CM $A$ is Hurwitz and the original system is guaranteed to be asymptotically stable. The goal here is to partition the network such that the *worst case* energy flow between the sub-networks is minimized. However, for generic nonlinear dynamical networks of the form , the concept of *energy-flow* along the edges is not well defined, but are usually informed by physical knowledge of the system (e.g. in power systems, the edge flow could be represented by the active power flowing from one node to another). It may be convenient to view the vector LFs as some form of energy-levels of the subsystems of the network, while the CS provides some bound on the rate of change of the energy levels of the subsystems. We use the vector LFs and their time derivatives to quantify the energy flow across edges as follows: Let us define the *‘power-flow’*, $\phi_{ij}(t)$, from node $j$ into the node $i$ along a directional edge $(i,j)$ in the dynamical network which admits vector LFs $V_i(x_i)\,\forall i$, as $$\begin{aligned} \phi_{ij}(t):=\nabla V_i^T g_{ij}(x_i,x_j)\,,\end{aligned}$$ and *energy-flow*, $\psi_{ij}(t_0,t_f)$, across the directional edge $(i,j)$ between time $t_0$ and $t_f$ is defined as the following integral, $$\begin{aligned} \psi_{ij}(t_0,t_f):=\int_{t_0}^{t_f}\left|\,\phi_{ij}(t)\,\right|\,dt\,.\end{aligned}$$ Note that the *power-flow* can have both positive and negative values, where the positive values are referred to as the *‘in-flow’* of power, while negative values are referred to as *‘out-flow’* of power. The *energy-flow* is the time integral of the absolute value of the power-flow. Computing the power and energy-flows requires solving the nonlinear dynamics for the given initial conditions. However, for a given set of possibly (or likely) initial conditions, we can compute upper bounds on the power and energy-flow using the vector LFs. Specifically, note that, since the dynamical equations are polynomial, and if the states are bounded, we can find finite positive scalars $\alpha_{ij}\in\mathbb{R}_{\geq 0}\,\forall (i,j)\in\mathcal{E}$, such that $$\begin{aligned} x\in\mathcal{D}\implies\left|\phi_{ij}(t)\right|\leq \alpha_{ij}\,V_j(x_j)\,~\,\forall (i,j)\in\mathcal{E}\,.\end{aligned}$$ Recall that $x_j\!=\!0$ implies $V_j(0)\!=\!0$ and $\phi_{ij}\!=\!0$. The scalars $\alpha_{ij}$ are bounded (since $\nabla V_i^T$ and $g_{ij}(\cdot)$ are polynomials), but not unique and depend on the domain of interest. Since the evolution of the vector LFs are governed by the CS, a closed-form expression can be found for the upper bound of the energy-flow in the network within a domain of interest. Specifically, let us define the observable (or, output) vector of the CS as $$\begin{aligned} y&=C^Tr\,,\quad C=\begin{bmatrix} c_1 & c_2 &\dots &c_l \end{bmatrix}\in\mathbb{R}_{\geq 0}^{m\times l}\end{aligned}$$ where each column of $C$ corresponds to the a different edge in the network. If the $k$-th edge $(i,j)\in\mathcal{E}$ is from the node $j$ to the node $i$, then the $k$-th column is given by $c_k=\alpha_{ij}e_j$ where $e_j$ denotes an $m$-dimensional *standard basis vector* with the $j$-th entry equal to $1$ and all other entries equal to $0$. Suppose that a nonlinear dynamical network is represented by the CS in some domain of interest $\mathcal{D}$ with a Hurwitz CM. Given a trajectory within $\mathcal{D}\,$, the net energy-flow in the network over time $t\in[0,\infty)$ is upper bounded by $-(C\,\mathbf{1}_l)^TA^{-1}v(0)$, while the energy-flow in the directional edge $(i,j)$ is upper bounded by $-\alpha_{ij}e_j^TA^{-1}v(0)$. The CS is a *positive system* with a CM that is Metzler (i.e. non-negative off-diagonal elements). If the CM is also Hurwitz, its inverse exists and has only non-positive entries [@Weissenberger:1973]. Therefore, the integral $\int_0^{\infty}\exp(At)dt$ is equal to $-A^{-1}$ with all non-negative entries. Note that, $$\begin{aligned} \psi_{ij}(0,\infty)&=\int_0^\infty\left|\phi_{ij}(t)\right|\,dt\leq \int_0^\infty \alpha_{ij}V_j(x_j(t))\,dt\\ &\leq \int_0^\infty\alpha_{ij}r_j(t)\,dt=\int_0^\infty \alpha_{ij}e_j^Tr(t)\,dt\\ &=\alpha_{ij}e_j^T\int_0^\infty e^{At}\,dt\,r(0)=-\alpha_{ij}e_j^TA^{-1}v(0)\end{aligned}$$ The total energy-flow across all edges is given by $$\begin{aligned} \sum_{(i,j)\in\mathcal{E}}\!\!\!\!\psi_{ij}(0,\infty)\leq\sum_{k=1}^l-c_k^TA^{-1}v(0)=-(C\,\mathbf{1}_l)^TA^{-1}v(0)\,.\!\!\!\!\!{\nobreak \ifvmode \relax \else \ifdim\lastskip<1.5em \hskip-\lastskip \hskip1.5em plus0em minus0.5em \fi \nobreak \vrule height0.75em width0.5em depth0.25em\fi}\end{aligned}$$ Note that, since $-A^{-1}$ has only non-negative entries, the upper bound on the energy flow in any edge is monotonically increasing with respect to the initial level-sets, $v(0)$, of the vector LFs. This is useful if we want to compute the *worst-case* energy flow in the network given some domain of interest. For example, if the domain of interest is defined in the form of $\mathcal{D}= \left\lbrace x\in\mathbb{R}^n\left| \,V_i(x_i)\leq \gamma_i~\forall i\right.\right\rbrace$, for some $\gamma_i\leq 1$, then the *worst-case* upper bound on the energy flow across the $k$-th edge is $-c_k^TA^{-1}\Gamma$, where $\Gamma=\begin{bmatrix} \gamma_1&\dots&\gamma_m \end{bmatrix}^T$. $K$-means Clustering on Comparison System ----------------------------------------- The formulation of energy-flow along the edges proposed here allows us to quantify the energy-flow pattern in the network given an initial condition. It is expected that, under different operating conditions, the energy-flow pattern of the network changes and henceforth the decomposition that minimizes the energy-flow between sub-networks may be different as well. While the exact operating conditions may not be known *apriori*, one can proceed with the analysis of the *worst-case* energy-flow within some domain of interest. For a domain of interest $\mathcal{D}\!=\! \left\lbrace x\in\mathbb{R}^n\left| \,v(t)\leq \Gamma~\forall t\right.\right\rbrace$, where $\Gamma\leq 1$ (element-wise), we use the *worst-case* upper bound on the energy flow across the edge in either direction to construct a symmetric *weighted adjacency matrix* of the network. Specifically, we construct the symmetric weighted adjacency matrix $W(\mathcal{G})=[w_{ij}]$ as, $$\begin{aligned} w_{ij}=\left\lbrace\begin{array}{c} \max\lbrace -\alpha_{ij}e_j^TA^{-1}\Gamma\,,\,-\alpha_{ji}e_i^TA^{-1}\Gamma\rbrace,\quad (i,j)\in\mathcal{E}\\ 0\,,\qquad\qquad \text{otherwise} \!\!\end{array}\!\!\right.\!\!\end{aligned}$$ Note, of course, that while this construction of weighted adjacency matrix is concerned with the *worst-case* energy-flow in a domain of definition, the formulation can be adapted to investigate the energy-flow under a given set of initial conditions, by replacing $\Gamma$ with $v(0)$. The final step involves choosing the desired number of sub-networks and applying the $K$-means clustering algorithm on the weighted adjacency matrix $W$. Simulation Results {#S:results} ================== In this section we present some numerical examples to illustrate how the proposed method can be used to compute the energy-flow across nonlinear dynamical networks under varying operating conditions, and decompose the network minimizing energy-flow between sub-networks. In the figures presented below, the width of the line between any two nodes represents the relative values of the energy flowing through the edge (i.e. thicker lines represent higher energy-flow). Lotka-Volterra System --------------------- Our first example is a Lotka-Volterra system that describes the evolution of population in a network of $16$ communities: $$\begin{aligned} \label{E:LV} \dot{x}_i&=\left(b_i-x_i\right)x_i-{\sum}_{j\in\mathcal{N}_i\backslash\lbrace i\rbrace}x_j\left(c_{ij}+d_{ij}x_i\right)\end{aligned}$$ where $i=1,2,\dots,16,$. We use a similar model as in [@Anderson:2012], with slightly modified network structure and interaction coefficients ($c_{ij}$ and $d_{ij}$) that are scaled down by a factor of $0.3$, in order to facilitate the construction of comparison systems over a wider range of operating conditions. After shifting the equilibrium point of the system to origin, the Lotka-Volterra system is expressed in the form . Quadratic vector LFs $V_i(x_i)$ are computed for each community, such that $\mathcal{R}_i=\left\lbrace x_i\left|\,V_i(x_i)\leq 1\right.\right\rbrace$ is an estimation of the region of attraction. Defining the domain of interest as $\mathcal{D}=\left\lbrace x\in\mathbb{R}^{16}\left|\,V_i(x_i)\leq \gamma_i~\forall i\right.\right\rbrace$, for some $\gamma_i\leq 1$, a CS is constructed based on which the energy-flow pattern of the network can be computed. Fig.\[F:LV\_low\_high\] shows calculated (based on the entries of the weighted adjacency matrix $W=[w_{ij}]$) and simulated (by randomly selecting an initial condition and evaluating the maximum value of $\psi_{ij}(0,\infty)$ over a number of scenarios) energy-flow patterns in the network for different domains of interest, and the associated decomposition into two sub-networks. When the domain of interest is defined to be a close neighborhood of the equilibrium point (by choosing $\gamma_i=0.01\,\forall i\,$), the simulated and calculated energy patterns look similar resulting in the same decomposition. When we choose $\gamma_i=0.6\,\forall i$, the calculated and simulated energy-flow patterns differ, although the decomposition remains the same. Note that, even when the energy-flow patterns differ in the calculation and simulation, the energy-flow between the sub-networks turns out to be small in both cases. Fig.\[F:LV\_gr\_cross\] shows the energy-flow patterns when the initial condition of the network, determined by the initial energy level $V_i(x_i(0))$ of the nodes, is varied. Specifically, we consider three different scenarios - 1) when the initial level-sets of the vector LFs of the nodes in group-$1$ are set at high value of $0.5$ and the nodes in group-$1$ are assigned a low value of $0.1$ (Fig.\[F:LV\_gr1\]), 2) when group-$2$ nodes are assigned high initial level-sets and group-$1$ nodes low initial level-sets (Fig.\[F:LV\_gr2\]), and 3) when the level-sets of the nodes at the interconnection of the two sub-networks are assigned high initial level-sets and others low initial level-sets (Fig.\[F:LV\_cross\]). We can see that in all three scenarios, the energy-flows between two sub-networks remain small, while the energy-flow patterns within each group varies. This shows that such a decomposition indeed favors a distributed analysis and control design, by dividing the network into weakly interacting subsystems. Network of Van der Pol Systems ------------------------------ Next we consider another example of a nonlinear network, composed of 9 Van der Pol systems from [@Kundu:2017Multiple], described by \[E:VP\] $$\begin{aligned} &\dot{x}_i=f_i(x_i)+\sum_{j\in\mathcal{N}_i\backslash\lbrace i\rbrace}g_{ij}(x_i,x_j)\\ &f_i(x_i)\!=\! \begin{bmatrix} x_{i,2}\\ \mu_i\,x_{i,2}(c_i^{(1)}\!\!-\!c_i^{(2)}x_{i,1}\!-\!x_{i,1}^2) \!-\! c_i^{(3)}x_{i,1}\end{bmatrix} \\ & g_{ij}(x_i,x_j)\!=\! \left[0\,,~\beta_{ij}^{(1)}\,x_{j,2} + \beta_{ij}^{(2)}\,x_{j,2}\,x_{i,1}\right]^T\,.\end{aligned}$$ where, $c_i^{(1)}\!=\!1\!-\!\left(0.5\,c_i^{(2)}\right)^2$, $c_i^{(3)}\!=\!1\!-\!{\sum}_{j\in\mathcal{N}_i\lbrace i\rbrace}({0.5\,\beta_{ij}^{(2)}c_i^{(2)}}\!-\!\beta_{ij}^{(1)})$, $\mu_i\,,\,\beta_{ij}^{(1)}$ and $\beta_{ij}^{(2)}$ are chosen randomly and $c_i^{(2)}$ are related to the equilibrium point before shifting. Vector LFs for the nine nodes are computed using sum-of-squares methods. Fig.\[F:VP\_gr\_cross\] shows the energy-flow patterns and decomposition for varying operating conditions. Specifically, we choose a domain of interest by selecting $\gamma_i=0.6\,\forall i$, and plot the worst-case energy flow pattern in Fig.\[F:VP\_high\], along with the decomposition. Keeping the same decomposition, we then monitor the change in energy-flow patterns as we vary the initial conditions across the network. In specific, we consider three scenarios: 1) choose high initial level-sets (equal to 0.6) for the group 1 nodes and low level-sets (equal to 0.1) for others (Fig.\[F:VP\_gr1\]), 2) high initial level-sets (=0.6) for group 2 nodes and low (=0.1) for others (Fig.\[F:VP\_gr2\]), and 3) high level-sets (=0.6) only for the nodes 1,5,7 and 8 that connect the two sub-networks. We observe that most of the energy-flow is contained within the sub-network, with minimal flow between the two sub-networks, implying that the decomposition yields a weakly interacting network. Conclusion {#S:concl} ========== In this article we considered the problem of decomposing a nonlinear network into weakly interacting subsystems, to facilitate distributed analysis and control design. Using a vector Lyapunov functions based approach we showed that the evolution of the energy levels of the nodes of the system can be modeled via a linear comparison system. Further, introducing a notion of power and energy flows between two nodes in a dynamical network, we proposed a method to compute the dynamic edge weights in the network. Finally, using spectral clustering techniques on the weighted adjacency matrix the nonlinear dynamical network is decomposed into weakly interacting subsystems. Sum-of-squares programming tools are used to demonstrate the working on the algorithm on two examples of nonlinear networks. Future work will investigate the applicability of this method to larger systems (interested readers are referred to the recent work in [@Ahmadi:2017]), and to real-world problems such as coherency detection in power systems. Acknowledgment {#acknowledgment .unnumbered} ============== This work was supported by the United States Department of Energy (contract DE-AC02-76RL01830) under the Control of Complex Systems Initiative at Pacific Northwest National Laboratory. [^1]: $A=[a_{ij}]$ is strictly diagonally dominant if $\sum_{j\neq i}\left|a_{ij}\right|<\left|a_{ii}\right|,\forall i$.

ArXiv

--- abstract: 'The generation of an undular bore in the vicinity of a wave-breaking point is considered for the integrable Kaup-Boussinesq shallow water system. In the framework of the Whitham modulation theory, an analytic solution of the Gurevich-Pitaevskii type of problem for a generic “cubic" breaking regime is obtained using a generalized hodograph transform, and a further reduction to a linear Euler-Poisson equation. The motion of the undular bore edges is investigated in detail.' author: - | G.A. El$^{\dagger}$, R.H.J. Grimshaw$^{\ddagger}$, and A.M. Kamchatnov$^{\S}$\ $^{\dagger}$ School of Mathematical and Informational Sciences,\ Coventry University, Priory Street, Coventry CV1 5FB, UK\ $^{\ddagger}$ Department of Mathematical Sciences, Loughborough University,\ Loughborough LE11 3T, UK\ $^{\S}$ Institute of Spectroscopy, Russian Academy of Sciences,\ Troitsk, Moscow Region, 142190 Russia title: 'Wave breaking and the generation of undular bores in an integrable shallow-water system' --- Introduction ============ It is well known that the solutions to an initial value problem for the non-dispersive shallow water equations may lead to wave-breaking after a finite time, when the first spatial derivatives blow up. After the wave breaking point, a formal solution becomes multi-valued and loses its physical meaning. The divergence of the spatial derivatives at the wave-breaking point suggests that dispersion effects described by terms with higher order spatial derivatives must be taken into account. Then these small dispersion effects lead to the onset of oscillations in the vicinity of the wave breaking point followed by the development of an undular bore, or in different terminology, a dissipationless shock wave. This physical picture has been put into mathematical form for waves described by the Korteweg-de Vries (KdV) equation by Gurevich and Pitaevskii (GP) in [@GP1] (see also Whitham in [@whitham2]). In the GP formulation, the region of oscillations is presented as a slowly modulated periodic wave solution of the KdV equation. The parameters of the wave change little on typical wavelength/period scales which permits one to apply the Whitham modulation theory ([@whitham1; @whitham2]). The resulting Whitham equations describe the slow evolution of the parameters in the undular bore. In the original paper [@GP1], two typical problems were considered. One problem is concerned with the description of decay of an initial discontinuity for the KdV equation and the exact $x/t$-similarity solution of this problem was constructed. Another problem corresponds to the (universal) initial stage of development of a bore when the solution of the dispersionless equation can be approximated locally by a properly chosen cubic curve. This problem was studied in [@GP1] numerically. An exact analytic solution to this problem was later obtained by Potëmin [@Potemin] using Krichever’s algebro-geometrical procedure for integration of the Whitham equations [@Kr88], [@DN]. Later, Pot[ë]{}min’s solution was put into the general context of Tsarev’s generalized hodograph transform [@Tsarev1] in [@KS; @GKE; @T], where the hodograph equations were reduced to the classical Euler-Poisson equation. However, the KdV equation describes unidirectional propagation of nonlinear dispersive waves. An integrable bi-directional analog of the KdV equation was derived by Kaup in [@kaup] using the Boussinesq approximation for shallow water waves [@whitham2]. Like the KdV equation, the Kaup-Boussinesq (KB) system is completely integrable and therefore a powerful inverse scattering transform method can be applied to its investigation. In particular, the multi-phase periodic solutions of the KB system were found in [@MY], the Whitham theory of modulations was applied in [@EGP] to the problem of the decay of an initial discontinuity, and a quasiclassical description of soliton trains arising from a large initial pulse was developed in [@KKU]. In this paper, we further extend the Gurevich-Pitaevskii theory to the case of bi-directional shallow water equations using the KB system, and construct an analytic solution to the Whitham-KB equations for the regime of generation of an undular bore in the vicinity of a breaking point. The obtained solution, along with its own significance in the representation of undular bores, will serve as an intermediate asymptotic in a more general formulation we are presently undertaking, in which small dissipation is taken into account. Periodic waves in the Kaup-Boussinesq system ============================================ In dimensionless units (see, e.g. [@EGP]) the KB system can be written in the form $$\begin{array}{l} h_{t}+(hu)_{x}+ \frac{1}{4}u_{xxx}=0, \\ u_{t}+uu_{x}+h_{x}=0, \end{array} \label{eq1}$$ where $h(x,t)$ denotes the height of the water surface above a horizontal bottom and $u(x,t)$ is related to the horizontal velocity field (at the leading order it is the depth-averaged horizontal field). The KB system (\[eq1\]) is completely integrable and can be represented as the compatibility condition of two linear equations [@kaup] $$\psi _{xx}=\mathcal{A} \psi , \qquad \psi _{t}=-\frac12\mathcal{B}_x\psi+\mathcal{B}\psi_x \label{eq2}$$ with $$\label{eq3} \mathcal{A}=\left({\lambda}-\frac12 u\right)^2-h,\qquad \mathcal{B}=-\left({\lambda}+\frac12 u\right).$$ Thus, the inverse scattering transform method can be applied for its investigation. In particular, the periodic solution of (\[eq1\]) can be obtained by the well-known finite-gap integration method (see, e.g. [@kamch2000]) in the following way. Let $\psi _{+}$ and $\psi _{-}$ be two basis solutions of the second order spatial linear differential equation (\[eq2\]). Then their product $$g=\psi _{+}\psi _{-} \label{eq4}$$ satisfies the third order equation $$g_{xxx}-2\mathcal{A}_xg-4\mathcal{A}g_x=0. \label{eq5}$$ Upon multiplication by $g$, this equation can be integrated once to give $$\frac12 gg_{xx}-\frac14 g_x^2-\mathcal{A}g^2=-P({\lambda}) \label{eq6}$$ where the integration constant $P(\lambda )$ can only depend on $\lambda $. The time dependence of $g(x,t)$ is determined by the equation $$g_{t}=\mathcal{B}g_x-\mathcal{B}_xg. \label{eq7}$$ This equation can readily be put in the form $$\left( \frac{1}{g}\right)_{t}=\left( \frac{\mathcal{B}}{g}\right)_{x}, \label{eq8}$$ which can in turn be considered as a generating function of an infinite sequence of conservation laws. The periodic solutions of the system (\[eq1\]) are distinguished by the condition that $P(\lambda )$ in (\[eq6\]) be a polynomial in $\lambda $. The one-phase periodic solution, which we are interested in, corresponds to the fourth degree polynomial $$P(\lambda )=\prod_{i=1}^{4}(\lambda -\lambda _{i})=\lambda ^{4}-s_{1}\lambda ^{3}+s_{2}\lambda ^{2}-s_{3}\lambda +s_{4}. \label{eq9}$$ Then we find from Eq. (\[eq6\]) that $g(x,t)$ is the first-degree polynomial, $$g(x,t)=\lambda -\mu (x,t), \label{eq10}$$ where $\mu (x,t)$ is connected with $u(x,t)$ and $h(x,t)$ by the relations $$u(x,t)=s_{1}-2\mu (x,t),\quad h(x,t)=\tfrac{1}{4}s_{1}^{2}-s_{2}-2\mu ^{2}+s_{1}\mu , \label{eq11}$$ which in turn follow from a comparison of the coefficients of $\lambda ^{i}$ on both sides of Eq. (\[eq6\]). The spectral parameter $\lambda $ is arbitrary and on substitution of $\lambda =\mu $ into Eq. (\[eq6\]) we obtain an equation for $\mu $, $$\mu _{x}=2\sqrt{ P(\mu )},$$ while a similar substitution into Eq. (\[eq7\]) gives $$\mu _{t}=-(\mu +\tfrac{1}{2}u)\mu _{x}=-\tfrac{1}{2}s_{1}\mu _{x}.$$ Hence, $\mu (x,t)$ as well as $u(x,t)$ and $h(x,t)$ depend only on the phase $$\theta =x-\tfrac{1}{2}s_{1}t, \label{eq12}$$ $$\label{eq12a} \hbox{so that} \quad V=\frac12 s_1=\frac12\sum_{i=1}^4{\lambda}_i$$ is the phase velocity of the nonlinear wave, and $\mu (\theta )$ is determined by the equation $$\mu _{\theta }=2\sqrt{ P(\mu )}. \label{eq13}$$ For the fourth degree polynomial (\[eq9\]) the solution of this equation is readily expressed in terms of elliptic functions. Let the zeros $\lambda _{i} $, $i=1,2,3,4,$ of the polynomial $P(\lambda )$ be real and ordered according to the rule $$\lambda _{1}\leq\lambda _{2}\leq\lambda _{3}\leq\lambda _{4}. \label{eq14}$$ Then the real variable $\mu $ oscillates in the interval where the expression under the square root in (\[eq13\]) is positive, $$\label{eq15} {\lambda}_2\leq\mu\leq{\lambda}_3.$$ Consequently the solution of Eq. (\[eq13\]) with the initial condition $\mu(0)={\lambda}_3$ is given by $$\label{eq16} \mu(\theta)=\frac{{\lambda}_3({\lambda}_4-{\lambda}_2)-{\lambda}_4({\lambda}_3-{\lambda}_2){\mathrm{sn}}^2 \left(\sqrt{({\lambda}_4-{\lambda}_2)({\lambda}_3-{\lambda}_1)}\,\theta,m\right)} {{\lambda}_4-{\lambda}_2-({\lambda}_3-{\lambda}_2){\mathrm{sn}}^2 \left(\sqrt{({\lambda}_4-{\lambda}_2)({\lambda}_3-{\lambda}_1)}\,\theta, m \right)},$$ where $$\label{eq17} m=\frac{({\lambda}_3-{\lambda}_2)({\lambda}_4-{\lambda}_1)}{({\lambda}_4-{\lambda}_2)({\lambda}_3-{\lambda}_1)}$$ is the modulus of the elliptic functions. An equivalent solution corresponding to the initial condition $\mu(0)={\lambda}_2$ is given by $$\label{eq18} \mu(\theta)=\frac{{\lambda}_2({\lambda}_3-{\lambda}_1)-{\lambda}_1({\lambda}_3-{\lambda}_2){\mathrm{sn}}^2 \left(\sqrt{({\lambda}_4-{\lambda}_2)({\lambda}_3-{\lambda}_1)}\,\theta,m\right)} {{\lambda}_3-{\lambda}_1-({\lambda}_3-{\lambda}_2){\mathrm{sn}}^2 \left(\sqrt{({\lambda}_4-{\lambda}_2)({\lambda}_3-{\lambda}_1)}\,\theta,m\right)}.$$ Substitution of (\[eq16\]) or (\[eq18\]) into (\[eq11\]) gives expressions for $u(\theta)$ and $h(\theta)$ in the periodic nonlinear wave. Its wavelength is given by $$\label{eq19} L=\int_{{\lambda}_2}^{{\lambda}_3}\frac{d\mu}{\sqrt{P(\mu)}}= \frac{2{\mathrm{K}}(m)}{\sqrt{({\lambda}_4-{\lambda}_2)({\lambda}_3-{\lambda}_1)}},$$ ${\mathrm{K}}(m)$ being the complete elliptic integral of the first kind. The soliton limit $(m=1)$ is obtained either for ${\lambda}_1={\lambda}_2$ or for ${\lambda}_3={\lambda}_4$. For ${\lambda}_1={\lambda}_2$ Eq. (\[eq16\]) gives $$\label{eq20} \mu(\theta)={\lambda}_4-\frac{({\lambda}_4-{\lambda}_1)({\lambda}_4-{\lambda}_3)} {{\lambda}_4-{\lambda}_1+({\lambda}_3-{\lambda}_1)/\cosh^2[\sqrt{({\lambda}_4-{\lambda}_1)({\lambda}_3-{\lambda}_1)}\,\theta]},$$ and for ${\lambda}_3={\lambda}_4$ Eq. (\[eq18\]) gives $$\label{eq21} \mu(\theta)={\lambda}_1+\frac{({\lambda}_2-{\lambda}_1)({\lambda}_4-{\lambda}_1)} {{\lambda}_2-{\lambda}_1+({\lambda}_4-{\lambda}_2)/\cosh^2[\sqrt{({\lambda}_4-{\lambda}_2)({\lambda}_4-{\lambda}_1)}\,\theta]}.$$ Their substitution into (\[eq11\]) yields the soliton solution of the KB system. In the opposite limit ${\lambda}_2 = {\lambda}_3$ $(m=0)$ both expressions (\[eq16\]) and (\[eq18\]) reduce to $$\label{eq22} \mu={\lambda}_2={\lambda}_3\,.$$ Thus the limit ${\lambda}_2 \to {\lambda}_3 $ yields sinusoidal waves. Whitham modulation equations for the Kaup-Boussinesq system =========================================================== The Whitham modulation equations describe the slow evolution of the parameters ${\lambda}_i,\,i=1,2,3,4,$ of a modulated nonlinear wave. They are $$\label{eq23} \frac{{\partial}{\lambda}_i}{{\partial}t}+v_i({\lambda})\frac{{\partial}{\lambda}_i}{{\partial}x}=0, \quad i=1,2,3,4,$$ where the Whitham velocities $v_i({\lambda})$ can be expressed in the form $$\label{eq24} v_i({\lambda})=\left(1-\frac{L}{{\partial}_iL}{\partial}_i\right)V,\quad {\partial}_i\equiv\frac{{\partial}}{{\partial}{\lambda}_i},\quad i=1,2,3,4,$$ where the phase velocity $V$ and the wavelength $L$ are given correspondingly by (\[eq12a\]) and (\[eq19\]). A simple calculation yields the explicit expressions [@EGP], $$\label{eq25} \begin{array}{l}\displaystyle{ v_1=\frac12\sum{\lambda}_i-\frac{({\lambda}_4-{\lambda}_1)({\lambda}_2-{\lambda}_1){\mathrm{K}}(m)} {({\lambda}_2-{\lambda}_1){\mathrm{K}}(m)+({\lambda}_4-{\lambda}_2){\mathrm{E}}(m)},}\\\displaystyle{ v_2=\frac12\sum{\lambda}_i-\frac{({\lambda}_3-{\lambda}_1)({\lambda}_2-{\lambda}_1){\mathrm{K}}(m)} {({\lambda}_2-{\lambda}_1){\mathrm{K}}(m)-({\lambda}_3-{\lambda}_1){\mathrm{E}}(m)},}\\\displaystyle{ v_3=\frac12\sum{\lambda}_i+\frac{({\lambda}_4-{\lambda}_3)({\lambda}_3-{\lambda}_2){\mathrm{K}}(m)} {({\lambda}_4-{\lambda}_3){\mathrm{K}}(m)-({\lambda}_4-{\lambda}_2){\mathrm{E}}(m)},}\\\displaystyle{ v_4=\frac12\sum{\lambda}_i+\frac{({\lambda}_4-{\lambda}_3)({\lambda}_4-{\lambda}_1){\mathrm{K}}(m)} {({\lambda}_4-{\lambda}_3){\mathrm{K}}(m)+({\lambda}_3-{\lambda}_1){\mathrm{E}}(m)},} \end{array}$$ where ${\mathrm{K}}(m)$ and ${\mathrm{E}}(m)$ are complete elliptic integrals of the first and second kind, respectively. In the limit ${\lambda}_1={\lambda}_2$ $(m=1)$ the Whitham velocities reduce to $$\label{eq26} v_1=v_2=\frac12\sum{\lambda}_i,\quad v_3=\frac12(3{\lambda}_3+{\lambda}_4), \quad v_4=\frac12({\lambda}_3+3{\lambda}_4);$$ in the limit ${\lambda}_3={\lambda}_4$ $(m=1)$ they reduce to $$\label{eq27} v_1=\frac12(3{\lambda}_1+{\lambda}_2),\quad v_2=\frac12({\lambda}_1+3{\lambda}_2), \quad v_3=v_4=\frac12\sum{\lambda}_i;$$ and in the limit ${\lambda}_2={\lambda}_3$ $(m=0)$ they reduce to $$\label{eq28} v_1=\frac12(3{\lambda}_1+{\lambda}_4),\quad v_2=v_3=\frac12\sum{\lambda}_i+ \frac{2({\lambda}_2-{\lambda}_1)({\lambda}_4-{\lambda}_2)}{{\lambda}_4+{\lambda}_1-2{\lambda}_2}, \quad v_4=\frac12({\lambda}_1+3{\lambda}_4).$$ Next we shall apply the Whitham theory to the description of the undular bore forming in the vicinity of a wave-breaking singularity. The Gurevich-Pitaevskii problem for the KB system ================================================= Wave breaking in the dispersionless limit ----------------------------------------- In the dispersionless limit, the KB system (\[eq1\]) reduces to well-known shallow water equations $$\label{eq30} h_t+(hu)_x=0,\qquad u_t+uu_x+h_x=0,$$ which can be transformed to the diagonal form $$\label{eq31} \frac{{\partial}{\lambda}_+}{{\partial}t}+\frac12(3{\lambda}_++{\lambda}_-)\frac{{\partial}{\lambda}_+}{{\partial}x}=0, \quad \frac{{\partial}{\lambda}_-}{{\partial}t}+\frac12({\lambda}_++3{\lambda}_-)\frac{{\partial}{\lambda}_-}{{\partial}x}=0,$$ $$\label{eq32} \hbox{where} \quad {\lambda}_{\pm}=\frac{u}2\pm\sqrt{h}$$ are the Riemann invariants of Eqs. (\[eq30\]). Initial data are given by two functions ${\lambda}_+(x,0)$ and ${\lambda}_-(x,0)$ determined by the initial distributions $h_0(x)$ and $u_0(x)$. The system (\[eq31\]) has two families of characteristics in the $(x,t)$ plane along which one of two Riemann invariants (either ${\lambda}_+$ or ${\lambda}_-$) is constant. The wave-breaking point corresponds to the moment when characteristics of one of the families begin to intersect, so that the corresponding Riemann invariant becomes a three-valued function in the physical plane. Let such an intersection occur for the characteristics transferring the values of ${\lambda}_+$. Then at the wave-breaking point the profile of ${\lambda}_+$ as a function of $x$ has a vertical tangent line and, hence, in vicinity of this point it varies very fast, whereas the second Riemann invariant varies with $x$ more slowly and may be considered here as a constant parameter: $$\label{eq33} {\lambda}_-={\lambda}_0={\rm const}.$$ Thus, in the vicinity of the breaking point we are dealing with a simple wave. The second equation in (\[eq31\]) is identically satisfied by Eq. (\[eq33\]). The first equation in (\[eq31\]) then has the well-known solution $$x-\frac12(3{\lambda}_++{\lambda}_-)t=f({\lambda}_+),$$ where $f({\lambda}_+)$ is an inverse function to an initial profile ${\lambda}_+(x,0)$. At the wave-breaking time, normalized here to be $t=0$, the function $x=f({\lambda}_+)$ must have an inflexion point with a vertical tangent line. Then in the vicinity of this point $f({\lambda}_+)$ can be approximated by a cubic function, $$\label{eq35} x-\frac12(3{\lambda}_++{\lambda}_-)t=-C({\lambda}_+- \overline{{\lambda}}_1)^3.$$ The KB system (\[eq1\]) is invariant with respect to Galilean transformation $$\label{eq36} x'=x-u_0t,\quad t'=t,\quad h=h',\quad u=u'+u_0,\quad {\lambda}_{\pm}={\lambda}_{\pm}'+\frac{u_0}2,$$ and scaling transformation $$\label{eq37} x=ax',\quad t=a^2t',\quad h=h'/a^2,\quad u=u'/a,\quad {\lambda}_{\pm}={\lambda}_{\pm}'/a.$$ With the aid of these transformations Eq. (\[eq35\]) can be cast into the form $$\label{eq38} x-\frac12(3{\lambda}_++{\lambda}_-)t=-{\lambda}_+^3,\quad {\lambda}_-=\lambda_0 \, .$$ where we have omitted the prime superscripts for notational convenience. It corresponds to the wave breaking picture shown in Fig. 1. ![Wave breaking of the water elevation in the dispersionless limit; ${\lambda}_-$ is taken equal to -10. []{data-label="figone"}](fig1.eps){width="8cm" height="5cm"} The actual solution of the KB system now consists of two parts. Following Gurevich and Pitaevskii [@GP1], we suppose that the region of oscillations can be approximated by a modulated periodic solution of the KB system. Its evolution is determined by the Whitham equations (\[eq23\]) and we have to find that solution which matches the solution (\[eq38\]) at the end points of the oscillatory region. One may say that this oscillatory region (the undular bore) “replaces” a non-physical multi-valued region of the solution (\[eq38\]). One should emphasize, however, that the boundaries of the undular bore [*do not coincide*]{} with the boundaries of formal multi-valued solution. Outside these boundaries, the solution approaches the dispersionless solution (\[eq38\]). Undular bore solution --------------------- We look for the solution of the Whitham equations (\[eq23\]) in the form $$\label{eq39} x-v_i({\lambda})t=w_i({\lambda}),\quad i=1,2,3,4,$$ where $v_i$ are the Whitham velocities (\[eq25\]). Since we consider the breaking of the Riemann invariant ${\lambda}_+$ and ${\lambda}_-<{\lambda}_+$, we take $$\label{eq40} {\lambda}_1={\lambda}_-={\lambda}_0=\mathrm{const}.$$ Then the limiting formulas (\[eq26\])-(\[eq28\]) show that if we find $w_i({\lambda})$ such that $$\label{eq41} \begin{split} w_4=-{\lambda}_4^3\quad \mathrm{at}\quad {\lambda}_2={\lambda}_3\quad (m=0),\\ w_2=-{\lambda}_2^3\quad \mathrm{at}\quad {\lambda}_4={\lambda}_3\quad (m=1), \end{split}$$ then Eqs. (\[eq39\]) determine the Riemann invariants in such a way that ${\lambda}_4={\lambda}_+$ at the trailing edge $x^-(t)$ where $m=0$, ${\lambda}_2={\lambda}_+$ at the leading edge $x^+(t)$ where $m=1$, and ${\lambda}_1={\lambda}_-={\lambda}_0$ everywhere. Thus, the plots of Riemann invariants ${\lambda}_2,{\lambda}_3,{\lambda}_4$ as functions of $x$ are joined into continuous curve whose upper and lower branches match with the solution (\[eq38\]) of the dispersionless equations (see Fig. 2). In the region $x^-(t)<x<x^+(t)$ there are four Riemann invariants which determine the modulated periodic solution representing the undular bore. At its trailing edge $x\to x^-(t)$ the amplitude of oscillations vanishes and at the leading edge $x\to x^+(t)$ the periodic solution transforms into a soliton train. ![Dependence of Riemann invariants ${\lambda}_2,{\lambda}_3,{\lambda}_4$ on $x$ at fixed value of time $t=1$ and with ${\lambda}_1=-10$. The dashed line shows the corresponding dependence of ${\lambda}_+$ for the solution of the KB equations in the dispersionless limit. []{data-label="figtwo"}](fig2.eps){width="8cm" height="5cm"} According to the generalized hodograph method [@Tsarev1; @Tsarev2], Eqs. (\[eq39\]) satisfy the Whitham equations (\[eq23\]) provided $w_i({\lambda})$ are the velocities of the flows $$\label{eq42} \frac{{\partial}{\lambda}_i}{{\partial}\tau}+w_i({\lambda})\frac{{\partial}{\lambda}_i}{{\partial}x}=0, \quad i=1,2,3,4,$$ commuting with (\[eq23\]), i.e $\partial_{t \tau}{\lambda}_i= \partial_{ \tau t}{\lambda}_i$. If we represent $w_i({\lambda})$ in the form analogous to Eqs. (\[eq24\]), $$\label{eq43} w_i({\lambda})=\left(1-\frac{L}{\partial_iL}\partial_i\right)W,\quad i=1,2,3,4,$$ then the condition of commutativity of the flows (\[eq23\]) and (\[eq43\]) reduces to the system of Euler-Poisson equations, exactly as happens in the KdV [@KS; @GKE; @T] and NLS [@EK95] cases, $$\label{eq44} \partial_i\partial_jW-\frac1{2({\lambda}_i-{\lambda}_j)}(\partial_iW-\partial_jW)=0,\quad i\neq j.$$ It is easy to check that this equation has a particular solution $W={\mathrm{const}}/\sqrt{P({\lambda})},$ $P({\lambda})=\prod({\lambda}-{\lambda}_i),$ which is sufficient for our purpose. We choose the normalization factor so that the coefficient before ${\lambda}^{-1}$ in the series expansion of $W$ in powers of ${\lambda}^{-1}$ be equal to the phase velocity of the periodic wave $s_1/2=V$. Thus, we obtain the sequence of $W^{(k)}$ defined by the generating function $$\label{eq45} W=\frac{{\lambda}^2}{\sqrt{P({\lambda})}}=\sum\frac{W^{(k)}}{{\lambda}^k}= 1+\tfrac12{s_1}\cdot\frac1{{\lambda}}+ \left(\tfrac38s_1^2-\tfrac12s_2\right)\cdot\frac1{{\lambda}^2} +\left(\tfrac5{16}s_1^3-\tfrac34s_1s_2+\tfrac12s_3\right) \cdot\frac1{{\lambda}^3}+\ldots.$$ Next a sequence of velocities of the commuting flows is given by $$\label{eq46} w_i^{(k)}({\lambda})=\left(1-\frac{L}{\partial_iL}\partial_i\right)W^{(k)},\quad i=1,2,3,4,$$ where $w_i^{(1)}=v_i$ coincide with the Riemann velocities (\[eq25\]). It is not difficult to find the limiting formulas analogous to (\[eq26\])-(\[eq28\]). In particular, we get at ${\lambda}_3={\lambda}_4$ $(m=1)$ $$\label{eq47} \begin{split} &\left.w_1^{(1)}\right|_{{\lambda}_3={\lambda}_4}=\left.w_2^{(1)}\right|_{{\lambda}_3={\lambda}_4} =\frac12({\lambda}_1+3{\lambda}_2), \\ &\left.w_1^{(2)}\right|_{{\lambda}_3={\lambda}_4}=\left.w_2^{(2)}\right|_{{\lambda}_3={\lambda}_4} =\frac38({\lambda}_1^2+2{\lambda}_1{\lambda}_2+5{\lambda}_2^2), \\ &\left.w_1^{(3)}\right|_{{\lambda}_3={\lambda}_4}=\left.w_2^{(3)}\right|_{{\lambda}_3={\lambda}_4} =\frac1{16}(5{\lambda}_1^3+9{\lambda}_1^2{\lambda}_2+15{\lambda}_1{\lambda}_2^3+35{\lambda}_2^3). \end{split}$$ Now we take such the linear combination $$w_2=a_0+a_1w_2^{(1)}+a_2w_2^{(2)}+a_3w_2^{(3)}$$ so that $w_2$ satisfies the condition (\[eq41\]). The coefficients $a_1,a_2,a_3,a_4$ depend on the constant Riemann invariant ${\lambda}_1={\lambda}_-={\lambda}_0$ and their values found in this way yield the required solution of the Whitham equations: $$\label{eq48} \begin{split} x-v_i({\lambda})t&=-\tfrac{16}{35}w_i^{(3)}({\lambda})+\tfrac8{35}{{\lambda}_0}w_i^{(2)}+ \tfrac2{35}{{\lambda}_0}^2v_i({\lambda})+\tfrac1{35}{{\lambda}_0}^3,\quad i=2,3,4;\\ {\lambda}_1&={{\lambda}_0}={\rm const}. \end{split}$$ These formulas define ${\lambda}_2,{\lambda}_3,{\lambda}_4$ implicitly as functions of $x$ and $t$ and give the solution of the Gurevich-Pitaevskii problem for the KB-Whitham system. It is interesting to note that, unlike the KdV case, this solution is not scale-invariant (we recall that in the counterpart KdV solution ${\lambda}_i=t^{-1/2}l_i(x/t^{3/2}), i=1,2,3$). This happens due to the presence of the fourth Riemann invariant $\lambda_1$ in the Whitham equations which is constant ($\lambda_0$) for the obtained solution and cannot be eliminated by simple Galilean transform. As a result, the solution (\[eq48\]) does not possess the scaling invariance required for a generalized similarity behaviour of $\lambda_j\, , \ j=2,3,4$. The only family of admissible such similarity ($x/t$) solutions is realized in the simplest case of the decay of an initial discontinuity studied in [@EGP]. Laws of motion at the trailing and leading edges of the oscillatory region -------------------------------------------------------------------------- Let us find the laws of motion at the leading and trailing edges of the undular bore. First we consider the leading edge ${\lambda}_3={\lambda}_4$ $(m=1)$, and define the small deviations ${\lambda}_3'$ and ${\lambda}_4'$ from the value ${\lambda}_3={\lambda}_4={\lambda}_4^+$: $$\label{eq49a} {\lambda}_3={\lambda}_4^++{\lambda}_3',\qquad {\lambda}_4={\lambda}_4^++{\lambda}_4' \,.$$ Then we seek the asymptotic expressions for formulas (\[eq48\]) with $i=3,4$ for small $|{\lambda}_3'|,|{\lambda}_4'|$: $$\label{eq49} x^++x'-(v_3^++v_3')t=w_3^++w_3',\quad x^++x'-(v_4^++v_4')t=w_4^++w_4',$$ where $x'$ denotes the space coordinate reckoned from its limiting value $x^+$ and $$\label{eq50} v_3^+=v_4^+=\tfrac12({\lambda}_1+{\lambda}_2+2{\lambda}_4),$$ $$\label{eq51} v_3'=-v_4'=-\frac12\left\{{\lambda}_3'\ln\left[\frac{-({\lambda}_2-{\lambda}_1){\lambda}_3'} {({\lambda}_4-{\lambda}_1)({\lambda}_4-{\lambda}_2)}\right]- {\lambda}_4'\ln\left[\frac{({\lambda}_2-{\lambda}_1){\lambda}_4'} {({\lambda}_4-{\lambda}_1)({\lambda}_4-{\lambda}_2)}\right]\right\},$$ $$\label{eq52} w_3^+=w_4^+=-\frac1{35}(5{\lambda}_2^3+6{\lambda}_2^2{\lambda}_4+8{\lambda}_2{\lambda}_4^2+16{\lambda}_4^2),$$ $$\label{eq53} \begin{split} w_3'=-w_4'=&\frac1{35}(3{\lambda}_2^2+8{\lambda}_2{\lambda}_4+24{\lambda}_4^2)\times\\ &\left\{{\lambda}_3'\ln\left[\frac{-({\lambda}_2-{\lambda}_1){\lambda}_3'} {({\lambda}_4-{\lambda}_1)({\lambda}_4-{\lambda}_2)}\right]- {\lambda}_4'\ln\left[\frac{({\lambda}_2-{\lambda}_1){\lambda}_4'} {({\lambda}_4-{\lambda}_1)({\lambda}_4-{\lambda}_2)}\right]\right\}, \end{split}$$ while ${\lambda}_1,{\lambda}_2,{\lambda}_4$ denote here their limiting values ${\lambda}_1^+,{\lambda}_2^+,{\lambda}_4^+$, correspondingly. Then subtraction of one equation (\[eq49\]) from the other yields at once the relationship $$\label{eq54} t=\frac2{35}(3{\lambda}_2^2+8{\lambda}_2{\lambda}_4+24{\lambda}_4^2).$$ On the other hand, the limiting formulas $$x^+-v_2^+t=w_2^+,\quad x^+-v_3^+t=w_3^+,$$ with $$\label{eq55} v_2^+=\tfrac12({\lambda}_1+3{\lambda}_2),\qquad w_2^+=-{\lambda}_2^3$$ and $v_3^+$, $w_3^+$ given by (\[eq50\]) and (\[eq52\]) give after subtraction the relationship $$\label{eq56} t=\frac2{35}(15{\lambda}_2^2+12{\lambda}_2{\lambda}_4+8{\lambda}_4^2).$$ Then equating of the right hand sides of (\[eq54\]) and (\[eq56\]) shows that at the leading edge we have $$\label{eq57} {\lambda}_4^+=-\frac34{\lambda}_2^+.$$ Substitution of this relation into (\[eq54\]) or (\[eq56\]) gives the dependence of ${\lambda}_2^+$ on $t$: $$\label{eq58} {\lambda}_2^+=-\left(\frac{5t}3\right)^{1/2}.$$ At last, the formula $x^+-v_2^+t=-({\lambda}_2^+)^3$ yields the law of motion for the leading edge: $$\label{eq59} x^+(t)=\frac12{\lambda}_0t+\frac16\sqrt{\frac53}\,t^{3/2}.$$ In a similar way we can consider the trailing edge ${\lambda}_3={\lambda}_2={\lambda}_2^-$ $(m=0)$ where we define $$\label{eq60} {\lambda}_2={\lambda}_2^-+{\lambda}_2',\qquad {\lambda}_3={\lambda}_2^-+{\lambda}_3'$$ and Eqs. (\[eq48\]) with $i=2,3$ reduce to $$\label{eq61} x^-+x'-(v_2^++v_2')t=w_2^++w_2',\quad x^-+x'-(v_3^++v_3')t=w_3^++w_3',$$ where $$\label{eq62} v_2^-=v_3^-=\frac{{\lambda}_1^2+4{\lambda}_1{\lambda}_2-8{\lambda}_2^2-2{\lambda}_1{\lambda}_4+4{\lambda}_2{\lambda}_4+{\lambda}_4^2} {2({\lambda}_1-2{\lambda}_2+{\lambda}_4)},$$ $$\label{eq63} v_2'=-v_3'=-\frac{{\lambda}_2'+3{\lambda}_3'}{2({\lambda}_1-2{\lambda}_2+{\lambda}_4)} (3{\lambda}_1^2-8{\lambda}_1{\lambda}_2+8{\lambda}_2^2+2{\lambda}_1{\lambda}_4-8{\lambda}_2{\lambda}_4+3{\lambda}_4^2),$$ $$\label{eq64} \begin{split} w_2^-=w_3^-=&[128{\lambda}_2^4-64{\lambda}_2^3{\lambda}_4-16{\lambda}_2^2{\lambda}_4^2-8{\lambda}_2{\lambda}_4^3-5{\lambda}_4^4\\ &-7{\lambda}_1(16{\lambda}_2^3-8{\lambda}_2^2{\lambda}_4-2{\lambda}_2{\lambda}_4^2-{\lambda}_4^3)]/(35({\lambda}_1-2{\lambda}_2+{\lambda}_4)) \end{split}$$ $$\label{eq65} \begin{split} w_2'=-w_3'=\frac{{\lambda}_2'+3{\lambda}_3'}{70({\lambda}_1-2{\lambda}_2+{\lambda}_4)^2}[&-384{\lambda}_2^4 +384{\lambda}_2^3{\lambda}_4-80{\lambda}_2^2{\lambda}_4^2-16{\lambda}_2{\lambda}_4^3-9{\lambda}_4^4\\ &+7{\lambda}_1^2(-24{\lambda}_2^2+8{\lambda}_2{\lambda}_4+{\lambda}_4^2)]. \end{split}$$ Then subtraction of one equation (\[eq61\]) from the other gives $$\label{eq66} \begin{split} t=\tfrac2{35}[&-384{\lambda}_2^4+384{\lambda}_2^3{\lambda}_4-80{\lambda}_2^2{\lambda}_4^2-16{\lambda}_2{\lambda}_4^3-9{\lambda}_4^4\\ &+7{\lambda}_1^2(-24{\lambda}_2^2+8{\lambda}_2{\lambda}_4+{\lambda}_4^2)]/(3{\lambda}_1^2-8{\lambda}_1{\lambda}_2+8{\lambda}_2^2 +2{\lambda}_1{\lambda}_4-8{\lambda}_2{\lambda}_4+3{\lambda}_4^2). \end{split}$$ Now the limiting formulas $$\label{eq67} x^--\frac12({\lambda}_1+3{\lambda}_4)t=-{\lambda}_4^3,\qquad x^--v_3^-t=-w_3^-$$ give $$\label{eq68} t=\frac2{35}\frac{(8{\lambda}_2-7{\lambda}_1)(8{\lambda}_2^2+4{\lambda}_2{\lambda}_4+3{\lambda}_4^2)-15{\lambda}_4^3} {4{\lambda}_2-3{\lambda}_1-{\lambda}_4}.$$ Equating the right hand sides of (\[eq66\]) and (\[eq68\]), we find the relationship between the values of ${\lambda}_2^-$ and ${\lambda}_4^-$ at the trailing edge: $$\label{eq69} \begin{split} 21{\lambda}_1^2({\lambda}_4+4{\lambda}_2)-10{\lambda}_1(20{\lambda}_2^2+2{\lambda}_2{\lambda}_4-{\lambda}_4^2) +16(8{\lambda}_2^3-{\lambda}_2^2{\lambda}_4-{\lambda}_2{\lambda}_4^2)+9{\lambda}_4^3=0. \end{split}$$ Given $t$ and ${\lambda}_1={\lambda}_0=\mathrm{const}$, we find ${\lambda}_2={\lambda}_2^-$ and ${\lambda}_4={\lambda}_4^-$ from (\[eq68\]) and (\[eq69\]) and then the law of motion of the trailing edge follows from $$\label{eq70} x^-=\frac12({\lambda}_1+3{\lambda}_4)t-{\lambda}_4^3.$$ ![Riemann invariant ${\lambda}_2$ as a function of ${\lambda}_4$ defined implicitly by Eq. (\[eq69\]). The plot corresponds to a fixed value of ${\lambda}_1=-10$. Dashed line shows the dependence according to asymptotic formula (\[eq73\]). []{data-label="figthree-a"}](fig3a.eps){width="8cm" height="5cm"} It is worth noticing that Eq. (\[eq57\]) coincides with the corresponding relation between the Riemann invariants at the leading edge in the solution of the Gurevich-Pitaevskii problem in the KdV equation case (see, e.g. [@kamch2000]). A similar relation ${\lambda}_2^-=-{\lambda}_4^-/4$ between the Riemann invariants of the KdV theory follows from (\[eq69\]) in the limit $|{\lambda}_1|=|{\lambda}_0|\to\infty$. In the next approximation we obtain $$\label{eq71} {\lambda}_2={\lambda}_3\cong-\frac14{\lambda}_4-\frac5{168}\frac{{\lambda}_4^2}{{\lambda}_1},$$ and similar expansion of Eq. (\[eq68\]) in powers of $1/{\lambda}_1$, $$\label{eq72} t\cong \frac13{\lambda}_4^2\left(1+\frac{10}{21}\frac{{\lambda}_4}{{\lambda}_1}\right),$$ yields with the same accuracy $$\label{eq73} {\lambda}_4\cong\sqrt{3t}-\frac57\frac{t}{{\lambda}_1},\qquad \sqrt{3t}\ll|{\lambda}_1|.$$ Dependence of ${\lambda}_2$ on ${\lambda}_4$ given by Eq. (\[eq69\]) with fixed value of ${\lambda}_1$ is illustrated in Fig. \[figthree-a\]. Substitution of Eq. (\[eq72\]) into Eq. (\[eq70\]) yields an approximate expression for the law of motion of the trailing edge; $$\label{eq74} x^-\cong\frac12{\lambda}_0t-\frac{3\sqrt{3}}2 t^{3/2}+\frac{75}{14} \frac{t^2}{{\lambda}_0},\qquad \sqrt{3t}\ll|{\lambda}_0|.$$ Thus, the analytic formulas (\[eq48\]) for the solution of the Whitham equations allowed us to find the main characteristics of the dissipationless shock. With the use of Eqs. (\[eq48\]) we can find ${\lambda}_2,{\lambda}_3,{\lambda}_4$ as functions of $x$ at given $t$, and their substitution into (\[eq16\]) and (\[eq11\]) yields the profiles of $u(x)$ and $h(x)$ in the undular bore. An example of such a profile of the water elevation is shown in Fig. \[figthree\]. ![Undular bore for the KB equations. The plot corresponds to the time $t=1$ and ${\lambda}_1=-10$. Dashed line shows the solution in the dispersionless limit. []{data-label="figthree"}](fig3.eps){width="8cm" height="5cm"} As we see, the non-physical solution obtained in the dispersionless limit is replaced by an undular bore. Its end points move according to the laws found above, so that the oscillatory region expands with time and its width grows mainly as $t^{3/2}$. Small amplitude oscillations are generated at the trailing edge (actually these oscillations represent gravity waves propagating into an undisturbed smooth region), and they transform gradually into solitons at the leading edge. Conclusion ========== In this article, the wave breaking (Gurevich-Pitaevskii) problem is solved for the shallow water waves described by the Kaup-Boussinesq system. This theory generalizes the KdV model to the case of a bi-directional wave propagation model, and so allows for the effects of wave interaction. As a result, the dispersionless theory includes two Riemann invariants and the breaking of the wave means breaking of one of these two invariants. Thus, the wave breaking picture depends on the value of an additional parameter: the value of the non-breaking Riemann invariant. Correspondingly, the Whitham theory for the KB system contains four Riemann invariants (instead of three in the KdV case) and, as a result, the solution of the Whitham equations for the Gurevich-Pitaevskii problem of the wave front breaking is parameterized by a constant value ${\lambda}_0$. The analytic solution of the Whitham equations obtained here allows one to investigate the form of an undular bore as a function of the parameter ${\lambda}_0$. In particular, the velocity of the trailing edge increases with decrease of $|{\lambda}_0|$ and in the limit $|{\lambda}_0|\to\infty$ the present theory reduces to known results of the KdV model. Our obtained solution can also be viewed as an intermediate asymptotic in a more general problem of the description of frictional shallow water undular bores where small dissipation is taken into account (see [@GP2; @AKN; @MG] for the results relevant to the KdV equation with weak dissipation). This problem will be the subject of a separate study. Acknowledgements {#acknowledgements .unnumbered} ---------------- This work was completed during stay of A.M.K. at Department of Mathematical Sciences, Loughborough University, UK. A.M.K. is grateful to the Royal Society for financial support. [99]{} A.V. Gurevich and L.P. Pitaevskii, “Nonstationary structure of a collisionless shock wave”, [*Zh. Eksp. Teor. Fiz.*]{} [**65**]{}, 590 (1973) \[[*Soviet Physics JETP,*]{} [**38**]{}, 291 (1973)\]. G.B. Whitham, “Non-linear dispersive waves”, [*Proc. Roy. Soc. London,*]{} [**283**]{}, 238 (1965). G.B. Whitham, [*Linear and Nonlinear Waves,*]{} Wiley–Interscience, New York, 1974. G.V. Potëmin, “Algebraic-geometrical construction of self-similar solutions of Whitham’s equations”, [*Usp. Mathem. Nauk*]{} [**43**]{}, 211 (1988) \[[*Russian Math. Surveys*]{} [**43**]{}, 252 (1988)\]. I.M. Krichever, The method of averaging for two-dimensional “integrable” equations. *Funct. Anal. Appl.* [**22,**]{} 200-213, (1988). S.P. Tsarev, On Poisson brackets and one-dimensional systems of hydrodynamic type [*Soviet Math. Dokl.* ]{} [**31,**]{} 488-491 (1985). B.A. Dubrovin and S.P. Novikov, Hydrodynamics of weakly deformed soliton lattices. Differential geometry and Hamiltonian theory. *Russian Math. Surveys* [**44,**]{} 35–124 (1989). D.J. Kaup, A higher order water-wave equation and method for solving it, [*Progr. Theor. Phys.*]{} [**54,**]{} 396–408 (1976). V.B. Matveev and M.I. Yavor, Almost periodical solutions of nonlinear hydrodynamic equation of Kaup, [*Ann. Inst. H. Poincaré,*]{} Sect. A [**31,**]{} 25–41 (1979). V.R. Kudashev and S.E. Sharapov, Inheritance of KdV symmetries under Whitham averaging and hydrodynamic symmetries of the Whitham equations. [*Theor. Math. Phys.*]{} [**87,**]{} 358–363 (1991). A.V.Gurevich, A.L. Krylov, and G.A. El, Riemann wave breaking in dispersive hydrodynamics. [*JETP Lett.*]{} [**54**]{} 102-107; Evolution of a Riemann wave in dispersive hydrodynamics. [*Sov. Phys. JETP*]{} [**74,**]{} 957–962 (1991). F.R. Tian, Oscillations of the zero dispersion limit of the Korteweg – de Vries equation. [*Comm. Pure Appl. Math.*]{} [**46,**]{} 1093–1129 (1993). G.A. El, R.H.J. Grimshaw, and M.V. Pavlov, Integrable shallow-water equations and undular bores, [*Stud. Appl. Math.*]{} [**106,**]{} 157–186 (2001). A.M. Kamchatnov, R.A. Kraenkel, and B.A. Umarov, Asymptotic soliton train solutions of Kaup-Boussinesq equations, [*Wave Motion,*]{} [**38,**]{} 355–365 (2003). A.M. Kamchatnov, [*Nonlinear Periodic Waves and Their Modulations—An Introductory Course*]{}, World Scientific, Singapore, 2000. S.P. Tsarev, “The geometry of Hamiltonian systems of hydrodynamic type. The generalized hodograph method,” [*Izv. Akad. Nauk,*]{} [**54**]{}, 1048 (1990) \[[*Math. USSR Izvestia,*]{} [**37**]{}, 397 (1991)\]. G.A. El and A.L. Krylov, General solution of the Cauchy problem for the defocusing NLS equation in the Whitham limit, [*Phys. Lett.*]{} A [**203,**]{} 77–82 (1995). A.V. Gurevich and L.P. Pitaevskii, Averaged description of waves in the Korteweg-de Vries-Burgers equation, [*Zh. Eksp. Teor. Fiz.*]{} [**93**]{}, 871 (1987) \[[*Soviet Physics JETP,*]{} [**66**]{}, 490 (1987)\]. V.V. Avilov, I.M. Krichever, and S.P. Novikov, Evolution of Wthitham zone in the theory of Korteweg-de Vries, [*Dokl. Akad. Nauk SSSR,*]{} [**295,**]{} 345-349 (1987) \[[*Sov. Phys. Dokl.,*]{} [**32,**]{} 564-566 (1987)\]. S. Myint and R.H.J. Grimshaw, The modulation of nonlinear periodic wavetrains by dissipative terms in the korteweg-de Vries equation, [*Wave Motion,*]{} [**22,**]{} 215-238 (1995).

ArXiv

--- abstract: 'Simultaneous localization and mapping (SLAM) has been extensively researched in past years particularly with regard to range-based or visual-based sensors. Instead of deploying dedicated devices that use visual features, it is more pragmatic to exploit the radio features to achieve this task, due to their ubiquitous nature and the widespread deployment of Wi-Fi wireless network. This paper presents a novel approach for collaborative simultaneous localization and radio fingerprint mapping (C-SLAM-RF) in large unknown indoor environments. The proposed system uses received signal strengths (RSS) from Wi-Fi access points (AP) in the existing infrastructure and pedestrian dead reckoning (PDR) from a smart phone, without a prior knowledge about map or distribution of AP in the environment. We claim a loop closure based on the similarity of the two radio fingerprints. To further improve the performance, we incorporate the turning motion and assign a small uncertainty value to a loop closure if a matched turning is identified. The experiment was done in an area of 130 meters by 70 meters and the results show that our proposed system is capable of estimating the tracks of four users with an accuracy of 0.6 meters with Tango-based PDR and 4.76 meters with a step counter-based PDR.' author: - | Ran Liu, Sumudu Hasala Marakkalage, Madhushanka Padmal, Thiruketheeswaran Shaganan,\ Chau Yuen, Yong Liang Guan, and U-Xuan Tan [^1] [^2] [^3] [^4] bibliography: - 'literatur.bib' title: '**Collaborative SLAM based on Wifi Fingerprint Similarity and Motion Information** ' --- Introduction {#Introduction} ============ With growing applications of the Internet of Things (IoT), recent research shows an increasing interest in indoor positioning due to the rapid demand of location-based services, such as indoor guidance and asset tracking [@Yassin_ieee_tutorials_2016; @he2016wi; @wifi_fusion; @localization_framework]. To perform indoor positioning, the knowledge of the existing infrastructure must be provided in advance (for example a map of the environment or locations of the beacons). In the scenario of emergency response in disaster areas or large scale environments, such kind of knowledge is not available or difficult to obtain beforehand, which makes the indoor positioning challenging. Therefore, recent researchers are focusing on developing efficient methods and technologies to simultaneously localize mobile devices (robots and smartphones) and generate a map of the environment [@montemerlo2002fastslam; @ferris2007wifi; @burgard2009comparison; @survey_av_localization]. The underlying problem is well known by the term, Simultaneous Localization and Mapping (SLAM). Extensive researches have been done with regard to visual-based [@Taketomi2017] or range-based sensors [@slam_trends_2016]. Loop closure detection is elementary to any SLAM system. It denotes a situation that the mobile device has entered a previously-visited location, which permits to correct the accumulated odometry error. In order to perform loop closure detection in SLAM, dedicated devices (i.e., laser range finders or cameras) are required to measure the similarity of observations by scan matching [@Lu_millos_1997] or feature matching [@Taketomi2017], which are usually computationally expensive. However, growing popularity of Wi-Fi wireless networks provide a new opportunity to detect loop closure and perform SLAM in a different way. Most existing buildings with Wi-Fi network deployed can be exploited for localization and mapping with low hardware requirements and computational cost (for example with the ubiquitous IoT devices like normal smartphones) due to their ubiquitous nature of in-built sensing capabilities [@Yassin_ieee_tutorials_2016; @ran_ieee_sensors2017]. The current signal-strength-based SLAM requires an analytical model to feature the radio signal distribution over distance [@ferris2007wifi; @Huang_wifi_slam_11]. However, it is not practical to build such a model due to many multiple path issues in uncontrolled environments. On the other hand, radio fingerprinting [@he2016wi; @pervasive_integration_imu_fp; @Robust_neighborhood_graphing], represents a location with a collection of radio signals from Wi-Fi access points, which is considered to be more robust against the signal distortions. Therefore, we adopt this technique to simultaneously determine the location of a user and create a radio map of the environment. In addition to the Wi-Fi network, a typical indoor environment consists of many landmarks, such as turnings, elevators, rooms, and doors, which can be also considered as features for the positioning of a device. These landmarks can be recognized through inertial sensors, which are available in most commercial off-the-shelf smartphones [@turn_detection_trans; @kaiser_elevator_detection_2016]. In contrast to large location uncertainty of radio fingerprints due to the distortion of signals, such kind of landmarks can better confine the location of a device and enhance the positioning accuracy of fingerprinting-based approaches [@Robust_neighborhood_graphing; @wang_motion_slam_iros]. In opposite to the feature map or occupancy map built by laser range finders or visual cameras, our goal is to build a map (in particular a radio map) with radio fingerprint as feature, and use that for the positioning. To ensure a good positioning accuracy in large scale environment, a fine-grained radio map is required [@Yassin_ieee_tutorials_2016; @he2016wi] and it will be time consuming to create such a map with a single user. Therefore, a low cost method (e.g., acquire fingerprints via crowdsensing by multiple users) to create the radio map is a necessity. This paper presents a system that fuses the pedestrian dead reckoning from a smartphone, and received signal strength (RSS) measurements from surrounding Wi-Fi access points (AP), to estimate the trajectory of multiple users and map the radio signals in unknown environment via a collaborative fashion, using graph SLAM technique. To further improve the accuracy, we incorporate the turning features and reduce the uncertainty of loops inferred based on radio fingerprints similarity. The proposed approach requires neither the map of the environment nor the locations of the access points. We tested the system under two different dead reckoning systems, one is based on Tango that has a high motion tracking accuracy through vision-based odometry, and the second one is based on step counter using on-board inertial sensors that has a poor motion tracking accuracy. By leveraging on in-built sensing capabilities from smart phones and crowdsensing nature, our system can generate a radio fingerprint map in a large indoor environment at low cost as compared with traditional site surveying methods. We summarize the contributions of this paper as: - We present a solution that incorporates Wi-Fi fingerprint and dead reckoning information for crowdsensing SLAM in unknown indoor environments; - We propose an algorithm that automatically learns a model to characterize the uncertainty of a loop based on the degree of similarity using the short term odometry measurement; - We integrate the turning features to further reduce the uncertainty of radio fingerprint-based loop closures and improve the overall accuracy; - We throughly evaluate our approach in one building at our campus with an area of approx. 9000 square meters with two different pedestrian dead reckoning systems. We organize the rest of this paper as follows. The related work is discussed in Section \[related\_work\]. Section \[system\_overview\] formulates the problem and explains the detail of the proposed system. Section \[experimental\_evaluations\] presents the experimental results. Conclusions with possible directions of future work are made in Section \[conclusions\]. Related Work {#related_work} ============ Over the past decades, indoor positioning shows a growing popularity due to the increasing demand of location-aware applications [@Yassin_ieee_tutorials_2016; @he2016wi]. A large number of researches have been performed regarding indoor positioning given a reference of the infrastructure (i.e., map of the environment or distributions of beacons). Obtaining and maintaining such kind of information is challenging, particularly in large scale environments [@wifi_positioning_challenge] or emergency response for example search and rescue in disaster scenes [@Liu_Relative_Globecom; @Liu_relative_positioning_icra]. A solution to this problem is SLAM (Simultaneous Localization and Mapping), which has been investigated extensively in robotics community. In this section, we present a summary of the related work in SLAM, using different kinds of techniques. Throughout the years, many techniques and algorithms have been proposed, mainly including filtering-based solutions (for example the Kalman filter [@montemerlo2002fastslam] and the particle filter [@Thrun_Probabilistic_robotics]) and graph-based solutions [@burgard2009comparison; @kuemmerle11icra]. Depending on the types of sensors used, one can classify the SLAM into laser-based SLAM, visual-SLAM, magnetic-SLAM, WifiSLAM, and FootSLAM. Laser-based SLAM uses laser-range finders to create a structural map of an environment. The detection of loop closure is achieved by scan matching. Visual-SLAM methods, utilize cameras like Kinect or Tango [@engelhard2011real] to construct a 3D model of the indoor scene. Bundle adjustment [@Triggs00bundleadjustment] is another popular technique for SLAM that uses visual images and has been used in commercialized SLAM systems such as Google’s Project Tango [@tango]. Magnetic-SLAM systems, exploit digital magnetic compass for localization and mapping of a device [@JUNG_sequence_mag_slam]. The loop closure is inferred by examining spatial similarity of a sequence of magnetic measurements. For example, authors in [@wang_motion_slam_iros] correlated motion patterns with the magnetic field to address the SLAM problem. A unique magnetic fingerprint may not be guaranteed due to the distortion of the environment, which makes this solution challenging for real applications. WifiSLAM [@ferris2007wifi; @Huang_wifi_slam_11] techniques use the radio signal and motion data of the device for localization and signal strength mapping in unknown environments. With SLAM technique, the hassle of site surveying can be avoided, and radio map can be created and updated conveniently whenever needed. For example, authors in [@ferris2007wifi] solved the WifiSLAM problem by mapping the high-dimensional signal strength into a two-dimensional latent space with a Gaussian process. Authors in [@Huang_wifi_slam_11] proposed a generalized and effective algorithm to solve the WifiSLAM using GraphSLAM algorithm. Both approaches assume the signal strength at two close locations are similar and the measurement likelihood can be modeled as a Gaussian process. In contrast, our approach does not require any model to describe the signal strength distribution, and the closeness of locations is determined by the similarity of the radio fingerprints. FootSLAM [@foot_slam_original] uses inertial-based measurements to determine the underlying building structure. No ranging or visual measurement were required; the only features used are the probability distributions of human motions at different locations. Several extensions, for example ActionSlam [@action_slam], incorporate location-orientated actions (for example entering elevators or door opening) as features to compensate for the IMU drifting error. Additional information [@foot_slam] can be further incorporated into FootSLAM, for example a prior map or signal strength from a Wi-Fi access point. Authors in [@floor_plan_generation_jsac] proposed SenseWit that utilizes inertial measurements to generate a floorplan by identifying featured locations (turning, water dispenser, and door) in indoor space. When the indoor environment becomes huge, generating the radio map with single mobile device becomes time consuming. The power of crowd comes into play in this scenario. Mobile crowdsensing is a popular computing paradigm, which enables ubiquitous devices to collect sensing data at large scales [@crowd_facility_mapping; @Smartphones_crowdsourcing; @localisation_database_crowdsourcing]. This technique can be utilized to unleash the potential of mobile phones of people who move inside the indoor environment [@faragher2012opportunistic]. Prior research have harnessed the power of crowdsensing to reconstruct indoor floor plans by combining user mobility traces, images of landmarks, and Wi-Fi fingerprints [@radu2013pazl; @gao2014jigsaw]. Localization by combination of 6-DOF gyro-odometry and Wi-Fi localization has been done in [@jirkuu2016wifi], using multiple robots. Authors in [@cooperative_fingerprint_localization] proposed an approach to utilize pairwise distance measures between users to reduce the positioning error in fingerprinting-based approaches. Our system combines crowdsensed RSS from Wi-Fi APs and dead reckoning information from a phone to localize a device and build a radio map of the environment. To improve the accuracy, we additionally incorporate turning features extracted from users’ trajectories into our system. Section \[system\_overview\] explains the details of the system implementation. Collaborative SLAM based on Pose Graph Optimization {#system_overview} =================================================== We present a novel approach that incorporates radio fingerprint measurement and motion information for collaborative SLAM in an unknown environment. The approach presented here does not require any prior knowledge about the map or the distribution of the access points nor does it need a labor-intensive phase to collect the measurements in the existing infrastructure. Our approach features a cost-effective alternative to estimate the trajectory of multiple users in unknown environments. A radio map is created simultaneously, which can be used as reference to localize other users afterwards. In our proposed collaborative framework, user walks in the environment and collects the radio measurements. Our approach merges the tracks from different users, performs loop closure detection, and optimizes the graph to generate a consistent radio fingerprint map. The data collected will be shared among all users through the server and each user will contribute to certain part of the map. The collaborative approach will accelerate the conventional way of map building. With our collaborative approach, the mapping of a building will become easier, since all users will participate in the map creation, which eliminates the expensive on-site survey phase in the conventional way of fingerprint map generation. The goal is to estimate the entire trajectories from observations (i.e., Wi-Fi observations and motion measurements) taken from different users at different times without a prior knowledge about the environment. The problem addressed here is known as SLAM, which has been well studied in the field of robotics [@RFM-SLAM; @Thrun_Probabilistic_robotics; @slam_trends_2016; @ceres-solver]. Among those, the graph-based approach, which formulates the problem as maximum likelihood estimation using pose graphs, is regarded as one of the most effective way to solve SLAM problem. Based on the raw sensor measurements, graph-based SLAM [@Thrun_Probabilistic_robotics] creates a graph, where nodes denote the poses of the users and edges decode the constraints between two nodes. The problem turns into graph optimization, which determines the best configuration of the poses by considering all constraints in the graph. Loop closure is important for any SLAM system and is considered as one of the main challenges in implementing a SLAM system in large-scale environment. It represents a situation that users have revisited a previously observed location. Since the odometry will inevitably drift for long term run, loop closure allows to correct the accumulated odometric errors and create a consistent map of the scene. The loop closure problem has been researched extensively using visual-based or ranging-based sensors [@Taketomi2017; @SujiwoATNE16], which are usually costly and computationally expensive. Instead of using the dedicated devices to perform loop closure detection, we focus on the radio fingerprints, which are available in existing indoor infrastructures and can be easily retrieved from every smartphone. Formally, let ${\mathbf{x}}_{1:T}^k=\{ {\mathbf{x}}_{1}^k,...,{\mathbf{x}}_{T}^k \}$ be the path of user $k$ we would like to estimate up to time $T$, where ${\mathbf{x}}_t^k= (x_t^k,y_t^k,\theta_t^k)$ represents the global 2D location and heading of the user at time $t$. Let ${\mathbf{z}}_{m,i}^{n,j}$ and $\Sigma_{m,i}^{n,j}$ represent the mean and covariance of a measurement (i.e., constraint or edge) between node ${\mathbf{x}}_{i}^m$ and ${\mathbf{x}}_{j}^n$. We use $\mathbb{C}$ to denote the set containing all pairs of constraints. $\hat{{\mathbf{z}}}_{m,i}^{n,j}({\mathbf{x}}_{i}^m,{\mathbf{x}}_{j}^n)$ is the prediction of a measurement based on the current configuration of node ${\mathbf{x}}_{i}^m$ and ${\mathbf{x}}_{j}^n$. The graph-based SLAM aims to find the best configuration ${\mathbf{x}}^*$ to meet the following criteria: $$\begin{split} {\mathbf{x}}^*= \operatorname*{arg\,min}_{{\mathbf{x}}} \sum_{({\mathbf{x}}_{i}^m,{\mathbf{x}}_{j}^n) \in \mathbb{C}} & {({\mathbf{z}}_{m,i}^{n,j}-\hat{{\mathbf{z}}}_{m,i}^{n,j}({\mathbf{x}}_{i}^m,{\mathbf{x}}_{j}^n))}^\intercal {\Sigma_{m,i}^{n,j}}^{-1} \\ &\times({\mathbf{z}}_{m,i}^{n,j}-\hat{{\mathbf{z}}}_{m,i}^{n,j}({\mathbf{x}}_{i}^m,{\mathbf{x}}_{j}^n)) \end{split} \label{eq:optimization}$$ In particular for graph-based SLAM, ${\mathbf{z}}_{m,i}^{n,j}$ is known as edge or constraint which represents a rigid-body transformation between node ${\mathbf{x}}_{i}^m$ and ${\mathbf{x}}_{j}^n$. The transformation is a $3 \times 1$ vector which encodes the 2D translation (i.e., $x_{m,i}^{n,j}$ and $y_{m,i}^{n,j}$) and the rotation $\theta_{m,i}^{n,j}$. The constraint can be either sequential odometry measurement (i.e., odometry-based constraint) or loop closure (observation-based constraint), which is determined by aligning the sensor observations at two non-consecutive poses. Since observations are usually erroneous, all constraints are additionally parameterized with a certain degree of uncertainty (i.e., $\Sigma_{m,i}^{n,j}$). For laser range finders, the transformation ${\mathbf{z}}_{m,i}^{n,j}$ can be determined by matching two scans [@Lu_millos_1997] using model-based registration, for example iterative closest point (ICP). Given a signal strength measurement from an AP, it is straightforward to know if an area has been visited by a user, since each reported RSS value is associated with a unique MAC address. However, estimating the precise transformation ${\mathbf{z}}_{m,i}^{n,j}$ between two observations turns out to be tricky, since radio signal neither reports distance nor bearing, and the detection range of an AP can be up to 50 meters, which is usually much larger than the accumulated error of a pedestrian dead reckoning system. The distance to an access point can be approximated by the signal strength via an analytical signal-to-distance model. This model is used by some researchers to address the SLAM problem [@Xiong2017ADG; @ferris2007wifi; @Huang_wifi_slam_11] and indoor positioning [@he2016wi; @Yassin_ieee_tutorials_2016]. However, obtaining such a model is usually not practical, as the propagation of signal will be distorted by many environmental factors (for example, multiple path or obstruction from obstacles). Instead of modeling them explicitly, this paper represents the location with radio fingerprint, which consists of the address of detected device and the measured signal strength. These fingerprints are location-dependent and are assumed to be unique to describe a location in an infrastructure. The closeness of two locations can be determined by comparing the degree of similarity of the fingerprints. We claim a loop closure if the similarity between two radio measurements at ${\mathbf{x}}_{i}^m$ and ${\mathbf{x}}_{j}^n$ reaches a threshold $\vartheta_s$. We then infer that their locations are the same and add a constraint ${\mathbf{z}}_{m,i}^{n,j}$, with all elements zero, to the graph. Actually, the two locations are unlikely to be exactly the same, which will produce a small amount of error to the loop closure. The error can be compensated by associating a covariance $\Sigma_{m,i}^{n,j}$ to the constraint. A choice of this can be a diagonal matrix by setting small values on the main diagonal. Our solution is a careful examination of the uncertainty of a loop based on the degree of similarity in a training phase. Based on training data, we automatically learn a nonparametric model to characterize the degree of similarity conditioned on the distance of two locations. ![ Illustration of our collaborative simultaneous localization and radio fingerprint mapping (C-SLAM-RF) system. The proposed system automatically creates a radio fingerprint map of an environment using the radio and odometry measurements from a group of users. The fingerprint-based and turning-based loops are identified and incorporated into a graph-based SLAM algorithm for optimization. []{data-label="fig:system_overview"}](images/slam_overview_update){width="49.00000%"} The uncertainty of the loop closure inferred from radio fingerprints is very high, we therefore exploit the turning features to improve the accuracy. We identify the turnings using motion information, match the turnings, and assign the loop closure with a small covariance if a match is found. Figure\[fig:system\_overview\] gives an overview of the system. We will describe the details of each component in our proposed solutions in the following subsections. We summarize the notations of the symbols used in this paper in Table\[notations\]. Symbol Meaning --------------------------------------------- --------- ${\mathbf{x}}_t^k$ ${\mathbf{x}}_{1:T}^k$ ${\mathbf{z}}_{m,i}^{n,j}$ $\Sigma_{m,i}^{n,j}$ ${\mathbf{F}}_t^k$ ${\mathbf{f}}_t^k$ ${\mathbf{s}}_{m,i}^{n,j}$ $d({\mathbf{x}}_i^m,{\mathbf{x}}_j^n)$ $\theta({\mathbf{x}}_i^m,{\mathbf{x}}_j^n)$ $\vartheta_{s}$ $\vartheta_{r}$ $r$ $\{s_k,d_k\}_{k=1}^{K}$ ${\mathbf{b}}(s,r)$ $var(d|s)$ $c({\mathbf{b}}(s,r))$ $a_{t}^k$ $s$ $R_t^k(\tau)$ $\mu_t^k(\tau)$ $\sigma_t^k(\tau)$ $c_t^k$ $w$ $\mathbb{C}_t^k$ $ \overline{\mathbb{C}}_t^k$ $\theta_{t}^{k^-}$, $\theta_{t}^{k^+}$ $\vartheta_{f}$ : Summary of important variables used in this paper[]{data-label="notations"} Radio Fingerprints and the Similarity {#sect_rf_fingerprint} ------------------------------------- Radio fingerprinting represents location with radio signals from radio-based sensors, for example Wi-Fi APs, Bluetooth beacons, and RFID tags. These fingerprints are robust against location-dependent distortions as compared to the model-based approaches, since the propagation of the radio signal in an environment is hard to predict due to the blockage of obstacles and multipath fading issue. This is quite similar to appearance-based approach, where the scene is represented by a number of visual features. Extracting visual features involves a large amount of computation, while this process can be ignored for the radio fingerprint, since the AP can be regarded as the unique feature for the positioning. We represent a fingerprint of user $k$ at time $t$ as a pair ${\mathbf{F}}_t^k={({\mathbf{f}}_t^k,{\mathbf{x}}_t^k)}$. ${\mathbf{x}}_t^k=(x_t^k,y_t^k,\theta_t^k)$ denotes the odometry at time $t$ when user $k$ traverses the environment. ${\mathbf{f}}_t^k$ represents the radio measurement at location ${\mathbf{x}}_t^k$, which consists of the RSS values from $L$ access points: ${\mathbf{f}}_t^k=\{ f_{t,1}^k,...,f_{t,L}^k \}$. Let $L_i^m$ and $L_j^n$ denote the number of detections in ${\mathbf{f}}_i^m$ and ${\mathbf{f}}_j^n$, respectively. $L_{m,i}^{n,j} = \left | {\mathbf{f}}_i^m \cap {\mathbf{f}}_j^n \right |$ is used to represent the common APs in both ${\mathbf{f}}_i^m$ and ${\mathbf{f}}_j^n$. The similarity function ${\mathrm{sim}}({\mathbf{F}}_{i}^m,{\mathbf{F}}_{j}^n)$ yields a positive value, representing the similarity between two radio measurements, namely ${\mathbf{f}}_{i}^m$ and ${\mathbf{f}}_{j}^n$. We apply the cosine similarity which has been extensively used in the literature [@RanArtur_IROS_2012] [@vorst2010isr]. $${\mathbf{s}}_{m,i}^{n,j}={\mathrm{sim}}({\mathbf{F}}_{i}^m,{\mathbf{F}}_{j}^n)= \frac{\sum_{l=1}^ {L_{m,i}^{n,j}} {f_{i,l}^m f_{j,l}^n}}{\sqrt{\sum_{l=1}^ {L_i^m} {\left(f_{i,l}^m\right)^{2}}}\sqrt{\sum_{l=1}^ {L_j^n} {\left(f_{j,l}^n\right)^{2}}}} \label{eq:similarity}$$ We refer the readers to [@he2016wi] [@vorst2011rfidta] for a comparison between different similarity measures. Pedestrian Dead Reckoning {#pdr} ------------------------- The spatial relationship between sequential poses in Equation \[eq:optimization\] can be determined by the odometry measurements, which is known as odometry-based constraint. Nowadays, smartphones are equipped with various types of sensors, including IMU sensor, camera, light sensor, and proximity sensor. This enables one to implement a variety of pedestrian dead reckoning systems using different techniques. We evaluated our system under two pedestrian dead reckoning systems (PDR): Tango-based PDR using visual-inertial odometry (VIO) and step counter-based PDR using accelerometers and compass. The goal is to compare the approach under various tracking systems with different tracking accuracies. Tango is developed by Google that uses visual-inertial odometry, to estimate the location of a device without GPS or any external referencing. It uses visual features with a combination of inertial measurements from accelerometer and gyroscope to track the movement of a device in 3D space. Lenovo Phab 2 Pro and Asus Zenfone AR are two examples of commercially available Tango phones. Alternatively, the IMU sensor embedded inside a phone can be used to implement a step counter-based dead reckoning. A typical IMU system is comprised of accelerometer, gyroscope, and magnetometer for motion or orientation sensing. Following Zee [@Zero_calibration], we implemented the step counting based on auto-correlation. Given the acceleration measurement $a_t^k$ of user $k$ at time $t$, the step counting is achieved by examining the periodic step patterns through normalized auto-correlation $R_t^k(\tau)$ for a lag $\tau$: $$R_t^k(\tau)= \frac{\sum_{l=0}^{l=\tau-1} \begin{bmatrix} \begin{array}{l} (a_{t+l}^k-\mu_t^k(\tau)) \\ (a_{t+l+\tau}^k-\mu_{t+\tau}^k(\tau)) \end{array} \end{bmatrix} } {\tau \sigma_t^k(\tau) \sigma_{t+\tau}^k(\tau)},$$ where $\mu_t^k(\tau)$ and $\sigma_t^k(\tau)$ are mean and standard deviation of the sequential acceleration measurements $\{a_{t}^k,a_{t+1}^k,...,a_{t+\tau-1}^k\}$. The algorithm first identifies an optimal $\tau_{opt}$ to maximize $R_t^k(\tau)$. Then $\tau_{opt}$ is used as a replacement of $\tau$ to count further steps. Similar to Zee[@Zero_calibration], the sampling rate of the IMU is 50Hz, we therefore set the initial searching window $\tau$ to $[40,100]$. During the walking of a person, we assume the phone is always held in front of him. Therefore, we use the magnetometer reading to approximate the orientation of the user. Let $c_t^k$ and $\theta_t^k$ be the step counting and the orientation of user $k$ at time $t$ respectively, the position (i.e., $x_t^k$ and $y_t^k$) of user $k$ is determined by: $$\begin{aligned} x_t^k=x_{t-1}^k+s\cdot(c_t^k-c_{t-1}^k) \cdot \cos (\theta_{t-1}^k) \\ y_t^k=y_{t-1}^k+s\cdot(c_t^k-c_{t-1}^k) \cdot \sin (\theta_{t-1}^k), \label{eq:pdr} \end{aligned}$$ where $s$ is the step length, which is assumed to be fixed throughout the experiments. The estimation of the pose in odometry frame works in recursive fashion. The initial values of $x$ and $y$ are set to zero. The initial headings of different PDR systems are treated differently. For Tango-based PDR, the initial heading of a user is set to zero. For step counter-based PDR, the heading is determined by the in-built magnetometer based on geographic cardinal directions. The magnetic materials in the building might affect the accuracy of the orientation estimation. In the future, we would like to compensate for the orientation by incorporating the gyroscope readings [@richard_ieeesensor_2016]. The odometry-based edge is determined based on the relative translation and rotation between the sequential odometry measurements. For a user $k$, the rigid-body transformation ($\Delta x_t^{k}$, $\Delta y_t^{k}$ and $\Delta \theta_t^{k}$) between the pose at time $t-1$ and time $t$ can be computed as: $$\begin{aligned} \begin{bmatrix} \Delta x_t^{k} \\ \Delta y_t^{k} \\ \Delta \theta_t^{k} \\ \end{bmatrix} = \begin{bmatrix} \cos( \theta_{t-1}^{k} ) & -\sin( \theta_{t-1}^{k} ) & 0\\ \sin( \theta_{t-1}^{k} ) & \cos( \theta_{t-1}^{k} ) & 0 \\ 0 & 0 & 1 \\ \end{bmatrix} ^\top \begin{bmatrix} x_t^{k}-x_{t-1}^{k} \\ y_t^{k}-y_{t-1}^{k} \\ \theta_t^{k}-\theta_{t-1}^{k} \\ \end{bmatrix} \end{aligned}$$ RSS Thresholding {#sect_rss_threshold} ---------------- The time required to compute the similarity in Equation (\[eq:similarity\]) grows linearly with the number of APs in the two fingerprints. The computational time can be significant in densely AP covered environment, which is the typical case in modern office or commercial buildings. A large amount of computational time can be saved if the size of the measurements can be reduced. Therefore, we propose to filter out the RSS observation with value below a threshold $\vartheta_r$. Thresholding prunes observations with small RSS values, which represent spurious readings due to multiple propagation issues in indoor environment. In addition, larger RSS values indicate a location close to the access point with more confidence. These measurements are expected to better confine the location of the user. In the experimental section, we show that thresholding technique can provide a better accuracy while consuming less computational time. Finding Loop Closure Candidates {#sect_finding_loop_candidate} ------------------------------- To find the observation-based edge (i.e., constraint) between the non-consecutive poses in Equation \[eq:optimization\], we need to perform loop closure detection. The observation-based edge consists of two different types of edges, namely similarity-based edge and turning-based edge. In our approach, each fingerprint carries the odometry information where the fingerprint is recorded (i.e., $x$, $y$, and the orientation $\theta$ of a user). We first compute the relative distance $d({\mathbf{x}}_i^m,{\mathbf{x}}_j^n)$ and orientation $\theta({\mathbf{x}}_i^m,{\mathbf{x}}_j^n)$ between the odometric poses of two fingerprints ${\mathbf{F}}_i^m$ and ${\mathbf{F}}_j^n$. If these values are smaller than pre-defined thresholds (50 meters and 0.3 radians for distance and orientation respectively), we compute the similarity ${\mathbf{s}}_{m,i}^{n,j}$ between them. We add a tuple $<{\mathbf{x}}_{i}^m,{\mathbf{x}}_{j}^n,{\mathbf{s}}_{m,i}^{n,j}>$ as a candidate of the loop closure if the similarity ${\mathbf{s}}_{m,i}^{n,j}$ exceeds a threshold $\vartheta_s$, which is one of the few parameters that has to be supplied by the user. The impact of $\vartheta_s$ on the performance is not too critical, as shown in our experiments. In most cases, $\vartheta_s=0.7$ gives good results. We reject the similarity with values smaller than $\vartheta_s$, to avoid false positive loop closures. To improve the accuracy of the system, we further check if this loop is a turning-based loop. We identify a turn by examining the orientation changes and checked the fitness of their respective tracks by a matching algorithm. If the fitness score is higher than a threshold, we consider this loop as a turning-based loop. The detail of the detection of turning-based loop can be found in Section \[turning\_indentification\_matching\]. ![Scatter plot of similarity versus distance from experimental data in building 1.[]{data-label="fig:scatter_plot"}](images/scatter_plot){width="45.00000%"} Model Training {#sect_model_training} -------------- To optimize Equation \[eq:optimization\], an uncertainty estimation of the constraint is necessary for all edges in SLAM graph. For odometry-based edges, the parameter is obtained from the motion model. We now need to derive a model to represent the uncertainty of the observation-based edges. Our solution is to train such model by passing over the observation data (i.e., odometry and radio fingerprints), which is recorded by the smart phone as shown in Figure \[fig:system\_overview\]. Our goal is to generate an uncertainty model to feature the distance variance of two radio fingerprints given their similarity. To build such a model, we need to know the true locations where the fingerprints are recorded or relative distance between the recorded positions, which is not possible without any external reference system. Although the error from odometry accumulates in the long term, it is sufficiently small for a short time of duration. In this work, we assume odometry is accurate enough for the distance traveled less than 100 meters, which is suitable for most inertial tracking platforms [@richard_ieeesensor_2016] [@visual_odom_zhang_ICRA] [@foot_mounted]. For example, authors in [@visual_odom_zhang_ICRA] evaluated the visual odometry with wide angle and fisheye cameras, and showed a relative positioning error of less than 1.4% with a distance of 538 meters traveled. Therefore, we compute the degree of similarity for close fingerprint pairs. These values are annotated with the distance between the two locations using the PDR system. As a result, we will get a set of $K$ training samples: $\{s_k,d_k\}_{k=1}^{K}$, where $s_k$ is the similarity and $d_k$ is the distance of the fingerprint pair. Figure \[fig:scatter\_plot\] shows the scatter plot of distance versus similarity in one of the buildings. We then train a model which features the variance of distance given a similarity by binning. That is, for a similarity value $s$, we compute the variance of the samples that sites in the small interval $r$ around $s$: $$var(d|s)=\frac{1}{ c ({\mathbf{b}}(s,r)) } \sum_{k \in {\mathbf{b}}(s,r)} { d_k}^2$$ where $c({\mathbf{b}})$ counts the number of samples in interval $r$. $var(d|s)$ denotes the variance of the distance $d$ given a similarity $s$. Although binning is a simple way for smoothing, the computation is efficient, since assigning the sample into a bin is straightforward. One example of the variance computed in two different buildings is shown in Figure \[fig:sim\_distance\]. The resulted variance is stored in a look up table, which could be used in the second stage of SLAM, as shown in Figure \[fig:system\_overview\]. Merging Tracks at Different Times {#track_merge} --------------------------------- To leverage the power of crowdsourcing, we utilize the measurements captured from different users to generate a radio map of the environment. This involves the loop closure detection between different users in order to correct their paths using the power of crowdsourcing. The tracks recorded from different users are based on different reference systems, for example different starting positions. The determination of orientation is different for the two pedestrian dead reckoning systems. The Tango-based PDR estimates the orientation by visual-inertial odometry based on the starting pose and the step counter-based PDR determines the orientation based on the compass readings and is not relevant to any starting position. Therefore, these trajectories are needed to be merged into the same coordinate system to guarantee a robust loop closure detection between different users. In this paper, we start the tracking by assuming all users passing by the same place. This is quite reasonable since users may pass through several key landmarks in an environment for example entrance of a building or the elevator. One might argue that in large buildings for example airport, not all users share the common place. It is possible to first build several sub-maps, and then merge them into a large and consistent map [@multi_SLAM_lidar] [@merge_multiple_map] [@Optimization_2Dmap]. An edge is added to connect the first nodes in different tracks. For the transformation matrix ${\mathbf{z}}_{m,i}^{n,j}$, we set $x_{m,i}^{n,j}$ and $y_{m,i}^{n,j}$ to zero and the covariance is obtained by checking the covariance table as detailed in Section \[sect\_model\_training\]. Due to the omnidirectional characteristics of antennas, the facing of a user has very little impact on the radio signals, therefore the radio fingerprint does not deliver any orientation information. This is the reason why Section \[sect\_model\_training\] does not model orientation variance with respect to the similarity. Based on two fingerprint observations, we have no knowledge about how accurate is the relative orientation between two poses, we therefore set $\theta_{m,i}^{n,j}$ to zero and give a very large covariance value (i.e., 1000) to the edge, meaning that we are not able to infer the relative angle from two radio fingerprint observations. For users starting from arbitrary locations, we refer to [@luo_localization] [@Zhou_activity_landmark] [@Shen_walkie] [@Li_crowdsourcing_traces] to merge the paths based on the radio measurements and activity landmarks in our future work. Turning Identification and Matching {#turning_indentification_matching} ----------------------------------- A typical indoor infrastructure often contains various landmarks, such as turnings, elevators, and staircases. These landmarks are unique in an existing infrastructure and can be used as a good feature for loop closure identification in a SLAM system. In this case, a loop closure can be claimed if two landmarks match each other. Due to the physical constraint of an environment, such kind of loop closure provides lower positioning uncertainty as compared to fingerprinting-based loops. For example, the size of an elevator is usually less than 3 meters, and the turning radius during human walking is smaller than a typical corridor width (i.e., 5 meters), while the positioning error using fingerprinting-based approach is usually larger than 5 meters. Here we only focus on the turning features in a trajectory, which is regarded as one of the most common indoor landmarks. A slide window is used to produce segmentations of the track. In particular, we define a segmentation $\mathbb{C}_t^k$ of user $k$ at time $t$ as a collection of sequential poses with a window size of $w$, i.e., $\mathbb{C}_t^k=\{{\mathbf{x}}_{t'}^k\,|\,-\frac{w}{2} \le t'-t \le \frac{w}{2} \}$. For each loop closure candidate $<{\mathbf{x}}_{i}^m,{\mathbf{x}}_{j}^n>$, we check if there are turnings at these poses by examining the orientation change in segment $\mathbb{C}_i^m$ and $\mathbb{C}_j^n$. If yes, we try to match the two segmentations $\mathbb{C}_i^m$ and $\mathbb{C}_j^n$ using ICP (Iterative Closest Point) [@icp_original]. If the fitness score (i.e., average of squared distances between the correspondence points) between $\mathbb{C}_i^m$ and $\mathbb{C}_j^n$ is smaller than a predefined threshold $\vartheta_f$, we regard loop $<{\mathbf{x}}_{i}^m,{\mathbf{x}}_{j}^n>$ as the turning-based loop. Otherwise, this loop closure is referred to as fingerprint-based loop, as shown in Figure \[fig:system\_overview\]. The transformation ${\mathbf{z}}_{m,i}^{n,j}$ of both types of loop is set to zero. We treat the covariance matrix $\Sigma_{m,i}^{n,j}$ differently: for fingerprint-based loop, the covariance along $x$ and $y$ can be found at the looking up table computed previously in Section \[sect\_model\_training\]; for turning-based loop, we set the covariance along $x$ and $y$ to 5.0, which is smaller than fingerprint-based loop (i.e., 8.0 in Figure \[fig:environment\] with the highest fingerprint similarity). The orientation covariance in $\Sigma_{m,i}^{n,j}$ is set to 1000, meaning that we do not have any knowledge about the orientation of the two poses by the radio observations or the matching of two turnings. The details of turning identification and matching will be described in subsequent parts: ![Example of the track (blue color) and turnings detected (red circles) with a window size of $20$ based on the approach described in Section \[turning\_indentification\_matching\]. In total, 111 turnings are identified. These turnings are further examined by turning matching module to find the potential turning-based loops.[]{data-label="fig:turning_detection"}](images/track_small_circles){width="45.00000%"} ![Example of two track segmentations (i.e., $\mathbb{C}_i^m$ and $\mathbb{C}_j^n$ in blue and red color, respectively), and the correspondence points found using ICP (green color). A fitness score is calculated to determine if this candidate is a turning-based loop.[]{data-label="fig:turning_matching"}](images/match_1408_823_38){width="49.00000%"} ### Turning Identification {#turning_indentification} We first segment the compass data and find the potential turnings [@turn_detection_trans]. For the pose ${\mathbf{x}}_t^k$, we calculate the mean orientations for the poses with timestamps smaller and larger than $t$ in the segmentation $\mathbb{C}_t^k$, i.e., $\theta_{t}^{k^-}= \mu \{\theta_{t'}^k | t'-t \le 0\}$ and $\theta_{t}^{k^+}= \mu \{\theta_{t'}^k | t'-t \ge 0\}$. If $|\theta_{t}^{k^+} - \theta_{t}^{k^-}|$ is higher than a threshold (for example $\frac{\pi}{3}$ as suggested in [@turn_detection_trans]), a turning is identified. The window size $w$ here has the impact on the performance of turning detection and we show its impact on the accuracy in the experimental section. One example of the track and turnings detected are shown in Figure \[fig:turning\_detection\]. A better approach to improve the accuracy of turning detection can be found in [@sun_corner_detection_2017]. ### Turning Matching {#turning_match} The ICP aims to find a transformation (translation and rotation) between two point clouds that minimizes the sum of the square distance between the correspondence points. This approach has been extensively used to match 2D laser scans in the field of robotics. To find the correct transformation using ICP, an appropriate initial transformation has to be provided, otherwise ICP will fall into the local minimum. Rather than using the global raw odometry for a segmentation $\mathbb{C}_t^k$, we use the relative translation between ${\mathbf{x}}_{t'}^k$ and ${\mathbf{x}}_t^k$, i.e., $\mathbb{\overline{C}}_t=\{ \mathbf{T}_{{\mathbf{x}}_{t}^k}^{{\mathbf{x}}_{t'}^k} \,|\,-\frac{w}{2} \le t'-t \le \frac{w}{2} \}$. Since the sampling rate of our pedestrian tracking system is too low (less than 1.0 HZ), we further interpolate the trajectory to get a large amount of locations for performing ICP. An illustration of the turning matching using ICP is shown in Figure \[fig:turning\_matching\]. We finally considers $<{\mathbf{x}}_{i}^m,{\mathbf{x}}_{j}^n>$ as a valid match (i.e., turning-based loop) if the fitness score is smaller than a threshold $\vartheta_f$. Robust loop closure detection is essential to a SLAM system, as incorrect loop closures will ruin the consistency of trajectory and the map. Other heuristic approach can be applied to further examine the loops and filter out the suspicious ones. Authors in [@GalvezIROS11], for example, proposed an approach to group the close loop closures and check the temporal and spatial consistency for robust loop closure detection. However, this technique goes beyond the scope of this paper, hence, we add the loop closures without performing futher consistency check. Pose Graph Optimization {#pose_graph_optimization} ----------------------- The Equation \[eq:optimization\], which represents a graph consisting of poses (i.e., nodes) and constraints (i.e., edges), is finally optimized through the pose graph optimization algorithm GraphSLAM. For the implementation, we choose Levenberg-Marquardt in g2o [@kuemmerle11icra], which is freely available and is one of the state-of-the-art SLAM algorithms[^5]. ![ Part of the estimated track over time using Tango-based pedestrian dead reckoning and fingerprint-based constraints inferred with a similarity threshold of $\vartheta_s=0.7$. []{data-label="fig:constraints"}](images/constrain){width="49.00000%"} Parameter Range Guidelines to choose the parameters ----------------- -------------- ------------------------------------- -- $\vartheta_{s}$ \[0, 1\] $\vartheta_{r}$ \[-90, -50\] $r$ \[0, 1\] $w$ \[0, 200\] $\vartheta_{f}$ \[0, 1\] [|c|c|c|c|c|c|c|c|c|]{} & &\ & ----------- Mean$\pm$ Std. Dev. ----------- & Max./Median & -------- Number of MAC -------- & ---------- Comput. time (s) ---------- & ----------- Mean$\pm$ Std. Dev. ----------- & Max./Median & -------- Number of MAC -------- & ---------- Comput. time (s) ---------- \ -90 & 1.39$\pm$1.51 & 7.86/1.37 & 179.23 & 24.50 & 14.95$\pm$7.09 & 37.72/14.85 & 129.36 & 426.37\ -80 & 1.26$\pm$1.21 & 6.30/1.19 & 137.81 & 19.53 & 5.91$\pm$3.76 & 20.87/5.14 & 70.08 & 212.07\ -75 & 1.06$\pm$0.49 & 4.53/0.94 & 81.29 & 12.31 & 5.40$\pm$3.60 & 21.52/4.56 & 43.36 & 127.67\ -70 & 0.77$\pm$0.46 & 3.79/0.73 & 47.97 & 5.73 & 5.19$\pm$3.48 & 21.64/4.47 & 25.49 & 72.34\ -65& 0.92$\pm$0.44 & 3.87/0.84 & 28.94 & 3.29 & 5.43$\pm$3.43 & 20.40/4.70 & 13.41 & 37.78\ -60& 1.24$\pm$0.89 & 4.91/1.13 & 16.22 & 1.76 & 5.95$\pm$3.82 & 22.78/5.06 & 6.08 & 20.36\ -50& 2.44$\pm$0.95 & 5.81/2.12 & 3.29 & 0.37& 11.51$\pm$5.82 & 35.03/11.52 & 0.67 & 3.08\ odom.& 8.03$\pm$7.80 & 31.55/4.22 & NA & NA & 15.39$\pm$7.89 & 41.23/14.58 & NA & NA\ Experimental Results {#experimental_evaluations} ==================== Experimental Details {#Implementation Detail} -------------------- We program two smart phones (Lenovo Phab 2 Pro with Android 6.0.1 and Sony Xperia Z3 with Android 5.0.2) to receive the signal strength from APs and perform pedestrian dead reckoning. In particular, the Lenovo phone uses the Tango for position tracking and the motion tracking data is recorded every five seconds due to its high tracking accuracy. We implement the step counting on a Sony phone and record the step counting and the compass readings every one second. To evaluate the performance of the proposed approach, we conducted experiments on the Level 3 of Building 2 at Singapore University of Technology and Design with a size of 130m$ \times $70m (see Fig. \[fig:floor\_plan\]). This environment is comprised of corridors, concrete walls, and wide open space. We asked a person to hold two phones (i.e., Lenovo Phab 2 Pro and Sony) and walk in the building along different paths with a regular walking speed. Four tracks were recorded at different times to show the power of the collaborative SLAM. For each track, the user started from the same position. The total distance traveled is 1906 meters with a duration of 2179 seconds and a number of 1702 unique MAC addresses are detected. This results in four log files with a duration of 687, 582, 439, and 471 seconds respectively. The step length $s$ is fixed to 0.7m throughout this paper. Fig. \[fig:environment\] shows a snapshot during the experiment. Ground Truth and Accuracy Comparisons {#ground_truth} ------------------------------------- To extract the ground truth as comparison, we optimized the track from Tango using GraphSLAM taken 3D point clouds as input. We implement loop closure detection based on the point cloud library (PCL) [@pcl_icra_2011]. We identify Harris keypoints in a pair of point clouds and compute the corresponding SHOT (Signature of Histograms of OrienTations) descriptors [@shot_feature]. We match these descriptors with k-nearest neighbors algorithm (k-NN) and find an initial transformation using SVD (singular value decomposition). The transformation is further refined by ICP. If the number of matched points exceeds a threshold (half size of the point cloud), a loop closure is confirmed and added to the graph as constraints. We treat this optimized path as the ground truth to evaluate the accuracy of our system. We show the accuracy by the root mean square error (RMSE) between the ground truth and the estimation. Our experiments show that we are able to achieve an accuracy of 0.6 meters with Tango-based PDR and 4.76 meters of a step counter-based PDR with a size of 130m$ \times $70m, as shown in Figure \[fig:floor\_plan\]. The optimized track is annotated with the radio measurement and can serve as the radio map for the positioning of another user. In Table \[important\_parameters\], we show the important parameters used in this paper and the remarks of how to choose these parameters. The final positioning error is calculated at the end of the process after the loop closure detection, turning matching, and pose graph optimization. Larger positioning errors are expected in a real time system, since we have to process the incoming data in a sequential way and provide pose estimation at regular intervals before loop closures are detected. To implement a practical indoor localization system, the mechanism to deal with the change of the radio environment is a necessity. The evolution of the radio environment (for example adding or removing the access points) can be examined by looking at the signal variance at similar locations, as shown in [@adaptive_fingerprint] [@self_updating_radio_maps] [@crowdsourced_signal_map] [@automated_construction] [@PF_radiomap_recalibrate]. Some MAC addresses might be static for a fixed duration of time, but are essentially mobile for example personal hotspots. A good way to address the mobile hotspots issue is to filter out the MAC addresses by the organizationally unique identifier (OUI), which is used to uniquely identify a vendor[^6]. The MAC address from a phone manufacturer should be removed from the detection list to prevent the uncertainty of incorporating the additional mobile hotspot observations. Another approach to filter out the mobile hotspots is to look at the spatial relations of the detected positions of a particular MAC address [@WOLoc] [@learning_wifi_hotspots] [@redundant_access_point_reduction]. Impact of Different RSS Threshold $\vartheta_r$ {#evaluation_rss_threshold} ----------------------------------------------- We examined the influence of RSS thresholding on the accuracy in this series of experiments. We set the similarity threshold $\vartheta_s=0.7$ and use a binning size $r=0.2$. We chose $\vartheta_r$ values between -90 and -50 to evaluate the mean accuracy, as listed in Table \[table\_rss\]. As compared to the raw odometry, our approach can effectively reduce the accumulated odometry error: with the setting of $\vartheta_r=-70$, our approach improves the positioning accuracy by 90.4% and 66.3% for Tango (from 8.03m to 0.77m) and step counter (from 15.39m to 5.19m) respectively. In addition, the accuracy of the PDR has a very high impact on the accuracy achieved with our SLAM system: Tango shows a good motion tracking performance and we achieved an accuracy of 0.77m with a threshold $\vartheta_r=-70$. The accuracy achieved with Tango is better than state-of-the-art fingerprinting-based approaches [@ran_ieee_sensors2017; @Yassin_ieee_tutorials_2016]. While the step counter results in a large amount of accumulated odometry error and the accuracy obtained with our approach is worse (5.19m). A further investigation to the PDR system will help to improve the accuracy, which will be one of our future work. One has to note that the accuracy is achieved without training as opposed to the fingerprinting-based approaches, where a time-consuming phase to collect and annotate the fingerprints is prerequisite to guarantee a good positioning accuracy. Table \[table\_rss\] also shows that we maintain a good accuracy with an RSS threshold between -75 and -65, while the computational time decreases considerably with the thresholding technique. As an example, for Tango-based system, a threshold of -70 reduces the computation time to 5.73 seconds as compared to a threshold of -90 (i.e., 24.28 seconds). At the same time, the accuracy even increases by 0.62 meters (error drops from 1.39m to 0.77m). A suitable threshold produces a good accuracy, as it will filter out the suspicious radio signals. However, a threshold larger than -65 leads to a bad result (for example, 1.24 meters of accuracy with $\vartheta_r=-60$ for Tango-based system). The ground truth, estimation, and odometry of individual tracks using two different pedestrian dead reckoning systems are visualized in Figure \[fig:example\_track\]. A part of estimated trajectory and the constraints inferred are shown in Figure \[fig:constraints\]. ------ --------------- ------ ---------------- ------- 0.95 5.63$\pm$6.76 18 13.86$\pm$7.67 2098 0.9 2.15$\pm$0.71 94 6.29$\pm$3.67 4241 0.8 1.06$\pm$0.51 384 5.49$\pm$3.61 10582 0.7 0.77$\pm$0.46 715 5.19$\pm$3.48 17020 0.6 0.81$\pm$0.76 1150 5.35$\pm$3.34 23859 0.4 0.80$\pm$0.79 2198 5.49$\pm$3.66 42625 0.2 1.06$\pm$0.89 3494 5.94$\pm$4.02 66423 0.1 1.09$\pm$0.95 4517 6.09$\pm$3.94 85834 ------ --------------- ------ ---------------- ------- : Evaluation of the proposed approach with two different pedestrian dead reckoning systems under different settings of similarity threshold $\vartheta_s$. The table shows the positioning accuracy (mean and standard deviation) in meters and the number of constraints inferred. []{data-label="table_sim"} ----- ------ ------ ------ ------ ------ ------ ------ -90 -80 -75 -70 -65 -60 -50 0.9 2.96 2.65 2.38 2.15 1.81 2.08 2.61 0.8 1.68 1.29 1.16 1.06 1.23 1.41 2.58 0.7 1.39 1.26 1.06 0.77 0.92 1.24 2.44 0.6 1.22 1.02 0.87 0.81 0.89 1.27 2.35 0.5 0.88 0.83 0.78 0.82 0.93 1.57 2.33 0.4 0.99 0.91 0.82 0.80 0.85 1.54 2.40 0.3 1.19 1.16 1.08 0.92 0.89 1.61 2.35 0.2 1.26 1.27 1.22 1.06 0.96 1.65 2.35 0.1 1.28 1.29 1.23 1.09 1.01 1.70 2.30 ----- ------ ------ ------ ------ ------ ------ ------ : Evaluation of the mean positioning accuracy under the impact of different settings of $\vartheta_s$ and $\vartheta_r$ for Tango-based approach. []{data-label="table:positioning_under_theta_s_theta_r"} Impact of Different Similarity Threshold $\vartheta_s$ {#evaluation_rss_threshold} ------------------------------------------------------ Next, we performed a series of experiments to examine the influence of accuracy with respect to different similarity thresholds $\vartheta_r$. We show the results in Table \[table\_sim\]. We fixed RSS threshold $\vartheta_r=-70$ and a binning size $r=0.2$. We increased the similarity threshold $\vartheta_s$ from 0.1 to 0.95 to evaluate the accuracy and the number of constraints inferred. From Table \[table\_sim\], we can observe that the number of constraints is different for Tango and step counter-based PDR due to different sampling rates of the device (5 seconds for Tango and 1 second for step counter): Tango-based PDR offers a small number of constraints as compared to step counter-based PDR. In addition, the threshold has a high impact on the accuracy and the number of constraints. Applying a high threshold will result in a small number of constraints and a decrease of the accuracy. For Tango-based system, we obtain a mean accuracy of 0.77m with $\vartheta_s=0.7$, which is an improvement of 86.3% as compared to the mean accuracy of 5.63m with $\vartheta_s=0.95$. Yet, such an improvement is at the expense of a higher number of constraints added (i.e., 715 constraints with $\vartheta_s=0.7$ as compared to 18 with $\vartheta_s=0.95$). But the accuracy does not get improved with a threshold smaller than 0.6. One reason could be because a low similarity value will always come along with a very large covariance, and has very less strength to correct the odometric error. A setting of $\vartheta_s=0.7$ seems to be a good trade off between the accuracy and the number of constraints inferred. Table \[table:positioning\_under\_theta\_s\_theta\_r\] showed the accuracy by jointly optimizing the parameters $\vartheta_s$ and $\vartheta_r$ for Tango-based approach. As can be seen from this table, a careful examination of $\vartheta_s$ and $\vartheta_r$ will improve the accuracy. A too large or too small will obviously deteriorate the performance of our approach. ------ --------------- ----------- --------------- ------------ 1.0 1.03$\pm$0.63 4.29/0.89 6.23$\pm$3.84 24.83/5.59 0.8 0.85$\pm$0.50 4.22/0.85 5.71$\pm$3.50 21.97/5.03 0.6 0.84$\pm$0.39 3.62/0.79 5.32$\pm$3.54 21.78/4.60 0.4 0.81$\pm$0.62 4.26/0.76 5.22$\pm$3.51 21.93/4.46 0.2 0.77$\pm$0.46 3.79/0.73 5.19$\pm$3.48 21.64/4.47 0.1 0.79$\pm$0.40 3.59/0.86 5.23$\pm$3.79 20.97/4.13 0.05 0.81$\pm$0.39 4.00/0.87 5.29$\pm$3.52 21.48/4.50 ------ --------------- ----------- --------------- ------------ : Evaluation of the positioning accuracy (mean, standard deviation, median, and maximum in meters) with two pedestrian dead reckoning systems under the impact of different configurations of binning size $r$. []{data-label="table_binning"} Impact of the Binning Size of Training {#evaluation_rss_threshold} -------------------------------------- Next, we examined the influence of accuracy with respect to various binning sizes $r$. We chose RSS threshold $\vartheta_r=-70$ and similarity threshold $\vartheta_s=0.7$. To evaluate accuracy under impact of different binning sizes, we set $r$ to the following values $r=\{0.05, 0.1, 0.2, 0.4, 0.6, 0.8, 1.0\}$. In our approach, the uncertainty model is trained with all the collected data. We show a comparison of the results in Table \[table\_binning\]. This table shows that the best choice of $r$ is $0.2$. The covariance estimated with a large $r$ is usually too large to compensate for the error from the odometry. Optimizing $r$ gives an improvement of 25.2% (0.77m and 1.03m for $r=0.2$ and $r=1.0$ respectively) for Tango-based PDR and 16.6% (5.19m and 6.23m for $r=0.2$ and $r=1.0$ respectively) for step counter-based PDR. Covariance of the loop is a key to optimize the pose graph, as it is the only information to measure how close the two locations are in a loop, therefore, a careful examination of the parameter will lead to an improvement of the accuracy. The covariance added here (see Figure \[fig:sim\_distance\]) is much smaller as compared to the accumulated odometry error (notice that the maximum positioning error of Tango and step counter in Table \[table\_rss\] are 31.55m and 41.23m respectively). This is why we are still able to correct the accumulated odometry error. The approach presented here provides a way to automatically calibrate the uncertainty model with the odometry measurement. The model generated in different environments might be slightly different, as shown in Figure \[fig:sim\_distance\]. To evaluate the accuracy under the impact of different similarity models, Table \[table:accuracy\_diff\_bld\] compared the results using the similarity models produced from two different buildings, namely Building1 (the one used for the verification of the positioning accuracy) and Building2 regarding to Tango-based approach. As can be seen from this table, the two models provide similar positioning accuracy, which proves our assumption that the similarity model can be applied to different environments. ------ ------ ------ ------ ------ ------ ------ ------ 0.05 0.1 0.2 0.4 0.6 0.8 1.0 Bld1 0.81 0.79 0.77 0.81 0.84 0.85 1.03 Bld2 0.84 0.81 0.78 0.85 0.86 0.91 1.07 ------ ------ ------ ------ ------ ------ ------ ------ : comparison of mean positioning accuracy under different similarity models trained at two different buildings regarding Tango-based pedestrian dead reckoning system. []{data-label="table:accuracy_diff_bld"} Impact of Turning Detection and Matching {#evaluation_motion_constraints} ---------------------------------------- We compared the accuracy with and without the integration of turning features in the next series of experiments. We fixed RSS threshold $\vartheta_r=-70$, similarity threshold $\vartheta_s=0.7$, and binning size $r=0.2$. We varied the setting of $w$ and $\vartheta_f$ to evaluate their impact on the accuracy. Figure \[fig:accuracy\_under\_turning\_detection\] shows a comparison of the results. This figure shows that the accuracy can be improved by additional integration of turning features: we obtain an improvement of 22.1% for the Tango-based pedestrian dead reckoning (from 0.77m to 0.6m with $w=40$ and $\vartheta_f=0.5$) and 8.3% for step counter-based dead reckoning (from 5.19m to 4.76m with $w=40$ and $\vartheta_f=0.5$). The improvement of Tango is slightly higher as compared to the step counter-based system. Since the odometry error of Tango is smaller than step counter, the turning feature here exhibits great capability to correct the drift error of odometry. However, for step counter-based PDR, the error is dominated by the odometry and the turning feature shows less improvement to the accuracy as compared to Tango-based PDR. The number of turnings detected with different settings of $w$ is shown in Table \[table:turnings\_detected\]. The constraints are mostly from the fingerprinting matching. As can be seen from Table \[table:turnings\_detected\], only 26 turnings are detected with $w=40$, which is much less than the number of fingerprinting-based constraints (715 for Tango-based approach with $\vartheta_s=0.7$ as shown in Table \[table\_sim\]). Due to the low sampling rate, Tango-based PDR leads to a small number of turnings as compared to the step counter-based PDR. A large value of $w$ leads to an increasing number of turnings detected. From Figure \[fig:accuracy\_under\_turning\_detection\], we can also observe that a window size $w=40$ leads to the best accuracy for both systems. A too high or too small $w$ obviously results in a less improvement to the accuracy. In addition, Figure \[fig:accuracy\_under\_turning\_detection\] shows that a too large or too small $\vartheta_f$ leads to a decrease of the accuracy. A suitable fitness threshold $\vartheta_f$ will help to remove the false turning-based loop closures. A small $\vartheta_f$ will not be able to identify true matched turnings and results in a small number of turning-based loops and therefore has less strength to improve the accuracy. A large $\vartheta_f$ produces too many false turning-based loops, which leads to a deterioration of the result. As an example, for step counter-based PDR with $w=40$, a setting of $\vartheta_f=0.5$ gives an accuracy of 4.76m, which produces an improvement of 3.8% and 7.0% when compared to $\vartheta_f=10.0$ (4.95m) and $\vartheta_f=0.02$ (5.12m), respectively. Window size $w$ ------------------------- --- ----- ------ ------ ------ Turnings (Tango) 0 6 26 45 78 Turnings (step counter) 0 423 1447 2666 3973 : comparison of number of turnings detected with two different pedestrian dead reckoning systems under various settings of window size $w$. []{data-label="table:turnings_detected"} Computational Time {#evaluation_rss_threshold} ------------------ Finally, we evaluated the time consumption at each stage of our approach. The results are listed in Table \[table:computational\_time\]. In our approach, we process the recordings in a batch matter, which means that we optimized the graph after all loop closures are identified. An Intel Core i5-4200M CPU with 2.5GHz frequency and 4GB RAM is used to process the measurements. We set $\vartheta_r=-70$, $\vartheta_s=0.7$, $r=0.2$, $w=40$, and $\vartheta_f=0.5$. As can be seen from Table \[table:computational\_time\], the entire data processing took 17.81 (6.08+5.73+5.93+0.07) seconds for tango-based dataset, which is almost 30 times faster than the data recording stage (544.75 seconds), while the processing time is much longer for step counter-based dataset (approx. 129.39 seconds, i.e., 23.58+72.34+31.82+1.65), due to its higher sampling rate. Additionally, optimization of the graph only took less than two seconds (0.07 and 1.65 seconds for Tango and step counter-based system respectively). In our current offline implementation, we compute the similarity of all pairs of Wi-Fi measurements in the entire dataset to find the potential loop closures. This module consumes too much time (72.34 seconds for the step counter version as shown in Table \[table:computational\_time\]), and thus cannot be used for real time pose estimation. However, we believe the computation can be further optimized to make online implementation possible, e.g. we only need to compare the current Wi-Fi measurement with the previous measurements for the loop closure detection (estimated to be 72.34/2179 = 0.03 seconds), which is less than the current Wi-Fi sampling rate of one second interval. Nonetheless, we believe further optimization is needed to ensure online implementation in real time. When more users are involved in the experiment, it will take longer time to run the algorithm due to the increasing number of nodes and the constraints in the graph. Still, our approach is efficient when compared to vision-based approaches, as vision-based approach requires heavy computational resources due to feature extraction and feature matching. In addition, there can be data association problems which will result in the loop closure failure. On other hand, the MAC address of the AP is unique. It is not necessary to run the optimization algorithm for each new Wi-Fi measurement. We suggest to run the optimization when a loop closure is detected or a certain amount of loop closures has been identified. Solutions to reduce the computational time can be found in [@life_long_SLAM] [@Long_term_mapping]. One might notice that the model training described in Section\[sect\_model\_training\] takes long time. But this phase can be performed offline, and the learned model can be saved and applied to other users or different environments once it is generated. [|c|c|c|]{} &\ & &\ Data recording (time per track) & 544.75 & 544.75\ ---------------------- Model training+ variance computation ---------------------- : Evaluation of the computational time (in seconds) in each stage using two different pedestrian dead reckoning systems. []{data-label="table:computational_time"} & 6.08 & 23.58\ Loop closure detection & 5.73 & 72.34\ Turning detection and matching & 5.93 & 31.82\ Pose graph optimization & 0.07 & 1.65\ Conclusions and Future Work {#conclusions} =========================== In this paper, we presented a novel approach for collaborative simultaneous localization and radio fingerprint mapping (C-SLAM-RF) in unknown environments. The proposed system makes use of a pedestrian dead reckoning system and the RSS measurement from surrounding wireless access points. We further incorporate the motion features to improve the accuracy of the system. The proposed approach does not require any knowledge of the map and locations of the access points. The performance of our approach is evaluated in a large scale environment under two pedestrian dead reckoning systems with different motion tracking accuracies. Our results reveal that the accuracy of a pedestrian tracking system plays an important role in the accuracy of our approach. We obtained an accuracy of 0.6m and 4.76m for Tango and step counter-based pedestrian dead reckoning systems, respectively. The quality of the radio map will increase with more users involved in collecting the measurements due to the crowdsourcing nature of the proposed approach. One of our future work is to enhance the accuracy of step counter-based PDR by stride length estimation and the fusion of gyroscope measurement. Another direction would be the evaluation of the indoor positioning accuracy by applying the radio map constructed from our SLAM system. [^1]: This work is partly supported by the National Science Foundation of China (No. 61601381, 61750110529, and 61701421) and the Sichuan Science and Technology Program (No. 2019YFH0161 and 2019JDTD0019). [^2]: R. Liu, S. H. Marakkalage, C. Yuen, and U-X. Tan are with the Engineering Product Development Pillar, Singapore University of Technology and Design, 8 Somapah Rd, Singapore, 487372. R. Liu is also with the School of Information Engineering, Southwest University of Science and Technology, 59 Qinglong Road, Mianyang, China, 621010. [{ran\_liu, yuenchau, uxuan\_tan}@sutd.edu.sg]{}. [^3]: M. Padmal and T. Shaganan are with the Department of Electronic and Telecommunication Engineering, University of Moratuwa, Sri Lanka, 10400. [^4]: Y. L. Guan is with the School of Electrical and Electronic Engineering, Nanyang Technological University, 50 Nanyang Avenue, Singapore, 639798 [eylguan@ntu.edu.sg]{}. [^5]: https://github.com/RainerKuemmerle/g2o [^6]: http://standards-oui.ieee.org/oui.txt

ArXiv

--- abstract: '[We extend the definitions of dyadic paraproduct and $t$-Haar multipliers to dyadic operators that depend on the complexity $(m,n)$, for $m$ and $n$ natural numbers. We use the ideas developed by Nazarov and Volberg to prove that the weighted $L^2(w)$-norm of a paraproduct with complexity $(m,n)$, associated to a function $b\in BMO^d$, depends linearly on the $A^d_2$-characteristic of the weight $w$, linearly on the $BMO^d$-norm of $b$, and polynomially on the complexity. This argument provides a new proof of the linear bound for the dyadic paraproduct due to Beznosova. We also prove that the $L^2$-norm of a $t$-Haar multiplier for any $t\in\mathbb{R}$ and weight $w$ is a multiple of the square root of the $C^d_{2t}$-characteristic of $w$ times the square root of the $A^d_2$-characteristic of $w^{2t}$, and is polynomial in the complexity. ]{}' address: - 'Instituto de Matemática, Universidade Federal do Rio Grande do Sul, Av Bento Gonçalves 9500, 91501-970, Caixa Postal 15080, Porto Alegre, RS, Brazil' - 'Department of Mathematics and Statistics, 1 University of New Mexico, Albuquerque, NM 87131-001, MSC01 1115' author: - JEAN CARLO MORAES - MARÍA CRISTINA PEREYRA title: '**WEIGHTED ESTIMATES FOR DYADIC PARAPRODUCTS AND $t$-HAAR MULTIPLIERS WITH COMPLEXITY $(m,n)$** ' --- [^1] Introduction ============ In the past decade, many mathematicians have devoted their attention to finding out how the norm of an operator $T$ on a weighted space $L^p(w)$ depends on the so called $A_p$-characteristic of the weight $w$. More precisely, is there some optimal growth function $\varphi: [0,\infty)\to \BBR$ such that for all functions $f\in L^p(w)$, $$\|Tf\|_{L^p(w)} \leq C_{p,T}\varphi([w]_{A_p} )\|f\|_{L^p(w)},$$ where $C_{p,T}>0$ is a suitable constant? The first result of this type was due to Buckley [@Bu] in 1993; he showed that $\varphi (t)= t^{1/(p-1)}$ for the Hardy-Littlewood maximal function. Starting in 2000, one at a time, some dyadic model operators and some important singular integral operators (Beurling, Hilbert and Riesz transforms) were shown to obey a linear bound with respect to the $A_2$-characteristic of $w $ in $L^2(w)$, meaning that for $p=2$, the function $\varphi (t) = t$ is the optimal one, see [@W; @W1; @HukTV; @PetV; @Pet2; @Pet3; @Be1]. These linear estimates in $L^2(w)$ imply $L^p(w)$-bounds for $1<p<\infty$, by the sharp extrapolation theorem of Dragičevič, Grafakos, Pereyra, and Petermichl,[@DGPPet]. All these papers used the Bellman function technique, see [@V] for more insights and references. The linear bound for $H$, the Hilbert transform, is based on a representation of $H$ as an average of Haar shift operators of complexity $(0,1)$, see [@Pet1]. Haar shift operators with complexity $(m,n)$ were introduced in [@LPetR]. Hytönen obtained a representation valid for *any* Calderón-Zygmund operator as an average of Haar shift operators of *arbitrary complexity*, paraproducts and their adjoints, and used this representation to prove the $A_2$-conjecture, see [@H]. Thus, he showed that for [*all*]{} Calderón-Zygmund operators $T$ in $\mathbb{R}^N$, and all weights $w\in A_p$, there is a constant $C_{p,N,T} >0$ such that, $$\|Tf\|_{L^p(w)}\leq C_{p,N,T} [w]_{A_p}^{\max\{1,{1}/{p-1}\}}\|f\|_{L^p(w)}.$$ See [@L1] for a survey of the $A_2$-conjecture including a rather complete history of most results that appeared up to November 2010, and that contributed to the final resolution of this mathematical puzzle. A crucial part of the proof was to obtain bounds for Haar shifts operators that depended linearly on the $A_2$-characteristic and at most polynomially on the complexity $(m,n)$. In 2011, Nazarov and Volberg [@NV] provided a beautiful new proof that still uses Bellman functions, although minimally, and that can be transferred to geometric doubling metric spaces [@NV1; @NRezV]. Treil [@T], independently [@HLM+] obtained linear dependence on the complexity. Similar Bellman function techniques have been used to prove the Bump Conjecture in $L^2$, see [@NRezTV]. It seems natural to study other dyadic operators with complexity $(m,n)$, and examine if we can recover the same dependence on the $A_2$-characteristic that we have for the original operator (the one with complexity $(0,0)$) times a factor that depends at most polynomially on the complexity of these operators. We will do this analysis for the dyadic paraproduct and for the $t$-Haar multipliers. For $b\in BMO^d$, a function of dyadic bounded mean oscillation, $m,n\in \mathbb{N}$, the [*dyadic paraproduct of complexity $(m,n)$*]{} is defined by, $$\pi_b^{m,n}f(x)= \sum_{L\in\mathcal{D}}\sum_{\substack{I\in\mathcal{D}_n(L)\\ J\in\mathcal{D}_m(L)}} c^L_{I,J}m_If\, \langle b,h_I\rangle h_J(x),$$ where $ |c^L_{I,J}| \leq {\sqrt{|I|\,|J|}}/{|L|}$, and $m_If$ is the average of $f$ on the interval $I$. Here $\mathcal{D}$ denotes the dyadic intervals, $|I|$ the length of interval $I$, $\mathcal{D}_m(L)$ denotes the dyadic subintervals of $L$ of length $2^{-m}|L|$, $h_I$ are the Haar functions, and $\langle f,g\rangle$ denotes the $L^2$-inner product on $\BBR$. We prove that the dyadic paraproduct of complexity $(m,n)$ obeys the same *linear bound* as obtained by Beznosova [@Be1] for the dyadic paraproduct of complexity $(0,0)$ (see [@Ch1] for the result in $\BBR^N$, $N>1$), multiplied by a factor that depends polynomially on the complexity. \[thm:paracplxty(m,n)\] If $w\in A^d_2$, $b\in BMO^d$, then $$\|\pi_b^{m,n}f\|_{L^2(w)}\leq C(m+n+2)^5[w]_{A^d_2}\|b\|_{BMO^d}\|f\|_{L^2(w)}.$$ Our proof of Theorem \[thm:paracplxty(m,n)\] shows how to use the ideas in [@NV] for this setting, explicitly displaying the dependence on $\|b\|_{BMO^d}$ and bypassing the more complicated Sawyer two-weight testing conditions present in other arguments [@HPzTV; @L1; @HLM+]. From our point view, this makes the proof more transparent. For $t\in\mathbb{R}$, $m,n\in \mathbb{N}$, and weight $w$, the [*$t$-Haar multiplier of complexity $(m,n)$*]{} is defined by $$T_{t,w}^{m,n}f(x)= \sum_{L\in\mathcal{D}} \sum_{\substack{I\in\mathcal{D}_n(L)\\ J\in\mathcal{D}_m(L)}} c^L_{I,J}\, \frac{w^t(x)}{(m_Lw)^t}\langle f,h_I\rangle h_J(x),$$ where $|c^L_{I,J}|\leq {\sqrt{|I|\, |J|}}/{|L|}$. When $(m,n)=(0,0)$ and $c^L_{I,J}=c^L=1$ for all $L\in\mathcal{D}$, we denote the corresponding Haar multiplier by $T^t_w$. In addition, if $t=1$, we denote the multiplier simply by $T_w$. A necessary condition for the boundedness of $T_w^t$ on $L^2(\BBR )$ is that $w\in C^d_{2t}$, that is, $$[w]_{C^d_{2t}}:=\sup_{I\in\mathcal{D}} \Big (\frac{1}{|I|}\int_I w^{2t}(x)dx \Big )\Big (\frac{1}{|I|}\int_I w(x)dx\Big )^{-2t}<\infty.$$ This condition is also sufficient for $t<0$ and $t > 1/2$. For $0\leq t \leq 1/2$ the condition $C^d_{2t}$ is always fulfilled; in this case, boundedness of $T^t_w$ is known when $w\in A^d_{\infty}$, see [@KP]. The Haar multipliers $T_w$ are closely related to the resolvent of the dyadic paraproduct [@P], and appeared in the study of Sobolev spaces on Lipschitz curves [@P3]. It was proved in [@P2] that the $L^2$-norm for the Haar multiplier $T_w$ depends linearly on the $C_2^d$-characteristic of the weight $w$. We show the following theorem that generalizes a result of Beznosova for $T_w^t$ [@Be Chapter 5]. If $w\in C^d_{2t}$ and $w^{2t}\in A_2^d$, then $$\|T_{t,w}^{m,n}f\|_2\leq C (m+n+2)^3[w]_{C_{2t}}^{\frac{1}{2}}[w^{2t}]^{\frac{1}{2}}_{A^d_2}\|f\|_2.$$ The condition $w\in C^d_{2t}$ is necessary for the boundedness of $T_{t,w}^{m,n}$ when $c^L_{I,J}= \sqrt{|I| |J|}/|L|$. The result is optimal for $T^{\pm 1/2}_w$, see [@Be; @P2] and [@BeMoP]. We expect that, for both the paraproducts and $t$-Haar multipliers with complexity $(m,n)$, the dependence on the complexity can be strengthened to be linear, in line with the best results for the Haar shift operators. However our methods yield polynomials of degree 5 and 3 respectively. To simplify notation, and to shorten the exposition we analyze the one-dimensional case. Some of the building blocks in our arguments can be found in the literature in the case of ${{\mathbb R}}^N$, or even in the geometric doubling metric space case. As we go along we will note where such results can be found. For a complete presentation of these results in the geometric doubling metric spaces (in particular in ${{\mathbb R}}^N$) see [@Mo1]. The paper is organized as follows. In Section 2 we provide the basic definitions and results that are used throughout this paper. In Section 3 we prove the lemmas that are essential for the main results. In Section 4 we prove the main estimate for the dyadic paraproduct with complexity $(m,n)$ and present a new proof of the linear bound for the dyadic paraproduct. In Section 5 we prove the main estimate for the $t$-Haar multipliers with complexity $(m,n)$, also discussing necessary conditions for these operators to be bounded in $L^p(\mathbb{R})$, for $1 < p < \infty$. .1in [**Acknowledgements:**]{} The authors would like to thank Carlos Pérez, Rafael Espinola and Carmen Ortiz-Caraballo for organizing the Doc- course: *Harmonic analysis, metric spaces and applications to PDE*, held in Seville, at the Instituto de Matemáticas de la Universidad de Sevilla (IMUS) during the Summer of 2011. We are grateful to our thoughtful referees who pointed out multiple ways for improving this paper. Preliminaries ============= Weights, maximal function and dyadic intervals ---------------------------------------------- A weight $w$ is a locally integrable function in $\mathbb{R}^N$ taking values in $(0,\infty)$ almost everywhere. The $w$-measure of a measurable set $E$, denoted by $w(E)$, is $ w(E)= \int_E w(x)dx.$ For a measure $\sigma$, $ \sigma(E) = \int_{E} d\sigma$, and $|E|$ stands for the Lebesgue measure of $E$. We define $m^{\sigma}_E f$ to be the integral average of $f$ on $E$, with respect to $\sigma$, $$m^{\sigma}_E f := \frac{1}{\sigma(E)} \int_E f(x) d\sigma.$$ When $d\sigma=dx$ we simply write $m_Ef$; when $d\sigma = v\,dx$ we write $m_E^v f$. Given a weight $w$, a measurable function $f: \BBR^N\to\BBC$ is in $L^p(w)$ if and only if $\|f\|_{L^p(w)}:= \left (\int_{\mathbb{R}} |f(x)|^pw(x)dx \right )^{1/p}<\infty$. For a weight $v$ we define the [*weighted maximal function of $f$*]{} by $$(M_vf)(x) := \sup_{Q \ni x} m_Q^v |f|,$$ where $Q$ is a cube in $\BBR^N$ with sides parallel to the axes. The operator $M_v$ is bounded in $L^q(v)$ for all $q>1$. Furthermore, $$\label{bddmaxfct} \|M_v f\|_{L^q(v)} \leq C_N q' \|f\|_{L^q(v)},$$ where $q'$ is the dual exponent of $q$, that is ${1}/{q} + {1}/{q'}=1$. A proof of this fact can be found in [@CrMPz1]. When $v=1$, $M_v$ is the usual Hardy-Littlewood maximal function, which we will denote by $M$. It is well-known that $M$ is bounded on $L^p(w)$ if and only if $w\in A_p$ [@Mu]. We work with the collection of all [*dyadic intervals*]{}, $\mathcal{D}$, given by: $\mathcal{D}= \cup_{n\in\mathbb{Z}}\mathcal{D}_n$, where $\mathcal{D}_n:=\{I \subset \mathbb{R} : I=[k2^{-n}, (k+1)2^{-n} ), \; k \in \mathbb{Z}\}.$ For a dyadic interval $L$ , let $\mathcal{D}(L)$ be the collection of its dyadic subintervals, $\mathcal{D}(L):=\{I \subset L : I \in \mathcal{D}\} ,$ and let $\mathcal{D}_n(L)$ be the $n^{th}$-*generation* of dyadic subintervals of $L$, $\mathcal{D}_n(L):=\{I \in \mathcal{D}(L) : |I|=2^{-n}|L| \}.$ Any two dyadic intervals $I, J \in \mathcal{D}$ are either disjoint or one is contained in the other. Any two distinct dyadic intervals $I, J \in \mathcal{D}_n$ are disjoint, furthermore $\mathcal{D}_n$ is a partition of $\mathbb{R}$, and $\mathcal{D}_n(L)$ is a partition of $L$. For every dyadic interval $I \in \mathcal{D}_n$ there is exactly one $\widehat{I} \in \mathcal{D}_{n-1}$, such that $I \subset \widehat {I}$; $\widehat{I}$ is called the *parent of $I$*. Each dyadic interval $I$ in $\mathcal{D}_n$ is the union of two disjoint intervals in $\mathcal{D}_{n+1}$, the right and left halves, denoted $I_+$ and $I_-$ respectively, and called the *children* of $I$. A weight $w$ is [*dyadic doubling*]{} if ${w(\widehat{I})}/{w(I)} \leq C\;$ for all $\; I \in \mathcal{D}$. The smallest constant $C$ is called the doubling constant of $w$ and is denoted by $D(w)$. Note that $D(w)\geq 2$, and that in fact the ratio between the length of a child and the length of its parent is comparable to one; more precisely, $ D(w)^{-1}\leq {w(I)}/{w(\widehat{I})}\leq 1- D(w)^{-1}$. Dyadic $A^d_p$, reverse Hölder $RH_p^d$ and $C_s^d$ classes ----------------------------------------------------------- A weight $w$ is said to belong to the [*dyadic Muckenhoupt $A_p^d$-class*]{} if and only if $$[w]_{A_p^d}:= \sup_{I\in \mathcal{D}} ({ m_Iw })(m_Iw^{\frac{-1}{p-1}})^{p-1} <\infty,\quad\quad\mbox{for}\quad 1< p<\infty ,$$ where $[w]_{A_p^d}$ is called the $A_p^d$-characteristic of the weight. If a weight is in $A_p^d$ then it is dyadic doubling. These classes are nested: $A_p^d\subset A_q^d$ for all $p\leq q$. The class $A^d_{\infty}$ is defined by $A^d_{\infty}:= \bigcup_{p>1}A_p^d$. A weight $w$ is said to belong to the [*dyadic reverse Hölder $RH_p^d$-class*]{} if and only if $$[w]_{RH_p^d}:= \sup_{I\in \mathcal{D}}(m_Iw^p)^{\frac{1}{p}}(m_Iw)^{-1}<\infty, \quad \quad\mbox{for}\quad 1<p<\infty,$$ where $[w]_{RH_p^d}$ is called the $RH_p^d$-characteristic of the weight. If a weight is in $RH_p^d$ then it is not necessarily dyadic doubling (in the non-dyadic setting reverse Hölder weights are always doubling). Also these classes are nested, $RH_p^d\subset RH_q^d$ for all $p\geq q$. The class $RH^d_1$ is defined by $RH^d_1:= \bigcup_{p>1}RH_p^d$. In the non-dyadic setting $A_{\infty}=RH_1$. In the dyadic setting the collection of dyadic doubling weights in $RH_1^d$ is $A_{\infty}^d$, hence $A_{\infty}^d$ is a proper subset of $RH_1^d$. See [@BeRez] for some recent and very interesting results relating these classes. Given $s \in\mathbb{R}$, a weight $w$ is said to satisfy the [*$C^d_s$-condition*]{} if $$[w]_{C^d_s}:= \sup_{I \in \mathcal{D}} \big (m_I w^s\big )\,\big ({ m_Iw }\big )^{-s} < \infty.$$ The quantity defined above is called the $C^d_s$-characteristic of $w$. The class $C^d_s$ was defined in [@KP]. Let us analyze this definition. For $0 \leq s \leq 1$, we have that any weight satisfies the condition with $C_s^d$-characteristic $1$, being just a consequence of Hölder’s Inequality (cases $s=0,1$ are trivial). When $s>1$, the condition is equivalent to the dyadic reverse Hölder condition and $[w]^{{1}/{s}}_{C^d_s} =[w]_{RH^d_s}.$ For $s<0$, we have that $ w \in C^d_s$ if and only if $w \in A^d_{1- 1/s}.$ Moreover $[w]_{C^d_s} = [w]^{-s}_{A^d_{1-1/s}}$. Weighted Haar functions ----------------------- For a given weight $v$ and an interval $I$ define the corresponding [*weighted Haar function*]{} by $$\label{def:Haarfunction} h^v_I(x)= \frac{1}{v(I)}\left ( \sqrt{\frac{v(I_-)}{v(I_+)}}\,\chi_{I_+}(x)-\sqrt{\frac{v(I_+)}{v(I_-)}} \,\chi_{I_-}(x)\right ),$$ where $\chi_I$ is the characteristic function of the interval $I$. If $v$ is the Lebesgue measure on $\mathbb{R}$, we will denote the [*Haar function*]{} simply by $h_I$. It is an important fact that $\{h^v_I\}_{I\in \mathcal{D}}$ is an orthonormal system in $L^2(v)$, with the inner product $\langle f, g\rangle_v = \int_{\mathbb{R}} f(x)\,\overline{g(x)}\,v(x) dx$. It is a simple exercise to verify that the weighted and unweighted Haar functions are related linearly as follows: \[whaarbasis\] For any weight $v$, there are numbers $\alpha_I^v$, $\beta^v_I$ such that $$h_I(x) = \alpha^v_I \,h^v_I(x) + \beta_I^v \,{\chi_I(x)}/{\sqrt{|I|}}$$ where [(i)]{} $|\alpha^v_I | \leq \sqrt{m_Iv},$ [ (ii)]{} $|\beta^v_I| \leq {|\Delta_I v|}/{m_Iv},$ $\Delta_I v:= m_{I_+}v - m_{I_-}v.$ For a weight $v$ and a dyadic interval $I$, ${|\Delta_I v|}/{m_I v} =2 \Big| 1 - {m_{I_-}v}/{m_{I}v} \Big| \leq 2.$ If the weight $v$ is dyadic doubling then we get an improvement on the above upper bound, ${|\Delta_I v|}/{m_I v} \leq 2\left (1-{2}/{D(v)}\right ).$ Dyadic BMO and Carleson sequences --------------------------------- A locally integrable function $b$ is a function of [*dyadic bounded mean oscillation*]{}, $b \in BMO^d$, if and only if $$\label{def:BMO} \|b\|_{BMO^d}:= \Big (\sup_{J \in \mathcal{D}} \frac{1}{|J|} \sum_{I \in \mathcal{D}(J)} | \langle b , h_I \rangle|^2 \Big )^{\frac{1}{2}}< \infty.$$ Note that if $\displaystyle{b_I := \langle b , h_I \rangle }$ then $|b_I|\,|I|^{-\frac{1}{2}} \leq \|b\|_{BMO^d}$, for all $ \; I \; \in \mathcal{D}$. If $v$ is a weight, a positive sequence $\{\alpha_I\}_{I\in \mathcal{D}}$ is called a [*$v$-Carleson sequence with intensity $B$*]{} if for all $J\in \mathcal{D}$, $$\label{def:vCarlesonseq} ({1}/{|J|}) \sum_{I \in \mathcal{D}(J)} {\lambda_I}\leq B\; m_Jv.$$ When $v=1$ we call a sequence satisfying for all ${J \in \mathcal{D}} $ a [*Carleson sequence with intensity*]{} $B$. If $b \in BMO^d$ then $\{ |b_I |^2\}_{I\in \mathcal{D}} $ is a Carleson sequence with intensity $\|b\|^2_{BMO^d} $. \[algcarseq\] Let $v$ be a weight, $\{\lambda_I\}_{I\in \mathcal{D}}$ and $\{\gamma_I\}_{I\in \mathcal{D}}$ be two $v$-Carleson sequences with intensities $A$ and $B$ respectively then for any $c, d >0$ we have that - $ \{c \lambda_I + d\gamma_I\}_{I\in \mathcal{D}}$ is a $v$-Carleson sequence with intensity $cA + dB$. - $\{ \sqrt{\lambda_I} \sqrt{\gamma_I}\}_{I\in \mathcal{D}}$ is a $v$-Carleson sequence with intensity $\sqrt{AB}$. - $ \{( c\sqrt{\lambda_I} + d \sqrt{\gamma_I})^{2}\}_{I\in \mathcal{D}}$ is a $v$-Carleson sequence with intensity $2c^2A+2d^2B$. The proof of these statements is quite simple. To prove the first one we just need properties of the supremum, for the second one we apply the Cauchy-Schwarz inequality, and the third one is a consequence of the first two statements combined with the fact that $2cd\sqrt{A}\sqrt{B} \leq c^2A + d^2B. $ Main tools ========== In this section, we state and prove the lemmas and theorems necessary to obtain the estimates for the paraproduct and the $t$-Haar multipliers of complexity $(m,n)$. The Weighted Carleson Lemma \[weightedCarlesonLem\], $\alpha$-Lemma \[alphacoro\] and Lift Lemma \[liftlem\] are fundamental for all our estimates. Carleson Lemmas --------------- We present some weighted Carleson lemmas that we will use. Lemma \[folklem\] was introduced and used in [@NV], it was called a folklore lemma in reference to the likelihood of having been known before. Here we obtain Lemma \[folklem\] as an immediate corollary of the Weighted Carleson Lemma \[weightedCarlesonLem\] and what we call the Little Lemma \[litlem\], introduced by Beznosova in her proof of the linear bound for the dyadic paraproduct. ### Weighted Carleson Lemma The Weighted Carleson Lemma we present here is a variation in the spirit of other weighted Carleson embedding theorems that appeared before in the literature [@NV; @NTV1]. All the lemmas in this section hold in ${{\mathbb R}}^N$ or even geometric doubling metric spaces, see [@Ch1; @NRezV]. \[weightedCarlesonLem\] Let $v$ be a dyadic doubling weight, then $\{\alpha_{L}\}_{L \in \mathcal{D}}$ is a $v$-Carleson sequence with intensity $B$ if and only if for all non-negative $v$-measurable functions $F$ on the line, $$\label{eqn:WCL} \sum_{L \in \mathcal{D}}\alpha_{L} \inf_{x \in L} F(x) \leq B \int_{\mathbb{R}}F(x) \,v(x)\,dx.$$ ($\Rightarrow$) Assume that $F \in L^1(v)$ otherwise the first statement is automatically true. Setting $\displaystyle{\gamma_L = \inf_{x \in L}F(x)}$, we can write $$\sum_{L \in \mathcal{D}} \gamma_L \alpha_L = \sum_{L \in \mathcal{D}} \int^{\infty}_{0} \chi(L,t)\,dt\;\alpha_L = \int_0^{\infty}\Big (\sum_{L\in\mathcal{D}} \chi (L,t)\,\alpha_L \Big )dt,$$ where $\chi(L,t)=1$ for $ t < \gamma_L$ and zero otherwise, and the last equality follows by the monotone convergence theorem. Define $E_t= \{ x \in \mathbb{R} \; :\; F(x)>t\}$. Since $F$ is assumed to be a $v$-measurable function, $E_t$ is a $v$-measurable set for every $t$. Moreover, since $F \in L^1(v)$ we have, by Chebychev’s inequality, that the $v$-measure of $E_t$ is finite for all real $t$. If $\;\chi(L,t)=1$ then $L \subset E_t$. Moreover, there is a collection of maximal disjoint dyadic intervals $\mathcal{P}_t$ that are contained in $E_t$. Then we can write $$\sum_{L \in \mathcal{D}} \chi(L,t) \alpha_L \leq \sum_{L \subset E_t} \alpha_L = \sum_{L \in \mathcal{P}_t} \sum_{I \in \mathcal{D}(L)} \alpha_I \leq B \sum_{L \in \mathcal{P}_t}v(L) \leq Bv(E_t),$$ where, in the second inequality, we used the fact that $\{ \alpha _I \}_{I \in \mathcal{D}}$ is a $v$-Carleson sequence with intensity $B$. Thus we can estimate $$\sum_{L \in \mathcal{D}} \gamma_L \alpha_L \leq B \int^{\infty}_{0} v(E_t ) dt = B \int_{\mathbb{R}} F(x)\,v(x)\,dx.$$ The last equality follows from the layer cake representation.\ ($\Leftarrow$) Assume (\[eqn:WCL\]) is true; in particular it holds for $F(x)={\chi_J(x)}/{|J|}$. Since $\inf_{x\in I} F(x)= 0$ if $I\cup J =\emptyset$, and $\inf_{x\in I} F(x)= {1}/{|J|}$ otherwise, $$\frac{1}{|J|}\sum_{I\in\mathcal{D}(J)} \alpha_I \leq \sum_{I\in\mathcal{D}} \alpha_I\inf_{x\in I} F(x) \leq \int_{\mathbb{R}} F(x)\,v(x)\,dx= m_J v.$$ ### Little Lemma The following Lemma was proved by Beznosova in [@Be1] using the Bellman function $\displaystyle{B(u,v,l)= u -{1}/{v(1+l)}}$. \[litlem\] Let $v$ be a weight, such that $v^{-1}$ is a a weight as well, and let $\{ \lambda_I \}_{I \in \mathcal{D}}$ be a Carleson sequence with intensity $B$. Then, $\{{ \lambda_I}/{m_Iv^{-1}} \}_{I \in \mathcal{D}}$ is a $v$-Carleson sequence with intensity $4B$, that is for all $J\in \mathcal{D}$, $$({1}/{|J|}) \sum_{I \in \mathcal{D}(J)} {\lambda_I}/{m_I{v^{-1}}}\leq 4B \; m_Jv.$$ For a proof of this result we refer [@Be Prop. 3.4], or [@Be1 Prop. 2.1]. For an ${{\mathbb R}}^N$ version of this result see [@Ch1 Prop 4.6]. Lemma \[litlem\] together with Lemma \[weightedCarlesonLem\] immediately yield the following: \[folklem\] Let $v$ be a weight such that $v^{-1}$ is also a weight. Let $\{\lambda_{J}\}_{J \in \mathcal{D}}$ be a Carleson sequence with intensity $B$, and let $F$ be a non-negative measurable function on the line. Then, $$\sum_{J \in \mathcal{D}}({ \lambda_{J} }/{m_J v^{-1}}) \, \inf_{x \in J} F(x) \leq C \;B \int_{\mathbb{R}}F(x)\,v(x)\,dx.$$ Note that Lemma \[litlem\] can be deduced from Lemma \[folklem\] with $F(x)=\chi_J(x)$. $\alpha$-Lemma -------------- The following lemma was proved by Beznosova for $\alpha={1}/{4}$ in [@Be], and by Nazarov and Volberg for $0<\alpha <1/2$ in [@NV], using the Bellman function $\displaystyle{B(u,v)= (uv)^{\alpha}}$. \[alphacoro\] Let $w\in A_2^d$ and then for any $\alpha \in (0,{1}/{2})$, the sequence $\{ \mu^{\alpha}_I \}_{I \in \mathcal{D}}$, where $$\mu^{\alpha}_I:= ({ m_Iw })^{\alpha}(m_I{w^{-1}})^{\alpha}|I| \bigg( {\frac{|\Delta_I w|^2}{(m_I w)^2}}+ {\frac{|\Delta_I w^{-1}|^2}{(m_I w^{\scriptscriptstyle{-1}})^2}}\bigg),$$ is a Carleson sequence with intensity $C_{\alpha}[w]_{A_2}^{\alpha}$, with $C_{\alpha}={72}/({\alpha-2\alpha^2})$. A proof of this lemma that works in ${{\mathbb R}}^N$ (for $\alpha=1/4$) can be found in [@Ch1 Prop. 4.8], and one that works in geometric doubling metric spaces can be found in [@NV1; @V]. The following lemmas simplify the exposition of the main theorems (this was pointed to us by one of our referees). We deduce these lemmas from the $\alpha$-Lemma. According to our kind anonymous referee, one can also deduce Lemma \[coro:referee\] from a pure Bellman-function argument without reference to the $\alpha$-Lemma. \[coro:referee\] Let $w\in A_2^d$ and let $\nu_I = |I| (m_Iw)^2(\Delta_I w^{-1})^2$. The sequence $\{\nu_I\}_{I\in\mathcal{D}}$ is a Carleson sequence with intensity at most $C[w]_{A_2^d}^2$ for some numerical constant $C$ *(*$C=288$ works*)*. Multiply and divide $\nu_I$ by $(m_I w^{-1})^2$ to get for any $0<\alpha <1/2$, $$\nu_I = |I| (m_Iw)^2 (m_I w^{-1})^2 {\big (|\Delta_I w^{-1}|/m_I w^{-1}\big )^2} \leq [w]_{A_2}^{2-\alpha}\mu^{\alpha}_I.$$ But $\{\mu^{\alpha}_I\}_{\mathcal{D}}$ is a Carleson sequence with intensity $C_{\alpha} [w]_{A_2}^{\alpha} $ by Lemma \[alphacoro\], therefore by Proposition \[algcarseq\](i) $\{\nu_I\}_{\mathcal{D}}$ is a Carleson sequence with intensity at most $C_{\alpha}[w]_{A_2^d}^2$ as claimed. It is well known that if $w\in A_2^d$ then $\{ |I||\Delta_Iw|^2/ (m_Iw)^2\}_{I\in \mathcal{D}}$ is a Carleson sequence with intensity $\log [w]_{A_2^d}$, see [@W]. This estimate together with Proposition \[algcarseq\](i), give intensities $[w]_{A_2^d}^{\alpha}\log [w]_{A_2^d}$ and $[w]_{A_2^d}^{2}\log [w]_{A_2^d}$ respectively for the sequences $\{\mu_I^{\alpha}\}_{I\in\mathcal{D}}$ and $\{\nu_I\}_{I\in\mathcal{D}}$. The lemmas show we can improve the intensities by dropping the logarithmic factor. Even more generally, we can show the following lemma, which extends the $\alpha$-Lemma \[alphacoro\] to the range $\alpha\geq 1/2$. It also refines it for the range $\alpha \in (1/4,1/2)$ and shows that the blow up of the constant $C_{\alpha}$ for $\alpha=1/2$ is an artifact of the proof. \[lem:A2square\] Let $w\in A_2^d$, $s>0$, and $$\tau^{s}_I := |I| (m_Iw)^{s}(m_Iw^{-1})^{s}\bigg( {\frac{|\Delta_I w|^2}{(m_I w)^2}}+ {\frac{|\Delta_I w^{-1}|^2}{(m_I w^{\scriptscriptstyle{-1}})^2}}\bigg).$$ Then for $0< \alpha < \min\{1/2, s\}$, the sequence $\{\tau^s_I\}_{I\in\mathcal{D}}$ is a Carleson sequence with intensity at most $C_{\alpha}[w]_{A_2^d}^s$ where $C_{\alpha}$ is the constant in Lemma \[alphacoro\] *(*when $s>1/4$ can choose $\alpha =1/4$ and $C_{\alpha}=576.$*)* Lift Lemma ---------- Given a dyadic interval $L$, and weights $u,v$, we introduce a family of stopping time intervals $\mathcal {ST}^m_L$ such that the averages of the weights over any stopping time interval $K \in \mathcal{ST}^m_L$ are comparable to the averages on $L$, and $|K|\geq 2^{-m}|L|$. This construction appeared in [@NV] for the case $u=w$, $v=w^{-1}$. We also present a lemma that lifts $w$-Carleson sequences on intervals to $w$-Carleson sequences on “$m$-stopping intervals". We present the proofs for the convenience of the reader. \[liftlem\] Let $u$ and $v$ be weights, $L$ be a dyadic interval and $m,n $ be fixed natural numbers. Let $\mathcal{ST}^m_L$ be the collection of maximal stopping time intervals $K \in \mathcal{D}(L)$, where the stopping criteria are either [*(i)*]{} $\; |\Delta_Ku|/m_Ku + |\Delta_Kv|/m_Kv \geq {1}/{(m+n+2)}$, or [*(ii)*]{} $\,|K| = 2^{-m}|L|$. Then for any stopping interval $K\in \mathcal{ST}^m_L$, $\, e^{-1}m_Lu \leq m_Ku \leq e\,m_Lu $, also $\, e^{-1}m_Lv \leq m_Kv \leq e\,m_Lv $. Note that the roles of $m$ and $n$ can be interchanged and we get the family $\mathcal{ST}^n_L$ using the same stopping condition (i) as above, but with (ii) replaced by $|K|=2^{-n}|L|$. Notice that $\mathcal{ST}^m_L$ is a partition of $L$ in dyadic subintervals of length at least $2^{-m}|L|$. Any collection of subintervals of $L$ with this property will be an *$m$-stopping time* for $L$. Let $K$ be a maximal stopping time interval; thus no dyadic interval strictly bigger than $K$ can satisfy either stopping criteria. If $F$ is a dyadic interval strictly bigger than $K$ and contained in $L$, then necessarily $ {|\Delta_F u|}/{m_F u}\leq (m+n+2)^{-1}$ and ${|\Delta_F v|}/{m_F v} \leq (m+n+2)^{-1}$. This is particularly true for the parent of $K$. Let us denote by $\widehat{K}$ the parent of $K$, then $ |m_Ku - m_{\widehat{K}}u| = {|\Delta_{\widehat{K}}u|}/{2} \leq {m_{\widehat{K}} u}/{2(m+n+2)}.$ So, $ m_{\widehat{K}} u \big(1 - {1}/{2(m+n+2)} \big) \leq m_Ku \leq m_{\widehat{K}} u \big(1 + {1}/{2(m+n+2)} \big).$ Iterating this process until we reach $L$, we will get that $$m_Lu \bigg(1 - \frac{1}{2(m+n+2)} \bigg)^m \leq m_Ku \leq m_Lu \bigg(1 + \frac{1}{2(m+n+2)} \bigg)^m.$$ Remember that $|K| = 2^{-j}|L|$ where $ 0 \leq j \leq m$ so we will iterate at most $m$ times. We can obtain the same bounds for $v$. These clearly imply the estimates in the lemma, since $\lim_{k\to\infty}(1+1/k)^k=e$. The following lemma lifts a $w$-Carleson sequence to $m$-stopping time intervals with comparable intensity. The lemma appeared in [@NV] for the particular stopping time $\mathcal{ST}^m_L$ given by the stopping criteria (i) and (ii) in Lemma \[liftlem\], and $w=1$. This is a property of any stopping time that stops once the $m^{th}$-generation is reached. \[corliftlemstop\] For each $L \in \mathcal{D}$, let $\mathcal{ST}^m_L$ be a partition of $L$ in dyadic subintervals of length at least $2^{-m}|L|$ (in particular it could be the stopping time intervals defined in Lemma \[liftlem\]). Assume $\{\nu_I \}_{I \in \mathcal{D}} $ is a $w$-Carleson sequence with intensity at most $A$, let $\nu^m_L := \sum _{K \in \mathcal{ST}^m_L} \nu_K$. Then $\{\nu_L^m\}_{L \in \mathcal{D}}$ is a $w$-Carleson sequence with intensity at most $(m+1)A$. In order to show that $\{\nu^m_L\}_{L\in \mathcal{D}}$ is a $w$-Carleson sequence with intensity at most $(m+1)A$, it is enough to show that for any $J \in \mathcal{D}$ $$\sum_{L \in \mathcal{D}(J)} \nu^m_L < (m+1)A \,w(J).$$ Observe that for each dyadic interval $K$ inside a fixed dyadic interval $J$ there exist at most $m+1$ dyadic intervals $L$ such that $K \in \mathcal{ST}^m_L$. Let us denote by $K^i$ the dyadic interval that contains $K$ and such that $|K^i|=2^i|K|$. If $K \in \mathcal{D}(J)$ then $L$ must be $K^0, K^1, ... \; \mbox{or} \; K^m$. We just have to notice that if $L=K^i$, for $i>m$ then $K$ cannot be in $\mathcal{ST}^{m}_L$ because $|K| < 2^{-m}|L|$. Therefore, $$\begin{aligned} \sum_{L \in \mathcal{D}(J)} \nu^m_L =& \sum_{L \in \mathcal{D}(J)} \sum_{K \in \mathcal{ST}^m_L} \nu_K = \sum_{K \in \mathcal{D}(J)} \sum_{L \in \mathcal{D}(J) s.t. K \in \mathcal{ST}^m_L} \nu_K\\ &\leq \sum_{K \in \mathcal{D}(J)} (m+1)\nu_k \leq (m+1)A \, w(J).\end{aligned}$$ The last inequality follows by the definition of $w$-Carleson sequence with intensity $A$. The lemma is proved. Paraproduct =========== For $b\in BMO^d$, and $m,n\in\mathbb{N}$, a [*dyadic paraproduct of complexity $(m,n)$*]{} is the operator defined by $$\label{paraproduct(m,n)} \big( \pi^{m,n}_{b} f \big)(x) := \sum_{L \in \mathcal{D}} \sum _{(I,J)\in \mathcal{D}^n_m(L)} c^L_{I,J} m_I f \langle b,h_I\rangle h_J(x),$$ where $|c^L_{I,J}| \leq {\sqrt{|I|\,|J|}}/{|L|}$ for all dyadic intervals $L$ and $(I,J)\in \mathcal{D}^n_m(L)$, where $\mathcal{D}^n_m(L)= \mathcal{D}_n(L)\times \mathcal{D}_m(L)$. A dyadic paraproduct of complexity $(0,0)$ is the usual dyadic paraproduct $\pi_b$ known to be bounded in $L^p(\mathbb{R})$ if and only if $b\in BMO^d$. A *Haar shift operator of complexity $(m,n)$*, $m,n\in\mathbb{N}$, is defined by $$\big( S^{m,n} f \big)(x) := \sum_{L \in \mathcal{D}} \sum_{(I,J)\in \mathcal{D}^n_m(L)} c^L_{I,J} \langle f,h_I\rangle h_J(x),$$ where $ |c^L_{I,J}| \leq {\sqrt{|I|\,|J|}}/{|L|}$. Notice that the Haar shift operators are automatically uniformly bounded on $L^2(\mathbb{R} )$, with operator norm less than or equal to one [@LPetR; @CrMPz]. The dyadic paraproduct of complexity $(m,n)$ is the composition of $S^{m,n}$ and $\pi_b$. Therefore, if $b\in BMO^d$ then $\pi_b^{m,n}$ is bounded in $L^2(\mathbb{R} )$, since $\pi_b^{m,n}=S^{m,n}\pi_b$, and both $\pi_b$ (the dyadic paraproduct) and $S^{m,n}$ (the Haar shift operators) are bounded in $L^2(\mathbb{R})$. Furthermore, $\pi_b$ and $S^{m,n}$ are bounded in $L^2(w)$ whenever $w\in A^d_2$. Both of them obey bounds on $L^2(w)$ that are linear in the $A_2$-characteristic of the weight, immediately providing a quadratic bound in the $A_2$-characteristic of the weight for $\pi_b^{m,n}$. We will show that in fact, the dyadic paraproduct of complexity $(m,n)$ obeys the same *linear bound* in $L^2(w)$ with respect to $[w]_{A_2^d}$ obtained by Beznosova [@Be1] for the dyadic paraproduct of complexity $(0,0)$, multiplied by a polynomial factor that depends on the complexity. The proof given by Nazarov and Volberg, in [@NV], of the fact that Haar shift operators with complexity $(m,n)$ are bounded in $L^2(w)$ with a bound that depends linearly on the $A^d_2$-characteristic of $w$, and polynomially on the complexity, works, with appropriate modifications, for the dyadic paraproducts of complexity $(m,n)$. Below we describe those modifications. Beforehand, however, we will present this new and conceptually simpler (in our opinion) proof for the linear bound on the $A^d_2$-characteristic for the dyadic paraproduct, which will allow us to highlight certain elements of the general proof without dealing with the complexity. Complexity $(0,0)$ {#sec:paraproduct} ------------------ The dyadic paraproduct of complexity $(0,0)$ is defined by $(\pi_b f)(x):= \sum _{I \in \mathcal{D}}c_I \; m_I f \; \langle b,h_I\rangle h_I(x)$, [where]{} $ |c_I| \leq 1$. It is known that $\pi_b$ obeys a linear bound in $L^2(w)$ both in terms of the $A^d_2$-characteristic of the weight $w$ and the $BMO$-norm of $b$. \[thmA2paraproduct\] There exists $C>0$, such that for all $ b \in BMO^d$ and for all $w\in A^d_2$, $$\|\pi_bf\|_{L^2(w)} \leq C [w]_{A^d_2} \|b\|_{BMO^d}\|f\|_{L^2(w)}.$$ Beznosova’s proof is based on the $\alpha$-Lemma, the Little Lemma (these were the new Bellman function ingredients that she introduced), and Nazarov-Treil-Volberg’s two-weight Carleson embedding theorem, which can be found in [@NTV]. Below, we give another proof of this result; this proof is still based on the $\alpha$-Lemma \[alphacoro\] (via Lemma \[coro:referee\]) however it does not make use of the two-weight Carleson embedding theorem. Instead we will use properties of Carleson sequences such as the Little Lemma \[litlem\], and the Weighted Carleson Lemma \[weightedCarlesonLem\], following the argument in [@NV] for Haar shift operators of complexity $(m,n)$. The extension of Theorem \[thmA2paraproduct\] to ${{\mathbb R}}^N$ can be found in [@Ch1], and the methods used there can be adapted to extend our proof to $\mathbb{R}^N$ even in the complexity $(m,n)$ case, see [@Mo1]. \[remark:constants\]Throughout the proofs a constant $C$ will be a numerical constant that may change from line to line. Fix $f \in L^2(w)$ and $g \in L^2(w^{-1})$. Define $b_I = \langle b , h_I \rangle$, then $\{b^2_I\}_{I\in\mathcal{D}}$ is a Carleson sequence with intensity $\|b\|_{BMO^d}^2$. By duality, it suffices to prove: $$\label{dualityparaproduct} |\langle \pi_b(fw), gw^{-1}\rangle | \leq C \|b\|_{BMO^d} [w]_{A_2} \|f\|_{L^2(w)}\|g\|_{L^2(w^{-1})}.$$ Note that $\langle \pi_b(fw), gw^{-1}\rangle |= \big \langle \sum_{I \in \mathcal{D}}c_I b_I { m_I (fw) }h_I,gw^{-1} \big \rangle . $ Write $h_I = \alpha_I h^{{w^{\scriptscriptstyle{-1}}}}_I + \beta_I {\chi_I}/{\sqrt{|I|}}$ where $\alpha_I = \alpha^{{w^{\scriptscriptstyle{-1}}}}_I$ and $\beta_I = \beta^{{w^{\scriptscriptstyle{-1}}}}_I$ as described in Proposition \[whaarbasis\]. Then $$\label{dualitysumpara} |\langle \pi_b(fw), gw^{-1}\rangle | \leq \sum_{I \in \mathcal{D}}|b_I| { m_I (|f|w)}\big |\big \langle gw^{-1}, {\alpha_I}h_I^{w^{-1}} + {\beta_I}\frac{\chi_I}{\sqrt{|I|}} \big \rangle \big |.$$ Use the triangle inequality to break the sum in (\[dualitysumpara\]) into two sums to be estimated separately, $| \langle \pi_b(fw), gw^{-1}\rangle | \leq \;\Sigma_1 + \Sigma_2 .$ Where, using the estimates $|\alpha_I |\leq \sqrt{{ m_I w^{-1}}} $, and $|\beta_I |\leq |\Delta_Iw^{-1}|/m_Iw^{-1}$, $$\begin{aligned} \Sigma_1 &:=& \sum_{I \in \mathcal{D}} |b_I| { m_I (|f|w)}|\langle gw^{-1},h_I^{w^{-1}} \rangle| \sqrt{{ m_I w^{-1}}} \\ \Sigma_2 & := & \sum_{I \in \mathcal{D}} |b_I| { m_I (|f|w)}|\langle gw^{-1}, \chi_I \rangle| {\frac{|\Delta_I w^{-1}|}{m_I w^{\scriptscriptstyle{-1}}}}\frac{1}{\sqrt{|I|}}.\end{aligned}$$ [**Estimating $\Sigma_1$:**]{} First using that ${{m_I(|f|w)}/{{ m_Iw }} \leq \inf_{x\in I} M_wf (x)}$, and that $\langle gv,f\rangle = \langle g,f\rangle_v$; second using the Cauchy-Schwarz inequality and ${ m_Iw }\, { m_I w^{-1}}\leq [w]_{A^d_2}$, we get\ $$\begin{aligned} \Sigma_1 &\leq & \sum_{I \in \mathcal{D}} |b_I| \frac{ \inf_{x \in I} M_wf (x) }{\sqrt{{ m_I w^{-1}}}} \big |\langle g, h_I^{w^{-1}} \rangle_{w^{-1}} \big | \; { m_I w^{-1}}\; { m_Iw }\\ & \leq & [w]_{A^d_2} \bigg (\sum_{I \in \mathcal{D}} |b_I|^2 \frac{\inf_{x \in I} M^2_wf (x)}{{ m_I w^{-1}}} \bigg)^{\frac{1}{2}}\bigg (\sum_{I \in \mathcal{D}}\big | \langle g, h_I^{w^{-1}} \rangle_{w^{-1}}\big |^2 \bigg)^{\frac{1}{2}}.\end{aligned}$$ Using Weighted Carleson Lemma \[weightedCarlesonLem\], with $F(x)=M^2_w f(x)$, $v=w$, and $\alpha_I={|b_I|^2}/{m_Iw^{-1}}$ (which is a $w$-Carleson sequence with intensity $4\|b\|_{BMO^d}^2$, according to Lemma \[litlem\] ), together with the fact that $\{h_I^{w^{-1}}\}_{I\in \mathcal{D}}$ is an orthonormal system in $L^2(w^{-1})$, we get $$\begin{aligned} \Sigma_1&\leq & 4[w]_{A^d_2} \|b\|_{BMO^d} \bigg( \int_{\mathbb{R}} M^2_{w}f(x) w(x) dx \bigg)^{\frac{1}{2}} \|g\|_{L^2(w^{-1})} \\ &\leq &C [w]_{A^d_2} \|b\|_{BMO^d} \|f\|_{L^2(w)} \|g\|_{L^2(w^{-1})}.\end{aligned}$$ In the last inequality we used the fact that $M_w$ is bounded in $L^2(w)$ with operator norm independent of $w$. .1in [**Estimating $\Sigma_2$:**]{} Using arguments similar to the ones used for $\Sigma_1$, we conclude that, $$\Sigma_2 = \sum_{I \in \mathcal{D}} |b_I| { m^{w}_I |f|}\; { m^{w^{\scriptscriptstyle{-1}}}_I |g|}\; \sqrt{\nu_I} \leq \sum_{I \in \mathcal{D}} |b_I| \sqrt{\nu_I} \; \inf_{x \in I} M_wf (x) M_{{w^{\scriptscriptstyle{-1}}}}g(x),$$ where $\nu_I= |I| (m_Iw)^2(\Delta_Iw^{-1})^2$ as defined in Lemma \[coro:referee\], and in the last inequality we used that for any $I \in \mathcal{D}$ and all $x \in I$, $${ m^{w}_I |f|}\,{ m^{w^{\scriptscriptstyle{-1}}}_I |g|}\leq M_wf(x)M_{{w^{\scriptscriptstyle{-1}}}}g(x).$$ Since $\{|b_I|^2\}_{I \in \mathcal{D}}$ and $\{\nu_I\}_{I \in \mathcal{D}}$ are Carleson sequences with intensities $\|b\|^2_{BMO^d}$ and $C[w]^2_{A^d_2}$, respectively, by Proposition \[algcarseq\], the sequence $\{|b_I| \sqrt{\nu_I}\}_{I \in \mathcal{D}}$ is a Carleson sequence with intensity $C\|b\|_{BMO^d}[w]_{A^d_2}$. Thus, by Lemma \[weightedCarlesonLem\] with $F(x)= M_wf(x)M_{{w^{\scriptscriptstyle{-1}}}}g(x)$, $\alpha_I= |b_I|\sqrt{\nu_I}$, and $v=1$, $$\Sigma_2 \; \leq \; C \|b\|_{BMO^d}[w]_{A^d_2} \int_{\mathbb{R}} M_{w}f(x) M_{{w^{\scriptscriptstyle{-1}}}}g(x) \,dx.$$ Using the Cauchy-Schwarz inequality and $w^{\frac{1}{2}}(x)w^{\frac{-1}{2}}(x)=1$ we get $$\begin{aligned} \Sigma_2 &\leq & C [w]_{A^d_2} \|b\|_{BMO^d} \Big( \int_{\mathbb{R}} M^2_{w}f (x)w(x)dx \Big)^{\frac{1}{2}}\Big( \int_{\mathbb{R}} M^2_{{w^{\scriptscriptstyle{-1}}}}g (x){w^{\scriptscriptstyle{-1}}}(x)dx \Big)^{\frac{1}{2}}\\ &= & C [w]_{A^d_2} \|b\|_{BMO^d} \|M_w f \|_{L^2(w)}\|M_{{w^{\scriptscriptstyle{-1}}}}g \|_{L^2({w^{\scriptscriptstyle{-1}}})} \\ &\leq& C [w]_{A^d_2} \|b\|_{BMO^d} \|f \|_{L^2(w)}\|g \|_{L^2({w^{\scriptscriptstyle{-1}}})}.\end{aligned}$$ These estimates together give (\[dualityparaproduct\]), and the theorem is proved. Complexity $(m,n)$ ------------------ In this section, we prove an estimate for the dyadic paraproduct of complexity $(m,n)$ that is linear in the $A_2$-characteristic and polynomial in the complexity. The proof will follow the general lines of the argument presented in Section \[sec:paraproduct\] for the complexity $(0,0)$ case, with the added refinements devised by Nazarov and Volberg [@NV], adapted to our setting, to handle the general complexity. \[theoparcompmn\] Let $ b \in BMO^d$ and $w\in A^d_2$. Then there is $C>0$ such that $$\|\pi^{m,n}_bf\|_{L^2(w)} \leq C (n+m+2)^5[w]_{A^d_2} \|b\|_{BMO^d}\|f\|_{L^2(w)}.$$ Fix $f \in L^2(w)$ and $g \in L^2(w^{-1})$, define $b_I = \langle b , h_I \rangle$ and let $C^n_m := (m+n+2)$. By duality, it is enough to show that $$| \langle \pi_b^{m,n}(fw), gw^{-1}\rangle | \leq C(C^n_m)^5 [w]_{A^d_2}\|b\|_{BMO^d} \|g\|_{L^2({w^{\scriptscriptstyle{-1}}})} \|f\|_{L^2(w)}.$$ We write the left-hand side as a double sum, that we will estimate as $$| \langle \pi_b^{m,n}(fw), gw^{-1}\rangle | \leq \sum_{L \in \mathcal{D}} \sum_{(I,J)\in \mathcal{D}^n_m(L)} |b_I| \frac{\sqrt{|I|\,|J|}}{|L|} { m_I (|f|w)}|\langle gw^{-1}, h_J \rangle|.$$ As before, we write $h_J = \alpha_J h^{{w^{\scriptscriptstyle{-1}}}}_J + \beta_J {\chi_J}/{\sqrt{|J|}}$, with $\alpha_J = \alpha_J^{w^{-1}}$, $\beta_J=\beta_J^{w^{-1}}$, and break the double sum into two terms to be estimated separately. Then $\,| \langle \pi_b^{m,n}(fw), gw^{-1}\rangle | \leq \Sigma_1^{m,n} + \Sigma_2^{m,n},$ where $$\begin{aligned} \Sigma_1^{m,n}& := \sum_{L \in \mathcal{D}} \sum_{(I,J)\in \mathcal{D}^n_m(L)} |b_I| \frac{\sqrt{|I|\,|J|}}{|L|} { m_I (|f|w)}|\langle g,h_J^{w^{-1}}\rangle_{w^{-1}} | \sqrt{{ m_J w^{-1}}},\\ \Sigma_2^{m,n} & := \sum_{L \in \mathcal{D}} \sum_{(I,J)\in \mathcal{D}^n_m(L)} |b_I| \frac{\sqrt{|I|}}{|L|} { m_I (|f|w)}|\langle gw^{-1}, \chi_J \rangle| {\frac{|\Delta_J w^{-1}|}{m_J w^{\scriptscriptstyle{-1}}}}.\end{aligned}$$ For a weight $v$, and a locally integrable function $\phi $ we define the following quantities, $$\begin{aligned} S^{v,m}_L \phi &:= \sum_{J \in \mathcal{D}_m(L)} |\langle \phi,h_J^v \rangle_v| \sqrt{m_Jv}\sqrt{|J|/|L|}, \label{def:SvmLphi}\\ R^{v,m}_L \phi &:= \sum_{J \in \mathcal{D}_m(L)} \frac{|\Delta_J v|}{m_Jv} m_J(|\phi|v)\;{|J|}/{\sqrt{|L|}},\label{def:RvmLphi}\\ Pb^{v,n}_L \phi &:= \sum_{I \in \mathcal{D}_n(L)} |b_I| \; m_I(|\phi|v) \sqrt{|I|/|L|}.\label{def:PbvnLphi}\end{aligned}$$ For $s=1,2$ and $w\in A_2^d$, we also define the following Carleson sequences (see Lemma \[corliftlemstop\] and Lemma \[lem:A2square\]): $\displaystyle{\mu^s_K := ({ m_Kw })^{s}({ m_K w^{\scriptscriptstyle{-1}}})^{s} \bigg( {\frac{|\Delta_K w^{-1}|^2}{(m_K w^{\scriptscriptstyle{-1}})^2}}+{\frac{|\Delta_K w|^2}{(m_K w)^2}}\bigg)|K|}$,\ with intensity $C[w]^{s}_{A^d_2}$, $\displaystyle{\mu^{m,s}_L := \sum_{K \in \mathcal{ST}^m_L} \mu_K} $, with intensity $C(m+1)[w]^{s}_{A^d_2}$, $\displaystyle{\mu^{n,s}_L := \sum_{K \in \mathcal{ST}^n_L} \mu_K} $, with intensity $C(n+1)[w]^{s}_{A^d_2}$, $\displaystyle{\mu^{b,s}_K := {|b_K|^2}({ m_Kw }\;{ m_K w^{\scriptscriptstyle{-1}}})^{s}}$, with intensity $\|b\|^2_{BMO^d}[w]^{s}_{A^d_2}$, and $\displaystyle{\mu^{b,n,s}_L := \sum_{K \in \mathcal{ST}^n_L} \mu^{b,s}_K} $, with intensity $(n+1)\|b\|^2_{BMO^d}[w]^{s}_{A^d_2}$.\ Note that $$\Sigma_1^{m,n} \leq \sum_{L \in \mathcal{D}} Pb^{w,n}_L f \; S^{{w^{\scriptscriptstyle{-1}}},m}_L g \quad \text{and} \quad \Sigma_2^{m,n} \leq \sum_{L \in \mathcal{D}} Pb^{w,n}_L f \; R^{{w^{\scriptscriptstyle{-1}}},m}_L g.$$ In order to estimate $\Sigma_1^{m,n}$ and $\Sigma_2^{m,n}$ we will use the following estimates for , $S^{{w^{\scriptscriptstyle{-1}}},m}_L g$, $R^{{w^{\scriptscriptstyle{-1}}},m}_L g$ and $Pb^{w,n}_L f$, $$S^{{w^{\scriptscriptstyle{-1}}},m}_L g \leq \Big( \sum_{J \in \mathcal{D}_m(L)} |\langle g,h_J^{{w^{\scriptscriptstyle{-1}}}} \rangle_{{w^{\scriptscriptstyle{-1}}}}|^2 \Big)^{\frac{1}{2}} ({ m_L w^{\scriptscriptstyle{-1}}})^{\frac{1}{2}}, \label{Sestpar}$$ $$R^{{w^{\scriptscriptstyle{-1}}},m}_L g \leq C\, C^n_m ({ m_Lw })^{\frac{-s}{2}}({ m_L w^{\scriptscriptstyle{-1}}})^{1-\frac{s}{2}} \inf_{x \in L} \big (M_{{w^{\scriptscriptstyle{-1}}}}(|g|^p)(x)\big )^{\frac{1}{p}} \sqrt{\mu^{m,s}_L}, \label{Restpar}$$ $$Pb^{w,n}_L f \leq C\, C^n_m ({ m_Lw })^{1-\frac{s}{2}}({ m_L w^{\scriptscriptstyle{-1}}})^{\frac{-s}{2}} \inf_{x \in L} \big (M_{w}(|f|^p)(x)\big )^{\frac{1}{p}}\nu_L^{n,s},\label{Pbestpar}$$ where $\nu_L^{n,s}=\|b\|_{BMO^d}\sqrt{\mu^{n,s}_L} + \sqrt{\mu^{b,n,s}_L}$, and $\displaystyle{p = 2 - (C_m^n)^{-1}}$ (note that $1<p<2$). In the proof it will become clear why this is a good choice; the reader is invited to assume first that $p=2$ and reach a point of no return in the argument. Estimate (\[Sestpar\]) is easy to show. We just use the Cauchy-Schwarz inequality and the fact that $\mathcal{D}_m(L)$ is a partition of $L$. $$S^{{w^{\scriptscriptstyle{-1}}},m}_L g \leq \Big( \sum_{J \in \mathcal{D}_m(L)} |\langle g ,h_J^{{w^{\scriptscriptstyle{-1}}}} \rangle_{{w^{\scriptscriptstyle{-1}}}}|^2 \Big)^{\frac{1}{2}} \big({ m_L w^{\scriptscriptstyle{-1}}}\big)^{\frac{1}{2}}.$$ Estimate (\[Restpar\]) was obtained in [@NV]. With a variation on their argument we prove estimate (\[Pbestpar\]) in Lemma \[Pblempar\]. Let us first use estimates , and to estimate $\Sigma_1^{m,n}$ and $\Sigma_2^{m,n}$.\ [**Estimate for $\Sigma_1^{m,n}$:**]{} Use estimates and with $s=2$, the Cauchy-Schwarz inequality and the fact that $\{h_J^{w^{-1}}\}_{J\in\mathcal{D}}$ is an orthonormal system in $L^2(w^{-1})$ and $\mathcal{D}=\cup_{L\in\mathcal{D}}\mathcal{D}_m(L)$. Then $$\Sigma_1^{m,n} \leq C\,C^n_m \Big(\sum_{L \in \mathcal{D}} \frac {(\nu^{n,2}_L)^2}{{ m_L w^{\scriptscriptstyle{-1}}}}\inf_{x \in L}\big( M_{w}(|f|^p)(x)\big)^{\frac{2}{p}}\Big)^{\frac{1}{2}} \|g\|_{L^2(w^{-1})}.$$ We will now use the Weighted Carleson Lemma \[weightedCarlesonLem\] with $F(x)= \big (M_w(|f|^p)(x)\big )^{2/p}$, $v=w$, and $\alpha_L={(\nu_L^{n,2} )^2 }/{m_Lw^{-1}}$. Recall that $\nu^{n,2}_L =\|b\|_{BMO^d}\sqrt{\mu^{n,2}_L} + \sqrt{\mu^{b,n,2}_L}$. By Proposition \[algcarseq\], $\{(\nu_L^{n,2})^2\}_{L\in\mathcal{D}}$ is a Carleson sequence with intensity at most $C\,C^n_m\|b\|^2_{BMO^d}[w]^{2}_{A^d_2}$. By Lemma \[litlem\], $\{(\nu_L^{n,2} )^2 /m_Lw^{-1} \}_{L\in\mathcal{D}}$ is a $w$-Carleson sequence with comparable intensity. Thus we will have that $$\begin{aligned} \Sigma_1^{m,n} \leq& \;C\,(C^n_m)^{\frac{3}{2}} [w]_{A^d_2}\|b\|_{BMO^d} \|g\|_{L^2({w^{\scriptscriptstyle{-1}}})} \Big \| M_{w}(|f|^p) \Big \|^{\frac{1}{p}}_{L^{\frac{2}{p}}(w)}\\ \leq& \; C \big[({2}/{p})'\big]^{\frac{1}{p}}(C^n_m)^{\frac{3}{2}} [w]_{A^d_2}\|b\|_{BMO^d} \|g\|_{L^2({w^{\scriptscriptstyle{-1}}})} \big \| \, |f|^p \big \|^{\frac{1}{p}}_{L^{\frac{2}{p}}(w)}\\ =& \;C (C_m^n)^{\frac{5}{2}} [w]_{A^d_2}\|b\|_{BMO^d} \|g\|_{L^2({w^{\scriptscriptstyle{-1}}})} \|f\|_{L^2(w)}.\end{aligned}$$ We used in the first inequality that $M_w$ is bounded in $L^q(w)$ for all $q>1$, more specifically we used that $\|M_w f\|_{L^q(w)} \leq C q' \|f\|_{L^q(w)}$. In our case ${q={2}/{p}}$ and $q'= {2}/{(2-p)}=2C^n_m$.\ [**Estimate for $\Sigma_2^{m,n}$:**]{} Use estimates and with $s=1$ in both cases, together with the facts that $(m_Iw\, m_Iw^{-1} )^{-1}\leq 1$, and that the product of the infimum of positive quantities is smaller than the infimum of the product. Then $$\Sigma_2^{m,n} \leq C(C^n_m)^2 \sum_{L \in \mathcal{D}} \nu^{n,1}_L\sqrt{\mu^{m,1}_L}\inf_{x \in L} \big (M_{w}(|f|^p)(x)\big )^{\frac{1}{p}} \big (M_{{w^{\scriptscriptstyle{-1}}}}(|g|^p)(x)\big )^{\frac{1}{p}}.$$ Since $(\nu^{n,1}_L)^2$ and $\mu^{m,1}_L$ have intensity at most $C(n+1)[w]_{A^d_2}\|b\|^2_{BMO}$ and $C(m+1)[w]_{A^d_2}$, by Proposition \[algcarseq\], we have that $\nu_L^{n,1} \sqrt{\mu^{m,1}_L}$ is a Carleson sequence with intensity at most $C\,C^m_n\|b\|_{BMO^d}[w]_{A^d_2}$. If we now apply Lemma \[weightedCarlesonLem\] with $F^p(x)= M_w(|f|^p)(x) M_{w^{-1}}(|g|^p)(x)$, $\alpha_L= \nu_L^{n,1}\sqrt{\mu_L^{m,1}}$, and $v=1$, we will have, by the Cauchy-Schwarz inequality and the boundedness of $M_v$ in $L^q(v)$ for $q=p/2>1$, $$\begin{aligned} \Sigma_2^{m,n} \leq& \; C(C^n_m)^3 [w]_{A^d_2} \|b\|_{BMO^d} \int_{\mathbb{R}} \big (M_{w}(|f|^p)(x)\big )^{\frac{1}{p}}\big (M_{{w^{\scriptscriptstyle{-1}}}}(|g|^p)(x)\big )^{\frac{1}{p}} dx\\ \leq& \; C(C^n_m)^3 [w]_{A^d_2} \|b\|_{BMO^d}\big\| M_{w}(|f|^p) \big \|^{\frac{1}{p}}_{L^{\frac{2}{p}}(w)}\big \|M_{{w^{\scriptscriptstyle{-1}}}}(|g|^p) \big\|^{\frac{1}{p}}_{L^{\frac{2}{p}}({w^{\scriptscriptstyle{-1}}})}\\ \leq& \; C\big[({2}/{p})'\big]^{\frac{2}{p}}(C^n_m)^3 [w]_{A^d_2} \|b\|_{BMO^d}\big \| |f|^p \big\|^{\frac{1}{p}}_{L^{\frac{2}{p}}(w)}\big \| |g|^p \big \|^{\frac{1}{p}}_{L^{\frac{2}{p}}({w^{\scriptscriptstyle{-1}}})}\\ =& \;C (C^n_m)^5 [w]_{A^d_2} \|b\|_{BMO^d}\|f\|_{L^2(w)} \|g\|_{L^2({w^{\scriptscriptstyle{-1}}})}.\end{aligned}$$ Together these estimates prove the theorem, under the assumption that estimate (\[Pbestpar\]) holds. Key Lemma --------- The missing step in the previous proof is estimate (\[Pbestpar\]), which we now prove. The argument we present is an adaptation of the argument used in [@NV] to obtain estimate (\[Restpar\]). \[Pblempar\] Let $b\in BMO^d$, and let $\phi$ be a locally integrable function. Then, $$Pb^{w,n}_L \phi \leq C\,C_m^n ({ m_Lw })^{1-\frac{s}{2}}({ m_L w^{\scriptscriptstyle{-1}}})^{\frac{-s}{2}} \inf_{x \in L}\big ( M_{w}(|\phi |^p)(x)\big )^{\frac{1}{p}} \nu_L^{n,s},$$ where $\nu_L^{n,s}=\|b\|_{BMO^d}\sqrt{\mu^{n,s}_L} + \sqrt{\mu^{b,n,s}_L}$, and $p=2-(C^n_m)^{-1}$. Let $\mathcal{ST}^n_L$ be the collection of stopping time intervals defined in Lemma \[liftlem\]. Noting that $\mathcal{D}_n(L)=\cup_{K\in\mathcal{ST}_L^n}\big (\mathcal{D}(K)\cap \mathcal{D}_n(L)\big )$, we get, $$Pb^{w,n}_L \phi = \sum_{K \in \mathcal{ST}^{n}_L} \sum_{I \in \mathcal{D}(K) \bigcap \mathcal{D}_n(L)} {|b_I|}\; m_I(|\phi |w) \sqrt{|I|/|L|}.$$ Note that if $K$ is a stopping time interval by the first criterion then $$\begin{aligned} Pb^{w,n}_L \phi & \leq & \|b\|_{BMO^d}\, m_K(|\phi |w) {|K|}/{\sqrt{|L|}} \\ & \leq & C_m^n \|b\|_{BMO^d}\,m_K (|\phi |w) (\sqrt{|K|/|L|})\sqrt{2\mu^s_K} \,({ m_Kw }\, { m_K w^{\scriptscriptstyle{-1}}})^{\frac{-s}{2}}.\end{aligned}$$ The first inequality is true because ${|b_I|}/{\sqrt{|I|}} \leq \|b\|_{BMO^d}$ and the second one because $$1 \leq C_m^n \bigg ({\frac{|\Delta_K w|}{m_K w}}+ {\frac{|\Delta_K w^{-1}|}{m_K w^{\scriptscriptstyle{-1}}}}\bigg )\sqrt{|K|}\leq C_m^n \sqrt{2\mu^{s}_K}({ m_Kw }\;{ m_K w^{\scriptscriptstyle{-1}}})^{\frac{-s}{2}}.$$ Now we use the fact, proved in Lemma \[liftlem\], that we can compare the averages of the weights on the stopping intervals with their averages in $L$, paying a price of a constant $e$, and continue estimating by $$\sqrt{2}C^m_n e^{s}\|b\|_{BMO^d}m_K (|\phi |w) \sqrt{|K|/|L|}\sqrt{\mu^s_K} ({ m_Lw }\, { m_L w^{\scriptscriptstyle{-1}}})^{\frac{-s}{2}}.$$ If $K$ is a stopping time interval by the second criterion, then the sum collapses to just one term $$\begin{aligned} \sum_{I \in \mathcal{D}(K) \bigcap \mathcal{D}_n(L)}|b_I| \; & m_I(|\phi | w){\sqrt{|I|/|L|}} \; = \; {|b_K|} \; m_K(|\phi |w) {\sqrt{|K|/|L|}} \\ =\; & \; m_K (|\phi |w) \sqrt{|K|/|L|}\sqrt{\mu^{b,s}_K}\, ({ m_Kw }\, { m_K w^{\scriptscriptstyle{-1}}})^{\frac{-s}{2}}\\ \leq\; & \; C^m_n e^{s}m_K (|\phi |w) \sqrt{|K|/|L|}\sqrt{\mu^{b,s}_K}\, ({ m_Lw }\, { m_L w^{\scriptscriptstyle{-1}}})^{\frac{-s}{2}}.\end{aligned}$$ Let $\Xi_1 (L): = \{ K \in \mathcal{ST}^n_L : K \text{ is a stopping time interval by criterion 1} \}$, and $\Xi_2(L) := \{ K \in \mathcal{ST}^n_L : K \text{ is a stopping time interval by criterion 2} \}.$ Note that $\Xi_1(L) \bigcup \Xi_2(L)$ is a partition of $L$. We then have, $$\label{PbLwn} Pb^{w,n}_L \leq \sqrt{2} C^m_n e^{s}\; ({ m_Lw }\, { m_L w^{\scriptscriptstyle{-1}}})^{\frac{-s}{2}} \Big(\|b\|_{BMO^d} \Sigma^1_{Pb} + \Sigma^2_{Pb} \Big),$$ where the terms $\Sigma^1_{Pb}$ and $\Sigma^2_{Pb}$ are defined as follows, $$\begin{aligned} \Sigma^1_{Pb} & := &\sum_{K \in \Xi_1(L)} m_K (|\phi |w){\sqrt{|K|/|L|}}\sqrt{\mu^s_K}, \\ \Sigma^2_{Pb} & := & \sum_{K \in \Xi_2(L)} m_K (|\phi |w) {\sqrt{|K/|L|}}\sqrt{\mu^{b,s}_K}.\end{aligned}$$ Now estimate $\Sigma^1_{Pb}$ using the Cauchy-Schwarz inequality, noting that we can move a power ${p}/{2} <1$ from outside to inside the sum, and that $\mu^{n,s}_L: = \sum_{I \in \mathcal{ST}^n_L} \mu^s_K \geq \sum_{I \in \Xi_1(L)} \mu^s_K$, $$\begin{aligned} \Sigma^1_{Pb} &\leq& \Big (\sum_{K \in \Xi_1(L)} (m_K (|\phi |w))^2 {|K|}/{|L|}\Big)^{\frac{1}{2}} \Big(\sum_{K \in \Xi_1(L)} \mu^s_K\Big)^{\frac{1}{2}} \nonumber \\ &\leq& \Big(\sum_{K \in \Xi_1(L)} (m_K (|\phi |w))^p \big({|K|}/{|L|}\big)^{\frac{p}{2}}\Big)^{\frac{1}{p}} \sqrt{\mu^{n,s}_L} \label{sigmapb}.\end{aligned}$$ By the second stopping criterion $|K|/{|L|}= 2^{-j}$ for $0 \leq j \leq m$, then $$\label{whatever} \big({|K|}/{|L|}\big)^{\frac{p}{2}} = 2^{-j + \frac{j}{2(m+n+2)}} < 2\cdot 2^{-j}= 2{|K|}/{|L|}.$$ Plugging into gives $$\Sigma^1_{Pb} \leq \Big(2\sum_{K \in \Xi_1(L)} (m_K (|\phi |w))^p {|K|}/{|L|}\Big)^{\frac{1}{p}}\sqrt{\mu^{n,s}_L}.$$ Use Hölder’s inequality inside the sum, then Lift Lemma \[liftlem\], to get $$\begin{aligned} \Sigma^1_{Pb} &\leq& \Big (\sum_{K \in \Xi_1(L)} (m_K (|\phi |^p w)) ({ m_Kw })^{p-1} {|K|}/{|L|}\Big)^{\frac{1}{p}}\sqrt{\mu^{n,s}_L} \\ &\leq& 2^{\frac{1}{p}} (e\,{ m_Lw })^{1-\frac{1}{p}} \Big(\frac{1}{|L|}\sum_{K \in \Xi_1(L)} \int_K |\phi (x)|^p w(x)\, dx \Big)^{\frac{1}{p}}\sqrt{\mu^{n,s}_L}.\end{aligned}$$ Observe that the intervals $K\in \Xi_1(L)$ are disjoint subintervals of $L$, therefore, $\sum_{K \in \Xi_1(L)} \int_K |\phi (x)|^p w(x)\, dx \leq \int_L |\phi (x)|^p w(x)\, dx$, thus, $$\Sigma^1_{Pb} \leq 2e\, { m_Lw }\inf_{x \in L} \big (M_w (|\phi |^p)(x) \big)^{\frac{1}{p}}\sqrt{\mu^{n,s}_L}. \label{S1Pb}$$ Similarly we estimate $\Sigma^2_{Pb}$, to get, $$\begin{aligned} \Sigma^2_{Pb} &\leq \Big(\sum_{K \in \Xi_2} (m_K (|\phi |w))^2 {|K|}/{|L|}\Big)^{\frac{1}{2}} \Big(\sum_{K \in \Xi_2} \mu^{b,s}_K\Big)^{\frac{1}{2}}\\ &\leq \; \Big(\sum_{K \in \mathcal{ST}^n_L} (m_K (|\phi |w))^p \big({|K|/|L|}\big)^{\frac{p}{2}}\Big)^{\frac{1}{p}} \sqrt{\mu^{b,n,s}_L}.\end{aligned}$$ Following the same steps as we did in the estimate for $\Sigma^1_{Pb}$, we will have $$\label{S2Pb} \Sigma^2_{Pb} \leq 2e \, { m_Lw }\inf_{x \in L} \Big(M_w (|\phi |^p)(x) \Big)^{\frac{1}{p}} \sqrt{\mu^{b,n,s}_L}.$$ Insert estimates (\[S1Pb\]) and (\[S2Pb\]) into (\[PbLwn\]). Altogether, we can bound $ Pb^{w,n}_L$ by $$C\, C^m_n e^{s+1}\; ({ m_Lw })^{1-\frac{s}{2}} ({ m_L w^{\scriptscriptstyle{-1}}})^{\frac{-s}{2}}\inf_{x \in L} \Big(M_w (|\phi |^p)(x) \Big)^{\frac{1}{p}}\nu^{n,s}_L.$$ The lemma is proved. In [@NV1], Nazarov and Volberg extend the results that they had for Haar shift operators in [@NV] to metric spaces with geometric doubling. One can extend Theorem \[theoparcompmn\] to this setting as well, see [@Mo1]. Haar Multipliers ================ For a weight $w$, $t\in\mathbb{R}$, and $m,n\in \mathbb{N}$, a [*$t$-Haar multiplier of complexity $(m,n)$*]{} is the operator defined as $$T^{m,n}_{t,w} f (x) := \sum_{L \in \mathcal{D}} \sum_{(I,J)\in \mathcal{D}^n_m(L)} c^L_{I,J}\, \Big(\frac{w(x)}{m_L w}\Big)^t \langle f,h_I\rangle h_J(x),$$ where $|c^L_{I,J} |\leq {\sqrt{|I|\,|J|}}/{|L|}$. Note that these operators have symbols, namely $c^L_{I,J} \big({w(x)}/{m_L w}\big)^t$, that depend on: the space variable $x$, the frequency encoded in the dyadic interval $L$, and the complexity encoded in the subintervals $I\in\mathcal{D}_n(L)$ and $J\in\mathcal{D}_m(L)$. This makes these operators akin to pseudodifferential operators where the trigonometric functions have been replaced by the Haar functions. Observe that $T_{t,w}^{m,n}$ is different from both $S^{m,n}T^t_w$ and $T^t_wS^{m,n}$ and, that, unlike $T^{m,n}_{t,w}$, both $S^{m,n}T^t_w$ and $T^t_wS^{m,n}$ obey the same bound that $T^t_w$ obeys in $L^2(\mathbb{R})$, because the Haar shift multipliers have $L^2$-norm less than or equal to one. Necessary conditions -------------------- Let us first show a necessary condition on the weight $w$ so that the Haar multiplier $T^{m,n}_{w,t}$ with $c^L_{I,J}=\sqrt{|I|\,|J|}/|L|$ is bounded on $L^p(\mathbb{R})$. This necessary $C^d_{tp}$-condition is the same condition found in [@KP] for the $t$-Haar multiplier of complexity $(0,0)$. \[neccondhaarmult\] Let $w$ be a weight, $m,n$ positive integers and $t$ a real number. If $T^{m,n}_{t,w}$ is the $t$-Haar multiplier with $c^L_{I,J}=\sqrt{|I|\,|J|}/|L|$ and is a bounded operator in $L^p(\mathbb{R})$, then $w$ is in $C^d_{tp}.$ Assume that $T^{m,n}_{t,w}$ is bounded in $L^p(\mathbb{R})$ for $1<p<\infty$. Then there exists $C>0$ such that for any $f \in L^p(\mathbb{R})$ we have $\|T^{m,n}_{t,w}f\|_{p}\leq C\|f\|_{p}. $ Thus for any $I_0 \in \mathcal{D}$ we should have $$\label{proofsufcond1} \|T^{m,n}_{t,w}h_{I_0}\|^p_{p}\leq C^p\|h_{I_0}\|^p_{p}.$$ Let us compute the norm on the left-hand side of (\[proofsufcond1\]). Observe that $$\label{proofsufcond3} T^{m,n}_{t,w}h_{I_0}(x) = \sum_{L \in \mathcal{D}} \sum_{(I,J)\in \mathcal{D}^n_m(L)}{\sqrt{|I|\,|J|}}/{|L|} \Big(\frac{w(x)}{m_L w}\Big)^t \langle h_{I_{0}},h_I\rangle h_J(x).$$ We have $\langle h_{I_{0}},h_I\rangle = 1$ if $I_{0}=I$ and $\langle h_{I_{0}},h_I\rangle = 0$ otherwise. Also, there exists just one dyadic interval $L_{0}$ such that $I_{0} \subset L_{0}$ and $|I_{0}| = 2^{-n}|L_{0}|$. Therefore we can collapse the sums in to just one sum, and calculate the $L^p$-norm as follows, $$\|T^{m,n}_{t,w}h_{I_0}\|^p_{p} = \int_{\mathbb{R}}\Big| \sum _{ J \in \mathcal{D}_m(L_{0})} {\sqrt{|I_{0}|\, |J|}}/{|L_{0}|}\Big(\frac{w(x)}{m_{L_{0}} w}\Big)^t h_J(x) \Big|^p dx.$$ Furthermore, since $\mathcal{D}_m(L_0)$ is a partition of $L_0$, the power $p$ can be put inside the sum, and we get, $$\label{proofsufcond4} \|T^{m,n}_{t,w}h_{I_0}\|^p_{p} = \big ({|I_{0}|^{\frac{p}{2}}}/{|L_{0}|^{p-1}}\big ) \big ({m_{L_{0}}w^{tp}}/{(m_{L_0}w)^{pt}}\big ).$$ Inserting $\|h_{I_0}\|^p_{p}=|I_{0}|^{1-\frac{p}{2}}$ and in , we will have that for any dyadic interval $I_{0}$ there exists $C$ such that $$\big ({|I_{0}|^{\frac{p}{2}}}/{|L_{0}|^{p-1}}\big ) \big ({m_{L_{0}}w^{tp}}/{(m_{L_0}w)^{pt}}\big ) \leq C^p |I_{0}|^{1-\frac{p}{2}}.$$ Thus, $ { m_{L_{0}}w^{tp}}/{(m_{L_0}w)^{pt}} \leq C^p |I_{0}|^{1-p}|L_{0}|^{p-1} = C^p 2^{n(p-1)}=:C_{n,p}.$ Now observe that this inequality should hold for any $L_{0} \in \mathcal{D}$, we just have to choose as $I_{0}$ any of the descendants of $L_{0}$ in the $n$-th generation, and that $n$ is fixed. Therefore, $${\; [w]_{C^d_{2t}}=\sup_{L \in \mathcal{D}} (m_{L}w^{tp} ) ({ m_Lw })^{-pt} \leq C_{n,p}.}$$ We conclude that $\; w \in C^d_{tp}$; moreover $\displaystyle{\;[w]_{C^d_{tp}} \leq 2^{n(p-1)} ||T^{m,n}_{t,w}||^p_{p}.}$ Sufficient condition -------------------- For most $t\in{{\mathbb R}}$, the $C^d_{2t}$-condition is not only necessary but also sufficient for a $t$-Haar multiplier of complexity $(m,n)$ to be bounded on $L^2(\mathbb{R} )$; this was proved in [@KP] for the case $m=n=0$. Here we are concerned not only with the boundedness but also with the dependence of the operator norm on the $C^d_{2t}$-constant. For the case $m=n=0$ and $t=1, \pm 1/2$ this was studied in [@P2]. Beznosova [@Be] was able to obtain estimates, under the additional condition on the weight: $w^{2t}\in A^d_{p}$ for some $p>1$, for the case of complexity $(0,0)$ and for all $t\in\mathbb{R}$. We generalize her results when $w^{2t}\in A^d_2$ for complexity $(m,n)$. Our proof differs from hers in that we are adapting the methods of Nazarov and Volberg [@NV] to this setting as well. Both proofs rely on the $\alpha$-Lemma (Lemma \[alphacoro\]) and on the Little Lemma (Lemma \[litlem\]). See also [@BeMoP]. \[sufcondhaarmult\] Let $t$ be a real number and $w$ a weight in $C^d_{2t}$, such that $w^{2t} \in A^d_2$. Then $T^{m,n}_{t,w}$, a $t$-Haar multiplier with depth $(m,n)$, is bounded in $L^2(\mathbb{R} )$. Moreover, $$\|T^{m,n}_{t,w}f \|_{2} \leq C(m+n+2)^3 [w]^{\frac{1}{2}}_{C^d_{2t}}[w^{2t}]^{\frac{1}{2}}_{A^d_2}\|f\|_2.$$ Fix $f,\; g \in L^2(\mathbb{R})$. By duality, it is enough to show that $$| \langle T^{m,n}_{t,w}f,g \rangle | \leq C (m+n+2)^3 [w]^{\frac{1}{2}}_{C^d_{2t}}[w^{2t}]^{\frac{1}{2}}_{A^d_2}\|f\|_{2} \|g\|_{2}.$$ The inner product on the left-hand-side can be expanded into a double sum that we now estimate, $$| \langle T^{m,n}_{t,w}f,g \rangle | \leq \sum_{L \in \mathcal{D}} \sum_{(I,J)\in \mathcal{D}^n_m(L)} ({\sqrt{|I|\,|J|}}/{|L|})\; \frac{ |\langle f,h_I\rangle|}{(m_L w)^t}\;| \langle gw^t , h_J \rangle|.$$ Decompose $h_J$ into a linear combination of a weighted Haar function and a characteristic function, $h_J = \alpha_J h^{w^{2t}}_J + \beta_J {\chi_J}/{\sqrt{|J|}}$, where $\alpha_J = \alpha^{w^{2t}}_J$, $\beta_J = \beta^{w^{2t}}_J$, $|\alpha_J|\leq \sqrt{m_Jw^{2t}}$, and $|\beta_J|\leq {|\Delta_J(w^{2t})|}/{m_Jw^{2t}}$. Now we break this sum into two terms to be estimated separately so that, $$| \langle T^{m,n}_{t,w}f,g \rangle | \leq \Sigma_3^{m,n} + \Sigma_4^{m,n},$$ where $$\begin{aligned} \Sigma_3^{m,n} :=& \sum_{L \in \mathcal{D}} \sum_{(I,J)\in \mathcal{D}^n_m(L)} \frac{\sqrt{|I|\,|J|}}{|L|} \frac{\sqrt{m_{J}(w^{2t})}}{(m_L w)^t} |\langle f,h_I\rangle| \; | \langle gw^{t} , h^{w^{2t}}_J \rangle |,\\ \Sigma_4^{m,n}:=& \sum_{L \in \mathcal{D}} \sum_{(I,J)\in \mathcal{D}^n_m(L)} \frac{|J| \sqrt{|I|} }{|L|(m_L w)^t} \frac{|\Delta_J (w^{2t})|}{m_{J}(w^{2t})} |\langle f,h_I\rangle| \; m_J( |g| w^t ) .\end{aligned}$$ Again, let $p = 2 - (C_n^m)^{-1}$, and define as in (\[def:SvmLphi\]) and (\[def:RvmLphi\]), the quantities $S_L^{v,m}\phi$ and $R_L^{v,m}\phi $, with $v=w^{2t}$. Let $$P^n_L \phi := \sum_{I \in \mathcal{D}_n(L)} \; |\langle f, h_I \rangle| {\sqrt{|I|/|L|}},$$ and $$\eta_I := m_I (w^{2t}) \; m_I (w^{-2t}) \bigg( \frac{|\Delta_I(w^{2t})|^2}{|m_I w^{2t}|^2} + \frac{|\Delta_I(w^{-2t})|^2}{|m_I w^{-2t}|^2} \bigg)|I|.$$ By Lemma \[lem:A2square\] with $s=1$, $\{\eta_I\}_{I\in\mathcal{D}}$ is a Carleson sequence with intensity $C[w^{2t}]_{A^d_2}$. Let ${\eta^m_L := \sum_{I \in \mathcal{ST}_L^m} \eta_I},$ where the stopping time $\mathcal{ST}_L^m$ is defined as in Lemma \[liftlem\] (with respect to the weight $w^{2t}$). By Lemma \[corliftlemstop\], $\{\eta_L^m\}_{L\in\mathcal{D}}$ is a Carleson sequence with intensity $C(m+1)[w^{2t}]_{A^d_2}$. Observe that on the one hand $\langle gw^t,h^{w^{2t}}_J\rangle = \langle gw^{-t},h^{w^{2t}}_J \rangle_{w^{2t}} $, and on the other $m_J(|g|w^t)=m_J(|gw^{-t}|w^{2t})$. Therefore, $$\Sigma_3^{m,n}=\sum_{L \in \mathcal{D}}{(m_Lw)^{-t}}S^{w^{2t},m}_L (gw^{-t}) \; P_L^nf ,$$ $$\Sigma_4^{m,n} = \sum_{L \in \mathcal{D}} {(m_Lw)^{-t}} R^{w^{2t},m}_L (gw^{-t}) \; P^n_L f.$$ Estimates (\[Sestpar\]) and (\[Restpar\]) with $s=1$ hold for $S^{w^{2t},m}_L (gw^{-t})$ and $R^{w^{2t},m}_L (gw^{-t})$, with $w^{-1}$ and $g$ replaced by $w^{2t}$ and $gw^{-t}$: $$\begin{aligned} S^{w^{2t},m}_L (gw^{-t}) & \leq& (m_L w^{2t})^{\frac{1}{2}}\Big(\sum_{J \in \mathcal{D}_m(L)} |\langle gw^{-t},h_J^{w^{2t}} \rangle_{w^{2t}} |^2\Big)^{\frac{1}{2}},\\ R^{w^{2t},m}_L (gw^{-t}) & \leq & C\,C^n_m (m_L w^{2t})^{\frac{1}{2}}(m_L w^{-2t})^{\frac{-1}{2}} F^{\frac12}(x) \sqrt{\eta^m_L},\end{aligned}$$ where $F(x)=\inf_{x \in L} \big ( M_{w^{2t}}(|gw^{-t}|^p)(x)\big )^{\frac{2}{p}}$. Estimating $P_L^n f$ is simple: $$P_L^n f \leq \Big( \sum_{I \in \mathcal{D}_n(L)} {|I|}/{|L|} \Big)^{\frac{1}{2}} \Big( \sum_{I \in \mathcal{D}_n(L)} |\langle f,h_I \rangle|^2 \Big)^{\frac{1}{2}} = \Big( \sum_{I \in \mathcal{D}_n(L)} |\langle f,h_I \rangle|^2 \Big)^{\frac{1}{2}}.$$ [**Estimating $\Sigma_3^{m,n}$:**]{} Plug in the estimates for $S^{w^{2t},m}_L (gw^{-t})$ and $P_L^nf $, observing that ${(m_Lw^{2t})^{\frac{1}{2}}}/{(m_L w)^t}\leq [w]_{C^d_{2t}}^{\frac12}$. Using the Cauchy-Schwarz inequality, we get, $$\begin{aligned} \Sigma_3^{m,n} & \; \leq \sum_{L \in \mathcal{D}} [w]_{C^d_{2t}}^{\frac12} \Big(\sum_{J \in \mathcal{D}_m(L)} |\langle gw^{-t},h_J^{w^{2t}} \rangle_{w^{2t}}|^2 \Big)^{\frac{1}{2}} \Big( \sum_{I \in \mathcal{D}_n(L)} |\langle f,h_I \rangle|^2 \Big)^{\frac{1}{2}} \\ & \leq \; [w]_{C^d_{2t}}^{\frac{1}{2}} \|f\|_{2} \| gw^{-t}\|_{L^2(w^{2t})} \; = \; [w]_{C^d_{2t}}^{\frac{1}{2}} \|f\|_{2}\|g\|_{2}.\end{aligned}$$ [**Estimating $\Sigma_4^{m,n}$:**]{} Plug in the estimates for $R^{w^{2t},m}_L (gw^{-t})$ and $P_L^nf $, where $F(x)=\big ( M_{w^{2t}}(|gw^{-t}|^p)(x)\big )^{2/p}$. Using the Cauchy-Schwarz inequality and considering again that ${(m_L w^{2t})^{\frac12}}/{(m_L w)^t}\leq [w]_{C^d_{2t}}^{\frac12}$, then $$\Sigma_4^{m,n} \leq C\, C^n_m [w]_{C^d_{2t}}^{\frac{1}{2}} \|f\|_2\Big(\sum_{L \in \mathcal{D}}\frac{\eta_L^m}{m_L w^{-2t}} \inf_{x \in L} F(x) \Big)^{\frac{1}{2}}.$$ Now, use the Weighted Carleson Lemma \[weightedCarlesonLem\] with $\alpha_L={\eta_L^m}/{m_L(w^{-2t} )}$ (which by Lemma \[litlem\] is a $w^{2t}$-Carleson sequence with intensity at most $C\,C^n_m[w^{2t}]_{A^d_2}$). Let $F(x)=\big ( M_{w^{2t}}|gw^{-t}|^p(x)\big )^{2/p}$, and $v=w^{2t}$, then $$\Sigma_4^{m,n} \leq C(C^n_m)^2 [w]_{C^d_{2t}}^{\frac{1}{2}} [w^{2t}]^{\frac{1}{2}}_{A^d_2} \|f\|_{2} \big \|M_{w^{2t}}(|gw^{-t}|^p)\big \|^{\frac{1}{p}}_{L^{\frac{2}{p}}(w^{2t})}.$$ Using , that is the boundedness of $M_{w^{2t}}$ in $L^{\frac2p}(w^{2t})$ for $2/p>1$, and $(2/p)'=2C^n_m$, we get $$\begin{aligned} \Sigma_4^{m,n} & \leq C( C^n_m)^2 (2/p)' [w]_{C^d_{2t}}^{\frac{1}{2}} [w^{2t}]^{\frac{1}{2}}_{A^d_2} \|f\|_{2}\big\| |gw^{-t}|^p \big\|^{\frac{1}{p}}_{L^{\frac{2}{p}}(w^{2t})}\\ &\leq C (C^n_m)^3 [w]_{C^d_{2t}}^{\frac{1}{2}} [w^{2t}]^{\frac{1}{2}}_{A^d_2} \|f\|_{2}\| g\|_{2}.\end{aligned}$$ The theorem is proved. [09]{} O. Beznosova, *Bellman functions, paraproducts, Haar multipliers and weighted inequalities*. PhD. Dissertation, University of New Mexico (2008). O. Beznosova, *Linear bound for the dyadic paraproduct on weighted Lebesgue space $L^2(w)$*. [J. Func. Anal.]{} [**255**]{} (2008), 994 –1007. O. Beznosova, J. C. Moraes, M. C. Pereyra, *Sharp bounds for $t$-Haar multipliers in $L^2$*. Submitted to the Proceedings of El Escorial 2012 to be published by Cont. Math. AMS. O. Beznosova, A. Reznikov, *Sharp estimates involving $A_\infty$ and $LlogL$ constants, and their applications to PDE.* Preprint (2011) available at arXiv:1107.1885 S. M. Buckley, *Estimates for operator norms on weighted spaces and reverse Jensen inequalities*. Trans. Amer. Math. Soc. [**340**]{} no. 1 (1993), 253–272. D. Chung, *Weighted inequalities for multivariable dyadic paraproducts*. Publ. Mat. [**55**]{}, no. 2 (2011), 475–499. D. Cruz-Uribe, SFO, J. Martell, C. Pérez, [*Sharp weighted estimates for classical operators*]{}. Advances in Mathematics, vol. 229, (2012), 408-441. D. Cruz-Uribe, J. M. Martell, C. Peréz, *Weights, extrapolation and the theory of Rubio the Francia*. Birkhäuser, (2011). O. Dragičevič, L. Grafakos, M. C. Pereyra, S. Petermichl, *Extrapolation and sharp norm estimates for classical operators in weighted Lebesgue spaces*. Publ. Mat. [**49**]{}, (2005), 73–91. S. Hukovic, S. Treil, and A. Volberg, *The Bellman functions and sharp weighted inequalities for for square function*. Oper. Theory Adv. Appl., [**113**]{} (2000), 97–113. T. Hytönen, *The sharp weighted bound for general Calderón-Sygmund operators*. Ann. Math. (2) [**175**]{}(3) (2012), 1473-1506. T. Hytönen, M. Lacey, H. Martkainen, T. Orponen, M. Reguera, E. Sawyer, I. Uriarte-Tuero, *Weak and strong type estimates for maximal truncations of Calderón-Zygmund operators on $A_p$ weighted spaces*. J. Anal. Math. [**118**]{}, 1 (2012), 177-220. T. Hytönen, C. Peréz, S. Treil, A. Volberg, *Sharp weighted estimates for dyadic shifts and the $A_2$ conjecture*. To appear J. Reine Angew. Math. N. H. Katz, M. C. Pereyra, *Haar multipliers, paraproducts and weighted inequalities*. Analysis of Divergence, [**10**]{}, 3, (1999), 145-170. M. Lacey, [*On the $A_2$ inequality for Calderón-Zygmund operators*]{}. To appear Cont. Math. Vol. M. Lacey, [*The linear bound in $A_2$ for Calderón-Zygmund operators: A survey*]{}. Marcinkiewicz Centenary volume, 97-114 Banach Center Publ. 95, Polish Acad. Sci. Inst. Math. Warsaw, 2011. M. Lacey, S. Petermichl, M. Reguera, *Sharp $A_2$ inequalities for Haar shift operators*. Math. Ann. [**348**]{} (2010), no. 1, 127–141. A. Lerner, *Sharp weighted norm inequalities for Littlewood-Paley operators and singular integrals*. Adv. Math. [**226**]{} (2011), 3912-3926. J. C. Moraes, *Weighted estimates for dyadic operators with complexity*. PhD Dissertation, University of New Mexico, 2011. J. C. Moraes, *Weighted estimates for dyadic operators with complexity in geometrically doubling metric spaces*. In preparation. B. Muckenhoupt, *Weighted norm inequalities for the Hardy–Littlewood maximal function*. Trans. Amer. Math. Soc. [**165**]{} (1972), 207–226. F. Nazarov, A. Reznikov, S. Treil, A. Volberg, *A Bellman function proof of the $L^2$ bump conjecture.* To appear J. Anal. Math. F. Nazarov, A. Reznikov, A. Volberg, *The proof of $A_2$ conjecture in a geometrically doubling metric space.* Preprint (2011) available at arXiv: 1106.1342v2. F. Nazarov, A. Volberg, *Bellman function, polynomial estimates of weighted dyadic shifts, and $A_2$ conjecture*. Preprint (2011). F. Nazarov, A. Volberg, *A simple sharp weighted estimate of the dyadic shifts on metric spaces with geometric doubling*. Int. Math. Res. Notices (2012) doi:10.1093/imrn/rns159. F. Nazarov, S. Treil and A. Volberg, *The Bellman functions and the two-weight inequalities for Haar multipliers*. J. Amer. Math. Soc., [**12**]{} (1999), 909–928. F. Nazarov, S. Treil and A. Volberg, *Two weight inequalities for individual Haar multipliers and other well localized operators.* Math. Res. Lett. [**15**]{} (2008), no.3, 583-597. M. C. Pereyra, *On the resolvents of dyadic paraproducts*. Rev. Mat. Iberoamericana [**10**]{}, 3, (1994), 627-664. M. C. Pereyra, *Haar multipliers meet Bellman function*. Rev. Mat. Iberoamericana [**25**]{}, 3, (2009), 799-840. M. C. Pereyra, *Sobolev spaces on Lipschitz curves*. Pacific J. Math. [**172**]{} (1996), no. 2, 553–589. C. Pérez, S. Treil, A. Volberg, *On $A_2$ conjecture and corona decomposition of weights*. Preprint (2010) available at arXiv: 1006.2630 S. Petermichl, *Dyadic shift and a logarithmic estimate for Hankel operators with matrix symbol.* C. R. Acad. Sci. Paris Sér. I Math. **330** (2000) \# 6, 455–460. S. Petermichl, *The sharp bound for the Hilbert transform on weighted Lebesgue spaces in terms of the classical $A_p$ characteristic*. Amer. J. of Math. [**129**]{} (2007), 1355–1375. S. Petermichl, *The sharp bound for the Riesz transforms*. Proc. Amer. Math. Soc. [**136**]{} (2008) 1237–1249. S. Petermichl, A. Volberg, *Heating of the Ahlfors-Beurling operator: Weakly quasiregular maps on the plane are quasiregular*. Duke Math J. [**112**]{} (2002), 281–305. S. Treil, [*Sharp $A_2$ estimates of Haar shifts via Bellman function.*]{} Preprint (2011) available at arXiv:1105.2252 A. Volberg, *Bellman function technique in Harmonic Analysis. Lectures of INRIA Summer School in Antibes, June 2011.* Preprint (2011) available at arXiv:1106.3899 J. Wittwer, *A sharp estimate on the norm of the martingale transform*. Math. Res. Letters, [**7**]{} (2000), 1–12. J. Wittwer, *A sharp estimate on the norm of the continuous square function*. Proc. Amer. Math. Soc. [**130**]{} (2002), no. 8, 2335–2342 (electronic). [^1]: The first author was supported by fellowship CAPES/FULBRIGHT, BEX 2918-06/4

ArXiv

--- abstract: 'Viro method plays an important role in the study of topology of real algebraic hypersurfaces. The $T$-primitive hypersurfaces we study here appear as the result of Viro’s combinatorial patchworking when one starts with a primitive triangulation. We show that the Euler characteristic of the real part of such a hypersurface of even dimension is equal to the signature of its complex part. We use this result to prove the existence of maximal surfaces in some three-dimensional toric varieties, namely those corresponding to Nakajima polytopes. In fact, these results belong to the field of tropical geometry and we explain how they can be understood tropically.' address: | Section de mathématiques\ Université de Genève\ case postale 64\ 2-4 rue du lièvre\ 1211 Genève 4\ Suisse author: - Benoit Bertrand title: 'Euler characteristic of primitive $T$-hypersurfaces and maximal surfaces' --- Introduction {#introduction .unnumbered} ============ The number of connected components of a real algebraic plane projective curve of degree $m$ is not more than $(m-1)(m-2)/2 + 1$. This bound was proved by Harnack who also showed that for any positive integer $m$ there exist curves of degree $m$ which are maximal in this sense (i.e. with $(m-1)(m-2)/2 + 1$ connected components). Harnack’s bound is generalized to the case of any real algebraic variety by the Smith-Thom inequality. Let $b_i(V;K)$ be the $i^{\mbox{\scriptsize th}}$ Betti number of a topological space $V$ with coefficients in a field $K$ (i.e. $b_i(V;K)=\dim_K(H_i(V;K))$). Denote by $b_*(V;K)$ the sum of the Betti numbers of $V$. Let $X$ be a complex algebraic variety equipped with an anti-holomorphic involution $c$. The real part ${{\mathbb{R}}}X$ of $X$ is the fixed point set of $c$. Then the Smith-Thom inequality states that $b_*({{\mathbb{R}}}X;{{\mathbb{Z}}}_2) \le b_*(X;{{\mathbb{Z}}}_2)$. A variety $X$ for which $b_*({{\mathbb{R}}}X;{{\mathbb{Z}}}_2)= b_*(X;{{\mathbb{Z}}}_2)$ is called a [*maximal*]{} variety or [*$M$-variety*]{}. The question “does a given family of real algebraic varieties contain maximal elements?” is one of the main problems in topology of real algebraic varieties. For the family of the hypersurfaces of a given degree in ${{\mathbb{R}}}P^d$ a positive answer is obtained in [@IteVir] using the combinatorial Viro method called $T$-construction (see [@vir3], [@vir6], [@ite1], and Theorem \[Tcons\]). The Viro patchworking theorem –and its combinatorial particular case– is essential in the study of existence of hypersurfaces with prescribed topology. Let $\Delta$ be a lattice polytope equipped with a regular (convex) triangulation $\tau$ and a distribution of signs $D$ at its vertices. From this data Viro’s combinatorial patchwork produces a real algebraic hypersurface in the toric variety $X_\Delta$ associated to $\Delta$ whose topology can be easily recovered from $\Delta, \, \tau $ and $D$ (see the above references and Section \[tcon\]). [[******]{}Primitive $T$-hypersurfaces]{} are those obtained by the combinatorial patchworking if one starts with a unimodular (primitive) triangulation $\tau$. Provided that $\Delta$ correspond to a nonsingular toric variety, we prove that for any primitive $T$-hypersurface $Z$, $$\label{sc} \chi({{\mathbb{R}}}Z) = \sigma(Z) \, ,$$ where $\chi({{\mathbb{R}}}Z)$ is the Euler characteristic of ${{\mathbb{R}}}Z$ and $\sigma(Z) := \sum_{p+q=0\> [2]} {(-1)}^p h^{p,q}(Z)$ is the signature of $Z$ if $\dim Z$ is even and $0$ otherwise (See Theorem \[theprop\]). This result can be naturally formulated in the tropical setting introduced by G. Mikhalkin. (See Section \[S:trophyp\] or [@Mikhmath.AG/0403015] and [@Mikh05] for definitions and propreties of tropical hypersurfaces.): The Euler characteristic of a compactified nonsingular real tropical hypersurface with Newton polytope $\Delta$ is equal to the signature of a complex algebraic hypersurface in $X_\Delta$ with Newton polytope $\Delta$. For primitive $T$-surfaces in the projective space ${{\mathbb{P}}}^3$ this result was first proven by I. Itenberg [@ite1] who used it to construct maximal $T$-surfaces in all degrees in ${{\mathbb{P}}}^3$. The existence of maximal surfaces in all degrees in the projective space had been proven in [@Vir79] by Viro using small perturbation technics. Using the combinatorial Viro method and equality (\[sc\]) we prove the existence of maximal surfaces in a family of toric varieties which includes ${{\mathbb{P}}}^3$. A polytope $\Delta$ in ${{\mathbb{R}}}^d$ is a [[******]{}Nakajima polytope]{} (see Figure \[figure:nakapol3\]) if either $\Delta$ is $0$-dimensional or there exists a Nakajima polytope $\bar{\Delta}$ in ${{\mathbb{R}}}^{d-1}$ and a linear function $f: {{\mathbb{R}}}^{d-1} \to {{\mathbb{R}}}$ , nonnegative on $\bar{\Delta}$, such that $f({{\mathbb{Z}}}^{d-1})\subset {{\mathbb{Z}}}$ and $\Delta=\{(x,x_d) \in \bar{\Delta}\times {{\mathbb{R}}}\mid 0 \le x_d \le f(x)\}$. We show that for any 3-dimensional Nakajima polytope $\Delta$ corresponding to a nonsingular toric variety there exists a maximal hypersurface in $X_\Delta$ with Newton polytope $\Delta$. I am grateful to Ilia Itenberg for his valuable advice and to Erwan Brugallé for his pertinent remarks. [**Organization of the material.**]{} We first introduce notation and describe combinatorial patchworking in Section \[prelim\]. In the same section, we describe briefly tropical hypersurfaces and state the results in the tropical setting. We also recall there facts about lattice triangulations and state Danilov and Khovanskii Formulae. Section \[sigmachi\] is devoted to the proof of equality (\[sc\]) for primitive hypersurfaces stated in Theorem \[theprop\]. In the third section we construct maximal surfaces in a family of toric varieties using equality (\[sc\]). We postpone some proofs of Lemmae needed in Section \[sigmachi\] to the appendix (Section \[appendix\]) where we also give some usefull combinatorial formulae (e.g. Lemma \[vanlint\]). Preliminaries {#prelim} ============= Toric varieties --------------- We fix here some conventions and notations, the construction of toric varieties we use is based on the one described in [@ful]. Fix an orthonormal basis of ${{\mathbb{R}}}^d$ and thus an inclusion ${{\mathbb{Z}}}^d \to {{\mathbb{R}}}^d$. This inclusion defines a lattice $N$ in ${{\mathbb{R}}}^d$. Denote by $M$ the dual lattice $\text{\rm Hom}(N, {{\mathbb{Z}}})$ of $N$. Here we consider polytopes in $M\otimes{{\mathbb{R}}}$. By [[******]{}polytope]{} we mean convex polytope whose vertices are integer (i.e. belong to the lattice $M$). Let $\Delta$ be a polytope in $M\otimes{{\mathbb{R}}}$. Let $p$ be a vertex of $\Delta$ and let $\Gamma_1, \cdots, \Gamma_k$ be the facets of $\Delta$ containing $p$. To $p$ we associate the cone $\sigma_p$ generated by the minimal integer inner normal vectors of $\Gamma_1, \cdots, \Gamma_k$. The inner normal fan $\mathfrak{E}_\Delta$ is the fan whose $d$-dimensional cones are the cones $\sigma_p$ for all vertices $p$ of $\Delta$. The toric variety $X_\Delta$ associated to $\Delta$ is the toric variety $X(\mathfrak{E}_\Delta)$ associated to the fan $\mathfrak{E}_\Delta$. Combinatorial patchworking {#tcon} -------------------------- Let us briefly describe the [*combinatorial patchworking*]{}, also called [*$T$-construction*]{} , which is a particular case of the Viro method. A more detailed exposition can be found in [@IteVir], [@vir6] or [@GKZ] p. 385. By a [[******]{}subdivision]{} of a polytope we mean a subdivision in convex polytopes whose vertices have integer coordinates. A subdivision $\tau$ of a $d$-dimensional polytope $\Delta$ is called [[******]{}convex]{} (or [[******]{}regular]{}) if there exists a convex piecewise-linear function $\Phi : \Delta \to {{\mathbb{R}}}$ whose domains of linearity coincide with the $d$-dimensional polytopes of $\tau$. A triangulation is a subdivision into simplices. Given a triple $(\Delta,\tau,D)$, where $\Delta$ is a polytope, $\tau$ a convex triangulation of $\Delta$, and $D$ a distribution of signs at the vertices of $\tau$ (i.e. each vertex is labelled with $+$ or $-$), the combinatorial patchworking, produces an algebraic hypersurface $Z$ in $X_\Delta$. Let $\Delta$ be a $d$-dimensional polytope in the positive orthant $({{\mathbb{R}}}^+)^d = \{(x_1, \ldots x_d) \in {{\mathbb{R}}}^d \mid\\ x_1 \geq 0, \ldots , x_d \geq 0\}$, and $\tau$ be a convex triangulation of $\Delta$. Denote by $ s_{(i)}$ the reflection with respect to the coordinate hyperplane $x_i=0 $ in ${{\mathbb{R}}}^d$. Consider the union $\Delta^*$ of all copies of $\Delta$ under the compositions of reflections $ s_{(i)} $ and extend $\tau$ to a triangulation $\tau^*$ of $\Delta^*$ by means of these reflections. We extend $D(\tau)$ to a distribution of signs at the vertices of $\tau^*$ using the following rule : for a vertex $a$ of $\tau^*$, one has ${\operatorname{sign}}(s_{(i)}(a))={\operatorname{sign}}(a)$ if the $i$-th coordinate of $a$ is even, and ${\operatorname{sign}}(s_{(i)}(a))=-{\operatorname{sign}}(a)$, otherwise. Let $\sigma$ be a $d$-dimensional simplex of $\tau^*$ with vertices of different signs, and $E$ be the piece of hyperplane which is the convex hull of the middle points of the edges of $\sigma$ with endpoints of opposite signs. We separate vertices of $\sigma$ labelled with $+$ from vertices labelled with $-$ by $E$. The union of all these hyperplane pieces forms a piecewise-linear hypersurface $ H^*$. For any facet $ \Gamma$ of $ \Delta^*$, let $N^\Gamma$ be a vector normal to $\Gamma$. Let $F$ be a face of $\Delta^*$ and $\Gamma_1, \dots, \Gamma_k$ be the facets containing $F$. Let $L$ be the linear space spanned by $N^{\Gamma_1}, \dots, N^{\Gamma_k}$. For any $v = (v_1, \dots, v_d) \in L \cap {{\mathbb{Z}}}^d$ identify $F$ with ${s_{(1)}}^{v_1} \circ {s_{(2)}}^{v_2} \circ \dots \circ {s_{(d)}}^{v_d}(F)$. Denote by $ \widetilde{\Delta}$ the result of the identifications. The variety $\widetilde{\Delta}$ is homeomorphic to the real part ${{\mathbb{R}}}X_\Delta$ of $X_\Delta$ (see, for example, [@GKZ] Theorem 5.4 p. 383 or [@stu] Proposition 2). Denote by $\widetilde{H}$ the image of $H^*$ in $ \widetilde{\Delta}$. Let $Q$ be a polynomial with Newton polytope $\Delta$. It defines a hypersurface $Z_0$ in the torus $({{\mathbb{C}}}^*)^d$ contained in $X_\Delta$. The closure $Z$ of $Z_0$ in $X_\Delta$ is the hypersurface defined by $Q$ in $X_\Delta$. We call $\Delta$ the [ [******]{}Newton polytope ]{} of $Z$. \[Tcons\] Under the hypotheses made above, there exists a hypersurface $Z$ of $X_\Delta$ with Newton polytope $\Delta$ and a homeomorphism $h: {{\mathbb{R}}}X_\Delta\to \widetilde{\Delta}$ such that $h({{\mathbb{R}}}Z) = \widetilde{H}$. A [[******]{}real algebraic $T$-hypersurface]{} is a hypersurface which can be obtained via the above construction. A $d$-dimensional simplex with integer vertices is called [[******]{}primitive]{} (or [[******]{}unimodular]{}) if its volume is equal to $\frac{1}{d !}$. A triangulation $\tau$ of a $d$-dimensional polytope is [*[**primitive**]{}*]{} if every $d$-simplex of the triangulation is primitive. A $T$-hypersurface $Z$ is [[******]{}primitive]{} if it can be constructed with the combinatorial patchworking using a primitive triangulation. Let $\Delta$ be a $d$-dimensional polytope and $\tau$ a (convex) triangulation of $\Delta$. The star, ${\operatorname{Star}}(v)$, of a vertex $v$ of $\tau$ is the union of all simplices of $\tau$ containing $v$. Suppose that $\tau$ is equipped with a distribution of signs. We say that $v$ is [[******]{}isolated]{} if all other vertices of ${\operatorname{Star}}(v)$ have the sign opposite to ${\operatorname{sign}}(v)$. \[sphere\] Let $p$ be an isolated vertex of the triangulation $\tau^*$ of $\Delta^*$. Assume that $p$ is in the interior of $\Delta^*$. Then the star of $p$ contains a connected component of $H^*$ homeomorphic to a sphere $S^{d-1}$. So, $p$ corresponds to a $(d-1)$-dimensional sphere of the real part of the hypersurface $Z$ obtained by Viro’s theorem. Notice that the triangulation $\tau^*$ induces a natural cell decomposition of $H^*$: each cell is the intersection of a simplex of $\tau^*$ with $H^*$. A $k$-dimensional simplex in ${{\mathbb{R}}}^d$ is called [[******]{}elementary]{} if the reductions modulo $2$ of its vertices generate a $k$-dimensional affine space over ${{\mathbb{Z}}}_2$. Our computation of the Euler characteristic of ${{\mathbb{R}}}Z$ uses the following proposition due to Itenberg (see [@ite1] Proposition 3.1). \[itenprop\] Let $s$ be an elementary $k$-simplex of a triangulation $\tau$ of a $d$-dimensional polytope. Assume that $s$ is contained in $j$ coordinate hyperplanes. Then the union $s^*$ of the symmetric copies of $s$ contains exactly $(2^d-2^{d-k})/2^j$ cells of dimension $k-1$ of the cell decomposition of $H^*$. Tropical hypersurfaces {#S:trophyp} ---------------------- Before stating the results we briefly recall some basic facts about tropical hypersurfaces. A detailed presentation can be found in [@Mikh05]. In this article we will use the definition of tropical varieties as images of algebraic varieties defined over the field of Puiseux series. Let ${\mathbb{K}}$ be the field of Puiseux series. An element of ${\mathbb{K}}$ is a series $g(t)=\sum_{r\in R} b_r t^r$ where $b_r \in \mathbb{C}$, $ R \subset \mathbb{Q}$ is bounded from below and contained in an arithmetic sequence. Consider the valuation ${\operatorname{val}}(g(t)):=\min \{r | b_r \neq 0 \}$. For convenience we choose Mikhalkin’s conventions and consider in fact minus the valuation, $v(g):=-{\operatorname{val}}(g)$. Let $f$ be a polynomial in ${\mathbb{K}}[z_1,\cdots,z_n] ={\mathbb{K}}[z]$. It is of the form $f(z)=\sum_{\omega \in A} \;c_\omega z^\omega$ with $A$ a finite subset of ${{\mathbb{Z}}}^n$ and $c_\omega \in {\mathbb{K}}^*$. Consider $Z_f:=\{z\in ({\mathbb{K}}^*)^n | f(z)=0 \}$ the zero set of $f$ in $({\mathbb{K}}^*)^n$ and the coordinatewise valuation (up to sign) $$\begin{array}{rrrl} V: &({\mathbb{K}}^*)^n &\longrightarrow &{{\mathbb{R}}}^n \\ & z & \longmapsto &(v(z_1), \dots, v(z_n)). \end{array}$$ The tropical hypersurface $Z_f^{{\operatorname{trop}}}$ associated to $f$ is the closure (in the usual topology) of the image under $V$ of the hypersurface $Z_f$: $$Z^{{\operatorname{trop}}}_f := \overline{V(Z_f)} \subset {{\mathbb{R}}}^n.$$ The following theorem, due to Kapranov, allows one to see tropical hypersurfaces as nonlinearity domains of some piecewise-linear convex functions. Put $$\begin{array}{rrrl} \nu: & A &\longrightarrow &{{\mathbb{R}}}\\ & \omega & \longmapsto & - v (c_\omega)\, , \end{array}$$ the Legendre transform $\mathcal{L}(\nu)$ of $\nu$ is the piecewise-linear convex function defined as follows: $$\begin{array}{rrrl} \mathcal{L}(\nu): &{{\mathbb{R}}}^n &\longrightarrow &{{\mathbb{R}}}\\ & x & \longmapsto & \max_{\omega\in A} (x \cdot \omega - \nu(\omega)) \end{array}$$ The tropical hypersurface $Z^{{\operatorname{trop}}}_f$ is the nonlinearity domain of $\mathcal{L}(\nu)$: $$\overline{V(Z_f)}=\mbox{corner locus}(x\mapsto \max_{\omega\in A} (x \cdot \omega + v(c_\omega)) ).$$ Another very nice fact is that one can naturally associate to a tropical hypersurface a dual subdivision of its Newton polytope. Let $ \Gamma$ be the convex hull in ${{\mathbb{R}}}^n\times{{\mathbb{R}}}$ of the points $(\omega,v(c_\omega))$ for all $\omega$ in $A$. Put $$\begin{array}{rrrl} \overline{\nu} : & \Delta &\longrightarrow &{{\mathbb{R}}}\\ & x & \longmapsto & \min \{y | (x,y) \in \Gamma\}. \end{array}$$ The linearity domains of $\overline{\nu}$ are the $n$-cells of a convex polyhedral subdivision $\tau$ of $\Delta$. The hypersurface $Z_f^{{\operatorname{trop}}}$ induces a subdivision $\Xi$ of ${{\mathbb{R}}}^n$. The subdivisions $\tau$ and $\Xi$ are dual in the following sense (see Figure \[F:lineetcubic\]). There is a one-to-one inclusion reversing correspondence $L$ between cells of $\Xi$ and cells of $\tau$ such that for any $\xi \in \Xi$, 1. $\dim L(\xi)={\operatorname{codim}}\xi$, 2. the affine supports of $L(\xi)$ and $\xi$ are orthogonal. ------------------------------------------------------------------------------------------------------------------------------------------------------------ -- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![A tropical line, a tropical cubic with the dual subdivision of its Newton polytope.[]{data-label="F:lineetcubic"}](droitepaisse.eps "fig:"){width="5cm"} ![A tropical line, a tropical cubic with the dual subdivision of its Newton polytope.[]{data-label="F:lineetcubic"}](cubic-et-dual-red-thick.eps "fig:"){width="5cm"} ------------------------------------------------------------------------------------------------------------------------------------------------------------ -- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- We say that $ Z^{{\operatorname{trop}}}_f $ is [**nonsingular**]{} if its dual subdivision $\tau$ is a primitive triangulation. 0.6cm [**Complex tropical hypersurfaces.**]{} In order to prove his correspondence theorem Mikhalkin introduces in [@Mikh05] the complexification of tropical curves. Let $g(t)=\sum_{r\in R} b_r t^r$ be an element of ${\mathbb{K}}^*$. Define the argument $\arg(g)$ to be the argument of the coefficient of the smallest power of $t$ appearing in $g$ (i.e. $\arg(g)=\arg(b_{{\operatorname{val}}(g)}$). Let ${\operatorname{Arg}}$ be the coordinatewise argument on $({\mathbb{K}}^*)^n$. Consider the “complexification” of the valuation $$\begin{array}{rrrl} {W}:= V \times {\operatorname{Arg}}: &({\mathbb{K}}^*)^n &\longrightarrow &{{\mathbb{R}}}^n \times (S^1)^n\\ & z & \longmapsto &((v(z_1), \dots, v(z_n)),(\arg(z_1), \dots, \arg(z_n))). \end{array}$$ One can define complex tropical hypersurfaces to be the closure of the image under ${W}$ of hypersurfaces in $({\mathbb{K}}^*)^n$ but it is also convenient to consider the composition of ${W}$ with the exponential. Namely, put $$\begin{array}{rrrl} {V_\mathbb{C}}: &({\mathbb{K}}^*)^n &\longrightarrow & ({{\mathbb{C}}}^*)^n\\ & z & \longmapsto &(e^{v(z_1)+i \arg(z_1)}, \dots, e^{v(z_n)+i \arg(z_n)}) \end{array}$$ We will call both homeomorphic objects complex tropical hypersurfaces and use one and the other alternatively depending on the context. The complex tropical hypersurface ${{\mathbb{C}}}Z_{f,{V_\mathbb{C}}}^{{\operatorname{trop}}}$ (resp. ${{\mathbb{C}}}Z^{{\operatorname{trop}}}_{f,{W}}$) associated to $f$ is the closure of the image under ${V_\mathbb{C}}$ (resp. ${W}$) of the hypersurface $Z_f$: $${{\mathbb{C}}}Z^{{\operatorname{trop}}}_{f,{V_\mathbb{C}}} := \overline{{V_\mathbb{C}}(Z_f)} \subset ({{\mathbb{C}}}^*)^n,$$ or alternatively $${{\mathbb{C}}}Z^{{\operatorname{trop}}}_{f,{W}} := \overline{{W}(Z_f)} \subset {{\mathbb{R}}}^n \times (S^1)^n.$$ For example, a complex tropical line is homeomorphic to a sphere with three punctures (see Figure \[F:lines\]). 0.6cm [**Real tropical Hypersurfaces.**]{} Assume from now on that $f=\sum_{\omega \in A} c_\omega z^\omega$ is real (i.e. all the coefficients $a_r$ of each series $c_\omega = \sum_{r\in R} a_r t^r $ are real). The real tropical hypersurface ${{\mathbb{R}}}Z_{f,{V_\mathbb{C}}}^{{\operatorname{trop}}}$ associated to $f$ is the intersection in $({{\mathbb{C}}}^*)^n$ of ${{\mathbb{C}}}Z_{f,{V_\mathbb{C}}}^{{\operatorname{trop}}}$ with $({{\mathbb{R}}}^*)^n$: $${{\mathbb{R}}}Z^{{\operatorname{trop}}}_{f,{V_\mathbb{C}}} := \overline{{V_\mathbb{C}}(Z_f)}\cap({{\mathbb{R}}}^*)^n,$$ or alternatively $${{\mathbb{R}}}Z^{{\operatorname{trop}}}_{f,{W}} := \overline{{W}(Z_f)}\cap ({{\mathbb{R}}}^n\times \{0,\pi\}^n).$$ ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![A complex tropical line, its tropical “projection” and a real tropical line in the four quadrants.[]{data-label="F:lines"}](complex-trop3.eps "fig:"){height="5cm"} ![A complex tropical line, its tropical “projection” and a real tropical line in the four quadrants.[]{data-label="F:lines"}](droite-reelle.eps "fig:"){height="5cm"} ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- -- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- On Figure \[F:lines\] we have pictured the intersection of a real tropical line with each of the four quadrants. The sign ${\operatorname{sign}}(c_\omega)$ of a coefficient $c_\omega = \sum_{r\in R} a_r t^r$ of $f$ is defined to be the sign of the first coefficient $a_{{\operatorname{val}}(c_\omega)}$ of $c_\omega$ (i.e. ${\operatorname{sign}}c_\omega$ $:= {\operatorname{sign}}a_{{\operatorname{val}}(c_\omega)}$). From now on we will only consider nonsingular tropical hypersurfaces. For $\epsilon \in {\{+1,-1\}}^n$, let $\mathcal{Q}_\epsilon$ be the orthant of $({{\mathbb{R}}}^*)^n$ which maps to the positive orthant $({{\mathbb{R}}}_+^*)^n$ under the isometry $\varphi_\epsilon: \mathcal{Q}_\epsilon \to ({{\mathbb{R}}}_+^*)^n$ defined by $\varphi_\epsilon(x_1, \cdots,x_n)=(\epsilon_1 x_1, \cdots,\epsilon_n x_n)$. If $Z^{{\operatorname{trop}}}_{f,{V_\mathbb{C}}}$ is nonsingular one can reconstruct ${{\mathbb{R}}}Z^{{\operatorname{trop}}}_f \cap \mathcal{Q}_\epsilon$ only from the data of $ Z^{{\operatorname{trop}}}_f $ and the collection of signs ${\operatorname{sign}}c_\omega$ of the coefficients of $f$ (see [@Mikhmath.AG/0403015] pp. 25 and 37, [@Vir01], and [@Mik00] Appendix for the case of amoebas). First one sees that we can restrict the study to the case of the first orthant $\mathcal{Q}$ since one can map any $\mathcal{Q}_\epsilon$ to $\mathcal{Q}$ by $\varphi_\epsilon$ and change $f$ to $f_\epsilon = f(\epsilon_1 x_1, \cdots,\epsilon_n x_n)$. Let us rather use the map ${W}$ (it suffices to exponentiate to switch back to the image under ${V_\mathbb{C}}$). Then $\mathcal{Q}$ corresponds to ${{\mathbb{R}}}^n\times \{(0, \cdots, 0)\}$. Consider the tropical hypersurface $ Z^{{\operatorname{trop}}}_f \subset {{\mathbb{R}}}^n$, the induced subdivision $\Xi_f$ of ${{\mathbb{R}}}^n$ and the dual subdivision $\tau_f$ of its Newton polytope $\Delta_f$. Let $D_f$ be the sign distribution at the vertices of $\tau_f$ such that a vertex $\omega$ is labelled with the sign ${\operatorname{sign}}(c_\omega)$ of the corresponding coefficient in $f$. Let $x$ be a point of $Z^{{\operatorname{trop}}}_f$. The point $(x,(0, \cdots, 0))$ is in ${{\mathbb{R}}}Z^{{\operatorname{trop}}}_{f,{W}} \cap ({{\mathbb{R}}}^n\times \{(0, \cdots, 0)\})$ if and only if $x$ belongs to the closure of an $(n-1)$-cell $\xi$ of $\Xi_f$ which is dual to an edge of $\tau_f$ whose extremities have opposite signs (see Figure \[F:RT\] for an example in the case of a curve). ![A real tropical cubic in the first quadrant, the underlying tropical curve (in dotted lines) and its dual subdivision.[]{data-label="F:RT"}](dessins/Ttrop-const-quad1.eps){width="4cm"} ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![An attempt at representing a real nonsingular tropical cubic with the underlying symmetrized tropical curve (in dotted lines) and corresponding $T$-curve.[]{data-label="F:patchtrop"}](Ttrop-const-6.eps "fig:"){width="6cm"} ![An attempt at representing a real nonsingular tropical cubic with the underlying symmetrized tropical curve (in dotted lines) and corresponding $T$-curve.[]{data-label="F:patchtrop"}](Tcourbe.eps "fig:"){width="6cm"} ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Then one sees that one can associate a unique primitive $T$-hypersurface to a nonsingular real tropical hypersurface. Namely take the $T$-hypersurface $H$ constructed from $\Delta_f, \tau_f$ and $D_f$. Clearly, there exists a homeomorphism $h:({{\mathbb{R}}}^*)^n \to (\Delta \setminus \partial\Delta) ^*$ such that $h({{\mathbb{R}}}Z^{{\operatorname{trop}}}_f)= H \cap (\Delta \setminus \partial\Delta)^*$. On Figure \[F:patchtrop\] we tried to illustrate this homeomorphism in the case of curves. The results in this paper are nicer and much easier to state when one compactifies the tropical hypersurfaces. Let us describe the natural compactification of a tropical hypersurface. Recall that ${{\mathbb{C}}}Z^{{\operatorname{trop}}}_{f,{V_\mathbb{C}}} = \overline{{V_\mathbb{C}}(Z_f)}$ is a subset of the torus $({{\mathbb{C}}}^*)^n$. Consider the usual compactification of $({{\mathbb{C}}}^*)^n$ into the toric variety $X_{\Delta_f}$ associated to the Newton polytope $\Delta_f$ of $f$. Let $\iota: ({{\mathbb{C}}}^*)^n \hookrightarrow X_{\Delta_f}$ be the corresponding inclusion. We define the compactification $\overline{{{\mathbb{C}}}Z^{{\operatorname{trop}}}_{f}}$ to be the closure of $\iota({{\mathbb{C}}}Z^{{\operatorname{trop}}}_{f,{V_\mathbb{C}}})$ in $X_{\Delta_f}$. Note that the stratification of $X_{\Delta_f}$ into orbits of the action of $({{\mathbb{C}}}^*)^n$ defines a natural stratification of $\overline{{{\mathbb{C}}}Z^{{\operatorname{trop}}}_{f}}$. We sum up natural maps in the following commutative diagram. $$\xymatrix{ {{\mathbb{R}}}Z^{{\operatorname{trop}}}_{f,{W}} \ar[r]^\sim\ar@{^{(}->}[d] & {{\mathbb{R}}}Z^{{\operatorname{trop}}}_{f,{V_\mathbb{C}}} \ar@{^{(}->}[r]\ar@{^{(}->}[d] & {({{\mathbb{R}}}^{\relax *})^n}\ar@{^{(}->}[d] \ar@{^{(}->}[r]^{\iota_{{\mathbb{R}}}} & {{\mathbb{R}}}X_{\Delta_f} \ar@{^{(}->}[d] \\ {{{\mathbb{C}}}Z^{{\operatorname{trop}}}_{f,{W}}} \ar[r]^\sim & {{{\mathbb{C}}}Z^{{\operatorname{trop}}}_{f,{V_\mathbb{C}}}} \ar@{^{(}->}[r] & ({{\mathbb{C}}}^{\relax *})^n\ar@{^{(}->}[r]^\iota & X_{\Delta_f} }$$ Define $\overline{{{\mathbb{R}}}Z^{{\operatorname{trop}}}_f}$ to be the intersection of $\overline{{{{\mathbb{C}}}Z^{{\operatorname{trop}}}_{f}}}$ with the real part ${{\mathbb{R}}}X_{\Delta_f}$ of $X_{\Delta_f}$. Clearly $\overline{{{\mathbb{R}}}Z^{{\operatorname{trop}}}_f}$ is also the closure of $\iota_{{\mathbb{R}}}({{\mathbb{R}}}Z^{{\operatorname{trop}}}_f)$ in ${{\mathbb{R}}}X_{\Delta_f}$. One can see that the natural stratification of $\overline{{{\mathbb{R}}}Z^{{\operatorname{trop}}}_f}$ induced by the torus action corresponds to the stratification of the $T$-hypersurface $H$ induced by the face complex of $\Delta_f$ which will be used in the proof of Theorem \[theprop\]. Tropical Statements {#S:tropstate} ------------------- We can now state our results in the tropical language. Let $\Delta$ be a polytope corresponding to a nonsingular toric variety. \[Th:sigmachitrop\] Let $\overline{{{\mathbb{R}}}Z^{{\operatorname{trop}}}_f}$ be a compactified nonsingular real tropical hypersurface with Newton polytope $\Delta_f=\Delta$ then $$\chi(\overline{{{\mathbb{R}}}Z^{{\operatorname{trop}}}_f}) = \sigma(Z),$$ where $Z$ is a smooth complex algebraic hypersurface in $X_{\Delta_f}$ with Newton polytope $\Delta_f$ and $\sigma(Z)= \sum_{p+q=0\> [2]} {(-1)}^p h^{p,q}(Z)$. Theorem \[Th:sigmachitrop\] is equivalent to Theorem \[theprop\] about Euler characteristic of primitive $T$-hypersurfaces and is one the of main results of this paper. One can define a maximal real tropical hypersurface in the following way. Let $f$ be a polynomial over ${\mathbb{K}}$ such that every coefficient of $f$ is a series with only real coefficients. Let $\Delta_f$ be the Newton polytope of $f$. The tropical hypersurface $\overline{{{\mathbb{R}}}Z^{{\operatorname{trop}}}_f}$ is [**maximal**]{} if $$b_*(\overline{{{\mathbb{R}}}Z^{{\operatorname{trop}}}_f};{{\mathbb{Z}}}_2) = b_*(\overline{{{\mathbb{C}}}Z^{{\operatorname{trop}}}_f};{{\mathbb{Z}}}_2).$$ We can then prove existence results for maximal real tropical surfaces. \[Th:P2trop\] Let $\Delta$ be a $3$-dimensional Nakajima polytope corresponding to a nonsingular toric variety $X_\Delta$. Then, there exists a maximal real tropical surface in $X_\Delta$ with the Newton polytope $\Delta$. Theorem\[Th:P2trop\] will be proved in Section \[Mhypex\] (See Theorem \[P2\]) as well as the following corollary (See Corollary \[hirz\]). For a nonnegative integer $\alpha$ and positive integers $m$ and $n$ denote by $\delta_\alpha^{m,n}$ the polygon having the vertices $(0,0)$, $(n + m \alpha,0)$, $(0,m)$, and $(n,m)$ in ${{\mathbb{R}}}^2$. The toric variety associated with $\delta_\alpha^{m,n}$ is a rational ruled surface $\Sigma_\alpha$. Consider now the truncated cylinder $\Delta^{l,m,n}_\alpha$ with base $\delta_\alpha^{m,n}$ whose vertices are $$\displaylines{(0,0,0),(n + m \alpha,0,0),(0,m,0),(n,m,0), \cr (0,0,l), (n + m \alpha,0,l),(0,m,l), \; \text{\rm and} \; (n,m,l),}$$ where $l$ is a positive integer. The toric variety $X_{\Delta^{l,m,n}_\alpha}$ associated with $\Delta^{l,m,n}_\alpha$ is $\Sigma_\alpha \times {{\mathbb{C}}}P^1$. Since $\Delta^{l,m,n}_\alpha$ is, up to exchanging two coordinates, a Nakajima polytope, we have the foolowing corollary (See Corollary \[hirz\]). \[Th:hirztrop\] For any nonnegative integer $\alpha$ and any positive integers $m$, $n$ and $l$, there exists a (real) maximal nonsingular tropical surface $\overline{{{\mathbb{R}}}Z^{{\operatorname{trop}}}_{f}}$ with the Newton polytope $\Delta^{l,m,n}_\alpha$ in ${{\mathbb{R}}}\Sigma_\alpha \times {{\mathbb{R}}}P^1$. Some facts about triangulations of lattice polytopes ---------------------------------------------------- ### Ehrhart Polynomial {#ehrpol} In order to compute both sides of equality (\[sc\]), namely the Euler characteristic of the real part of a primitive $T$-hypersurface and the signature of its complex part, we use some combinatorics of triangulations of lattice polytopes. We first recall the results of E. Ehrhart on the number of integer points in a lattice polytope. Ehrhart showed that the number of integer points in the multiple $\lambda \cdot \Delta$ of a (not necessarily convex) polytope $\Delta$, where $\lambda$ is a positive integer, is a polynomial in $\lambda$ (see [@Ehr] or [@Ehr67]). We denote by $l(\Delta)$ and $l^*(\Delta)$ the numbers of integer points in $\Delta$ and in the interior of $\Delta$, respectively. Let $\Delta$ be a polytope with integer vertices. Then, the numbers $l(\lambda \cdot \Delta)$ and $l^*(\lambda \cdot \Delta)$ are polynomials in $\lambda$ of degree $\dim \Delta$. Denote them respectively by $Ehr_\Delta(\lambda)$ and $Ehr^*_\Delta(\lambda)$. They satisfy the reciprocity law $$(-1)^{dim \Delta} Ehr_\Delta(-\lambda)=Ehr^*_\Delta(\lambda) .$$ One often considers the Ehrhart series $$SE_p(t)=\sum_{\lambda=0}^{\infty} Ehr_\Delta(\lambda) t^{\lambda}.$$ Put $Q_\Delta(t) =(1-t)^{d+1} SE_p(t)$, where $d$ is the dimension of $\Delta$. We define the numbers $\Psi_j$ to be the coefficients of $Q_\Delta(t)$ : $$Q_\Delta(t)= \sum_{j=0}^\infty \Psi_j t^j .$$ In fact, $Q_\Delta(t)$ is a polynomial of degree $d$ (see [@bri] or lemma \[psi\]). Let $a_i^{\Delta}$ be the coefficient of $\lambda^i$ in $Ehr_\Delta(\lambda)$. The following formula can be found in Section 4.1 of [@DCZ] (See also the Appendix). \[psi\] One has $$\Psi_j= \sum_{i=0}^d (\sum_{n=0}^j (-1)^{j-n}{\binom{d+1}{j-n}} n^i) a_i^{\Delta},$$ and $\Psi_j= 0$ for $ j \ge d+1$. ### Number of simplices in a primitive triangulation {#nbs} Another usefull formula expresses the numbers of simplices of any dimension in the primitive triangulation $\tau$ of $\Delta$ in terms of the coefficients of the Ehrhart polynomial of $\Delta$. In the case of a primitive triangulation, these numbers happen not to depend on the primitive triangulation chosen. This statement can be found in [@Dai] and in Appendix \[ksimplices\] we provide a proof in the case of convex triangulations. Let $nbs_{r}^{F}$ be the number of $r$-dimensional simplices in a primitive triangulation of $F$ which are contained in the interior of $F$, and let $i$ be the dimension of $F$. Let $S_2$ be the Stirling number of the second kind defined by $S_2(i,j)=1/(j)! \sum_{k=0}^{j} (-1)^{j-k}{\binom{j}{k}} k^i$. \[nbsrformula\] Under the above hypotheses we have, $$N_{r}^{F}=\sum_{l=r+1}^{i+1} (-1)^{i-l+1}\cdot r! S_2(l,r+1) \cdot a^{F}_{l-1}.$$ Danilov and Khovanskii formulae {#S:DK} ------------------------------- V. Danilov and A. Khovanskii [@DanKho] computed the Hodge numbers of a smooth hypersurface in a toric variety $X_\Delta$ in terms of the polytope $\Delta$. Recall that, for a face $F$ of $\Delta$, the coefficient of the term of degree $l$ of the Ehrhart polynomial of $F$ is denoted by $a_l^F$. Let $\mathcal{F}_i(\Delta)$ be the set of $i$-dimensional faces of $\Delta$, and $f_i(\Delta)$ be the cardinality of $\mathcal{F}_i(\Delta)$. A $d$-dimensional [[******]{}simple]{} polytope is one for which each vertex is contained in exactly $d$ edges. \[DK\] Let $\Delta$ be a simple polytope of dimension $d$, and $Z$ be a smooth or quasi-smooth[^1] algebraic hypersurface with Newton polytope $\Delta$ in $X_\Delta$. Then, for $p\neq \frac{d-1}{2}$ $$\begin{aligned} h^{p,p}(Z) & = & {(-1)}^{p+1} \sum_{i=p+1}^{d}{(-1)}^i {\binom{i}{p+1}} f_i(\Delta)\\ h^{\frac{d-1}{2},\frac{d-1}{2}}(Z) & = & {(-1)}^{\frac{d+1}{2}} \sum_{i=\frac{d+1}{2}}^{d}{(-1)}^i {\binom{i}{\frac{d+1}{2}}} f_i(\Delta) - \sum_{i=\frac{d+1}{2}}^{d} \sum_{F \in \mathcal{F}_i(\Delta)} {(-1)}^i \Psi_{\frac{d+1}{2}}(F)\\ h^{p,d-1-p}(Z) & = & (-1)^{d} \sum_{i=p+1}^{d} \sum_{F \in \mathcal{F}_i(\Delta)}{(-1)}^i \Psi_{p+1}(F)\\ h^{p,q} & = & 0 \mbox{ if } q \neq p \; \mbox{or} \; p \neq d-1-p,\end{aligned}$$ where $\Psi_{p+1}(F)=\sum_{\alpha=1}^{i+1}\sum_{a=0}^{p+1} (-1)^a {\binom{i+1}{a}}(p+1-a)^{\alpha -1} a^{F}_{\alpha -1 } $. Primitive $T$-hypersurfaces {#sigmachi} =========================== Statement --------- Let $\Delta$ be a $d$-dimensional polytope corresponding to a nonsingular toric variety. Let $Z$ be a primitive $T$-hypersurface with Newton polytope $\Delta$ in $X_\Delta$, and $\tau$ be a primitive triangulation from which one can construct $Z$ via the combinatorial Viro method. Notice that since $X_\Delta$ is smooth the $T$-hypersurface $Z$ is also smooth. Define $\sigma(Z)$ by $$\sigma(Z):= \sum_{p+q=0\> [2]} {(-1)}^p h^{p,q}(Z).$$ If $d$ is odd and $Z$ is smooth, the intersection form $$\iota: H_{d-1}(Z;{{\mathbb{Z}}})/tors \times H_{d-1}(Z;{{\mathbb{Z}}})/tors \rightarrow {{\mathbb{Z}}}$$ is a symmetric bilinear form and its signature is called the signature of $Z$ and is equal to $\sigma(Z)$. (See, for example, [@GH] Chapter 0, Section 7 p. 126). We have the following statement. \[theprop\] If $Z$ is a primitive real algebraic $T$-hypersurface in a nonsingular toric variety $X_\Delta$, then $$\chi({{\mathbb{R}}}Z) = \sigma(Z),$$ where $\chi({{\mathbb{R}}}Z)$ is the Euler characteristic of the real part ${{\mathbb{R}}}Z$ of $Z$. \[quasismooth\] For simplicity we state Theorem \[theprop\] for a nonsingular ambiant toric variety. In fact, we only need to require that the polytope $\Delta$ be simple (see Subsection \[S:DK\]). In this case $X_\Delta$ is quasi-smooth (i.e. locally isomorphic to a toric variety defined by a simplicial cone; see [@Dan78] Section 14). Then $Z$ is also quasi-smooth (see [@Dan78] Section 13 ), the Hodge numbers can be calculated by Danilov and Khovanskii formulae (see Theorem \[DK\]) and the proof is verbatim. The signature $\sigma(Z)$ is the same for any generic hypersurface $Z$ with Newton polytope $\Delta$ in $X_\Delta$ but if $Z$ is not a primitive $T$-hypersurface $\chi({{\mathbb{R}}}Z)$ needs not be equal to the signature (and very often is not). Recall that $l^*(\Delta)$ is the number of integer points in the interior of $\Delta$. Here and further on we write $b_i(X)$ for $b_i(X;{{\mathbb{Z}}}_2)$. \[max\] Assume that $\Delta$ is a $3$-dimensional simple polytope and $Z$ is a primitive T-surface in $X_\Delta$ with Newton polytope $\Delta$. If the number $b_0({{\mathbb{R}}}Z)$ of connected components of ${{\mathbb{R}}}Z$ is at least $l^*(\Delta) + 1$, then $b_0({{\mathbb{R}}}Z)=l^*(\Delta) + 1$ and $Z$ is maximal. First, note that $l^*(\Delta)=h^{2,0}(Z)$. Then $\chi({{\mathbb{R}}}Z) \ge 2h^{2,0}(Z) + 2 - b_1({{\mathbb{R}}}Z)$. Now using the equalities $\chi({{\mathbb{R}}}Z) = \sigma( Z)=2h^{2,0}(Z) + 2 - h^{1,1}(Z)$ one gets $b_1({{\mathbb{R}}}Z) \ge h^{1,1}(Z)$. Furthermore $h^{1,0}(Z)=h^{0,1}(Z)=0$ (see Theorem \[DK\]), and thus $b_*(Z; {{\mathbb{C}}}) =2h^{2,0}(Z) + 2 + h^{1,1}(Z)$. Hence $b_*({{\mathbb{R}}}Z; {{\mathbb{Z}}}_2) \geq b_*(Z; {{\mathbb{C}}}) = b_*(Z; {{\mathbb{Z}}}_2)$. The Smith-Thom inequality implies that $b_*({{\mathbb{R}}}Z;{{\mathbb{Z}}}_2) = b_*(Z;{{\mathbb{Z}}}_2)$, and thus, that $b_1({{\mathbb{R}}}Z)=h^{1,1}(Z)$ and $b_0({{\mathbb{R}}}Z)=h^{2,0}(Z)+1$.[$\Box$ ]{} Proof of Theorem \[theprop\] ----------------------------- If $d$ is even, Theorem \[theprop\] is straightforward. Indeed, in this case $Z$ is a (smooth) odd dimensional hypersurface, so $\chi({{\mathbb{R}}}Z)=0$. On the other hand, we have the equality $h^{p,q}(Z)=h^{d-1-p,d-1-q}(Z)$ for any $p$ and $q$, and $d-1$ is odd. Thus, $\sigma(Z)= \sum_{p+q=0\> [2]} {(-1)}^p h^{p,q}(Z) = 0$. Assume now that $d$ is odd. \[exprsigchi\] We have $$\begin{aligned} \chi({{\mathbb{R}}}Z) = \sum_{i=1}^{d}\sum_{F\in \mathcal{F}_i(\Delta)} \sum_{l=2}^{i+1} \chi_{l,i+1} a_{l-1}^F & \mbox{and} & \sigma(Z) = \sum_{i=1}^{d}\sum_{F\in \mathcal{F}_i(\Delta)} \sum_{l=2}^{i+1} \sigma_{l,i+1} a_{l-1}^F,\\\end{aligned}$$ where $$\begin{aligned} \chi_{l,i+1} & = & (-1)^{i-l+1} \sum_{k=1}^{i} \frac{(2^i-2^{i-k})}{k+1} \sum_{m=0}^{k+1} (-1)^{m}{\binom{k+1}{m}} m^l,\\ \sigma_{l,i+1} & = & \sum_{p=0}^{d-1} {(-1)}^i \sum_{q=0}^{p+1} (-1)^{p+1-q} {\binom{i+1}{q}}(p+1-q)^{l -1}.\end{aligned}$$ The triangulation $\tau^*$ induces a cell decomposition $\mathcal{D}$ of $\widetilde{H}$. Let $\widetilde{I_F}$ be the image in $\widetilde{\Delta}$ of the union of the symmetric copies of the interior of a face $F$ and $\mathcal{D}(\widetilde{I_F})$ be the set of cells of $\mathcal{D}$ contained in $\widetilde{I_F}$. Put $\chi_F = \sum_{\delta \in \mathcal{D}(\widetilde{I_F}) } (-1)^{dim(\delta)}$. Then $\chi({{\mathbb{R}}}Z)=\sum_{i=1}^{d}\sum_{F\in \mathcal{F}_i(\Delta)} \chi_F$. According to Proposition \[itenprop\], if $Q$ is a $k$-simplex of $\tau$ contained in $j$ coordinate hyperplanes then the union of the symmetric copies of $Q$ contains exactly $(2^d-2^{d-k})/2^j$ cells of dimension $k-1$. An $i$-face $F$ of $\Delta^*$ contained in $j$ coordinate hyperplanes is identified with $2^{d-i-j}-1$ other copies of $F$ when passing from $\Delta^*$ to $\widetilde{\Delta}$. Thus, the number of $(k-1)$-cells in $\mathcal{D}(\widetilde{I_F})$ is equal to $\frac{(2^d-2^{d-k})}{2^{d-i}} N^F_k$, where $N_k^F$ is the number of $k$-simplices in the interior of $F$. Thus, $$\chi_F = \sum_{k=1}^{\dim F} (-1)^{k-1} (2^{\dim F} -2^{\dim F-k}) N_{k}^F \; .$$ According to Proposition \[nbsrformula\] of Section \[nbs\] $$N_{k}^{F}=\sum_{l=k+1}^{\dim F+1} k!S_2(l,k+1)(-1)^{\dim F-l+1}a^{F}_{l-1},$$ where $S_2(i,j)=1/j! \sum_{k=0}^{j}(-1)^{j-k}{\binom{j}{k}} k^i$ is the Stirling number of the second kind. Then one has $$\begin{aligned} \chi_F & = & \sum_{k=1}^{\dim F} (-1)^{k-1} (2^{\dim F} -2^{\dim F-k}) N_{k}^F \\ & = & \sum_{k=1}^{\dim F} (-1)^{\dim F -1} \frac{2^{\dim F} -2^{\dim F-k}}{k+1} \sum_{l=k+1}^{\dim F +1} (-1)^l a_{l-1}^F \sum_{m=0}^{k+1}(-1)^m {\binom{k+1}{m}} m^l \; . \end{aligned}$$ By Lemma \[vanlint\], Equality (\[nbsurj\]) (see Appendix), the sum on $l$ can be taken from $l=2$ instead of $l=k+1$ and one gets $$\chi_F = \sum_{l=2}^{\dim F +1} (-1)^l a_{l-1}^F \sum_{k=1}^{\dim F} (-1)^{\dim F-1} \frac{2^{\dim F} -2^{\dim F -k}}{k+1} \sum_{m=0}^{k+1}(-1)^m {\binom{k+1}{m}} m^l \; ,$$ which finishes the proof of the formula for $\chi({{\mathbb{R}}}Z)$. To compute $\sigma(Z)$, we use the Danilov and Khovanskii formulae (see Theorem \[DK\]). One obtains the following expression for $\sigma (Z)$: $$\begin{aligned} \sigma(Z) & = & \sum_{p=0}^{d-1} \sum_{i=p+1}^{d}{(-1)}^i \sum_{F\in \mathcal{F}_i(\Delta)} \Bigg(- {\binom{i}{p+1}} + {(-1)}^{p+1} \sum_{l=1}^{i+1}\sum_{q=0}^{p+1} (-1)^q {\binom{i+1}{q}}(p+1-q)^{l -1} a^{F}_{l -1} \Bigg).\end{aligned}$$ Consider $\sigma(Z)$ as an affine polynomial in the variables $a^{F}_{l}$. Denote by $\sigma^{cst}$ the constant term of $\sigma(Z)$, and by $\sigma^1$ the sum of monomials in variables $a^{F}_{0}$. We have $$\sigma^{cst}= - \sum_{p=0}^{d-1} \sum_{i=p+1}^{d}{(-1)}^i {\binom{i}{p+1}} f_i(\Delta)$$ and $$\sigma^1 = \sum_{p=0}^{d-1} \sum_{i=p+1}^{d}{(-1)}^i {(-1)}^{p+1} \sum_{F\in\mathcal{F}_i(\Delta)} \sum_{q=0}^{p+1} (-1)^q {\binom{i+1}{q}} a^F_0 .$$ One has $\sigma^{cst}+\sigma^1=0$. Since $ a^0_F= 1$, we obtain $$\begin{aligned} \sigma^1 &= &\sum_{p=0}^{d-1} \sum_{i=p+1}^{d}{(-1)}^i {(-1)}^{p+1} \sum_{F\in\mathcal{F}_i(\Delta)} \sum_{q=0}^{p+1} (-1)^q {\binom{i+1}{q}} \\ & = & \sum_{p=0}^{d-1} \sum_{i=p+1}^{d}{(-1)}^i {(-1)}^{p+1} f_i(\Delta) \sum_{q=0}^{p+1} (-1)^q {\binom{i+1}{q}}.\end{aligned}$$ Furthermore, $\sum_{q=0}^{p+1} (-1)^q {\binom{i+1}{q}} = (-1)^{p+1}{\binom{i}{p+1}}$, and thus, $$\begin{aligned} \sigma^1 & = &\sum_{p=0}^{d-1} \sum_{i=p+1}^{d}{(-1)}^i {\binom{i}{p+1}} f_i(\Delta). \end{aligned}$$ [$\Box$]{} For any $p \ge i$, $\sum_{q=0}^{p+1} (-1)^q {\binom{i+1}{q}}(p+1-q)^{l -1} =\sum_{q=0}^{i+1} (-1)^q {\binom{i+1}{q}}(p+1-q)^{l -1}$ which is zero by Lemma \[vanlint\] since $l\le i+1$. Thus one gets $$\label{sli} \sigma_{l,i+1} = \sum_{p=0}^{i-1} {(-1)}^i {(-1)}^{p+1} \sum_{q=0}^{p+1} (-1)^q {\binom{i+1}{q}}(p+1-q)^{l -1}.$$ [$\Box$]{} We can then prove the following lemma. \[sigchirec\] For $2 \le l \le i+1$, one has the following equalities $$\begin{aligned} \sigma_{l,i+2}& = &-2\sigma_{l,i+1},\\ \chi_{l,i+2} & = & -2 \chi_{l,i+1}.\end{aligned}$$ To prove the first equality, write ${\binom{i+2}{q}}={\binom{i+1}{q-1}} + {\binom{i+1}{q}}$ to get $$\displaylines{ \sigma_{l,i+2} = {(-1)}^{i+1}\sum_{p=0}^{i} {(-1)}^{p+1} \sum_{q=1}^{p+1} (-1)^q {\binom{i+1}{q-1}}(p+1-q)^{l -1}\cr + {(-1)}^{i+1}\sum_{p=0}^{i} {(-1)}^{p+1} \sum_{q=0}^{p+1} (-1)^q {\binom{i+1}{q}}(p+1-q)^{l -1}.}$$ Notice that $\sum_{q=1}^{1} (-1)^q {\binom{i+1}{q-1}}(p+1-q)^{l -1}$ is $0$. Similarly, in the second term,\ $\sum_{q=0}^{i+1} (-1)^q {\binom{i+1}{q}}(i+1-q)^{l -1}$ makes no contribution, by Lemma \[vanlint\]. Then, $$\displaylines{ \sigma_{l,i+2} = {(-1)}^{i+1}\sum_{p=1}^{i} {(-1)}^{p+1} \sum_{q=1}^{p+1} (-1)^q {\binom{i+1}{q-1}}(p+1-q)^{l -1}\cr + {(-1)}^{i+1}\sum_{p=0}^{i-1} {(-1)}^{p+1} \sum_{q=0}^{p+1} (-1)^q {\binom{i+1}{q}}(p+1-q)^{l -1}.}$$ So, with the changes of indices $ c=q-1$ and $ r=p-1$ one gets $$\begin{aligned} \sigma_{l,i+2}&=& {(-1)}^{i+1}\sum_{r=0}^{i-1} {(-1)}^{r} \sum_{c=0}^{r+1} (-1)^{c+1} {\binom{i+1}{c}}(r+1-c)^{l -1} \\ & & + {(-1)}^{i+1}\sum_{p=0}^{i-1} {(-1)}^{p+1} \sum_{q=0}^{p+1} (-1)^q {\binom{i+1}{q}}(p+1-q)^{l -1}\\ &=& -2 \sigma_{l,i+1}.\end{aligned}$$ To prove the second equality stated in the lemma, use Lemma \[vanlint\] to get rid of the term with $k=i+1$ in the expression of $\chi_{l,i+2}\,$: $$\begin{aligned} \chi_{l,i+2} & = & (-1)^{i-l+2} \sum_{k=1}^{i+1} \frac{(2^{i+1}-2^{i+1-k})}{k+1} \sum_{m=0}^{k+1} (-1)^{m}{\binom{k+1}{m}} m^l \\ & = & 2 (-1)^{i-l} \sum_{k=1}^{i} \frac{(2^{i}-2^{i-k})}{k+1} \sum_{m=0}^{k+1} (-1)^{m}{\binom{k+1}{m}} m^l \,,\end{aligned}$$ which finishes the proof. [$\Box$ ]{} \[sigeqchi\] For $2 \le l \le i+1 \le d+1$, we have $\chi_{l,i+1}=\sigma_{l,i+1}$. By Lemma \[sigchirec\] it is enough to prove the equality $\chi_{l,l}=\sigma_{l,l}$. By Formula (\[sli\]) one has, $$\begin{aligned} \label{sll} \sigma_{l,l} & = & \sum_{p=0}^{l-2} {(-1)}^{l-1} \sum_{q=0}^{p+1} (-1)^{p+1-q} {\binom{l}{q}}(p+1-q)^{l -1}\label{sll1}\\ & = &\sum_{b=0}^{l} {(-1)}^{l-1} \sum_{q=0}^{l-b} (-1)^b {\binom{l}{q}}(b)^{l-1}\label{sll2}\\ & = & {(-1)}^{l-1} \sum_{b=1}^{l}\sum_{j=b}^{l} (-1)^{b} {\binom{l}{j}}b^{l -1},\label{sll3}\end{aligned}$$ where we pass from (\[sll1\]) to (\[sll2\]) by setting $b=p+1-q$ and from (\[sll2\]) to (\[sll3\]) by putting $j=l-q$. One sees that $$\begin{aligned} \chi_{l,l} &= & \sum_{k=1}^{l-1} \frac{2^{l-1}-2^{l-1-k}}{k+1} \sum_{m=0}^{k+1} (-1)^{m}{\binom{k+1}{m}} m^l \\ &= &\sum_{k=0}^{l-1} \frac{2^{l-1}-2^{l-1-k}}{k+1} \sum_{m=1}^{k+1} (-1)^{m}{\binom{k+1}{m}} m^l \,, \end{aligned}$$ just noticing that the change of range does not affect the sum. Since ${\binom{k+1}{m}}\frac{m^l}{k+1}= {\binom{k}{m-1}} m^{l-1}$, $$\begin{aligned} \chi_{l,l}&= &\sum_{k=0}^{l-1} (2^{l-1}-2^{l-1-k}) \sum_{m=1}^{k+1} (-1)^{m}{\binom{k}{m-1}} m^{l-1} \\ & = & 2^{l-1} \sum_{k=0}^{l-1}\sum_{m=1}^{k+1} (-1)^{m}{\binom{k}{m-1}} m^{l-1} - \sum_{k=0}^{l-1} 2^{l-1-k}\sum_{m=1}^{k+1} (-1)^{m}{\binom{k}{m-1}} m^{l-1}\\ & = & 2^{l-1}\sum_{m=1}^{l} (-1)^{m} m^{l-1}\sum_{k=m-1}^{l-1}{\binom{k}{m-1}} - \sum_{m=1}^{l}(-1)^{m} m^{l-1} \sum_{k=m-1}^{l-1} 2^{l-1-k} {\binom{k}{m-1}}\\ & = & 2^{l-1}\sum_{m=1}^{l} (-1)^{m} m^{l-1}{\binom{l}{m}} - \sum_{m=1}^{l}(-1)^{m} m^{l-1} \sum_{j=m}^{l}{\binom{l}{j}},\\\end{aligned}$$ since $\sum_{k=m-1}^{l-1}{\binom{k}{m-1}}={\binom{l}{m}}$ and $\sum_{k=m-1}^{l-1} 2^{l-1-k} {\binom{k}{m-1}}$ by Lemma \[binrmq\]. By Lemma \[vanlint\], $ \sum_{m=1}^{l} (-1)^{m} m^{l-1}{\binom{l}{m}} = 0 $ thus $$\chi_{l,l}= - \sum_{m=1}^{l}(-1)^{m} m^{l-1} \sum_{j=m}^{l}{\binom{l}{j}}\,$$ and $\chi_{l,l} = (-1)^l \sigma_{l,l}$. This is the desired equality for $l$ even. For an odd $l > 2$ we use the symmetry of the expression of $\sigma_{l,l}$ and write $$\displaylines{\sigma_{l,l} = \sum_{p=0}^{\frac{l-3}{2}} \Big({(-1)}^{p+1} \sum_{q=0}^{p+1} (-1)^q {\binom{l}{q}}(p+1-q)^{l -1} \cr + (-1)^{l-1-p} \sum_{r=0}^{l-1-p} (-1)^r {\binom{l}{r}}(l-1-p-r)^{l-1}\Big).}$$ With the change of index $q=l-r$ in the second term of the sum and noticing that the contribution for $q=p+1$ is zero, one gets $$\sigma_{l,l} = \sum_{p=0}^{\frac{l-3}{2}} {(-1)}^{p+1} \sum_{q=0}^{l}(-1)^q{\binom{l}{q}} (p+1-q)^{l-1}.$$ The right hand side of the last equality is zero by Lemma \[vanlint\], and thus, $\sigma_{l,l} = \chi_{l,l} = 0$. [$\Box$ ]{} According to Lemma \[sigeqchi\] the coefficients of $\sigma(Z)$ and $\chi({{\mathbb{R}}}Z)$ in the expressions of Lemma \[exprsigchi\] are equal, and thus, $\sigma(Z) = \chi({{\mathbb{R}}}Z)$.[$\Box$ ]{} Nakajima polytopes and M-surfaces {#Mhypex} ================================= In this section we give examples of families of $M$-surfaces obtained by $T$-construction as hypersurfaces of $3$-dimensional toric varieties. Newton polytopes without maximal hypersurfaces {#Mhypconterex} ---------------------------------------------- We first explain that Itenberg-Viro’s theorem (see [@IteVir]) on the existence of $M$-hypersurfaces of any degree in the projective spaces of any dimension cannot be generalized straightforwardly to all projective toric varieties. Indeed, in dimensions greater or equal to $3$, there exist polytopes $\Delta$ such that no hypersurface in $X_\Delta$ with Newton polytope $\Delta$ is maximal (See Proposition \[nomax2\]). Note that this does not mean that the toric variety $X_\Delta$ does not admit $M$-hypersurfaces. Clearly, if $\Delta$ is an interval $[a, b]$ in ${{\mathbb{R}}}$, where $a$ and $b$ are nonnegative integers, then there exists a maximal $0$-dimensional subvariety in ${{\mathbb{C}}}P^1 = X_\Delta$ with the Newton polygon $\Delta$. If $\Delta$ is a polygon in the first quadrant of ${{\mathbb{R}}}^2$, then again there exists a maximal curve in $X_\Delta$ with the Newton polygon $\Delta$. Such a curve can be constructed by the combinatorial patchworking: it suffices to take as initial data a primitive convex triangulation of $\Delta$ equipped with the following distribution of signs: an integer point $(i, j)$ of $\Delta$ gets the sign “-” if $i$ and $j$ are both even, and gets the sign “+”, otherwise (cf., for example, [@Iterag], [@IteVir2], [@Haa]). However, in dimension $3$ there are polytopes $\Delta$ such that no surface in $X_\Delta$ with the Newton polytope $\Delta$ is maximal (See Proposition \[nomax\]). Let $k$ be a positive integer number, and $\Delta_k$ be the tetrahedron in ${{\mathbb{R}}}^3$ with vertices $(0,0,0), (1,0,0), (0,1,0)$, and $(1,1,k)$. Note that the only integer points of $\Delta_k$ are its vertices. \[nomax\] For any odd $k \geq 3$ and any even $k \geq 8$, there is no maximal surface in $X_{\Delta_k}$ with the Newton polytope $\Delta_k$. The proof of the above Proposition can be found in [@these] and [@Moi1]. There, this family of examples is generalized to higher dimensional polytopes and the following proposition is proved. \[nomax2\] For any integer $d \ge 3$ there exist $d$-dimensional polytopes $\Delta$ such that no hypersurface in $X_\Delta$ with the Newton polytope $\Delta$ is maximal. Construction of $M$-surfaces {#S:msurf} ---------------------------- Using Corollary \[max\] of Theorem \[theprop\] we prove the existence of maximal surfaces in all nonsingular toric varieties corresponding to Nakajima polytopes of dimension $3$ (see Figure \[figure:nakapol3\]). A Nakajima polytope (See Introduction) is a [[******]{} nondegenerate]{} if it is $0$-dimensional or if $\bar{\Delta}$ is nondegenerate and $f$ is positive on $\bar{\Delta}$. \[P2\] Let $\Delta$ be a $3$-dimensional Nakajima polytope corresponding to a nonsingular toric variety $X_\Delta$. Then, there exists a maximal surface in $X_\Delta$ with the Newton polytope $\Delta$. Let us first assume that $\Delta$ is nondegenerate. Decompose $\Delta$ into slices $T_k = \Delta \cap \{ (x,y,z)\in ({{\mathbb{R}}}^+)^3 ,k-1 \le x \le k \}$. Let $s_k$ be the section $\Delta \cap \{ (x,y,z)\in ({{\mathbb{R}}}^+)^3, x = k\}$. Denote by $B^{0,k}$, $B^{1,k}$, $B^{2,k}$, and $B^{3,k}$ the vertices of $s_k$ (see Figure \[figure:nakapol32\]). We subdivide the slices $T_k$ in two cones and two joins. Take the cones of apex $B^{0,2k}$ and $B^{0,2k+2}$ on $s_{2k+1}$ and the cones of apex $B^{3,2k-1}$ and $B^{3,2k+1}$ on $s_{2k}$. Take primitive convex triangulations of the sections $s_{k}$. They induce a primitive triangulation of the cones. The joins $[B^{0,2k},B^{1,2k}] * [B^{1,2k+1}, B^{3,2k+1}]$, $[B^{0,2k},B^{2,2k}] * [B^{2,2k+1}, B^{3,2k+1}]$, $[B^{0,2k+2},B^{1,2k+2}] * [B^{1,2k+1}, B^{3,2k+1}]$, $[B^{0,2k+2},B^{2,2k+2}] * [B^{2,2k+1}, B^{3,2k+1}]$ are also naturally primitively triangulated which gives a primitive convex triangulation of $\Delta$. Take the following distribution of signs at the integer points of $\Delta$: [*any integer point $(i,j,k)$ gets “-” if $j$ and $k$ are both odd, and it gets “+”, otherwise.*]{} The polytope $\Delta$ is now equipped with a convex primitive triangulation and a sign distribution, so we can apply the $T$-construction as in Section \[tcon\]. Let $p$ be an integer interior point of $\Delta$ and ${\operatorname{Star}}(p)$ be its star. Assume that $p$ belongs to a section $s_{l}$. Then, ${\operatorname{Star}}(p)$ has two vertices $c_1$ and $c_2$ outside $s_{l}$. They are the apices of the two cones over $s_{l}$. They have the same parity (i.e. their coordinates have the same reduction modulo $2$) and the distribution of signs depends only on the parity thus, in each octant, their symmetric copies carry the same sign. Consider an octant where the symmetric copy $q_0$ of $p$ is isolated in $s_{l}$ (i.e. all vertices of ${\operatorname{Star}}(q_0) \cap s_{l}$ except $q_0$ carry the sign opposite to the sign of $q_0$). Then, either in this octant the symmetric copies of $c_1$ and $c_2$ carry the sign opposite to the sign of $q_0$ (and, hence, $q_0$ is isolated), or $r(q_0)$ is isolated, where $r$ is the reflection with respect to the coordinate plane $x = 0$. Thus, for each integer interior point $p$ of $\Delta$, there exists a symmetric copy $q$ of $p$ such that $q$ is surrounded by a sphere $S^2(p) = {\operatorname{Star}}(q) \cap H^*$ (see Remark \[sphere\]). Moreover at least one component of $H^*$ intersects the coordinate planes. Thus, the $T$-surface constructed has at least $l^*(\Delta) +1$ connected components, and Corollary \[max\] shows that this surface is maximal. The degenerate case splits into two subcases (Since we only consider here Nakajima polytopes corresponding to nonsingular toric varieties). Either the Nakajima polytope is a truncated cylinder over a triangle corresponding to the projective plane, or it is a tetrahedron corresponding to the projective $3$-space. The existence of $M$-surfaces in the latter case was proved by Viro in [@Vir79]. In the former case the Nakajima polytope $\Delta$ is the convex hull of the triangles $((0,0,0), (m,0,0), (0,m,0))$ and $((0,0,l), (0,m,l+me), (m,0,l+mf))$ for some integers $m$, $l$, $e$ and $f$ (see Figure \[figure:nakatri2\]). Decompose $\Delta$ into slices $T_k = \Delta \cap \{ (x,y,z)\in ({{\mathbb{R}}}^+)^3 ,k-1 \le x \le k \}$. Let $s_k$ be the section $\Delta \cap \{ (x,y,z)\in ({{\mathbb{R}}}^+)^3, x = k\}$. We triangulate $\Delta \cap \{ (x,y,z)\in ({{\mathbb{R}}}^+)^3 ,y \le m-1 \}$ using the same triangulation as in the nondegenerate case. Take the cone over $s_{m-1}$ of apex $(m,0,0)$ if $m$ is even or $(m,0,l+mf)$, otherwise. Take in the cone the triangulation induced by the triangulation of $s_{m-1}$. Subdivide the only remaining non-primitive tetrahedron into primitive ones in the unique possible way. Take the same distribution of signs that in the nondegenerate case. Then, as in the nondegenerate case, any interior point has an isolated symmetric copy. There is also a component intersecting the coordinate planes. Thus, the surface is maximal. [$\Box$ ]{} Note that Theorem \[P2\] produces, in particular, $M$-surfaces in $\Sigma_\alpha \times {{\mathbb{C}}}P^1$. For a nonnegative integer $\alpha$ and positive integers $l$, $m$ and $n$ let $\delta_\alpha^{m,n}$ and $\Delta^{l,m,n}_\alpha$ be the polytopes defined in Section \[S:tropstate\]. Recall that the toric variety $X_{\Delta^{l,m,n}_\alpha}$ is isomorphic to $\Sigma_\alpha \times {{\mathbb{C}}}P^1$. Since $\Delta^{l,m,n}_\alpha$ is a Nakajima polytope, the following statement is a corollary of Theorem \[P2\]. \[hirz\] For any nonnegative integer $\alpha$ and any positive integers $m$, $n$ and $l$, there exists a maximal surface in $\Sigma_\alpha \times {{\mathbb{C}}}P^1$ with the Newton polytope $\Delta^{l,m,n}_\alpha$.[$\Box$ ]{} Appendix ======== Usefull combinatorial lemmae ---------------------------- We insert here the proofs of lemmae that we use quite often, mainly in Section \[sigmachi\]. \[vanlint\] Let $l$ and $i$ be nonnegative integers. Then, for $l + 1 \le i$, one has (See [@LiW] p. 71) $$\label{vl} \sum_{q=0}^{i} (-1)^q {\binom{i}{q}} q^{l} \;= \; \sum_{q=0}^{i} (-1)^q {\binom{i}{q}}(i-q)^{l} = 0$$ and, as a consequence, for any integer $p$, $$\label{nbsurj} \sum_{q=0}^{i} (-1)^q {\binom{i}{q}}(p-q)^{l} = 0.$$ Formula \[vl\] can be found in [@LiW] p. 71. We prove Equality (\[nbsurj\]) using (\[vl\]). Write $p-q = (i-q)+(p-i)$ and then, $$\begin{aligned} (p-q)^l & = & \sum_{m=0}^{l} {\binom{l}{m}} (i-q)^{l-m}(p-i)^m .\\\end{aligned}$$ So, $$\begin{aligned} \sum_{q=0}^{i} (-1)^q {\binom{i}{q}}(p-q)^{l}& = & \sum_{q=0}^{i} (-1)^q {\binom{i}{q}} \sum_{m=0}^{l} {\binom{l}{m}} (i-q)^{l-m}(p-i)^m\\ & = & \sum_{m=0}^{l} {\binom{l}{m}}(p-i)^m \Bigg(\sum_{q=0}^{i} (-1)^q {\binom{i}{q}} (i-q)^{l-m} \Bigg)\end{aligned}$$ The last term is zero by Equality \[vl\]. [$\Box$ ]{} [**Proof of lemma \[psi\]**]{}. From the definitions we see that $$Q_\Delta(t) = (\sum_{k=0}^{d+1} {\binom{d+1}{k}} t^k (-1)^k) \cdot \sum_{n=0}^\infty (\sum_{i=0}^d a_i n^i) t^n,$$ and thus, $\Psi_j = \sum_{i=0}^d (\sum_{n=0}^j (-1)^{j-n}{\binom{d+1}{j-n}} n^i) a_i^\Delta$. Let $A_j = \sum_{n=0}^j (-1)^{j-n}{\binom{d+1}{j-n}} n^i$.\ If $j \ge d+1$, then $$A_j=\sum_{k=0}^{d+1} (-1)^k {\binom{d+1}{k}} (j-k)^i,$$ which is zero according to Lemma \[vanlint\] Equality \[nbsurj\]. [$\Box$ ]{} \[binrmq\] One has $\sum_{n=0}^{p} 2^{p-n} {\binom{n}{k}} = \sum_{l=k+1}^{p+1} {\binom{p+1}{l}}$. We use the fact that $\sum_{n=k}^{p} {\binom{n}{k}} = {\binom{p+1}{k+1}}$ and write $$\begin{aligned} \sum_{n=0}^{p} 2^{p-n}{\binom{n}{k}} & = & \sum_{n=k}^{p} 2^{p-n}{\binom{n}{k}}\\ & = & \sum_{n=k}^{p} {\binom{n}{k}} + \sum_{i=0}^{p} (2^i)\sum_{n=k}^{p-i}{\binom{n}{k}}\\ & = & {\binom{p+1}{k+1}} + \sum_{i=0}^{p-k}(2^i) \sum_{i=0}^{p-k} {\binom{p+1-i}{k+1}}\\ & = & {\binom{p+1}{k+1}} + \sum_{n=k+1}^{p+1} 2^{p+1-n}{\binom{n}{k+1}},\\\end{aligned}$$ and the result follows by induction.[$\Box$ ]{} Formula for the number of $k$-simplices {#ksimplices} --------------------------------------- In order to prove the formula of Proposition \[nbsrformula\] we first check that it is true for all primitive simplices. ### Formula for a primitive simplex Let $s_i$ be a primitive simplex of dimension $i$. Then $$Ehr_{s_i}(\lambda) = \frac{1}{i !}(\lambda + 1) \dots (\lambda + i)=\frac{1}{i !} \sum_{j=1}^{i+1}{(-1)}^{i+1-j}S_1(i+1,j) \lambda^{j-1},$$ where $S_1$ is the first Stirling number defined by the formula $\sum_{m=0}^n S_1(n,m)x^m= x (x-1) \dots (x-n+1)$ (see [@BV]). Thus $$a^{s_i}_{j-1}=\frac{(-1)^{i+1-j}}{i !}S_1(i+1,j).$$ It remains to show that $$N_{r}^{s_i}=\frac{r !}{i!} \sum_{l=r+1}^{i+1} S_2(l,r+1).S_1(i+1,l).$$ We have $\sum_{l=r+1}^{i+1} S_2(l,r+1).S_1(i+1,l)=\delta_{r,i}$ where $\delta_{r,i}$ is the Kronecker index (see [@LiW] p. 107), and thus, $$\frac{r !}{i!}\sum_{l=r+1}^{i+1} S_2(l,r+1).S_1(i+1,l) =\delta_{r,i}$$ which is exactly the number of $r$-dimensional simplices contained in the interior of $s_i$ (zero if $r \neq i$ and one othewise). ### Proof for a general polytope A triangulation $\tau$ of a polytope of dimension $d$ is called [[******]{} shellable]{} if there exists a numbering of its $d$-simplices $s_1, s_2 , \dots ,s_k$ such that for $i \in \{2, \dots, k\}$ $$s_i \cap \cup_{j=1}^{i-1} s_j$$ is a nonempty union of $(d-1)$-simplices of $\tau$ homeomorphic to a $(d-1)$-ball. This numbering is called a [*shelling*]{} of $\tau$. Every convex triangulation of a polytope is shellable. The formula of Proposition \[nbsrformula\] holds for a point (with the convention that the point is in its interior). We now assume that the formula is true in all dimensions less than $d$. Let $\Delta$ be a $d$-dimensional polytope, and $\tau$ be a primitive triangulation of $\Delta$. Fix $s_1,s_2,\dots, s_t$ a shelling of $\tau$, and put $U_j=\cup_{i=1}^j s_j$. Note that $U_j$ need not to be convex. Assume that the formula is true for $U_j$. Note that the formula also holds for $U_j\cap s_{j+1}$ which is $(d-1)$-dimensional, eventhough it is not a polytope. We know that the formula is also true for $s_{j+1}$. The numbers $N_r$ satisfy the relation $$N_r^{U_j} + N_r^{s_{j+1}} + N_r^{U_j \cap s_{j+1}} = N_r^{U_{j+1}}.$$ This follows immediately from the fact that $ \stackrel{\circ}{U_j} \sqcup \stackrel{\circ}{s_{j+1}} \sqcup \stackrel{\circ}{ ( U_j \cap s_{j+1})} = \stackrel{\circ}{U_{j+1}}$, where $\stackrel{\circ}{U}$ stands for the interior of $U$. [$\Box$ ]{} By induction hypothesis, $$\begin{aligned} N_r^{U_j \cap s_{j+1}} & = & \sum_{l=r+1}^{d} {\operatorname{N}}(r+1,l) (-1)^{d-l} a_{l-1}^{U_j \cap s_{j+1}},\\\end{aligned}$$ and since $ a_{d}^{U_j \cap s_{j+1}}=0$, we can also write $$\begin{aligned} N_r^{U_j \cap s_{j+1}} & = & \sum_{l=r+1}^{d+1} {\operatorname{N}}(r+1,l) (-1)^{d-l} a_{l-1}^{U_j \cap s_{j+1}}.\\\end{aligned}$$ Then $N_r^{U_{j+1}} = \sum_{l=r+1}^{d+1} {\operatorname{N}}(r+1,l)(-1)^{d-l+1}.(a^{U_j}_{l-1}+ a^{s_{j+1}}_{l-1}- a^{U_j \cap s_{j+1}}_{l-1})$, and $a^{U_j}_{l-1}+ a^{s_{j+1}}_{l-1}- a^{U_j \cap s_{j+1}}_{l-1}$ is precisely the coefficient $a^{U_{j+1}}_{l-1}$ in the Ehrhart polynomial of $U_{j+1}$ (see [@Ehr67] for Ehrhart polynomials of general polyhedra). This completes the proof of the formula. [$\Box$]{} [VLW92]{} Beno[î]{}t Bertrand. Ph.d. thesis: Maximal hypersurfaces and complete intersections in toric varieties. 2003. http://www.unige.ch/math/folks/bertrand. Benoît Bertrand. Asymptotically maximal families of hypersurfaces in toric varieties. , 118:49–70, 2006. M. Brion. Points entiers dans les polytopes convexes. , 1994. Michel Brion and Mich[è]{}le Vergne. Lattice points in simple polytopes. , 10(2):371–392, 1997. D. Dais. ber unimodulare, koh[ä]{}rente [T]{}riangulierungen von [G]{}itterpolytopen. [B]{}eispiele und [A]{}nwendungen. , 2000. V. I. Danilov. The geometry of toric varieties. , 33(2(200)):85–134, 247, 1978. Dimitrios I. Dais, Christian Haase, and G[ü]{}nter M. Ziegler. All toric local complete intersection singularities admit projective crepant resolutions. , 53(1):95–107, 2001. V. Danilov and A. Khovanskii. Newton polyhedra and an algorithm for computing [H]{}odge-[D]{}eligne numbers. , pages 279–298, 1987. E. Ehrhart. Sur un problème de géométrie diophantienne linéaire. [I]{}. [P]{}olyèdres et réseaux. , 226:1–29, 1967. E. Ehrhart. Un th[é]{}or[è]{}me arithmo-g[é]{}om[é]{}trique et ses g[é]{}n[é]{}ralisations. , pages 33–34, 1994. W. Fulton. . Princeton University Press, 1993. P. Griffiths and J. Harris. . John Willey and Sons, 1978. I. Gelfand, M. Kapranov, and A. Zelevinsky. . Springer-Verlag, 1994. Bertrand Haas. . 1998. Ilia Itenberg. Counter-examples to [R]{}agsdale conjecture and [$T$]{}-curves. In [*Real algebraic geometry and topology (East Lansing, MI, 1993)*]{}, volume 182 of [*Contemp. Math.*]{}, pages 55–72. Amer. Math. Soc., Providence, RI, 1995. I. Itenberg. Topology of real algebraic [T]{}-surfaces. , 10, no. supp.:131–152, 1997. Ilia Itenberg and Oleg Viro. Patchworking algebraic curves disproves the [R]{}agsdale conjecture. , 18(4):19–28, 1996. I. Itenberg and O. Viro. Maximal real algebraic hypersurface of projective space. , 2005. G. Mikhalkin. Real algebraic curves, the moment map and amoebas. , 151(1):309–326, 2000. Grigory Mikhalkin. Amoebas of algebraic varieties and tropical geometry. In [*Different faces of geometry*]{}, Int. Math. Ser. (N. Y.), pages 257–300. Kluwer/Plenum, New York, 2004. Grigory Mikhalkin. Enumerative tropical algebraic geometry in [${{\mathbb{R}}}^2$]{}. , 18(2):313–377 (electronic), 2005. Bernd Sturmfels. Viro’s theorem for complete intersections. , 21(3):377–386, 1994. O. Ja. Viro. Construction of [$M$]{}-surfaces. , 13(3):71–72, 1979. Oleg Viro. Gluing of plane algebraic curves and construction of curves of degree 6 and 7 (lnm 1060). , pages 187–200, 1984. Oleg Viro. Dequantization of real algebraic geometry on logarithmic paper. In [*European Congress of Mathematics, Vol. I (Barcelona, 2000)*]{}, volume 201 of [*Progr. Math.*]{}, pages 135–146. Birkhäuser, Basel, 2001. O. Viro. Patchworking real algebraic varieties. , 2004. J. H. Van Lint and R. M. Wilson. . Cambridge university press, 1992. G. Ziegler. . Springer-Verlag, 1995. [^1]: See [@Dan78] and Remark \[quasismooth\]

ArXiv

--- abstract: 'We investigate the plasmonic nanohelix as an individual subwavelength element for locally probing light chirality. We show that an hybrid nanoantenna combining a carbon-gold core-shell helix and a plasmonic nanoaperture transmits circularly polarized light with the same handedness as the helix and blocks the other. Such an assymmetric response is spatially localized, spectrally broadband and background-free. Finally, we demonstrate the possibility to engineer an individual plasmonic helix at the apex of a sharp tip typically used in scanning near-field microscopies, thus opening the prospect of moveable local probes for high resolution sensing and mapping of light chirality and chiroptical forces.' address: 'FEMTO-ST Institute UMR 6174, Univ. Bourgogne Franche-Comte –CNRS – Besancon, France' author: - 'Mengjia Wang, Roland Salut, Miguel Angel Suarez, Nicolas Martin and Thierry Grosjean\*' title: Individual plasmonic helix for probing light chirality --- Chiral metamaterials and nanophotonics have recently attracted much interest for their unique ability to enhance and control inherently weak chiral light-matter interaction [@hentschel:sciadv17; @schaferling:prx12]. Research in these domains is essentially relying upon the ability to produce and control spinning light through the design of specific subwavelength systems. Characterizing light chirality is crucial to assess and improve such designs. To that end, theoretical modeling has been preparing for a long time with the development of rigorous numerical methods which allow full-wave light field simulation around arbitrarily complex nanodevices. From an experimental point-of-view, the chiroptical properties of nanostructures and metamaterials are usually detected via ensemble-averaged measurements with large, at best diffraction-limited illumination [@poulikakos:nl18; @meinzer:prb12]. Any individual chiral molecule or nanostructure is potentially a local probe of light chirality as it interacts differently to right and left spinning optical fields. However, reaching completely detangled information from the different handedness of light seems to be challenging at the scale of an individual subwavelength chiral element. Background problems and modest chiroptical responses may impede the development of individual nanoscale chiral probes without any complex signal post-processing involving bulky optics. Aperture nanoantennas have been recently demonstrated to accurately probe light fields at subdiffraction scale [@taminiau:nl07; @burresi:science09; @mivelle:nl14; @murphy:ox08; @grosjean:ox10; @vo:ox12; @lindquist:scirep13; @xie:nl17]. These subwavelength structures have been integrated at the apex of different kinds of tips of scanning near-field microscopies to be approached and raster-scanned across a sample while picking up detailed and background-free near-field optical information. The local detection of light chirality with individual subwavelength aperture nanoantennas has been hardly addressed. Strong chiroptical responses have been obtained from helical nanoantennas [@kaschke:nanophotonics16], leading to giant circular dichroism [@gibbs:apl13; @esposito:acsphot14; @esposito:natcom15], strongly polarization-dependent transmission [@gansel:sci09; @kosters:acsphot17] and superchiral light [@schaferling:acsphot14]. Core-shell nanohelices have been recently proposed to enhance chiroptical effects at visible frequencies by manipulating surface plasmons [@kosters:acsphot17]. Individual resonant core-shell helices have been shown to sustain chiral dipoles associated with resonances [@wozniak:ox18; @wozniak:optica19]. In a non-resonant operation, the carbon-gold core-shell plasmonic helix has led to the concept of helical traveling-wave nano-antenna (HTN) enabling subwavelength polarization optics and a new regime in polarization control [@wang:lsa19]. In this paper, we investigate the carbon-gold core-shell helix as a broadband subwavelength probe of light chirality. When coupled to a nanoaperture in an HTN configuration [@wang:lsa19], a left-handed carbon-gold core-shell is shown to transmit light of left circular polarization and blocks the other. A differential transmission larger than 0.96 is obtained for wavelengths ranging from 1.47 to 1.65 $\mu$m. Finally, we show that individual carbon-gold core-shell helices can be fabricated at the apex of a tip used in scanning near-field microscopy. This opens the prospect of on-tip local probes for measuring light chirality around nanostructures via direct optical detection or chiroptical force sensing [@zhao:natnano17].\ We fabricated individual carbon-gold core-shell helices to operate at telecommunication wavelengths ($\lambda$ around 1.55 $\mu$m). The carbon helical skeleton is sculpted by focused ion beam induced deposition (FIBID) [@esposito:nl16; @wang:lsa19]. Metal coating is realized by sputter-depositing a thin layer of gold onto the carbon core with a metal target tilted with an angle of 80$^{\circ}$ from the helix axis and rotated during the deposition time with a constant rotating speed of 2 rev.min$^{-1}$. We combined these two techniques to fabricate plasmonic helices consisting of a 105-nm diameter carbon wire wound up in the form of a four-turn corkscrew-type structure and covered with a 25-nm thick gold layer. The resulting helix has a 505-nm outer diameter and is 1.66-$\mu$m high. It is positioned on a cylindrical pedestal of 105 nm diameter and 100 nm high carbon whose lateral side is also coated with a 25-nm thick gold layer. ![(a) Scanning electron micrograph of a gold-coated carbon helix coupled to a rectangle nanoaperture: HTN configuration. (b) Scanning electron micrograph of a cross-section of the helix initially embedded in platinum by FIBID. (c) Reconstructed 3D image of the fabricated gold-coated carbon helix via FIB/SEM tomography of the platinum embedded nanostructure. Yellow: gold; black: carbon.[]{data-label="fig:fab1"}](fabrication.eps){width="0.7\linewidth"} We considered this plasmonic helix as a part of a non-resonant helical traveling-wave nanoantenna [@wang:lsa19]. To this end, the core-shell structure is engineered onto a 100-nm thick gold layer deposited onto a glass substrate and a 370 nm-by-40 nm rectangle nano-aperture is engraved by focused ion beam (FIB) milling in the flat gold layer right at the helix pedestal (Fig. \[fig:fab1\](a)). To evaluate the thickness of the gold layer deposited onto the helical carbon core, we perform tomography of the helix by alternating FIB slicing and SEM imaging of the nanostructure. To this end, the helix is embedded in platinum deposited by FIBID. We see in the helix cross-section of Fig. \[fig:fab1\](b) that the gold layer (white rings in the image) is not homogeneously distributed all around the carbon wire (dark circular regions in the SEM image). Figure \[fig:fab1\](c) shows a 3D reconstruction of the gold-coated carbon helix revealing a bottom side that is not fully covered with gold. The tilt angle of the gold target of 80$^{\circ}$ creates a shadowing effect in the gold coating process which prevents a a uniform gold layer to be deposited all around the helical wire. Non homogeneously-distributed metal coating around the carbon core mat spectrally redshift the helix optical response [@wang:lsa19]. We performed three-dimensional FDTD simulations using commercial software (Fullwave). The antenna is modeled as a wounded cylindrical carbon wire coated with a 25-nm thick gold layer. The helix is placed next to a rectangle aperture in an extended gold slab lying on a glass substrate. The geometrical parameters of the simulated HTN are those of the structure introduced above. The numerical parameters of the simulations can be found in [@wang:lsa19]. The helix is illuminated with a circularly polarized gaussian beam (beam waist: 2.3 $\mu$m) propagating along the helix axis. Because we consider this hybrid nanoantenna as an chiral optical probe working in collection mode, we calculate the transmitted light in the substrate for two incoming circular polarizations of opposite handedness. An overview is shown in Fig. \[fig:simu\](a). The antenna response at $\lambda$=1.57 $\mu$m is reported in Figs. \[fig:simu\] (b) and (c) for the left and right circular polarizations, respectively. ![(a) Description of the numerical study with the FDTD method. A focused beam is projected onto the helix. The two right and left circular polarizations are considered in two distinct simulations. (b) and (c) Cross section of the probe-based HTN configuration in the (x0z)-plane with incoming right and left circular polarizations, respectively ($\lambda$=1.57 $\mu$m). Scale bars: 200 nm. (d) Spectrum of the differential transmission $\Delta$T of the probe used in collection mode (simulations in pulsed regime followed by Fourier transform).[]{data-label="fig:simu"}](simulations2.eps){width="0.65\linewidth"} The HTN shows an extreme polarization-dependent response as the two right and left circular polarization states induce transmission (Fig. \[fig:simu\] (b)) and no transmission (Fig. \[fig:simu\] (c)) of light into the substrate through the nanoaperture. To quantify such an assymmetric response to the different polarization handedness, we used the differential transmission $\Delta T=(T_{LCP}-T_{RCP})/(T_{LCP}+T_{RCP})$ [@kosters:acsphot17]. $T_{LCP}$ and $T_{RCP}$ are the transmission spectra of the HTN used in light collection mode for incoming right and left circular polarizations, respectively. Figure \[fig:simu\] (d) reports on a calculated differential transmission peaking at 1 when $\lambda$=1.57 $\mu$m and remaining larger than 0.99 for wavelengths ranging from 1.48 to 1.7 $\mu$m. Simulations thus predict an extremely assymmetric response of the nanoprobe regarding the right and left circular polarization states, over a broad spectral bandwidth. To experimentally verify the polarization-dependent properties of the HTN, we characterized the transmission of the nanoantenna both in the scattering and collection modes represented in Fig. \[fig:exp\] (a). To this end, linearly polarized light from a tunable laser (Yenista) passes through a rotating quarter-wave plate (QWP; AHWP05M-1600, Thorlabs) before being focused with a (25X, 0.4) microscope objective onto the HTN. The quarter-wave plate is mounted onto a motorized stage (PRM1Z8, Thorlabs) to be accurately rotated with respect to the incident linear polarization direction. The light transmitted through the nanoantenna is detected by imaging the plasmonic structure with an (50X, 0.65) infrared objective from Olympus coupled to an infrared camera (GoldEye model G-033, Allied Vision Technologies GmbH). In the scattering and collection modes of the HTN, the rectangle nano-aperture and the helix are selectively illuminated, respectively. In Fig. \[fig:exp\] (b) the spectra are obtained by studying the nanoantenna transmission as a function of the wavelength of the incoming light. At each wavelength, the transmitted intensity is measured for two orthogonal orientations of the quarter-wave plate leading to circular right and left polarizations (fast axis of the waveplate at $\pm$45$^{\circ}$ relative to the incident polarization direction). Then, the differential transmission $\Delta T$ is deduced from these measurements. ![(a) Schematics of the two experimentally investigated operations of the HTN, involving light wave propagation in two opposite directions along the helix axis. The collection mode of interest involves helix illumination (red arrow) whereas the scattering mode implies aperture illumination from the backside (blue arrow). (b) Experimental spectra of the differential transmission $\Delta T$ in both the scattering and collection modes of the HTN. Figure inset: transmission through the HTN used in collection mode at $\lambda$=1.55 $\mu$m as a function of the incoming polarization (i.e., the angle $\theta$ between the fast axis of the rotating quarter-wave plate and the direction of the incoming linear polarization).[]{data-label="fig:exp"}](experiments.eps){width="0.7\linewidth"} We see that $\Delta T$ is almost equal to zero in the scattering mode of the HTN. This can be easily explained by the dipolar nature of the rectangle nano-aperture which imposes a linearly polarized end-firing of the plasmonic helix regardless of the incoming polarization. The nano-aperture acts as a nanoscale linear polarization filter. In collection mode, $\Delta$T remains larger than 0.96 over the whole spectral bandwidth of the laser. This result agrees well with the numerical predictions in Fig. \[fig:simu\]. It confirms that the HTN develops a strongly assymmetric response with regards to polarization handedness when used as a local probe of light. We also analyze the light collected by the HTN while rotating the quarter-wave plate by one turn. The typical two-lobe pattern of the resulting diagram evidences that in its light collection mode, the HTN blocks the optical waves whose handedness is opposite to the helix, regardless the incoming polarization. ![Gold-coated carbon helix at the apex of a tip used in scanning near-field microscopy. (a) Large view of the helix carbon skeleton sculpted by FIBID at the very tip. (b) Finalized on-tip gold-coated carbon helix. Scale bars: 1 $\mu$m.[]{data-label="fig:tip"}](fab_tip2.eps){width="0.7\linewidth"} Local probing of light chirality requires moveable nanoprobes. We finally investigate the possibility to integrate an individual gold-coated carbon helix at the apex of the sharp tips used in scanning near-field microscopies. To this end, we considered a typical model of a sharp dielectric near-field tip. We metal-coated the tip with a thin layer of aluminum to ensure charge removal during FIBID process. In the case of tips with nanoscale apex such as those used in atomic force microscopy, a flat of 200-300 nm width can be engineered by FIB at the very tip prior to FIBID to ensure a reliable fabrication process of the nanostructures. We performed on-tip fabrication of the above introduced carbon-gold core-shell helix with the same parameters of the FIBID and metal coating as in the previous on-chip fabrication [@wang:lsa19]. Figure \[fig:tip\] shows SEM images of a resulting tip-integrated structure before (Figs. \[fig:tip\](a)) and after (Figs. \[fig:tip\](b)) metal coating. Changing the helix holder does not affect much the fabrication process. FIBID keeps its high level of accuracy in the definition of the helix geometry and metal coating still leads to smooth gold layers compatible with plasmonics. These images demonstrate the on-tip integration of an individual plasmonic helix aimed at probing light chirality.\ To conclude, we investigate the core-shell nanohelix as an individual subwavelength element for locally probing light chirality. By coupling an helix to a nano-aperture engraved in an opaque metallic layer, we obtained background-free light collection on the subwavelength scale. Probing light chirality can thus be realized without any signal post-processing involving bulky optical components. Such a local probe can be developed directly onto a photodetector or at the apex of a fiber tip used in scanning near-field optical microscopy. We study here the gold-coated carbon helix as a non-resonant “travelling-wave” nano-antenna [@balanis:book; @kraus:book; @wang:lsa19]. In a reciprocal approach of Schaferling’s et al. [@schaferling:acsphot14], tuning the helix to the resonant mode [@balanis:book; @kraus:book; @wozniak:ox18; @wozniak:optica19] by downscaling the structure would lead to chiral dipoles capable to locally characterize superchiral optical fields and effects. Coupling these resonant helical structures to coaxial nanoapertures may enable direct background-free local probing of light super-chirality. Coaxial nanoapertures sustain the radially-polarized TEM plasmon mode [@banzer:ox10; @baida:apb07] which can be coupled to the helix chiral dipole. We finally show the possibility to fabricate a single plasmonic helix at the apex of tips used in near-field microscopies. Beyond direct measurement of light chirality, the integration of a plasmonic helix at the end of an AFM tip should enable the local measurement of enantioselective optical forces [@zhao:natnano17]. Plasmonic helices may thus pave the way for a new generation of nanoprobes for locally measuring light chirality and chiroptical effects as well as opto-mechanic, opto-acoustic and magneto-optic phenomena. Funding {#funding .unnumbered} ======= Region Bourgogne Franche-Comte; EIPHI Graduate School (ANR-17-EURE-0002); Agence Nationale de la Recherche: ANR-18-CE42-0016 [10]{} Mario Hentschel, Martin Sch[ä]{}ferling, Xiaoyang Duan, Harald Giessen, and Na Liu. Chiral plasmonics. , 3(5):e1602735, 2017. Martin Sch[ä]{}ferling, Daniel Dregely, Mario Hentschel, and Harald Giessen. Tailoring enhanced optical chirality: design principles for chiral plasmonic nanostructures. , 2(3):031010, 2012. Lisa V Poulikakos, Prachi Thureja, Alexia Stollmann, Eva De Leo, and David J Norris. Chiral light design and detection inspired by optical antenna theory. , 18(8):4633–4640, 2018. Nina Meinzer, Euan Hendry, and William L Barnes. Probing the chiral nature of electromagnetic fields surrounding plasmonic nanostructures. , 88(4):041407, 2013. T. Taminiau, R. Moerland, F. Segerink, L. Kuipers, and N. Van Hulst. $\lambda$/4 resonance of an optical monopole antenna probes by single molecule fluorescence. , 7:28, 2007. M. Burresi, D. van Oosten, T. Kampfrath, H. Schoenmaker, R. Heideman, A. Leinse, and L. Kuipers. Probing the magnetic field of light at optical frequencies. , 326:550–553, 2009. Mathieu Mivelle, Thomas S van Zanten, and Maria F Garcia-Parajo. Hybrid photonic antennas for subnanometer multicolor localization and nanoimaging of single molecules. , 14(8):4895–4900, 2014. N. Murphy-DuBay, L. Wang, E. C. Kinzel, S. M. V. Uppuluri, and X. Xu. Nanopatterning using nsom probes integrated with high transmission nanoscale bowtie aperture. , 16(4):2584–2589, 2008. T. Grosjean, I. A. Ibrahim, M. A. Suarez, G. W. Burr, M. Mivelle, and D. Charraut. Full vectorial imaging of electromagneticlight at subwavelength scale. , 18(6):5809–5824, 2010. Thanh-Phong Vo, M Mivelle, S Callard, A Rahmani, F Baida, D Charraut, A Belarouci, D Nedeljkovic, C Seassal, GW Burr, and T Grosjean. Near-field probing of slow bloch modes on photonic crystals with a nanoantenna. , 20(4):4124–4135, 2012. Nathan C Lindquist, Timothy W Johnson, Prashant Nagpal, David J Norris, and Sang-Hyun Oh. Plasmonic nanofocusing with a metallic pyramid and an integrated c-shaped aperture. , 3:1857, 2013. Zhihua Xie, Yannick Lefier, Miguel Angel Suarez, Mathieu Mivelle, Roland Salut, Jean-Marc Merolla, and Thierry Grosjean. Doubly resonant photonic antenna for single infrared quantum dot imaging at telecommunication wavelengths. , 17(4):2152–2158, 2017. Johannes Kaschke and Martin Wegener. Optical and infrared helical metamaterials. , 5(4):510–523, 2016. JG Gibbs, AG Mark, S Eslami, and P Fischer. Plasmonic nanohelix metamaterials with tailorable giant circular dichroism. , 103(21):213101, 2013. Marco Esposito, Vittorianna Tasco, Massimo Cuscuna?, Francesco Todisco, Alessio Benedetti, Iolena Tarantini, Milena De Giorgi, Daniele Sanvitto, and Adriana Passaseo. Nanoscale 3d chiral plasmonic helices with circular dichroism at visible frequencies. , 2(1):105–114, 2014. Marco Esposito, Vittorianna Tasco, Francesco Todisco, Massimo Cuscun[à]{}, Alessio Benedetti, Daniele Sanvitto, and Adriana Passaseo. Triple-helical nanowires by tomographic rotatory growth for chiral photonics. , 6:6484, 2015. Justyna K Gansel, Michael Thiel, Michael S Rill, Manuel Decker, Klaus Bade, Volker Saile, Georg von Freymann, Stefan Linden, and Martin Wegener. Gold helix photonic metamaterial as broadband circular polarizer. , 325(5947):1513–1515, 2009. Dolfine Kosters, Anouk De Hoogh, Hans Zeijlemaker, Hakk[i]{} Acar, Nir Rotenberg, and L Kuipers. Core–shell plasmonic nanohelices. , 4(7):1858–1863, 2017. Martin Sch[ä]{}ferling, Xinghui Yin, Nader Engheta, and Harald Giessen. Helical plasmonic nanostructures as prototypical chiral near-field sources. , 1(6):530–537, 2014. Pawel Wozniak, Israel De Leon, Katja Hoeflich, Caspar Haverkamp, Silke Christiansen, Gerd Leuchs, and Peter Banzer. Chiroptical response of a single plasmonic nanohelix. , 26(15):19275–19293, 2018. Pawe[ł]{} Wo[ź]{}niak, Israel De Le[ó]{}n, Katja H[ö]{}flich, Gerd Leuchs, and Peter Banzer. Interaction of light carrying orbital angular momentum with a chiral dipolar scatterer. , 6(8):961, 2019. Mengjia Wang, Roland Salut, Huihui Lu, Miguel-Angel Suarez, Nicolas Martin, and Thierry Grosjean. Controlling light polarization by swirling surface plasmons. , 2018. Yang Zhao, Amr AE Saleh, Marie Anne Van De Haar, Brian Baum, Justin A Briggs, Alice Lay, Olivia A Reyes-Becerra, and Jennifer A Dionne. Nanoscopic control and quantification of enantioselective optical forces. , 12(11):1055, 2017. Marco Esposito, Vittorianna Tasco, Francesco Todisco, Massimo Cuscunà, Alessio Benedetti, Mario Scuderi, Giuseppe Nicotra, and Adriana Passaseo. Programmable extreme chirality in the visible by helix-shaped metamaterial platform. , 16(9):5823–5828, 2016. C.A. Balanis. . John Wiley & Sons, New-York, 1997. John D Kraus, Ronald J Marhefka, and Ahmad S Khan. . Tata McGraw-Hill Education, 2006. Peter Banzer, Jochen Kindler, Susanne Quabis, Ulf Peschel, and Gerd Leuchs. Extraordinary transmission through a single coaxial aperture in a thin metal film. , 18(10):10896–10904, 2010. Fadi I Baida. Enhanced transmission through subwavelength metallic coaxial apertures by excitation of the TEM mode. , 89(2-3):145–149, 2007.

ArXiv

--- abstract: | Finite groups are said to be isospectral if they have the same sets of element orders. A finite nonabelian simple group $L$ is said to be almost recognizable by spectrum if every finite group isospectral to $L$ is an almost simple group with socle isomorphic to $L$. It is known that all finite simple sporadic, alternating and exceptional groups of Lie type, except $J_2$, $A_6$, $A_{10}$ and $^3D_4(2)$, are almost recognizable by spectrum. The present paper is the final step in the proof of the following conjecture due to V.D. Mazurov: there exists a positive integer $d_0$ such that every finite simple classical group of dimension larger than $d_0$ is almost recognizable by spectrum. Namely, we prove that a nonabelian composition factor of a finite group isospectral to a finite simple symplectic or orthogonal group $L$ of dimension at least 10, is either isomorphic to $L$ or not a group of Lie type in the same characteristic as $L$, and combining this result with earlier work, we deduce that Mazurov’s conjecture holds with $d_0=60$. **Keywords.** Simple group, symplectic group, orthogonal group, element orders, spectrum of a group. **2010 MSC.** 20D06, 20D60. address: - | Sobolev Institute of Mathematics and Novosibirsk State University,\ Ac. Koptyuga 4, Novosibirsk, 630090, Russia. - | Sobolev Institute of Mathematics and Novosibirsk State University,\ Ac. Koptyuga 4, Novosibirsk, 630090, Russia. author: - 'Mariya A. Grechkoseeva' - 'Andrey V. Vasil$''$ev' title: | On the structure of finite\ groups isospectral to finite simple groups --- [^1] Introduction ============ The [*spectrum*]{} $\omega(G)$ of a finite group $G$ is the set of element orders of $G$. Groups are [*isospectral*]{} if they have the same spectra. Given a finite group $G$ with nontrivial soluble radical, one can construct infinitely many different finite groups isospectral to $G$ [@94Shi; @98Maz.t]. In contrast, there is a conjecture due to V.D. Mazurov that in general the set of groups isospectral to a finite nonabelian simple group $L$ is finite and consists of groups closely related to $L$. More precisely, Mazurov conjectured[^2] that if $L$ is an alternating group of sufficiently large degree or a simple group of Lie type of sufficiently large Lie rank, and $G$ is a group isospectral to $L$, then $G$ is an almost simple group with socle isomorphic to $L$. For brevity we refer to a nonabelian simple group $L$ such that every finite group isospectral to $L$ is an almost simple group with socle isomorphic to $L$, as [*almost recognizable by spectrum*]{}. Gorshkov [@13Gor.t] proved that all alternating groups of degree at least 11 are almost recognizable by spectrum, and recently Vasil’ev and Staroletov [@14VasSt.t] completed the investigation of almost recognizability of exceptional groups. Thus Mazurov’s conjecture has been reduced to the following: there exists a positive integer $d_0$ such that every simple classical group of dimension larger than $d_0$ is almost recognizable. The main purpose of the present paper is to prove this conjecture with $d_0=60$. As a result, we establish the following theorem (our notation for nonabelian simple groups follows [@85Atlas]). \[t:main\] Let $L$ be one of the following nonabelian simple groups: 1. \[t:sporadic\] a sporadic group other than $J_2$; 2. \[t:alternating\] an alternating group $A_n$, where $n\neq6,10$; 3. \[t:exceptional\] an exceptional group of Lie type other than ${}^3D_4(2)$; 4. \[t:linear\] $L_n(q)$, where $n\geqslant45$ or $q$ is even; 5. \[t:unitary\] $U_n(q)$, where $n\geqslant45$, or $q$ is even and $(n,q)\neq (4,2),(5,2)$; 6. \[t:BnCn\] $S_{2n}(q),O_{2n+1}(q)$, where either $q$ is odd and $n\geqslant28$, or $q$ is even and $n\geqslant20$; 7. \[t:Dn\] $O^+_{2n}(q)$, where either $q$ is odd and $n\geqslant31$, or $q$ is even and $n\geqslant20$; 8. \[t:2Dn\] $O^-_{2n}(q)$, where either $q$ is odd and $n\geqslant30$, or $q$ is even and $n\geqslant20$. Then every finite group isospectral to $L$ is isomorphic to some group $G$ with $L\leqslant G\leqslant {\operatorname{Aut}}L$. In particular, there are only finitely many pairwise nonisomorphic finite groups isospectral to $L$. Theorem \[t:main\] is undoubtedly a sum of efforts by numerous mathematicians, and a comprehensive list of references covering its proof is too long to be given here. But for every series of simple groups mentioned in Theorem \[t:main\], we cite the work in which the proof for this series was completed, and as a rule this work includes a survey of previous investigations. Thus, see already mentioned [@13Gor.t] and [@14VasSt.t] for alternating and exceptional groups. See [@98MazShi] for sporadic groups, [@15Vas Theorem 1] for linear and unitary groups in odd characteristic, and [@08VasGr.t] and [@11Gr1.t Corollary 2] for linear and unitary groups in characteristic 2. The present paper is concerned with symplectic and orthogonal groups. By [@15Vas Theorem 2 and Proposition 6], if $L$ is a finite simple symplectic or orthogonal group as in (\[t:BnCn\])–(\[t:2Dn\]) of Theorem \[t:main\] and $G$ is a finite group with $\omega(G)=\omega(L)$, then $G$ has only one nonabelian composition factor and this factor $S$ is a symplectic or orthogonal group having the same underlying characteristic as $L$. If $S\simeq L$, then the soluble radical of $G$ is trivial [@15Gr Theorem 1.1], and hence $G$ is an almost simple group with socle isomorphic to $L$, as required. If $S\not\simeq L$, then by [@09VasGrMaz.t Theorem 3] there are at most two possibilities for $S$ (see Lemma \[l:previous\] below). We eliminate these possibilities by the following theorem, and thus Theorem \[t:main\] follows. \[t:quasi\] Let $q$ be a power of a prime $p$, $L$ one of the groups $S_{2n}(q)$, where $n\geqslant 2$ and $(n,q)\not\in\{(2,2),(2,3)\}$, $O_{2n+1}(q)$, where $n\geqslant3$, or $O^\pm_{2n}(q)$, where $n\geqslant 4$, and let $G$ be a finite group with $\omega(G)=\omega(L)$. Suppose that some nonabelian composition factor $S$ of $G$ is a group of Lie type over a field of characteristic $p$. Then either $S\simeq L$ or one of the following holds: 1. \[i:s4\] $L=S_4(q)$, where $q\neq3^{2k+1}$, and $S=L_2(q^2)$; 2. \[i:s8\] $L\in\{O_9(q), S_8(q)\}$ and $S=O^-_8(q)$; 3. \[i:s6\] $\{L,G\}=\{O^+_8(2), S_6(2)\}$; 4. \[i:o7\] $\{L,G\}=\{O^+_8(3), O_7(3)\}$. We conclude the introduction with a discussion of nonabelian simple groups that are genuinely not almost recognizable by spectrum. The groups $J_2$, $A_6$, $A_{10}$, $^3D_4(2)$, $U_4(2)\simeq S_4(3)$, and $U_5(2)$ are not almost recognizable by spectrum; moreover, each of them is isospectral to a group with nontrivial soluble radical (see [@98MazShi; @91BrShi; @98Maz.t; @13Maz.t]). All the groups $L_2(q)$, $L_3(q)$ and $U_3(q)$, except for $L_2(9)$, $L_3(3)$, $U_3(5)$ and $U_3(p)$ where $p$ is a Mersenne prime and $p^2-p+1$ is a prime too, are almost recognizable (see [@94BrShi], [@00MazXuCao.t; @04Zav] and [@00MazXuCao.t; @06Zav.t] respectively). The question is still open for $L_4(q)$ and $U_4(q)$ with $q$ odd: the group $L_4(13^{24})$ was proved to be not almost recognizable, and it was conjectured that there are infinitely many $q$ such that $L_4(q)$ is not almost recognizable [@08Zav2]. Thus, the exceptions in (\[t:sporadic\])–(\[t:exceptional\]) of Theorem \[t:main\] are necessary, while the condition $n\geqslant 45$ in (\[t:linear\]) and (\[t:unitary\]) can probably be replaced by $n\geqslant 5$. The simple group $S_4(q)$ is almost recognizable if and only if $q=3^m$ with $m>1$ odd; moreover, the case (\[i:s4\]) of Theorem \[t:quasi\] is possible [@00MazXuCao.t; @02Maz.t]. It is known that $\omega(O^+_8(2))=\omega(S_6(2))$ and $\omega(O^+_8(3))=\omega(O_7(3))$ and there are no other finite groups with such spectra [@97Maz.t; @97ShiTan]. The group $S_8(2)$ is not almost recognizable [@06MazMog]. In the last section we generalize this result to all groups $S_8(q)$ with $q$ even, showing the case (\[i:s8\]) of Theorem \[t:quasi\] is possible when $q$ is even. We suspect that it is also possible when $q$ is odd (and this is proved for $L=O_9(q)$ in [@14GrSt]). Thus it seems likely that all the exceptions appearing in Theorem \[t:quasi\] can be realised. On the other hand, we believe that all other symplectic and orthogonal groups are almost recognizable by spectrum. To summarize, we propose the following \[conj:main\] Let $L$ be one of the following groups: 1. $L_n(q)$, where $n\geqslant5$; 2. $U_n(q)$, where $n\geqslant5$ and $(n,q)\neq(5,2)$; 3. $S_{2n}(q)$, where $n\geqslant3$, $n\neq4$ and $(n,q)\neq(3,2)$; 4. $O_{2n+1}(q)$, where $q$ is odd, $n\geqslant3$, $n\neq4$ and $(n,q)\neq(3,3)$; 5. $O_{2n}^\varepsilon(q)$, where $n\geqslant4$ and $(n,q,\varepsilon)\neq (4,2,+),(4,3,+)$. Then every finite group isospectral to $L$ is isomorphic to some group $G$ with $L\leqslant G\leqslant {\operatorname{Aut}}L$. In fact, to prove Conjecture \[conj:main\] it suffices to show that the only nonabelian composition factor of a group isospectral to a simple classical group $L$ under consideration is not a group of Lie type whose underlying characteristic differs from that of $L$. The existing generic proof of this assertion recently obtained in [@15Vas] requires that the dimension of $L$ is as large as stated in Theorem \[t:main\]. It is worth also mentioning that the assertion was proved in some special cases, in particular for many classical groups with disconnected prime graph (see [@09VasGorGr.t; @09AleKon.t; @09Kon; @09HeShi; @10SheShiZin.t; @12HeShi; @12GrLyt.t; @12ForIraAha; @13ForIraAha] for recent research in this area). Preliminaries ============= By $(a_1,a_2,\dots,a_k)$ and $[a_1,a_2,\dots,a_k]$ we denote respectively the greatest common divisor and least common multiple of positive integers $a_1,a_2,\dots,a_k$. If $a$ is a positive integer and $r$ is a prime then $\pi(a)$ denotes the set of prime divisors of $a$ and $(a)_r$ denotes the highest power of $r$ that divides $a$. \[l:zsig\] Let $q\geqslant 2$ and $n\geqslant 3$ be integers with $(q,n)\neq (2,6)$. There exists a prime $r$ such that $r$ divides $q^n-1$ but does not divide $q^i-1$ for $i<n$. With notation of Lemma \[l:zsig\], we call a prime $r$ a [*primitive prime divisor*]{} of $q^n-1$ and denote it by $r_n(q)$. By $R_n(q)$ we denote the set of all primitive prime divisors of $q^n-1$. Observe that $R_n(q)\subseteq R_n(q^k)$ if $k$ is coprime to $n$, and $R_{nk}(q)\subseteq R_n(q^k)$ for all $n$ and $k$. Given a group $G$, we set $\pi(G)=\pi(|G|)$ and define the [*prime graph*]{} $GK(G)$ as follows: its vertex set is $\pi(G)$ and two different primes $r$ and $s$ are adjacent if and only if $G$ has an element of order $rs$. We use standard graph-theoretic terminology: a [*coclique*]{} of a graph is a set of pairwise nonadjacent vertices; a [*neighbourhood*]{} of a vertex $v$ of a graph is the subgraph consisting of all vertices adjacent to $v$ and all edges connecting two such vertices. \[l:structure\] Let $L$ be a finite nonabelian simple group of Lie type other than $L_3(3)$, $U_3(3)$, $S_4(3)\simeq U_4(2)$ and let $G$ be a finite group with $\omega(G)=\omega(L)$. Then the following hold. 1. There is a nonabelian simple group $S$ such that $S\leqslant \overline G=G/K\leqslant {\operatorname{Aut}}S,$ where $K$ is the soluble radical of $G$. 2. If $\rho$ is a coclique of size at least $3$ in $GK(G)$, then at most one prime of $\rho$ divides $|K|\cdot|\overline{G}/S|$. 3. If $r\in\pi(G)$ is not adjacent to $2$ in $GK(G)$, then $r$ is coprime to $|K|\cdot|\overline{G}/S|$. If there is a coclique of size 3 in $GK(L)$, then the assertion is the main theorem of [@05Vas.t] supplemented with [@09VasGor.t] and [@05VasVd.t Theorem 7.1]. If there are no cocliques of size 3 in $GK(L)$, then we have that $GK(L)$ is disconnected by [@11VasVd.t; @81Wil; @89Kon.t]. Then the Gruenberg–Kegel theorem [@81Wil Theorem A] implies that either (i) and (iii) holds true for $G$, or $G$ is a Frobenius or 2-Frobenius group ($G$ is called 2-Frobenius if $G=ABC$, where $A$, $AB$ are normal in $G$, $B$ is normal in $BC$, and $AB$ and $BC$ are Frobenius groups). Simple groups of Lie type that can be isospectral to a Frobenius or 2-Frobenius group are described in [@03Ale.t], and these groups are precisely $L_3(3)$, $U_3(3)$ and $S_4(3)\simeq U_4(2)$. We say that a finite group $H$ is a (proper) [*cover*]{} of a finite group $G$ if there is a (nontrivial) normal subgroup $K$ of $H$ such that $H/K\simeq G$. \[l:reduction\] Let $A$ and $B$ be finite groups. The following are equivalent. 1. $\omega(H)\not\subseteq\omega(B)$ for any proper cover $H$ of $A$; 2. $\omega(H)\not\subseteq\omega(B)$ for any split extension $H=K:A$, where $K$ is a nontrivial elementary abelian group. \[l:action\] Let $G$ be a finite group, $K$ a normal subgroup of $G$ and $G/K~$ a Frobenius group with kernel $N$ and cyclic complement $C.$ If $(|N|, |K|)=1$ and $N$ is not contained in $KC_G(K)/K$, then $r|C|\in\omega(G)$ for some $r\in\pi(K)$. \[l:hh\] Let $S$ be a finite simple group of Lie type over a field of characteristic $p$ and let $S$ act faithfully on a vector space $V$ over a field of characteristic $r$, where $r\neq p$. Let $H=V\rtimes S$ be a natural semidirect product of $V$ by $S$. Suppose that $s$ is a power of $r$ and some proper parabolic subgroup $P$ of $S$ contains an element of order $s$. If the unipotent radical of $P$ is abelian, or both $p$ and $r$ are odd, or $p=2$ and $r$ is not a Fermat prime, or $r=2$ and $p$ is not a Mersenne prime, then $rs\in\omega(H)$. We conclude with several lemmas on spectra of symplectic and orthogonal groups. If $p\in\pi(G)$, then $\omega_{p'}(G)$ denotes the subset of $\omega(G)$ consisting of numbers coprime to $p$. In Lemmas \[l:spectrum\_s\_odd\] and \[l:spectrum\_s\_2\], $\pm$ in $[a_1\pm 1,\dots, a_s\pm 1]$ means that we can choose $+$ or $-$ for every entry independently. In Lemma \[l:semi\], for brevity, we write $\varepsilon$ instead of $\varepsilon1$ for $\varepsilon\in\{+,-\}$. \[l:spectrum\_s\_odd\] Let $L$ be one of the simple groups $S_{2n}(q)$ or $O_{2n+1}(q)$, where $n\geqslant 2$ and $q$ is a power of an odd prime $p$. Let $d=1$ if $L=S_{2n}(q)$ or $n=2$, and let $d=2$ if $L=O_{2n+1}(q)$ with $n\geqslant 3$. Then $\omega(L)$ consists of all divisors of the following numbers: 1. $(q^n\pm1)/2$; 2. $[q^{n_1}\pm 1,\dots,q^{n_s}\pm 1]$, where $s\geqslant 2$, $n_i>0$ for all $1\leqslant i\leqslant s$ and $n_1+\dots+n_s=n$; 3. $p^k(q^{n_1}\pm 1)/d$, where $k,n_1>0$ and $p^{k-1}+1+2n_1=2n$; 4. $p^k[q^{n_1}\pm 1,\dots, q^{n_s}\pm1]$, where $k>0$, $s\geqslant 2$, $n_i>0$ for all $1\leqslant i\leqslant s$ and $p^{k-1}+1+2(n_1+\dots+n_s)=2n$; 5. $p^k$ if $2n=p^{k-1}+1$ for some $k > 0$. \[l:spectrum\_s\_2\] Let $L=S_{2n}(q)$, where $n\geqslant 2$ and $q$ is even. Then $\omega(L)$ consists of all divisors of the following numbers: 1. $[q^{n_1}\pm 1,\dots,q^{n_s}\pm 1]$, where $s\geqslant 1$, $n_i>0$ for all $1\leqslant i\leqslant s$ and $n_1+\dots+n_s=n$; 2. $2[q^{n_1}\pm 1,\dots,q^{n_s}\pm 1]$, where $s\geqslant 1$, $n_i>0$ for all $1\leqslant i\leqslant s$ and $n_1+\dots+n_s=n-1$; 3. $2^k[q^{n_1}\pm 1,\dots, q^{n_s}\pm1]$, where $k\geqslant 2$, $s\geqslant 1$, $n_i>0$ for all $1\leqslant i\leqslant s$ and $2^{k-2}+1+n_1+\dots+n_s=n$; 4. $2^k$ if $n=2^{k-2}+1$ for some $k\geqslant 2$. \[l:adj\_p\] Let $L=O_{2n}^\pm(q)$, where $n\geqslant 4$ and $q$ is a power of a prime $p$. If $r\in \pi(L)\cap R_k(q)$, where $k$ is odd and $k>n-2$, or $k$ is even and $k/2>n-2$, then $rp\not\in\omega(L)$. \[l:semi\] Let $L=O_{2n}^\varepsilon(q)$, where $n\geqslant 4$, $\varepsilon\in\{+,-\}$ and $q$ is a power of a prime $p$. Then $\omega_{p'}(L)$ consists of all divisors of the following numbers: 1. $(q^n-\varepsilon)/(4,q^n-\varepsilon)$; 2. $[q^{n_1}-\delta, q^{n_2}-\varepsilon\delta]/d$, where $\delta\in\{+,-\}$, $n_1,n_2>0$, $n_1+n_2=n$; $d=2$ if $(4,q^n-\varepsilon)=4$, $(q^{n_1}-\delta)_2=(q^{n_2}-\varepsilon\delta)_2$, and $d=1$ otherwise; 3. $[q^{n_1}-\delta_1, q^{n_2}-\delta_2, \dots, q^{n_s}-\delta_s]$, where $s\geqslant 3$, $\delta_i\in\{+,-\}$, $n_i>0$ for all $1\leqslant i\leqslant s$, $n_1+\dots+n_s=n$ and $\delta_1\delta_2\dots\delta_s=\varepsilon$. The assertion follows, for example, from [@07ButGr.t Theorem 6]. \[l:spectrum\_o8p\] Let $L=O_8^+(q)$, where $q$ is a power of a prime $p$. Then $\omega(L)$ consists of all divisors of the following numbers: 1. $(q^4-1)/(2,q-1)^2$, $(q^3\pm1)/(2,q-1)$, $q^2-1$, $p(q^2\pm1)/(2,q-1)$; 2. $p^2(q\pm 1)/(2,q-1)$ if $p=2,3$; 3. $25$ if $p=5$; 4. $8$ if $p=2$. The assertion follows from [@10But.t Corollaries 4 and 9]. \[l:spectrum\_o8m\] Let $L=O_8^-(q)$, where $q$ is even. Then $\omega(L)$ consists of all divisors of the following numbers: $q^4\pm1$, $(q^2\pm q+1)(q^2-1)$, $2(q^2+1)(q\pm 1)$, $4(q^2-1)$, and $8$. The assertion follows from [@10But.t Corollary 4]. \[l:diff\] Let $q$ be a power of a prime $p$. 1. \[i:bncn\_diff\] Let $q$ be odd and $n\geqslant 3$. Let $r=r_{2n-2}(q)$ if $q^{n-1}\equiv 1\pmod 4$ and $r=r_{n-1}(q)$ if $q^{n-1}\equiv -1\pmod 4$. Then $2pr\in\omega(S_{2n}(q))\setminus \omega(O_{2n+1}(q))$. 2. \[i:dnbn\] If $n\geqslant 4$ and $(n,q)\neq (4,2)$, then $pr_{2n-2}(q)\in\omega(O_{2n+1}(q))\setminus\omega({}O^-_{2n}(q))$. If $n\geqslant 4$ is even, then $pr_{n-1}(q)\in\omega(O_{2n+1}(q))\setminus\omega({}O^-_{2n}(q))$. 3. \[i:s6o8\] If $q>3$, then $(q^4-1)/(2,q-1)^2\in\omega(O^+_8(q))\setminus\omega(S_6(q))$. 4. \[i:o8s6\] If $q$ is odd, then $p(q^2+1)\in\omega(S_{6}(q))\setminus\omega(O^+_{8}(q))$. 5. \[i:bncn\]$\omega(O_{2n+1}(q))\subseteq\omega(S_{2n}(q))$ for all $n\geqslant 2$. 6. \[i:orth\] $\omega(O_{2n-1}(q))\subseteq \omega(O_{2n}^\pm(q))\subseteq \omega(O_{2n+1}(q))$ for all $n\geqslant 3$. (\[i:bncn\_diff\]) Let $q^{n-1}\equiv \varepsilon \pmod 4$. Then $2pr$ divides $p(q^{n-1}+\varepsilon)$, and in particular it belongs to $\omega(S_{2n}(q))$ by Lemma \[l:spectrum\_s\_odd\]. Suppose that $2pr\in \omega(O_{2n+1}(q))$. It follows by Lemma \[l:spectrum\_s\_odd\] that $2pr$ divides either $p^k(q^{n_1}\pm 1)/2$ for some $k\geqslant 1$ and $n_1>0$ with $p^{k-1}+1+2n_1=2n$, or $p^k[q^{n_1}\pm 1,\dots,q^{n_s}\pm 1]$ for some $k\geqslant 1$, $s\geqslant 2$, $n_1,\dots,n_s>0$ with $p^{k-1}+1+2n_1+\dots+2n_s=2n$. In fact, by the definition of primitive divisor, it cannot divide a number of the latter form since all $n_1,\dots,n_s$ are less than $n-1$. And if $2pr$ divides $p^k(q^{n_1}+\tau)/2$ with $p^{k-1}+1+2n_1=2n$ and $\tau=\pm1$, then by the definition of primitive divisor we have that $k=1$, $n_1=n-1$ and $\tau=\varepsilon$. But then $(q^{n_1}+\tau)/2$ is odd, a contradiction. (\[i:dnbn\]) See [@05VasVd.t Proposition 3.1]. (\[i:s6o8\]) Let $a=(q^4-1)/(2,q-1)^2$. Lemma \[l:spectrum\_o8p\] implies that $a\in\omega(O^+_8(q))$. By Lemmas \[l:spectrum\_s\_odd\] and \[l:spectrum\_s\_2\], the orders of semisimple elements of $S_6(q)$ are precisely divisors of $(q^3\pm1)/(2,q-1)$, $(q^2+1)(q\pm1)/(2,q-1)$ and $q^2-1$. It is clear that $a$ divides none of $(q^3\pm1)/(2,q-1)$ and $q^2-1$. Furthermore, $$a=(q^2+1)\cdot \frac{(q+1)}{(2,q-1)}\cdot\frac{(q-1)}{(2,q-1)}.$$ Since $q>2$, both $(q-1)/(2,q-1)$ and $(q+1)/(2,q-1)$ are greater than one, and so $a$ does not divide $(q^2+1)(q\pm1)/(2,q-1)$ either. (\[i:o8s6\])–(\[i:bncn\]) See Lemmas \[l:spectrum\_s\_odd\] and \[l:spectrum\_o8p\]. (\[i:orth\]) It is well known that $O_{2n-1}(q)<O_{2n}^\pm(q)<O_{2n+1}(q)$, and the assertion follows. \[l:s6\] Let $q$ be even, $S=S_6(q)$ and $L=O_8^+(q)$. Suppose that $V$ is a nontrivial $S$-module over a field of characteristic $2$ and $H$ is a natural semidirect product of $V$ and $S$. Then $\omega(H)\not\subseteq \omega(L)$. In the proof of Lemma 4.1 in [@15Gr], it was established that $\omega(H)$ contains at least one of the numbers $16$, $24$, and $2(q^2+q+1)$. By Lemma \[l:spectrum\_o8p\], none of these numbers belong to $\omega(L)$. Proof of Theorem \[t:quasi\] ============================ As we mentioned in the Introduction, the starting point for our proof of Theorem \[t:quasi\] is the following assertion. \[l:previous\] Let $q$ be a power of a prime $p$, $L$ one of the groups $S_{2n}(q)$, where $n\geqslant 2$ and $(n,q)\not\in\{(2,2),(2,3)\}$, $O_{2n+1}(q)$, where $n\geqslant3$, or $O^\pm_{2n}(q)$, where $n\geqslant 4$, and let $G$ be a finite group with $\omega(G)=\omega(L)$. Suppose that some nonabelian composition factor $S$ of $G$ is a group of Lie type over a field of characteristic $p$. If $S\not\simeq L$, then one of the following holds: 1. $L=S_4(q)$ and $S=L_2(q^2)$; 2. $\{L, S\}\subseteq \{S_6(q), O_7(q), O^+_8(q)\}$; 3. $\{L, S\}\subseteq \{S_{2n}(q), O_{2n+1}(q), {}O^-_{2n}(q)\}$ and $n\geqslant 4$; 4. $L=O^+_{2n}(q)$, $S\in\{S_{2n-2}(q), O_{2n-1}(q)\}$, and $n\geqslant 6$ is even. The assertion is a combination of [@09VasGrMaz.t Theorem 3] and [@12Sta.t Theorem 1]. Let $q=p^m$ and let $L$ be one of the groups $S_{2n}(q)$, where $n\geqslant 2$ and $(n,q)\not\in\{(2,2),(2,3)\}$, $O_{2n+1}(q)$, where $n\geqslant3$, or $O^\pm_{2n}(q)$, where $n\geqslant 4$. Let $G$ be a finite group with $\omega(G)=\omega(L)$. By Lemma \[l:structure\], we have $$S\leqslant G/K\leqslant {\operatorname{Aut}}S,$$ where $K$ is the soluble radical of $G$ and $S$ is a nonabelian simple group. Suppose that $S$ is a group of Lie type over a field of characteristic $p$. By Lemma \[l:previous\], either $S\simeq L$, or $L$ and $S$ are as in the conclusion of Lemma \[l:previous\]. Let $L=S_4(q)$, where $q>3$. If $q=3^{m}$ and $m$ is odd, then $G\simeq L$ by [@02Maz.t], and so $S\simeq L$. If $q\neq 3^{m}$ with $m$ odd, then $S=L_2(q^2)$ by Lemma \[l:previous\], and this is (\[i:s4\]) of the conclusion of Theorem \[t:quasi\]. To examine the other cases, we need some auxiliary results. \[l:out\] Let $L\neq S_4(q)$. If $k>2$ and $r\in R_{mk}(p)$, then $r\not\in\pi({\operatorname{Out}}S)$. Since $S$ is a symplectic or orthogonal group over a field of order $q=p^m$, it follows that $\pi({\operatorname{Out}}S)\subseteq \{2,3\}\cup\pi(m)$. By Fermat’s Little Theorem $mk$ divides $r-1$ and so $r>3$ and $r>m$. \[l:bncn\] If $L=S_{2n}(q)$, where $n\geqslant 3$ and $q$ is odd, then $S\neq O_{2n+1}(q)$. Assume the contrary. It follows by [@15Gr Proposition 1.3] that $K=1$ in this case, and hence $S\leqslant G\leqslant {\operatorname{Aut}}S$. Let $r=r_{2(n-1)m}(p)$ if $q^{n-1}\equiv 1 \pmod 4$ and $r=r_{(n-1)m}(p)$ otherwise. By Lemma \[l:diff\], there is an element $g$ of order $2pr$ in $G\setminus S$. By Lemma \[l:out\], we have $r\not\in\pi(G/S)$. Suppose that $p\in\pi(G/S)$. Then $G$ contains a field automorphism of $S$ of order $p$. By [@98GorLySol Proposition 4.9.1(a)], the centralizer of this automorphism in $S$ includes $O_{2n+1}(q_0)$, where $q=q_0^p$. Therefore $pr_{2n}(q_0)$ and $p[q_0+1, q_0^{n-1}\pm 1]$ lie in $\omega(G)$. If $p$ does not divide $n$, then $r_{2n}(q_0)\in R_{2n}(q)$ and so $pr_{2n}(q_0)\not\in\omega(L)$ by [@05VasVd.t Proposition 3.1]. Let $p$ divide $n$. Then $r_{2n-2}(q_0)\in R_{2n-2}(q)$. By the definition of primitive divisor and Lemma \[l:spectrum\_s\_odd\], it follows that $p(q_0+1)r_{2n-2}(q_0)\in\omega(L)$ if and only if $q_0+1$ divides $q^{n-1}+1$, which implies that $n$ is even. But then $q_0+1$ does not divide $q^{n-1}-1$, and by the same reasoning we conclude that $p(q_0+1)r_{n-1}(q_0)\not\in\omega(L)$. In any case, $\omega(G)\not\subseteq\omega(L)$, a contradiction. Thus $p,r\not\in\pi(G/S)$, and it follows that $g^{pr}\not\in S$ and $g^2\in S$. Therefore $G\setminus S$ contains an involution $t=g^{pr}$ such that $pr\in\omega(C_S(t))$. Suppose that $t\in \operatorname {Inndiag}S$. Then $\operatorname {Inndiag}S\leqslant G$ and since $\operatorname {Inndiag}S\simeq SO_{2n+1}(q)$, it follows that $q^n+1\in\omega(G)\setminus\omega(L)$, which is impossible. Therefore, by [@98GorLySol Proposition 4.9.1(d)], we have that $t$ is a field automorphism, and hence $m$ is even. In particular $q\equiv 1\pmod 4$ and $r=r_{2(n-1)m}(p)$. By [@98GorLySol Prop 4.9.1(a,b)], the centralizer $C_S(t)$ can be embedded into $SO_{2n+1}(q_0)$, where $q=q_0^2$, and so $C_S(t)$ has no elements of order $r$, a contradiction. \[l:action2\] Suppose $\pi(G)$ contains four different primes $p$, $s$, $r$, and $w$ with the following properties: 1. $S$ contains a Frobenius subgroup $F$ whose kernel is a $p$-group and whose complement has order $s$; 2. $r\in \pi(K)$; 3. $\{s,r,w\}$ is a coclique in $GK(G)$. Then $p\in\pi(K)$ and $pt\in \omega(G)$ for every $t\in\pi(S)\setminus\{p\}$. Since $\{s,r,w\}$ is a coclique in $GK(G)$ and $r\in \pi(K)$, it follows by Lemma \[l:structure\](ii) that $s\not\in\pi(K)$. Construct a normal $r$-series of $K$ as follows: $$1=R_0\leqslant K_1\leqslant R_1\leqslant ...\leqslant K_{l-1}\leqslant R_{l-1}\leqslant K_l\leqslant R_l=K,$$ where $K_i/R_{i-1}=O_{r'}(K/R_{i-1})$ and $R_i/K_i=O_r(K/K_i)$ for $1\leqslant i\leqslant l$. Suppose first that $K/K_l\neq 1$ and let $\tilde K=G/K_l$ and $\widetilde G=G/K_l$. Since the group $C_{\widetilde G}(\tilde K)\tilde K/\tilde K$ is a normal subgroup of $\tilde G/\tilde K \simeq G/K$, it either contains $S$ or is trivial. In the former case $C_{\widetilde G}(\tilde K)$ has an element of order $s$, and so $rs\in\omega(G)$, contrary to (iii). In the latter case we apply Lemma \[l:action\] to the Frobenius group $F$, and again obtain $rs\in\omega(G)$. Now suppose that $K=K_l$ and let $\tilde R=R_{l-1}/K_{l-1}$, $\tilde K=K/K_{l-1}$, and $\tilde G=G/K_{l-1}$. Since $O_{r'}(\tilde K)=1$, it follows by [@56HalHig Lemma 1.2.3] that $C_{\tilde K}(\tilde R)\leqslant \tilde R$. Furthermore, we may assume that $C_{\widetilde G}(\tilde R)\leqslant \tilde K$ as above, and therefore $C_{\widetilde G}(\tilde R)\leqslant \tilde R$. If $p$ does not divide $|\tilde K|$, then $(|\tilde K|,|F|)=1$ and the Schur–Zassenhaus theorem implies that $\tilde G$ has a subgroup isomorphic to $F$. In this case we apply Lemma \[l:action\] and deduce that $rs\in\omega(G)$, a contradiction. Thus $p$ divides $|\tilde K|$. Let $\tilde P$ be a Sylow $p$-subgroup of $\tilde K$, $\tilde Z=Z(\tilde P)$ and $\tilde N=N_{\tilde G}(\tilde P)$. By the Frattini argument, $\tilde G/\tilde K=\tilde N\tilde K/\tilde K$, and therefore $\tilde N$ has an element $g$ of order $s$. Furthermore, since $C_{\tilde N}(\tilde Z)$ is normal in $\tilde N$, this group either is contained in $\tilde K$ or has $S$ as a section. In the former case $g$ does not centralize $\tilde Z$, and then $\tilde Z=[\tilde Z,\langle g\rangle]\times C_{\tilde Z}(g)$ yields $[\tilde Z,\langle g\rangle]\neq 1$. Thus $[\tilde Z,\langle g\rangle]\rtimes\langle g\rangle$ is a Frobenius group that acts on $\tilde R$ faithfully, and applying Lemma \[l:action\] once again, we have $rs\in\omega(G)$, a contradiction. Thus $C_{\tilde N}(\tilde Z)$ has $S$ as a section, and so $pt\in\omega(G)$ for every $t\in\pi(S)\setminus\{p\}$, as required. \[l:adj\_s\] Let $L=S_{2n}(q)$ or $L=O_{2n+1}(q)$, where $n\geqslant 5$. Suppose that $r\in\pi(L)\cap R_k(q)$ with $k>2$, $b\in\omega_{p'}(L)$, and $r$ divides $b$. 1. If $k=2n$ then $b$ divides $(q^n+1)/(2,q-1)$. 2. If $k=2n-2$ then $b$ divides $[q^{n-1}+1,q\pm1]$. 3. If $n$ is even and $k=n-1$ then $b$ divides $[q^{n-1}-1,q\pm1]$. The assertion follows from the definition of primitive prime divisor and Lemmas \[l:spectrum\_s\_odd\] and \[l:spectrum\_s\_2\]. \[l:adj\_o\] Let $L=O_{2n}^+(q)$, where $n\geqslant 6$ is even. Suppose that $r\in\pi(L)\cap R_k(q)$ with $k>2$, $b\in\omega_{p'}(L)$, and $r$ divides $b$. 1. If $k=2n-2$ then $b$ divides $q^{n-1}+1$. 2. If $k=n-1$ then $b$ divides $q^{n-1}-1$. By the definition of primitive prime divisor and Lemma \[l:semi\], it follows that $b$ divides $[q^{n-1}+1,q+1]=q^{n-1}+1$ in (i) and $[q^{n-1}-1,q-1]=q^{n-1}-1$ in (ii). Now we return to the proof of Theorem \[t:quasi\]. \[l:o8\] If $L,S\in\{S_6(q), O_7(q), O^+_8(q)\}$, then either $S\simeq L$, or $\{L,G\}=\{O^+_8(2), S_6(2)\}$, or $\{L,G\}=\{O^+_8(3), O_7(3)\}$. Assume that $S\not\simeq L$. If $q=2$ or $q=3$, then respectively $\{L,G\}=\{S_6(2),O_8^+(2)\}$ or $\{L,G\}=\{O_7(3),O_8^+(3)\}$ by [@97ShiTan; @02Maz.t]. Let $q>3$. Since $\omega(S)\subseteq \omega(L)$, it follows from Lemma \[l:diff\](iii,v) that $S\neq O_8^+(q)$. Furthermore, if $q$ is odd then $S\neq S_6(q)$ by Lemma \[l:diff\](i,iv). Also by Lemma \[l:bncn\], if $L=S_6(q)$, then $S\neq O_7(q)$. Thus $L=O^+_8(q)$ and $S=O_7(q)$ (if $q$ is even we write $O_7(q)$ instead of $S_6(q)$). In particular, $\pi(S)=\pi(L)$. By Lemmas \[l:spectrum\_s\_odd\], \[l:spectrum\_s\_2\], and \[l:spectrum\_o8p\], the orders of semisimple elements of $L$ and $S$ are precisely the divisors of $$\label{e:sso8}(q^4-1)/(2,q-1)^2, (q^3\pm1)/(2,q-1), q^2-1$$ and $$\label{e:sso7}(q^2+1)(q\pm1)/(2,q-1), (q^3\pm1)/(2,q-1), q^2-1$$ respectively. In particular, if $rr_6(q)\in\omega(L)$ for some $r\in\pi(L)$, then $r$ divides $(q^3+1)/(2,q-1)$, and if $rr_3(q)\in\omega(L)$ for some $r\in\pi(L)$, then $r$ divides $(q^3-1)/(2,q-1)$. Since $(q^3+1)/(2,q-1)$ and $(q^3-1)/(2,q-1)$ are coprime, it follows that $r_6(q)$ and $r_3(q)$ are not adjacent and have disjoint neighbourhoods in $GK(L)$. Furthermore, $pr_6(q),pr_3(q)\not\in\omega(L)$ by Lemma \[l:adj\_p\]. We claim that $K\neq 1$ yields $\omega(G)\not\subseteq\omega(L)$. By Lemma \[l:reduction\] we may assume that $K$ is an elementary abelian $r$-group for some prime $r$ and $G$ has a subgroup that is isomorphic to a semidirect product of $K$ and $S$. Also we may assume that $S$ acts on $K$ faithfully. Otherwise $S$ centralizes $K$, and hence all primes of $\pi(S)=\pi(L)=\pi(G)$ other than $r$ are adjacent to $r$ in $GK(G)$, contrary to the fact that $r_3(q)$ and $r_6(q)$ have disjoint neighbourhoods in $GK(L)$. Let $r\neq p$. Since $r_3(q)$ divides the order of a proper parabolic subgroup of $S$ with Levi factor of type $A_2$ and $pr_3(q)\not\in\omega(S)$, the group $S$ has a Frobenius subgroup whose kernel is a $p$-group and whose complement has order $r_3(q)$. Furthermore, $S$ has a Frobenius subgroup with kernel of order $q^2$ and cyclic complement of order $(q^2-1)/(2,q-1)$ by [@11Gr Lemma 5]. Applying Lemma \[l:action\], we conclude that $G$ has elements of orders $rr_3(q)$ and $r(q^2-1)/(2,q-1)$. If $r(q^2-1)/(2,q-1)\in\omega(L)$, then keeping in mind that $q>3$ and consulting (\[e:sso8\]), we see that either $r$ divides $(q^2+1)/(2,q-1)$ or $r=2$. In the former case $rr_3(q)\not\in\omega(L)$. Let $r=2$. Then the highest power of $2$ in $\omega(L)$ is equal to $(q^2-1)_2$. Let us consider a parabolic subgroup $P$ of $SO_7(q)$ with Levi factor of type $B_2$. The Levi factor of $P$ is $A\times B$ where $A\simeq GL_1(q)$, $B\simeq SO_5(q)$ and both $A$ and $B$ contain elements with spinor norm a non-square (see [@90KlLie p. 98]). Since $SO_5(q)$ has elements of order $q^2-1$, it follows that $P\cap S$ also has elements of order $q^2-1$. Furthermore, the Chevalley commutator formula implies that the unipotent radical of $P$ is abelian (see, for example, [@92RRS Lemma 2.2]). Applying Lemma \[l:hh\] we obtain that $2(q^2-1)_2\in\omega(G)\setminus\omega(L)$. Now let $r=p$. If $p=2$ then we apply Lemma \[l:s6\]. If $p$ is odd then $pr_3(q)\in\omega(G)$ by [@15Gr Lemma 3.2], and so $\omega(G)\not\subseteq\omega(L)$. Thus $K=1$ and hence $S\leqslant G\leqslant {\operatorname{Aut}}S$. Consulting (\[e:sso8\]) and (\[e:sso7\]), we see that $r_{4m}(p)(q^2-1)/(2,q-1)$ lies in $\omega(G)\setminus\omega(S)$. Since $r_{4m}(p)\not\in\pi(G/S)$ by Lemma \[l:out\], it follows from (\[e:sso7\]) that at least one of the numbers $(q+1)/(2,q-1)$ and $(q-1)/(2,q-1)$ belongs to $\omega(G/S)$. The group ${\operatorname{Out}}S$ is a direct product of ${\operatorname{Outdiag}}S$ of order $(2,q-1)$ and a cyclic group of order $m$. If $q$ is odd then $q^3+1\in\omega(\operatorname{Inndiag} S)\setminus\omega(L)$, and so $G\cap \operatorname{Inndiag} S=S$ for both even and odd $q$. Therefore the exponent of $G/S$ divides $m$. However $(q\pm1)/(2,q-1)=(p^m\pm1)/(2,p-1)>m$ for $q>3$, a contradiction. If $L,S\in \{S_{2n}(q), O_{2n+1}(q), {}O^-_{2n}(q)\}$, where $n\geqslant 4$, then either $S\simeq L$, or $L\in\{O_{9}(q),S_{8}(q)\}$ and $S={}O^-_{8}(q)$. Assume that $S\not\simeq L$. Since $\omega(S)\subseteq\omega(L)$, it follows by Lemma \[l:diff\](i,vi) that $S\neq S_{2n}(q)$ and by Lemma \[l:diff\](ii) that $(L,S)\neq (O_{2n+1}(q), O_{2n}^-(q))$. Also $(L,S)\neq (S_{2n}(q), O_{2n+1}(q))$ by Lemma \[l:bncn\]. Therefore $L\in\{O_{2n+1}(q),S_{2n}(q)\}$ and $S={}O^-_{2n}(q)$. If $n=4$, the proof is complete. Suppose that $n\geqslant 5$. Let $s=r_{2n-2}(q)$. Then $s$ divides the order of a parabolic subgroup of $S$ with Levi factor of type ${}^2D_{n-1}$ and $ps\not\in\omega(S)$ by Lemma \[l:adj\_p\]. So $S$ has a Frobenius subgroup $F$ whose kernel is a $p$-group and whose complement has order $s$. For every $n$ we will find $r,w\in \pi(L)$ such that $r\in\pi(K)$ and $\{s,r,w\}$ is a coclique in $GK(L)$, and then will apply Lemma \[l:action2\] to the group $F$ and the primes $r$ and $w$. Let $n$ be odd and let $r=r_{nm}(p)$. Since $r\in\pi(L)\setminus\pi(S)$ and $r\not\in\pi({\operatorname{Out}}S)$ by Lemma \[l:out\], it follows that $r\in\pi(K)$. Exploiting Lemma \[l:adj\_s\], we deduce that $\{s,r,r_{2n}(q)\}$ is a coclique in $GK(L)$: both $r$ and $s$ do not divide $q^n+1$, and $r$ does not divide $q^{n-1}+1$ nor $q\pm 1$. Let $n$ be even. Let $r_1=r_{(n-2)m}(p)$, $r_2=r_{(n+2)m}(p)$ if $(n,q)\neq (8,2)$ and $r_1=r_3(2)$, $r_2=r_5(2)$ otherwise. Then $r_1r_2\in\omega(L)\setminus\omega(S)$ by Lemmas \[l:spectrum\_s\_2\], \[l:spectrum\_s\_odd\] and \[l:semi\]. Furthermore, $r_1,r_2\not\in\pi({\operatorname{Out}}S)$ by Lemma \[l:out\]. Therefore at least one of $r_1$ and $r_2$ divides $|K|$. Denote this number by $r$. Since $n\geqslant 6$, both $n+2$ and $n-2$ do not divide $2n-2$, and so $r$ is not adjacent to $s$ nor $r_{n-1}(q)$ in $GK(L)$ by Lemma \[l:adj\_s\]. By the same lemma, $s$ and $r_{n-1}(q)$ are not adjacent either, and $\{s,r, r_{n-1}(q)\}$ is the desired coclique. Applying Lemma \[l:action2\], we conclude that $pt\in \omega(G)$ for all $t\in\pi(S)$, $t\neq p$. But $r_{2n}(q)\in\pi(S)$ and $pr_{2n}(q)\not\in\omega(L)$ by Lemmas \[l:spectrum\_s\_odd\] and \[l:spectrum\_s\_2\], a contradiction. If $L= O^+_{2n}(q)$, where $n\geqslant 6$ is even, then $S\not\in\{S_{2n-2}(q), O_{2n-1}(q)\}$. Assume the contrary. Let $s=r_{n-1}(q)$ and $w=r_{2n-2}(q)$. Then $s$ divides the order of a parabolic subgroup of $S$ with Levi factor of type $A_{n-2}$ and $ps\not\in\omega(S)$, therefore, $S$ contains a Frobenius group whose kernel is a $p$-group and whose complement has order $s$. Furthermore, Lemma \[l:adj\_o\] implies that $s$ and $w$ are not adjacent in $GK(L)$. Let $r_1=r_{(n-2)m}(p)$, $r_2=r_{(n+2)m}(p)$ if $(n,q)\neq (8,2)$, and $r_1=r_3(2)$, $r_2=r_5(2)$ otherwise. Then $r_1r_2\in\omega(L)\setminus\omega(S)$ by Lemmas \[l:spectrum\_s\_2\], \[l:spectrum\_s\_odd\] and \[l:semi\], and $r_1,r_2\not\in\pi({\operatorname{Out}}S)$ by Lemma \[l:out\]. Thus at least one of the numbers $r_1$ and $r_2$ divides $|K|$. Denote this number by $r$. Both $n+2$ and $n-2$ do not divide $2n-2$, and so $\{s, r, w\}$ is a coclique in $GK(L)$ by Lemma \[l:adj\_o\]. By Lemma \[l:action2\], we have $pr_{2n-2}(q)\in \omega(G)$. On the other hand, $pr_{2n-2}(q)\not\in \omega(L)$ by Lemma \[l:adj\_p\], a contradiction. We have considered all the cases in the conclusion of Lemma \[l:previous\], and so the proof of Theorem \[t:quasi\] is complete. New examples of non-quasirecognizable simple groups =================================================== A finite nonabelian simple group $L$ is said to be [*quasirecognizable by spectrum*]{} if every finite group isospectral to $L$ has only one nonabelian composition factor and this factor is isomorphic to $L$. Clearly quasirecognizability is a necessary condition for being almost recognizable. Recall that for even $q$, the simple group $S_{2n}(q)$ is equal to $Sp_{2n}(q)$ and the simple group $O^\varepsilon_{2n}(q)$ is $\Omega^\varepsilon_{2n}(q)$, a subgroup of index 2 in $GO_{2n}^\varepsilon(q)$ (see, for example, [@85Atlas p. xii]). Mazurov and Moghaddamfar [@06MazMog] noted that $\omega(Sp_8(2))=\omega(GO_8^-(2))$, and thereby $Sp_8(2)$ is not quasirecognizable by spectrum. We generalize this result to all even $q$. \[l:GO8\] Let $q$ be even. Then $\omega(Sp_8(q))=\omega(GO_8^-(q))$ and in particular $Sp_8(q)$ is not quasirecognizable by spectrum. Denote $Sp_8(q)$, $GO_8^-(q)$, and $\Omega_8^-(q)$ by $L$, $G$, and $S$ respectively. Since $G$ is a subgroup in $GO_9(q)\simeq Sp_8(q)$, it follows that $\omega(G)\subseteq\omega (L)$. Thus it suffices to show that $\omega(L)\setminus\omega(S)\subseteq\omega(G)$. It follows from Lemma \[l:spectrum\_s\_2\] that $\omega(L)$ consists of all divisors of the numbers $q^4\pm 1$, $(q^2\pm q+1)(q^2-1)$, $2(q^3\pm1)$, $2(q^2+1)(q\pm 1)$, $4(q^2\pm 1)$, and $8(q\pm 1)$. Supplementing this by Lemma \[l:spectrum\_o8m\], we see that every number in $\omega(L)\setminus\omega (S)$ is a divisor of one of the numbers $2(q^3\pm1)$, $4(q^2+1)$, and $8(q\pm 1)$. It is well known that $GO_{2n}^-(q)$ has a subgroup of the form $GO_{2k}^\varepsilon(q)\times GO_{2n-2k}^{-\varepsilon}(q)$ for every $1\leqslant k\leqslant n-1$ and $\varepsilon\in\{+,-\}$. Thus $G$ contains $GO_6^+(q)\times GO_2^-(q)$, $GO_4^+(q)\times GO_4^-(q)$, and $GO_2^+(q)\times GO_6^-(q)$. It can be verified by means of [@GAP] that $4\in\omega(GO_4^+(2))$. Since $GO_4^+(2)\leqslant GO_4^+(q)$ and $SL_2(q^2)<GO_4^-(q)$, it follows that $4(q^2+1)\in\omega(G)$. Since $q^3-\varepsilon \in\omega(GO_6^\varepsilon(q))$ and $GO_2^\varepsilon(q)$ is dihedral of order $2(q-\varepsilon)$, we conclude that $2(q^3\pm 1)\in\omega(G)$. Furthermore, $S_8\simeq GO_6^+(2)\leqslant GO_6^+(q)$, and hence $8(q+1)\in\omega(G)$. It remains to establish that $GO_6^-(q)$ has an element of order 8. The group $GO_6^-(q)$ is isomorphic to a split extension of $SU_4(q)$ by a graph automorphism of order 2. We identify $SU_4(q)$ with $$H=\{A\in SL_4(q^2)\mid AJ\overline A^\top=J\},$$ where $\overline{(a_{ij})}=(a_{ij}^q)$ and $J$ is the $4\times 4$ matrix with 1’s on the antidiagonal and 0’s elsewhere. Then $\gamma$ defined by $A^\gamma=\overline A$ is a graph automorphism of $H$. Choose $t\in GF(q^2)$ such that $t^q\neq t$ and let $$B= \begin{pmatrix} 1& t & 0& 0\\ 0& 1 & 1 & 0 \\ 0& 0 & 1 & t^q \\ 0& 0 & 0 & 1 \\ \end{pmatrix}.$$ It is easy to verify that $B\in H$ and $$(B\gamma)^4= \begin{pmatrix} 1& 0 & 0& t^2+t^{2q}\\ 0& 1 & 0 & 0 \\ 0& 0 & 1 & 0 \\ 0& 0 & 0 & 1 \\ \end{pmatrix}.$$ Since $t^2+t^{2q}\neq 0$, the order of $B\gamma$ is equal to 8, and hence $8\in\omega(GO_6^-(q))$. It is worth noting that another example of a finite group isospectral to $S_8(q)$ with $q$ even and having $O_8^-(q)$ as a composition factor was constructed in [@14GrSt]. [**Acknowledgments.**]{} We are grateful to the referee for his careful reading of the article and helpful comments and suggestions. [10]{} M. R. Aleeva, [On finite simple groups with the set of element orders as in a Frobenius group or a double Frobenius group]{}, *Math. Notes* **73** (2003), no. 3, 299–313. O. A. Alekseeva and A. S. Kondrat’ev, On recognizability of some finite simple orthogonal groups by spectrum, *Proc. Steklov Inst. Math.* **266** (2009), no. suppl. 1, S10–S23. R. Brandl and W.J. Shi, [Finite groups whose element orders are consecutive integers]{}, *J. Algebra* **143** (1991), no. 2, 388–400. R. Brandl and W.J. Shi, [The characterization of $PSL(2,q)$ by its element orders]{}, *J. Algebra* **163** (1994), no. 1, 109–114. A. A. Buturlakin, [Spectra of finite symplectic and orthogonal groups]{}, *Siberian Adv. Math.* **21** (2011), no. 3, 176–210. A. A. Buturlakin and M. A. Grechkoseeva, [The cyclic structure of maximal tori of the finite classical groups]{}, *Algebra Logic* **46** (2007), no. 2, 73–89. J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker, and R. A. Wilson, [*Atlas of finite groups*]{}, Clarendon Press, Oxford, 1985. M. Foroudi Ghasemabadi, A. Iranmanesh, and N. Ahanjideh, Characterizations of the simple group [${}^2D_n(3)$]{} by prime graph and spectrum, *Monatsh. Math.* **168** (2012), no. 3-4, 347–361. M. Foroudi Ghasemabadi, A. Iranmanesh, and N. Ahanjideh, 2-recognizability of the simple groups [$B_n(3)$]{} and [$C_n(3)$]{} by prime graph, *Bull. Iranian Math. Soc.* **39** (2013), no. 6, 1273–1281. The GAP Group, [GAP – Groups, Algorithms, and Programming, Version 4.7.5]{}, 2014. D. Gorenstein, R. Lyons, and R. Solomon, [*The classification of the finite simple groups. Number 3*]{}, [Mathematical Surveys and Monographs]{}, vol. 40.3, American Mathematical Society, Providence, RI, 1998. I. B. Gorshkov, [Recognizability of alternating groups by spectrum]{}, *Algebra Logic* **52** (2013), no. 1, 41–45. M. A. Grechkoseeva, [On element orders in covers of finite simple classical groups]{}, *J. Algebra* **339** (2011), 304–319. M. A. Grechkoseeva, [On spectra of covers of finite simple classical groups]{}, *Dokl. Math.* **84** (2011), no. 1, 464–466. M. A. Grechkoseeva, [On element orders in covers of finite simple groups of Lie type]{}, *J. Algebra Appl.* **14** (2015), 1550056 \[16 pages\]. M. A. Grechkoseeva and D. V. Lytkin, [Almost recognizability by spectrum of finite simple linear groups of prime dimension]{}, *Siberian Math. J.* **53** (2012), no. 4, 645–655. M. A. Grechkoseeva and A. M. Staroletov, [Unrecognizability by spectrum of finite simple orthogonal groups of dimension nine]{}, *Sib. [É]{}lektron. Mat. Izv.* **11** (2014), 921–928. P. Hall and G. Higman, [On the $p$-length of $p$-soluble groups and reduction theorem for Burnside’s problem]{}, *Proc. London Math. Soc.* **6** (1956), no. 3, 1–42. H. He and W. Shi, [A note on the adjacency criterion for the prime graph and characterization of $C_p(3)$]{}, *Algebra Colloq.* **19** (2012), no. 3, 553–562. H. He and W. J. Shi, Recognition of some finite simple groups of type ${D}_n(q)$ by spectrum, *Int. J. Algebra Comput.* **19** (2009), no. 5, 681–698. P. Kleidman and M. Liebeck, [*The subgroup structure of the finite classical groups*]{}, [London Mathematical Society Lecture Note Series]{}, vol. 129, Cambridge University Press, Cambridge, 1990. A. S. Kondrat’ev, [On prime graph components of finite simple groups]{}, *Math. USSR-Sb.* **67** (1990), no. 1, 235–247. A. S. Kondrat’ev, [Recognition by spectrum of the groups $^2D_{2^m+1}(3)$]{}, *Sci. China Ser. A* **52** (2009), no. 2, 293–300. V. D. Mazurov, [Characterizations of finite groups by sets of orders of their elements]{}, *Algebra and Logic* **36** (1997), no. 1, 23–32. V. D. Mazurov, [Recognition of finite groups by a set of orders of their elements]{}, *Algebra and Logic* **37** (1998), no. 6, 371–379. V. D. Mazurov, [Recognition of finite simple groups $S_4(q)$ by their element orders]{}, *Algebra Logic* **41** (2002), no. 2, 93–110. V. D. Mazurov, [Unrecognizability by spectrum for a finite simple group $^3D_4(2)$]{}, *Algebra Logic* **52** (2013), no. 5, 400–403. V. D. Mazurov and A. R. Moghaddamfar, [The recognition of the simple group $S_8(2)$ by its spectrum]{}, *Algebra Colloq.* **13** (2006), no. 4, 643–646. V. D. Mazurov and W.J. Shi, [A note to the characterization of sporadic simple groups]{}, *Algebra Colloq.* **5** (1998), no. 3, 285–288. V. D. Mazurov, M.C. Xu, and H.P. Cao, [Recognition of the finite simple groups $L_3(2^m)$ and $U_3(2^m)$ by their element orders]{}, *Algebra and Logic* **39** (2000), no. 5, 324–334. R. Richardson, G. Röhrle, and R. Steinberg, [Parabolic subgroups with [A]{}belian unipotent radical]{}, *Invent. Math.* **110** (1992), no. 3, 649–671. R. Shen, W.J. Shi, and M. R. Zinov’eva, [Recognition of simple groups $B_p(3)$ by the set of element orders]{}, *Siberian Math. J.* **51** (2010), no. 2, 244–254. W.J. Shi, [The characterization of the sporadic simple groups by their element orders]{}, *Algebra Colloq.* **1** (1994), no. 2, 159–166. W.J. Shi and C.Y. Tang, [A characterization of some orthogonal groups]{}, *Progr. Natur. Sci.* **7** (1997), no. 2, 155–162. A. M. Staroletov, [On recognition by spectrum of the simple groups $B_3(q)$, $C_3(q)$ and $D_4(q)$]{}, *Siberian Math. J.* **53** (2012), no. 3, 532–538. A. V. Vasil’ev, [On connection between the structure of a finite group and the properties of its prime graph]{}, *Siberian Math. J.* **46** (2005), no. 3, 396–404. A. V. Vasil’ev, [On finite groups isospectral to simple classical groups]{}, *J. Algebra* **423** (2015), 318–374. A. V. Vasil’ev and I. B. Gorshkov, [On recognition of finite simple groups with connected prime graph]{}, *Siberian Math. J.* **50** (2009), no. 2, 233–238. A. V. Vasil’ev, I. B. Gorshkov, M. A. Grechkoseeva, A. S. Kondrat’ev, and A. M. Staroletov, [On recognizability by spectrum of finite simple groups of types $B_n$, $C_n$, and $^2D_n$ for $n=2^k$]{}, *Proc. Steklov Inst. Math.* **267** (2009), no. suppl. 1, 218–233. A. V. Vasil’ev and M. A. Grechkoseeva, [Recognition by spectrum for finite simple linear groups of small dimensions over fields of characteristic $2$]{}, *Algebra Logic* **47** (2008), no. 5, 314–320. A. V. Vasil’ev, M. A. Grechkoseeva, and V. D. Mazurov, [On finite groups isospectral to simple symplectic and orthogonal groups]{}, *Siberian Math. J.* **50** (2009), no. 6, 965–981. A. V. Vasil’ev and A. M. Staroletov, [Almost recognizability of simple exceptional groups of Lie type]{}, *Algebra Logic* **53** (2015), no. 6, 433–449. A. V. Vasil’ev and E.P. Vdovin, [An adjacency criterion for the prime graph of a finite simple group]{}, *Algebra Logic* **44** (2005), no. 6, 381–406. A. V. Vasil’ev and E. P. Vdovin, [Cocliques of maximal size in the prime graph of a finite simple group]{}, *Algebra Logic* **50** (2011), no. 4, 291–322. J. S. Williams, [Prime graph components of finite groups]{}, *J. Algebra* **69** (1981), 487–513. A. V. Zavarnitsine, [Recognition of the simple groups $L_3(q)$ by element orders]{}, *J. Group Theory* **7** (2004), no. 1, 81–97. A. V. Zavarnitsine, [Recognition of the simple groups $U_3(q)$ by element orders]{}, *Algebra Logic* **45** (2006), no. 2, 106–116. A. V. Zavarnitsine, [Exceptional action of the simple groups ${L}_4(q)$ in the defining characteristic]{}, *Sib. [É]{}lektron. Mat. Izv.* **5** (2008), 65–74. K. Zsigmondy, [Zur Theorie der Potenzreste]{}, *Monatsh. Math. Phys.* **3** (1892), 265–284. [^1]: The work is supported by the Russian Foundation for Basic Research (project 13-01-00505). [^2]: in his talk on International Algebraic Conference in St. Petersburg, September 24–29, 2007.

ArXiv

--- abstract: 'AGILE is an innovative, cost-effective gamma-ray mission proposed to the ASI Program of Small Scientific Missions. It is planned to detect gamma-rays in the 30 MeV–50 GeV energy band and operate as an [**Observatory**]{} open to the international community. Primary scientific goals include the study of AGNs, gamma-ray bursts, Galactic sources, unidentified gamma-ray sources, solar flares, and diffuse gamma-ray emission. AGILE is planned to be operational during the year 2001 for a 3-year mission. It will ideally ‘fill the gap’ between EGRET and GLAST, and support ground-based multiwavelength studies of high-energy sources.' author: - 'M.TAVANI$^{1,2}$, G.BARBIELLINI$^3$, P.CARAVEO$^1$, S.DI PIPPO$^4$, S.MEREGHETTI$^1$, A.MORSELLI$^5$, A.PELLIZZONI$^1$, A.PERRINO$^5$, P.PICOZZA$^5$, P.SCHIAVON$^3$, S.SEVERONI$^5$, F.TAVECCHIO$^1$, A.VACCHI$^3$, S.VERCELLONE$^1$' title: 'AGILE: A GAMMA-RAY MISSION FOR A LIGHT IMAGING DETECTOR$^{(\star)}$' --- \#1 \#2 [[*Mem. Soc. Astron. It.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*The Messenger*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Astron. Nach.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Astron. Astrophys.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Astron. Astrophys. Lett.*]{} [**\#1**]{}, L\#2]{} \#1 \#2 [[*Astron. Astrophys. Rev.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Astron. Astrophys. Suppl. Ser.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Astron. J.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Ann. Rev. Astron. Astrophys.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Astrophys. J.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Astrophys. J. Lett.*]{} [**\#1**]{}, L\#2]{} \#1 \#2 [[*Astrophys. J. Suppl.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Astrophys. Space Sci.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Adv. Space Res.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Bull. Astron. Inst. Czechosl.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*J. Quant. Spectrosc. Radiat. Transfer*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Mon. Not. R. Astr. Soc.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Mem. R. Astr. Soc.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Phys. Rev. Lett.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Phys. Lett. Rev.*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Publ. Astron. Soc. Japan*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Publ. Astr. Soc. Pacific*]{} [**\#1**]{}, \#2]{} \#1 \#2 [[*Nature*]{} [**\#1**]{}, \#2]{} epsf.sty psfig.sty   \#1 \#2 \#3 \#4 =3em =1 by 0em \#1, [*\#2*]{}, [**\#3**]{}, \#4. \#1 =-3em =1 by 0em \#1. Introduction ============= The gamma-ray mission AGILE ([*Astro-rivelatore Gamma a Immagini LEggero*]{}) is currently in a study phase for the Italian Space Agency (ASI) program of Small Scientific Missions. AGILE ideally conforms to the [*faster, cheaper, better*]{} philosophy adopted by space agencies for scientific missions. Gamma-ray detection by AGILE is based on silicon tracking detectors developed for space missions by INFN and Italian University laboratories during the past years (Barbiellini et al., 1995; Morselli et al., 1995). AGILE is both very light ($\sim 60$ kg) and highly efficient in detecting and monitoring gamma-ray sources in the energy range 30 MeV–50 GeV. The accessible field of view is unprecedently large (${\stackrel{>}{\sim}}1/5$ of the whole sky) because of state-of-the-art readout electronics and segmented anticoincidence system. AGILE was selected by ASI (1997 December) for a phase A study to be completed within the end of 1998. The goal is to achieve an on-axis sensitivity comparable to that of EGRET on board of CGRO (a smaller background resulting from an improved angular resolution more than compensates the loss due to a smaller effective area) and a better sensitivity for large off-axis angles (up to $\sim 60^{\circ}$). Planned to be operational during the year 2001 for a 3-year mission, AGILE will ideally ‘fill the vacuum’ between the end of EGRET operations and GLAST. AGILE’s data will provide crucial support for ground-based observations and several space missions including AXAF, INTEGRAL, XMM, ASTRO-E, SPECTRUM-X. ——————————————————————————-\ ($\star$) Adapted from a paper presented at the Conference [*Dal nano- al Tera-eV: tutti i colori degli AGN*]{}, Rome 18-21 May 1998, to be published by the [*Memorie della Societa’ Astronomica Italiana*]{}. Fig. 1 (left panel) shows the baseline instrument layout. We refer to a companion paper for more details on the instrument (Morselli et al., 1998). Spectral information ($\Delta \, E/E \sim 1$) will be obtained by multiple scattering of created pairs in tungsten-silicon planes (for energies less than $\sim 500$ MeV) and by the use of a mini-calorimeter. We are also studying the possibility of adding an ultra-light coded mask imaging system sensitive in the energy band $\sim $10–40 keV on top of AGILE. Super-AGILE is an innovative concept[^1], combining silicon technology to simultaneously detect gamma-rays and hard X-rays with accurate imaging. Scientific Objectives ===================== Table 1 summarizes the expected performance of AGILE vs. that of EGRET. Because of the large field of view ($\sim 0.8 \,\pi$ sr) AGILE will discover a large number of gamma-ray transients, monitor known sources, and allow rapid multiwavelength follow-up observations because of a dedicated data analysis and alert program. Fig. 1 (right panel) shows the off-axis response to gamma-ray detection of AGILE and EGRET. We summarize here AGILE’s scientific objectives. $\bullet$ [*Active Galactic Nuclei*]{}. For the first time, simultaneous monitoring of a large number of AGNs per pointing will be possible. Several outstanding issues concerning the mechanism of AGN gamma-ray production and activity can be addressed by AGILE including: (1) the study of transient vs. low-level gamma-ray emission and duty-cycles; (2) the relationship between the gamma-ray variability and the radio-optical-X-ray-TeV emission; (3) the correlation between relativistic radio plasmoid ejections and gamma-ray flares. A program for joint AGILE and ground-based monitoring observations is being planned. On the average, AGILE will achieve deep exposures of AGNs and substantially improve our knowledge on the low-level emission as well as detecting flares. We conservatively estimate that for a 3-year program AGILE will detect a number of AGNs 2–3 times larger than that of EGRET. A companion paper presents the impact of AGILE on the study of AGNs (Mereghetti et al., 1998). Super-AGILE will monitor, for the first time, simultaneous AGN emission in the gamma-ray and hard X-ray ranges. $\bullet$ [*Diffuse Galactic and extragalactic emission*]{}. The AGILE good angular resolution and large average exposure will further improve our knowledge of cosmic ray origin, propagation, interaction and emission processes. We also note that a joint study of gamma-ray emission from MeV to TeV energies is possible by special programs involving AGILE and new-generation TeV observatories of improved angular resolution. $\bullet$ [*Gamma-ray pulsars*]{}. will contribute to the study of gamma-ray pulsars in several ways: (1) improving photon statistics for gamma-ray period searches by dedicated observing programs with long observation times of 1-2 months per source; (2) detecting possible secular fluctuations of the gamma-ray emission from neutron star magnetospheres; (3) studying unpulsed gamma-ray emission from plerions in supernova remnants and searching for time variability of pulsar wind/nebula interactions, e.g., as in the Crab nebula (de Jager et al., 1996). $\bullet$ [*Galactic sources, new transients*]{}. A large number of gamma-ray sources near the Galactic plane are unidentified, and sources such as 2CG 135+1 or transients (e.g., GRO J1838-04) can be monitored on timescales of months/years. Also Galactic X-ray jet sources (such as Cyg X-3, GRS 1915+10, GRO J1655-40 and others) can produce detectable gamma-ray emission for favorable jet geometries, and a TOO program is planned to follow-up new discoveries of [*micro-quasars*]{}. $\bullet$ [*Gamma-ray bursts*]{}. About ten GRBs have been detected by EGRET’s spark chamber during $\sim 7$ years of operations (Schneid et al., 1996a). This number appears to be limited by the EGRET FOV and sensitivity and not by the GRB emission mechanism. GRB detection rate by AGILE is expected to be a factor of $\sim 5$ larger than that of EGRET, i.e., ${\stackrel{>}{\sim}}$5–10 events/year). The small AGILE deadtime (${\stackrel{>}{\sim}}100 $ times smaller than that of EGRET) allows a better study of the initial phase of GRB pulses (for which EGRET response was in many cases inadequate). The remarkable discovery by EGRET of ‘delayed’ gamma-ray emission up to $\sim 20$ GeV from GRB 940217 (Hurley et al., 1994) is of great importance to model burst acceleration processes. AGILE is expected to be highly efficient in detecting photons above 10 GeV because of limited backscattering. Super-AGILE will be able to locate GRBs within a few arcminutes, and will systematically study the interplay between hard X-ray and gamma-ray emissions. $\bullet$ [*Solar flares*]{}. During the last solar maximum, solar flares were discovered to produce prolonged high-intensity gamma-ray outbursts (e.g., Schneid et al., 1996b). AGILE will be operational during part of the next solar maximum and several solar flares may be detected. Particularly important for analysis will be the flares simultaneously detected by AGILE and HESSI (sensitive in the band 20 keV–20 MeV). Mission ======= AGILE is planned to be integrated with a spacecraft of the MITA class currently being developed by Gavazzi Space with the support of ASI. AGILE’s pointing is obtained by a three-axis stabilization system with an accuracy near 0.5–1 degree. Pointing reconstruction reaching an accuracy of 1–2 arcmin is obtained by star trackers. The downlink telemetry rate is planned to be $\sim 500 \; \rm kbit \, s^{-1}$, and is adequate for AGILE and Super-AGILE for a single contact per orbit. The ideal orbit is equatorial (550 – 650 km). The AGILE mission is being planned as an [**Observatory**]{} open to the international scientific community. Planning of pointed observations, quicklook and standard data analysis results will be available to the community through a Guest Observer Program. The AGILE mission emphasizes a rapid response to the detection of gamma-ray transients. The AGILE Science Support Group will help coordinating multiwavelength observations of gamma-ray sources, and will stimulate investigations of observational and theoretical nature on gamma-ray sources detected by AGILE. 111= 222= 111 222 Barbiellini G. [[et al.]{}]{}: 1995, [*Nucl. Instrum. & Methods*]{} [**354**]{}, 547 de Jager O.C. [[et al.]{}]{}: 1996, 457 253 Hurley K., [[et al.]{}]{}: 1994, 372 652 Mereghetti S. [[et al.]{}]{}: 1998, these Proceedings Morselli A. [[et al.]{}]{}: 1995, in [*XXIV Int. Cosmic Ray Conf.*]{}, [**3**]{}, p. 669 Morselli A. [[et al.]{}]{}: 1998, these Proceedings Schneid E.J. [[et al.]{}]{}: 1996a, in [*AIP Conf. Proc.*]{}, [**384**]{}, p. 253 Schneid E.J. [[et al.]{}]{}: 1996b, 120 299 Thompson D.J. [[et al.]{}]{}: 1993, 86 657 [^1]: Developed in collaboration with E. Costa, M. Feroci, L. Piro and P. Soffitta

ArXiv

--- address: | Institut de Física d’Altes Energies (IFAE), Universitat Autònoma de Barcelona,\ E-08193 Bellaterra (Barcelona), Spain author: - 'C. BIGGIO' title: 'DIVERGENCES AND SYMMETRIES IN HIGGS-GAUGE UNIFICATION THEORIES' --- Introduction: Why Studying Higgs-Gauge Unification Theories? {#sec:intro} ============================================================ One of the possible motivations for studying Higgs-gauge unification theories in extra dimensions [@Randjbar-Daemi:1982hi]$^\textrm{-}$[@Scrucca:2003ut] is the so-called little hierarchy problem.[@Giudice:2003nc] If one considers the Standard Model (SM) as an effective theory valid up to a certain scale $\Lambda_{SM}$ and calculates the radiative corrections to the Higgs mass, he finds that these diverge quadratically and that, in order to avoid a fine-tuning of the parameters, $\Lambda_{SM}$ must be smaller than 1 TeV. This means that new physics must enter into the game at the scale of 1 TeV to regularize the ultraviolet behaviour. On the other hand this new physics can be parametrized by adding to the SM lagrangian non-renormalizable operators suppressed by powers of $\Lambda_{LH}$. From the non-observation of dimension-six four-fermion operators at LEP a lower limit on $\Lambda_{LH}$ of 5-10 TeV has been derived.[@PDG] This one order of magnitude discrepancy between the theoretically required upper limit and the experimental lower limit is called little hierarchy problem. Up to now the best solution to the little (and grand) hierarchy problem is supersymmetry (SUSY). In the supersymmetric extensions of the SM quadratic divergences are absent so that the supersymmetric model can be extended up to $M_{Pl}$ without the need of other new physics: this solves the hierarchy problem. The $\Lambda_{SM}$ is now identified with the mass of the supersymmetric particles, which can be $O(1\textrm{ TeV})$. Moreover, if R-parity is conserved, it induces a suppression in the loop corrections to four-fermions operators which results in the relation $\Lambda_{LH}\sim 4\pi \Lambda_{SM}$, that precisely solves the little hierarchy problem. However SUSY has not yet been discovered and this reintroduces a small amount of fine-tuning in the theory. For this and other reasons we think it can be worthwhile looking for alternative solutions to the little hierarchy problem. One alternative solution is given by the so-called Higgs-gauge unifications theories, in the context of theories in extra dimensions compactified on orbifolds. If we consider a gauge field in $D$ dimensions, its components can be split into two parts, according to the transformation properties under the four-dimensional ($4D$) Lorentz group: $A^A_M = (A^A_\mu , A^A_i)$, where $A^A_\mu$ is a $4D$ Lorentz vector while $A^A_i$ are $4D$ Lorentz scalars. The latter can be identified with the Higgs fields and they can acquire a non-vanishing vacuum expectation value through the Hosotani mechanism.[@Hosotani:1983xw] The good feature of this kind of constructions is that the Higgs mass in the bulk is protected from quadratic divergences by the higher-dimensional gauge invariance and only finite corrections $\propto (1/R)^2$, where $R$ is the compactification radius, can appear. So the picture is the following: we have the $4D$ SM valid up to the scale $\Lambda_{SM}\sim 1/R$ that can be $O(1\textrm{ TeV})$, then we have a non-renormalizable $D$-dimensional theory valid up to a certain scale $\Lambda_D$ which can be greater or equal to $10\textrm{ TeV}$ and then we have the ultraviolet completion of the theory. As we see also in this case the little hierarchy problem is solved. In this talk we will deal with some specific features of these theories and, in particular, we will discuss how the mass protection given by the higher-dimensional gauge invariance can be spoiled at the fixed points of the orbifold and under which conditions it can be restored. Symmetries at the Fixed Points of an Orbifold and Allowed Localized Terms {#sec:main} ========================================================================= Gauge Theories on Orbifolds {#sub:gaugeteo} --------------------------- We begin by considering a gauge theory coupled to fermions in a $D$-dimensional ($D=d+4>4$) space-time parametrized by coordinates $x^M=(x^\mu, y^i)$ where $\mu=0,1,2,3$ and $i=1,\dots ,d$. The lagrangian is $${\mathcal L}_{D}=-\frac{1}{4} {F}_{MN}^A{F}^{AMN} +i{\overline \Psi}\Gamma_D^M D_M{\Psi}, \label{bulk-lagr}$$ with $F^A_{MN}=\partial_M A_N^A - \partial_N A_M^A - g f^{ABC} A_M^B A_N^C$, $D_M=\partial_M-igA_M^AT^A$ and where $\Gamma_D^M$ are the $\Gamma$-matrices corresponding to a $D$-dimensional space-time. This lagrangian is invariant under the gauge group $\mathcal{G}$ and of course under the $D$-dimensional Lorentz group $SO(1,D-1)$. Now we compactify the extra dimensions on an orbifold.[@orbifold] Firstly we build up a $d$-dimensional torus $T^d$ by identifying $(x^\mu, y^i)$ with $(x^\mu, y^i + u^i)$, with $\vec{u}$ belonging to a $d$-dimensional lattice $\Lambda^d$. Then we act on the torus with the element $k$ of the group $\mathbb G$ generated by a discrete subgroup of $SO(d)$ that acts crystallographically on the torus lattice and by discrete shifts that belong to it. The orbifold is finally defined by the identification $(x^\mu, y^i)=(x^\mu, (P_k\ \vec{y})^i + u^i)$, where $P_k$ is the rotation associated to the element $k\in \mathbb G$ . This group acts non-freely on the torus, i.e. it leaves some points invariant: these are called fixed points. The construction of the orbifold $S^1/Z_2$ in five dimensions is depicted in Fig. \[disegno\]: by identifying $y$ with $y+2 n \pi R$ (with $n$ integer) we are left with a segment of length $2\pi R$ and the extrema identified which can be represented by a circle; then by identifying $y$ with $-y$ we end up with a segment of length $\pi R$, with two fixed points in $y_f=0,\pi R$, which is the orbifold. Now that we have defined the orbifold we have to specify how it acts on the fields. If $\phi_{\mathcal R}$ is a generic field transforming as an irreducible representation ${\mathcal R}$ of the gauge group $\mathcal G$, then the orbifold action is defined by $$k\cdot\phi_{\mathcal R}(y) =\lambda^k_{\mathcal R}\otimes\mathcal P^k_\sigma\, \phi_{\mathcal R}(k^{-1}\cdot y) \label{orbfield}$$ where $\lambda^k_{\mathcal R}$ is acting on gauge and flavor indices and $\mathcal P^k_\sigma$, where $\sigma$ refers to the field spin, on Lorentz indices. Splitting the action of the orbifold in this way is particularly useful since while $\mathcal P^k_\sigma$ is fixed by requiring the invariance of the lagrangian under this transformation (in particular we obtain $\mathcal P_0^k=1$ for scalar fields and $\mathcal P_1^k=P_k$ for gauge fields), $\lambda^k_{\mathcal R}$ is unconstrained and it can be used to break symmetries.[^1] Now we consider the gauge symmetry breaking realized by the orbifold at the generic fixed point $y_f$. First of all we have to understand why we are looking precisely there. The reason is that new lagrangian terms localized at the fixed points can be generated by bulk radiative corrections [@ggh] if they are compatible with the existing symmetries. Since we are interested in the stability of the Higgs mass under radiative corrections, the knowledge of these symmetries can tell us if our theory is stable or not. In general the orbifold action breaks the gauge group in the bulk $\mathcal G=\{T^A\}$ to a subgroup $\mathcal H_f=\{T^{a_f}\}$, at the fixed point $y_f$, defined by the generators of $\mathcal G$ which commute with $\lambda^k_{\mathcal R}$, i.e. $[\lambda^k_{\mathcal R},T^{a_f}_{\mathcal R}]=0$. This condition must be satisfied by any irreducible representation $\mathcal R$ of $\mathcal G$. The symmetry breaking pattern also defines which fields are non-zero at $y_f$. In this case they will be $A^a_\mu$, which are the gauge bosons of the unbroken gauge group $\mathcal H_f$, and $A^{\hat{a}}_i$, for some $i$ and $\hat{a}$, with zero modes, plus some derivatives of non-invariant fields without zero modes. Since also the parameters which define the gauge transformation transform under the orbifold action, the derivatives of some of them which are invariant define a set of local transformations that are called $\mathcal K$-transformations.[@vonGersdorff:2002us] Eventually at the fixed point $y_f$ there are two symmetries remnant of the original gauge symmetry $\mathcal G$: $\mathcal H_f$ and $\mathcal K$. All this is summarized in Fig. \[disegno\]. Allowed Localized Lagrangian Terms {#sub:symmfix} ---------------------------------- Now that we have discussed the main features of gauge symmetry breaking on orbifold, we are ready to write down the most general $4D$ effective lagrangian. This is given by the integral over the extra coordinates of the $D$-dimensional lagrangian plus the terms localized at the fixed points which are compatibles with the symmetries: $$\mathcal L_4^{eff}=\int d^d y \bigl[ \mathcal L_D+\sum_{f}\delta^{(d)}(y-y_f)\,\mathcal L_f \bigr]. \label{eff-lag}$$ How are exactly these $\mathcal L_f$? Before answering to this question we have to list the symmetries holding at the fixed points. These are the orbifold group \[$\mathbb G_f$\], the $4D$ Lorentz group \[$SO(1,3)$\], the residual gauge group \[$\mathcal H_f$\] and the residual local symmetry \[$\mathcal K$\]. The $\mathcal K$-symmetry is very important since it forbids the appearance of direct mass terms like $\Lambda^2 A^{\hat{a}}_i A^{\hat{b}}_j$ in the case in which $A^{\hat{a}}_i$ is $\mathbb G_f$-invariant. Anyway the previously listed symmetries allow localized terms as $(F^a_{\mu\nu})^2$, which corresponds to a localized kinetic term for $A_\mu^a$, and $F^a_{\mu\nu}\tilde{F}^{a\mu\nu}$ which is a localized anomaly. Moreover if for some $(i,j)$ $F^a_{ij}$ and $A_i^{\hat a}$ are orbifold invariant (this is model-dependent), $(F^a_{ij})^2$ and $(F^{\hat a}_{i\mu})^2$ are also allowed, giving rise respectively to localized quartic couplings and kinetic terms for $A_i^{\hat a}$. All these operators are dimension-four, that is they renormalize logarithmically. However if $\mathcal H_f$ contains a $U(1)$ factor $$F_{ij}^\alpha=\partial_i A_j^\alpha-\partial_jA_i^\alpha-gf^{\alpha \hat b \hat c}A_i^{\hat b}A_j^{\hat c}, \label{fij}$$ where $\alpha$ is the $U(1)$ quantum number, is invariant under all the above discussed symmetries and can be generated by bulk radiative corrections at the fixed points. This means that we expect both a tadpole for the derivatives of odd fields and a mass term for the even fields. Since these operators are dimension-two, their respective renormalizations will lead to quadratic divergences, making the theory ultraviolet-sensitive. Apart from the $5D$ case where the term $F_{ij}$ does not exist, for $D\ge 6$ it does and its generation has been confirmed by direct computation in 6D orbifold field [@vonGersdorff:2002us]$^\textrm{-}$[@Scrucca:2003ut] and 10D string [@GrootNibbelink:2003gb] theories. Of course if these divergent localized mass terms were always present, Higgs-gauge unification theories would not be useful in order to solve the little hierarchy problem. One way out can be that local tadpoles vanish globally, but this requires a strong restriction on the bulk fermion content.[@Scrucca:2003ut] A more elegant and efficient solution, based on symmetry arguments, has been presented in Ref. \[12\] and will be discussed in the following. The Residual $O_f$ Symmetry {#sub:of} --------------------------- When we discussed the symmetry breaking induced by the orbifold, we did not consider the $D$-dimensional Lorentz group. When compactifying a $d$-dimensional space to a smooth Riemannian manifold (with positive signature), at each point a tangent space can be defined and the orthogonal transformations acting on it form the group $SO(d)$.[@Witten] When the orbifold group acts on the manifold, in the same way as the gauge group $\mathcal G$ is broken down to $\mathcal H_f$, the internal rotation group $SO(d)$ is broken down to a subgroup $\mathcal O_f$ defined by the generators of $SO(d)$ which commutes with $\mathcal P^k_\sigma$, i.e. $[\mathcal P^k_\sigma,\mathcal O_f]=0$. This means that the original Lorentz group $SO(1,D-1)$ is firstly broken down to $SO(1,3)\otimes SO(d)$ (where $SO(d)$ must be understood as acting on the tangent space) by the smooth compactification and then it is definitively broken down to $SO(1,3)\otimes \mathcal O_f$ by the orbifold action. All this is outlined in Fig. \[disegno\]. We have then identified an additional symmetry that the lagrangian $\mathcal L_f$ at the fixed point $y_f$ must conserve. Summarizing, the invariances that we have to take into account are the following: 4D Lorentz invariance \[$SO(1,3)$\], invariance under the action of the orbifold group \[$\mathbb G_f$\], usual 4D gauge invariance \[$\mathcal H_f$\], remnant of the bulk gauge invariance \[$\mathcal K$\] and remnant of the invariance under rotations of the tangent space \[$\mathcal O_f$\]. Now the question is: can this $\mathcal O_f$ forbid the appearance of the tadpole (or, equivalently, of the divergent mass term for the Higgs)? If $\mathcal O_f$ contains among its factors at least one $SO(2)$, then a corresponding Levi-Civita tensor $\epsilon^{ij}$ exists, such that the lagrangian term $\epsilon^{ij} F_{ij}^{(\alpha)}$ is also $\mathcal O_f$-invariant. In this case tadpoles are allowed. On the other side, if $\mathcal O_f$ is given by a product of $SO(p_i)$ with $p_i > 2\ \forall i$, then the Levi-Civita tensor has $p_i$ indices and only invariants constructed using $p_i$-forms are allowed. Since $F_{ij}^\alpha$ has two indices this means that in this case tadpoles are not allowed. We have then found a sufficient condition for the absence of localized tadpoles which precisely is that the smallest internal subgroup factor be $SO(p)$ with $p>2$. Evidently $\mathcal O_f$ is orbifold-dependent; in Ref. \[12\] we analyzed the case of the orbifold $T^d/Z_N$ for $d$ even. In this case the generator of the orbifold group is given by $P_N={\rm diag}(R_1,\dots,R_{d/2})$, where $R_i$ is the discrete rotation in the $(y_{2i-1},y_{2i})$-plane. If $\mathbb Z_{N_f}$ is the orbifold subgroup which leaves invariant the point $y_f$, it can be shown that if $N_f>2$ then $\mathcal O_f=\bigotimes_{i=1}^{d/2}SO(2)_i$, where $SO(2)_i$ is the $SO(2)\subseteq SO(d)$ that acts on the $(y_{2i-1},y_{2i})$-subspace. Then in every subspace $\epsilon^{IJ}$ exists and we expect a tadpoles appearance at the fixed points $y_f$ of the form $$\sum_{i=1}^{d/2}\mathcal C_i\sum_{I,J=2i-1}^{2i}\epsilon^{IJ}F^\alpha_{IJ}\;\delta^{(d/2)}(y-y_f). \label{tadpolos}$$ On the contrary if $N_f=2$ then the generator of the orbifold subgroup is the inversion $P_2=-\mathbf{1}$ that obviously commutes with all the generators of $SO(d)$ so that we have $\mathcal O_f=SO(d)$. In this case the Levi-Civita tensor is $\epsilon^{i_1\dots i_d}$ and only a $d$-form can be generated linearly in the localized lagrangian. Therefore tadpoles are only expected in the case of $d=2$ $(D=6)$. This last comment also apply to the case of $\mathbb Z_2$ orbifolds in arbitrary dimensions (even or odd), since the orbifold generator is always $P=-\mathbf{1}$ and then the internal rotation group is always $\mathcal O_f=SO(d)$. In Ref. \[12\] we explicitly checked this result at one- and two-loops for the orbifold $T^d/Z_2$ for any $D$. Conclusions {#sec:conclu} =========== In orbifold field theories the SM Higgs field can be identified with the internal components of gauge fields. Then the higher-dimensional gauge invariance prevents the Higgs from acquiring a quadratically divergent mass term in the bulk, while at the fixed points a remnant of bulk gauge symmetry after symmetry breaking forbids the appearance of direct mass terms. Still, if the residual gauge symmetry contains a $U(1)$ factor, the corresponding field strength for the $4D$ scalar fields is invariant under the orbifold action, the $4D$ Lorentz symmetry, the residual gauge invariance and the residual local symmetry, so that it can be radiatively generated at the fixed points. This is a dimension-two operator and gives rise to a quadratically divergent mass for the Higgs. However we showed that another symmetry must be considered. Indeed, when compactifying on an orbifold, the internal rotation group acting on the tangent space that can be defined at each point of a smooth manifold is broken down at the fixed points, since there a tangent space cannot be defined. How the breaking is realized depends on the particular orbifold but in general a group $\mathcal O_f$, subgroup of the internal rotation group, will survive and then it shall be taken into account when looking for lagrangian terms that can be radiatively generated. Actually this residual symmetry can forbid the appearance of dangerous divergent terms. Indeed if $\mathcal O_f$ contains among its factors at least one $SO(2)$, then a Levi-Civita tensor $\epsilon^{ij}$ exists and the previously mentioned invariant field strength will be generated at the considered fixed point in the form $\epsilon^{ij}F^{(\alpha)}_{ij}$. On the contrary if $\mathcal O_f$ is given by a product of $SO(p_i)$ with each $p_i>2$, then only invariants constructed with $p_i$-forms can be generated and our dangerous term will not appear. What we have found is then another sufficient condition for the absence of localized tadpoles. Also we have shown that in the case of the orbifolds $T^d/Z_N$ ($N>2$, $d$ even) $\mathcal O_f$ is a product of $SO(2)$ groups (at least at the $Z_N$-fixed points) and then divergent terms will always be allowed. On the other side for the orbifolds $T^d/Z_2$ (any $d$) $\mathcal O_f$ always coincides with the whole $SO(d)$ and then tadpoles will never appear for $d>2$ ($D>6$). Acknowledgments {#acknowledgments .unnumbered} =============== I would like to thank Mariano Quirós for the pleasant collaboration on which this talk is based. I also would like to thank the organizers of the “XL Rencontres de Moriond” for the enjoyable atmosphere of this interesting conference. References {#references .unnumbered} ========== [99]{} S. Randjbar-Daemi, A. Salam and J. Strathdee, Nucl. Phys. B [**214**]{} (1983) 491; I. Antoniadis and K. Benakli, Phys. Lett. B [**326**]{} (1994) 69 \[arXiv:hep-th/9310151\]; H. Hatanaka, T. Inami and C. S. Lim, Mod. Phys. Lett. A [**13**]{} (1998) 2601 \[arXiv:hep-th/9805067\]; H. Hatanaka, Prog. Theor. Phys.  [**102**]{} (1999) 407 \[arXiv:hep-th/9905100\]; M. Kubo, C. S. Lim and H. Yamashita, arXiv:hep-ph/0111327; G. R. Dvali, S. Randjbar-Daemi and R. Tabbash, Phys. Rev. D [**65**]{} (2002) 064021 \[arXiv:hep-ph/0102307\]; I. Antoniadis, K. Benakli and M. Quiros, Nucl. Phys. B [**583**]{} (2000) 35 \[arXiv:hep-ph/0004091\]; I. Antoniadis, K. Benakli and M. Quiros, New Jour. Phys.  [**3**]{} (2001) 20 \[arXiv:hep-th/0108005\]. G. von Gersdorff, N. Irges and M. Quiros, Phys. Lett. B [**551**]{} (2003) 351 \[arXiv:hep-ph/0210134\]. C. Csaki, C. Grojean and H. Murayama, Phys. Rev. D [**67**]{} (2003) 085012 \[arXiv:hep-ph/0210133\]. C. A. Scrucca, M. Serone, L. Silvestrini and A. Wulzer, JHEP [**0402**]{} (2004) 049 \[arXiv:hep-th/0312267\]; A. Wulzer, arXiv:hep-th/0405168. R. Barbieri and A. Strumia, arXiv:hep-ph/0007265; G. F. Giudice, Int. J. Mod. Phys. A [**19**]{} (2004) 835 \[arXiv:hep-ph/0311344\]. S. Eidelman et al., “Review of Particle Physics”, Phys. Lett. B [**592**]{} (2004) 1. Y. Hosotani, Phys. Lett. B [**126**]{} (1983) 309; Y. Hosotani, Annals Phys.  [**190**]{} (1989) 233; N. Haba, Y. Hosotani and Y. Kawamura, Prog. Theor. Phys.  [**111**]{} (2004) 265 \[arXiv:hep-ph/0309088\]; N. Haba, M. Harada, Y. Hosotani and Y. Kawamura, Nucl. Phys. B [**657**]{} (2003) 169 \[Erratum-ibid. B [**669**]{} (2003) 381\] \[arXiv:hep-ph/0212035\]; Y. Hosotani, S. Noda and K. Takenaga, Phys. Rev. D [**69**]{} (2004) 125014 \[arXiv:hep-ph/0403106\]; Y. Hosotani, S. Noda and K. Takenaga, arXiv:hep-ph/0410193. L. J. Dixon, J. A. Harvey, C. Vafa and E. Witten, Nucl. Phys. B [**261**]{} (1985) 678; L. J. Dixon, J. A. Harvey, C. Vafa and E. Witten, Nucl. Phys. B [**274**]{} (1986) 285. C. Biggio, arXiv:hep-ph/0312209. H. Georgi, A. K. Grant and G. Hailu, Phys. Lett. B [**506**]{} (2001) 207 \[arXiv:hep-ph/0012379\]; H. Georgi, A. K. Grant and G. Hailu, Phys. Rev. D [**63**]{} (2001) 064027 \[arXiv:hep-ph/0007350\]; R. Contino, L. Pilo, R. Rattazzi and E. Trincherini, Nucl. Phys. B [**622**]{} (2002) 227 \[arXiv:hep-ph/0108102\]; G. von Gersdorff, N. Irges and M. Quiros, Nucl. Phys. B [**635**]{} (2002) 127 \[arXiv:hep-th/0204223\]. S. Groot Nibbelink, H. P. Nilles, M. Olechowski and M. G. A. Walter, Nucl. Phys. B [**665**]{} (2003) 236 \[arXiv:hep-th/0303101\]; S. Groot Nibbelink and M. Laidlaw, JHEP [**0401**]{}, 004 (2004) \[arXiv:hep-th/0311013\]; S. Groot Nibbelink and M. Laidlaw, JHEP [**0401**]{} (2004) 036 \[arXiv:hep-th/0311015\]. C. Biggio and M. Quiros, Nucl. Phys. B [**703**]{} (2004) 199 \[arXiv:hep-ph/0407348\]. See e.g. M. B. Green, J. H. Schwarz and E. Witten, “Superstring Theory. Vol. 2: Loop Amplitudes, Anomalies And Phenomenology,” Cambridge, Uk: Univ. Pr (1987) (Cambridge Monographs On Mathematical Physics). [^1]: For a review on symmetry breaking on orbifolds see for instance Ref. \[9\] and references therein.

ArXiv

--- abstract: 'The most recent observational results on the search for high redshift field ellipticals are reviewed in the context of galaxy formation scenarios. The perspectives for Large Binocular Telescope (LBT) observations are also discussed.' author: - 'Andrea Cimatti Osservatorio Astrofisico di Arcetri, Largo Fermi 5, I-50125,Firenze, Italy' title: 'Searching for high-$z$ field ellipticals: successes and problems' --- INTRODUCTION ============ The question on the formation of the present-day massive spheroidals is one of the most debated issues of galaxy evolution and it is strongly linked to the general problem of structure formation in the universe (see [@renz] for a recent review). In one scenario, massive spheroidals are formed at early cosmological epochs (e.g. $z>3$) through the “monolithic” collapse of the whole gas mass[@egg; @lar]. Such a formation would be characterized by an episode of intense star formation, followed by a passive evolution (or pure luminosity evolution, PLE) of the stellar population to nowadays. In marked contrast, the hierarchical scenarios predict that massive spheroidals are the product of rather recent merging of pre-existing disk galaxies taking place mostly at $z<1$[@k96; @bau]. In hierarchical scenarios, fully assembled massive field spheroidals at $z>1$ are rare objects[@k98], and the spheroids of cluster ellipticals were assembled before those of field ellipticals[@bau]. From an observational point of view, a direct way to test the above scenarios is to search for massive field ellipticals at $z>1$ and to compare their number with the model predictions (see the introduction of [@schade] for a recent review on observational tests). HOW TO FIND $z>1$ ELLIPTICALS ? =============================== Since the near-IR light is a good tracer of the galaxy mass[@gava; @k98], $K$-band imaging provides an important possibility to perform surveys aimed at selecting massive ellipticals at high-$z$. A galaxy with a stellar mass of about $10^{11}$ M$_{\odot}$ is expected to have $18<K<20$ for $1<z<2$[@k98], thus implying that moderately deep $K$-band surveys can efficiently select massive galaxies. A first selection criterion to find ellipticals at $z>1$ is to apply a color threshold to $K$-band selected galaxies. In the framework of passive evolution, such a threshold is set by the colors expected for a galaxy at $z>1$ formed at a given $z_{f}$. For instance, according to the Bruzual & Charlot (1999) spectral synthesis models ($Z=Z_{\odot}$, Salpeter IMF), a very red color of $R-K>5.3$ would allow to select $z\geq1$ passively evolving galaxies formed at $z_{f}>2$ ($H_0$=50 km s$^{-1}$ Mpc$^{-1}$, $\Omega_0=0.1-1.0$), thus allowing to search for elliptical candidates formed at early epochs. However, searches based on color selection criteria gave discrepant results: some works showed that the number of such red galaxies is lower compared to the predictions of PLE[@zepf; @fra; @bar], whereas others did not confirm such a deficit up to redshifts of about two[@tot; @ben; @bro; @schade; @scod]. In order to avoid the possible biases (e.g. star formation) present in the color selection technique, another approach is to derive the fraction of ellipticals by taking spectra of all the galaxies in $K$-selected samples irrespectively of colors[@cowie96; @cohen; @eisen; @cim2001] (see also [http://www.arcetri.astro.it/$\sim$k20/]{}). This method allows to overcome the putative problem of ellipticals missed because bluer than the adopted color threshold due to a low level of residual star formation[@jim]. Finally, the third possibility to find $z>1$ ellipticals is to select galaxies according to their morphology and surface brightness profiles (with or without an associated color selection criterion). This approach was adopted for example by [@schade] for $0.2<z<1.0$, and by [@fra; @ben; @bro; @menan; @treu; @morio] for $z>1$. While the results seem to agree with no or little number density evolution for early type galaxies at $z<1$, the analysis of the $z>1$ samples led again to discrepant results, thus making the question on $z>1$ ellipticals even more controversial. RECENT IMAGING AND SPECTROSCOPY RESULTS ======================================= After several small field surveys ($\sim$1-60 arcmin$^2$, see references in previous section) leading to discrepant results, the most recent success was provided by a wide field survey for extremely red objects (EROs)[@daddi1]. Such a survey (the widest so far: 700 arcmin$^2$ to $K<18.8$, with a sub-area of 447 arcmin$^2$ to $K<19.2$) provided a complete sample of about 400 objects with $R-K>5$ suitable for reliably constraining the number density of high-$z$ elliptical candidates. The main results of such a survey are the detection of strong angular clustering of EROs (an order of magnitude larger than that of field galaxies; see Fig. 1-2), and the accurate estimate of the surface density of elliptical candidates at $z>1$. The observed clustering can explain the previous discrepant results on the surface density of $z>1$ ellipticals as due to strong field-to-field variations (the “cosmic variance”), and it suggests that most EROs are ellipticals rather than dust reddened starbursts (see [@daddi1] for more details). Finally, even in the conservative case where up to 70% of EROs are [*not*]{} ellipticals, the observed surface density (complemented by the results of [@tho]) is in good agreement with the predictions of PLE (Fig. 3), suggesting that most field ellipticals were fully assembled at least by $z=2.5$[@daddi2]. This result does not imply that the formation of massive spheroidals occurred necessarily through a “monolithic collapse” scenario, but it simply constrains the epoch when the formation took place, and it implies that, if ellipticals formed through merging, this occurred mostly at $z>2.5$. [c]{} [c]{} [c]{} In addition to wide field imaging, the existence of galaxies with the colors expected in the case of passive evolution and with de Vaucouleurs $r^{1/4}$ surface brightness profiles consistent with being dynamically relaxed spheroidals at $z>1$ has significantly grown thanks to HST deep imaging (e.g [@ben; @sti; @morio]). The spectroscopic confirmation of high-$z$ ellipticals is extremely challenging because of their faintness both in the optical and in the near-IR, and because of the few characteristic spectral features present in their spectra: mainly the strong 4000 Å  continuum break, the weaker breaks at 2600 Å, 2900 Å  and 3260 Å, and a handful of absorptions detectable with the present-day largest telescopes and very long integration times (Fig. 4). Moreover, for $1.3<z<1.9$, the 4000 Å  break falls in a critical spectral region where the optical and near-IR spectrographs are less efficient and the atmosphere severely hampers the observations. Despite such difficulties, the Keck telescopes and the ESO VLT are confirming the existence of $z>1$ passively evolving ellipticals with old ages ($1-4$ Gyr) consistent with being formed at remote cosmological epochs[@spi; @liu; @soi; @cim99; @cim2001] (Fig. 4(a)), as well as ellipticals displaying a low level of star formation indicated by the possible detection of weak \[OII\]$\lambda$3727 emission (see Fig. 4(b); see also[@schade]). [c]{} PROBLEMS ======== The major problem affecting the statistical studies of $z>1$ ellipticals is the “pollution” of color-selected samples by a fraction of star forming galaxies reddened by strong dust extinction[@c98; @d99; @smail; @and; @gear]. The fraction of dusty EROs is currently unknown, but its accurate estimate is crucial to infer a reliable surface density of high-$z$ ellipticals “cleaned” by the contamination of dusty systems. This goal can be reached observing complete samples of EROs with optical+near-IR spectroscopy (when feasible), deep HST imaging and submm photometry. Although based on small and/or incomplete samples, recent submm and HST observations suggested that the dusty galaxies are probably segregated among the reddest EROs with $R-K>7$ or $I-K>6$[@c98; @d99; @smail; @gear; @morio]. The other major problem in the identification of $z>1$ ellipticals is that a large fraction of EROs are beyond the spectroscopic limits of the present largest telescopes (e.g. $R>25$, $K>19-20$). This strongly limits our ability to spectroscopically confirm the nature of a high-$z$ elliptical candidate. PROSPECTS FOR THE LBT ===================== The Large Binocular Telescope will play an important role in the study of high-$z$ ellipticals. As a single-dish telescope, deep spectroscopy in the ranges 0.7-1.0$\mu$m and 0.9-1.8$\mu$m will be possible with the MODS and the LUCIFER spectrographs respectively. This will allow us to enlarge the samples of spectroscopically identified high-$z$ ellipticals and, for instance, to infer their luminosity function and 3D clustering at $z>1$. As a diffraction limited telescope, the LBT will provide high angular resolution near-IR images that will be crucial in deriving the surface brightness profiles and, for instance, to extend the Kormendy relation and to attempt the Tolman test at $z>1$. Finally, in case of objects unfeasible with spectroscopy, the combination of high resolution imaging and optical + near-IR photometry (possibly done with [*special medium band*]{} filters) will allow us first to confirm the spheroidal nature of the elliptical candidates, and then to reliably estimate their photometric redshifts. I am grateful to Alvio Renzini for reading this manuscript and for the constructive comments, and to Shri Kulkarni for “politely” pointing out the “high quality” of the spectra. [99]{} Renzini A., Cimatti A. 2000, in “The Hy-Redshift Universe: Galaxy Formation and Evolution at High Redshift", ed. A.J. Bunker & W.J.M. van Breugel, A.S.P. Conf. Series Vol. 193, in press (astro-ph/9910162) Eggen O.J., Lynden-Bell D., Sandage A. 1962, ApJ, 136, 748 Larson R.B. 1974, MNRAS, 173, 671 Kauffmann G. 1996, MNRAS, 281, 487 Baugh C.M., Cole S., Frenk C.S. 1996, MNRAS, 283, 1361 Gavazzi G., Pierini D., Boselli A. 1996, A&A, 312, 397 Kauffmann G., Charlot S. 1998, MNRAS, 297, L23 Schade D. et al. 1999, ApJ, 525, 31 Zepf S.E. 1997, Nature, 390, 377 Franceschini A. et al. 1998, ApJ, 506, 600 Barger A.J. et al. 1999, AJ, 117, 102 Totani T., Yoshii J. 1997, ApJ, 501, L177 Benitez N. et al. 1999, ApJ, 515, L65 Broadhurst T.J., Bouwens R.J. 1999, ApJ, 530, L53 Scodeggio M., Silva D.R. 2000, A&A, 359, 953 Jimenez R. et al. 1999, MNRAS, 305, L16 Cowie L.L. et al. 1996, AJ, 112, 839 Cohen J.G. et al. 1999, ApJ, 512, 30 Eisenhardt P. et al. 2000, in “The Birth of Galaxies", Xth Rencontres de Blois, ed. B. Guiderdoni et al. (astro-ph/0002468) Cimatti et al. 2001, in preparation (see [http://www.arcetri.astro.it/$\sim$k20/]{}) Menanteau F. et al. 1999, MNRAS, 309, 208 Treu T., Stiavelli M. 2000, ApJ, 524, L27 Moriondo G., Cimatti A., Daddi E. 2000, A&A, in press Daddi E. et al. 2000, A&A, in press Thompson D. et al. 1999, ApJ, 523, 100 Daddi E., Cimatti A., Renzini A. 2000, submitted Stiavelli M. et al. 1999, A&A, 343, L25. Spinrad H. et al. 1997, ApJ, 484, 581 Liu M.C. et al. 2000, AJ, 119, 2556 Soifer B.T. 1999, AJ, 118, 2065 Cimatti A. et al. 1999, A&A, 352, L45 Cimatti A., Andreani P., Röttgering H., Tilanus R. 1998, Nature, 392, 895 Dey A. et al. 1999, ApJ, 519, 610 Smail I. et al. 1999, MNRAS 308, 1061 Andreani P., Cimatti A., Loinard L., Röttgering H.J.A. 2000, A&A, 354, L1 Gear W.K. et al. 2000, MNRAS, in press (astroph/0007054)

ArXiv

--- abstract: 'We review some examples of heterotic/type II string duality which shed light on the infrared dynamics of string compactifications with N=2 and N=1 supersymmetry in four dimensions.' --- [H]{} Ł[[L]{}]{} ¶[[P]{}]{} =msbm10 at 11pt \#1 =10000 6.0in 8.0in =-.6truein =-.2truein **Duality and $4d$ String Dynamics[^1]** Shamit Kachru [^2] *Harvard University, Cambridge, MA 02138, USA* [*Presented at the Workshop “Frontiers in Quantum Field Theory”*]{} [*in Honor of the 60th birthday of Keiji Kikkawa*]{} [*Osaka, Japan December 1995*]{} 1.0 truein =1 cm Introduction ============ 0.4cm I am very happy to have the opportunity to speak about strong/weak coupling duality on this occasion honoring the 60th birthday of Professor Keiji Kikkawa. His own foundational work on T-duality [@KY], the worldsheet analogue of S-duality, was in many ways instrumental in inspiring the recent developments in nonperturbative string theory. Strong-weak coupling dualities now allow us to determine the strong coupling dynamics of string vacua with $N \geq 4$ supersymmetry in four dimensions [@Schwarz]. It is natural to ask if this progress in our understanding of string theory can be extended to the more physical vacua with less supersymmetry. For N=2 theories in four dimensions, quantum corrections significantly modify the mathematical structure of the moduli space of vacua, as well as the physical interpretation of its apparent singularities. This was beautifully demonstrated in the field theory case in [@SeiWit] and it has more recently become possible to compute the exact quantum moduli spaces for N=2 string compactifications as well [@KV; @FHSV]. This constitutes the subject of the first part of my talk. Of course, the case of most physical interest is $N \leq 1$ theories. In the second part of my talk, I discuss examples of dual heterotic/type II string pairs where the heterotic theory is expected to exhibit nonperturbative dynamics which may fix the dilaton and break supersymmetry [@KS]. The type II dual manages to reproduce the qualitative features expected of the heterotic side at tree level. It is to be hoped that further work along similar lines will result in a better understanding of supersymmetry breaking in string theory. The first part of this talk is based on joint work with C. Vafa, and the second part of this talk is based on joint work with E. Silverstein. N=2 Gauge Theory and String Compactifications ============================================= Recall that the N=2 gauge theory with gauge group $SU(2)$ is the theory of a single N=2 vector multiplet consisting of a vector $A^{\mu}$, two Weyl fermions $\lambda$ and $\psi$, and a complex scalar field $\phi$, all in the adjoint representation of $SU(2)$. In N=1 language, this is a theory of an N=1 vector multiplet $(\lambda, A^{\mu})$ coupled to an N=1 chiral multiplet $(\phi, \psi)$. The scalar potential of the theory is determined by supersymmetry to be $$V(\phi) = {1\over g^{2}} [\phi, \phi^{+}]^{2}$$ We see that $V$ vanishes as long as we take $\phi = diag (a,-a)$, so there is a moduli space of classical vacua parameterized by the gauge invariant parameter $u = tr(\phi^{2})$. At generic points in this moduli space ${\cal M}_{v}$ of vacua, there is a massless N=2 U(1) vector multiplet $A$. The leading terms in its effective lagrangian are completely determined in terms of a single holomorphic function $F(A)$, the prepotential: $$L \sim \int d^{4}\theta {\partial F \over {\partial A}}\bar A + \int d^{2}\theta {\partial^{2} F \over{\partial A^{2}}}W_{\alpha}W^{\alpha} + c.c.$$ The first term determines, in N=1 language, the Kahler potential (and hence the metric on ${\cal M}_{v}$) while the second term determines the gauge coupling as a function of moduli. In [@SeiWit] the exact form of $F$ including instanton corrections was determined. In addition, the masses of all of the BPS saturated particles were computed. This was reviewed in great detail in several other talks at this conference, so I will not repeat the solution here. It will suffice to say that the crucial insight is that the singular point $u=tr(\phi^{2}) = 0$ where $SU(2)$ gauge symmetry is restored in the classical theory splits, in the quantum theory, into two singular points $u = \pm \Lambda^{2}$, where a monopole and a dyon become massless. In this talk our interest is not really in $N=2$ gauge theories but in the string theories which reduce to $N=2$ gauge theories in the infrared. There are two particularly simple classes of $d=4, N=2$ supersymmetric string compactifications. One obtains such theories from Type II (A or B) strings on Calabi-Yau manifolds, and from heterotic strings on $K_{3} \times T^{2}$ (with appropriate choices of instantons on the $K_3$). Here we briefly summarize some basic properties of these theories. Type IIA strings on a Calabi-Yau threefold $M$ give rise to a four-dimensional effective theory with $n_v$ vector multiplets and $n_{h}$ hypermultiplets where $$n_{v} = h^{1,1}(M), ~~n_{h} = h^{2,1}(M) + 1$$ The $+1$ in $n_{h}$ corresponds to the fact that for such type II string compactifications, the $\it dilaton$ is in a hypermultiplet. The vector fields in such a theory are Ramond-Ramond U(1)s, so there are no charged states in the perturbative string spectrum. Furthermore, because of the theorem of de Wit, Lauwers, and Van Proeyen [@dLVP] which forbids couplings of vector multiplets to neutral hypermultiplets in N=2 effective lagrangians, the dilaton does not couple to the vector moduli. This means that there are no perturbative or nonperturbative corrections to the moduli space of vector multiplets. On the other hand the moduli spaces of hypermultiplets are expected to receive highly nontrivial corrections, including “stringy” corrections with $e^{-1/g}$ strength [@BBS]. One interesting feature of the moduli spaces of vector multiplets in such theories is the existence of conifold points at finite distance in the moduli space. At such points the low energy effective theory becomes singular (e.g., the prepotential develops a logarithmic singularity) [@CDGP]. This phenomenon is reminiscent of the singularities in the prepotential which occur at the “massless monopole” points in the Seiberg-Witten solution of N=2 gauge theory, singularities which are only present because one has integrated out a charged field which is becoming massless. In the case at hand, in fact, one can show that there are BPS saturated states (obtained by wrapping 2-branes around collapsing 2-cycles) which become massless and which are charged under (some of) the Ramond-Ramond $U(1)$s [@Strominger]. These explain the singularity in the prepotential. In fact at special such points, where enough charged fields (charged under few enough $U(1)$s) become massless, one can give them VEVs consistent with D and F flatness. This results in new “Higgs branches” of the moduli space. These new branches correspond to string compactifications on different Calabi-Yau manifolds, topologically distinct from $M$ [@GMS], and there is evidence that all Calabi-Yau compactifications may be connected in this manner [@Cornell; @Texas]. The other simple way of obtaining an N=2 theory in four dimensions from string theory is to compactify the heterotic string (say $E_8\times E_8$) on $K_3 \times T^2$. Because of the Bianchi identity $$dH = Tr(R\wedge R) - Tr(F\wedge F)$$ one must embed 24 instantons in the $E_8 \times E_8$ in order to obtain a consistent theory. An $SU(N)$ k-instanton on $K_3$ comes with $Nk + 1 - N^2$ hypermultiplet moduli (where $k\geq 2N$), and $K_3$ comes with 20 hypermultiplet moduli which determine its size and shape. Embedding an $SU(N)$ instanton in $E_8$ breaks the observable low energy gauge group to the maximal subgroup of $E_8$ which commutes with $SU(N)$ ($E_7$ for N=2, $E_6$ for N=3, and so forth). In addition, there are three $U(1)$ vector multiplets associated with the $T^2$. Their scalar components are the dilaton $S$ and the complex and kahler moduli $\tau$ and $\rho$ of the torus (both of which live on the upper half-plane $H$ mod $SL(2,Z)$). At special points in the moduli space the $U(1)^2$ associated with $\tau$ and $\rho$ is enhanced to a nonabelian gauge group: $$\tau = \rho \rightarrow SU(2)\times U(1),~~ \tau=\rho=i \rightarrow SU(2)^2, ~~\tau = \rho = 1/2 + i{\sqrt 3}/2 \rightarrow SU(3)$$ Because the dilaton lives in a vector multiplet in such compactifications, the moduli space of vectors is modified by quantum effects. On the other hand, the moduli space of hypermultiplets receives neither perturbative nor nonperturbative corrections. An interesting feature of the heterotic ${\cal M}_{v}$ is the existence of special points where the classical theory exhibits an enhanced gauge symmetry (as described above for the compactification on $T^2$). Sometimes by appropriate passage to a Higgs or Coulomb phase, such enhanced gauge symmetry points link moduli spaces of N=2 heterotic theories which have different generic spectra (for some examples see [@KV; @AFIQ]). It is natural to conjecture that such transitions connect all heterotic N=2 models, in much the same way that conifold transitions connect Calabi-Yau compactifications of type II strings. N=2 String-String Duality ========================= &gt;From the brief description of heterotic and type II N=2 vacua in the previous section, it is clear that a duality relating the two classes of theories would be extremely powerful. If one were to find a model with dual descriptions as a compactification of the Type IIA string on $M$ and the heterotic string on $K_3 \times T^{2}$, one could compute the exact prepotential for ${\cal M}_{v}$ from the Type IIA side (summing up what from the heterotic perspective would be an infinite series of instanton corrections). Similarly, one would get exact results for ${\cal M}_{h}$ from the heterotic side – this would effectively compute the $e^{-1/g}$ corrections expected from the IIA perspective. In fact, such a duality has been found to occur in several examples in [@KV; @FHSV]. One of the simplest examples is as follows. Consider the heterotic string compactified to eight dimensions on $T^{2}$ with $\tau = \rho$. Further compactify on a $K_{3}$, satisfying the Bianchi identity for the $H$ field by embedding $c_{2}=10$ $SU(2)$ instantons in each $E_8$ and a $c_{2}=4$ $SU(2)$ instanton into the “enhanced” $SU(2)$ arising from the $\tau=\rho$ torus. After Higgsing the remaining $E_7$ gauge groups one is left with a generic spectrum of 129 hypermultiplets and 2 vector multiplets. The 2 vectors are $\tau$ and the dilaton $S$ – when $\tau = i$, one expects an $SU(2)$ gauge symmetry to appear (the other $SU(2)$ factor that would normally appear there has been broken in the compactification process). This tells us that if there is a type IIA dual compactification on a Calabi-Yau $M$, then the Betti numbers of $M$ must be $$h_{11}(M) = 2, ~~h_{21}(M) = 128$$ There is a known candidate manifold with these Betti numbers – the degree 12 hypersurface in $WP^{4}_{1,1,2,2,6}$ defined by the vanishing of $p$ $$p = z_{1}^{12} + z_{2}^{12} + z_{3}^{6} + z_{4}^{6} + z_{5}^{2} + ....$$ This manifold has in fact been studied intensively as a simple example of mirror symmetry in [@Hosono; @Morrison]. The mirror manifold $W$ has $h_{11}(W) = 128, h_{21}(W) = 2$. The conjecture that IIA on $M$ is equivalent to the heterotic string described above implies that IIB on $W$ is also equivalent to that heterotic string. The structure of the moduli space of vector multiplets of the heterotic string should be $\it exactly$ given by the classical (in both sigma model and string perturbation theory) moduli space of complex structures of $W$. The mirror manifold can be obtained by orbifolding $p=0$ by the maximal group of phase symmetries which preserves the holomorphic three-form [@GP]. Then the two vector moduli are represented by $\psi$ and $\phi$ in the polynomial $$p = z_{1}^{12} + z_{2}^{12} + z_{3}^{6} + z_{4}^{6} + z_{5}^{2} - 12 \psi z_{1}z_{2}z_{3}z_{4}z_{5} - 2\phi z_{1}^{6} z_{2}^{6}$$ It is also useful, following [@Hosono], to introduce $$x = {-1\over 864} {\phi\over \psi^{6}}, ~~y = {1\over \phi^{2}}$$ These are the convenient “large complex structure” coordinates on the moduli space of vector multiplets for the IIB string. In order to test our duality conjecture, we should start by checking that the IIB string reproduces some qualitative features that we expect of the heterotic ${\cal M}_{v}$. For example, $\tau = i$ for weak coupling $S\rightarrow \infty$ is an $SU(2)$ point. There should therefore be a singularity of ${\cal M}_{v}$ at this point which splits, as one turns on the string coupling, to $\it two$ singular points (where monopoles/dyons become massless), as in the case of pure $SU(2)$ gauge theory. The “discriminant locus” where the IIB model becomes singular is given by $$(1-x)^{2} - x^{2} y = 0$$ So we see that as a function of $y$ for $y \neq 0$ there are two solutions for $x$ and as $y \rightarrow 0$ they merge to a single singular point $x=1$. This encourages us to identify $x=1, y=0$ with $\tau =i, S \rightarrow \infty$ of the heterotic string – the $SU(2)$ point. The metric on the moduli space for $y$ at $y=0$ and $S$ at weak coupling also agree if one makes the identification $y\sim e^{-S}$. There is also a remarkable observation in [@Morrison] that the mirror map, restricted to $y=0$, is given by $$x = {j(i)\over j(\tau_{1})}$$ where $\tau_{1}$ is one of the coordinates on the Kahler cone of $M$. Here $j$ is the elliptic j-function mapping $C$ onto $H/SL(2,Z)$. This tells us that the classical heterotic $\tau$ moduli space, which is precisely $H/SL(2,Z)$, is embedded in the moduli space of $M$ at weak coupling precisely as expected from duality. In fact using the uniqueness of special coordinates up to rotations, one can find the exact formula expressing the IIB coordinates $(x,y)$ in terms of the heterotic coordinates $(\tau,S)$. Of course with this map in hand there are now several additional things one can check. The tests which have been performed in [@KV; @KLT; @AGNT; @KKLMV] include 1\) A matching of the expected loop corrections to the heterotic prepotential with the form of the tree-level exact Calabi-Yau prepotential. 2\) A test that the g-loop F-terms computed by the topological partition functions $F_g$ on the type II side (which include e.g. $R^{2}$ and other higher derivative terms) are reproduced by appropriate (one-loop!) computations on the heterotic side. 3\) A demonstration that in an appropriate double-scaling limit, approaching the $\tau = i$, $S \rightarrow \infty$ point of the heterotic string while taking $\alpha^{'} \rightarrow 0$, the IIB prepotential reproduces the exact prepotential of $SU(2)$ gauge theory (including Yang-Mills instanton effects) computed in [@SeiWit]. These tests give very strong evidence in favor of the conjectured duality. Given its veracity, what new physics does the duality bring into reach? $\bullet$ One now has examples of four-dimensional theories with exactly computable quantum gravity corrections. In the example discussed above, the Seiberg-Witten prepotential which one finds in an expansion about $\tau = i, S \rightarrow \infty$ receives gravitational corrections which are precisely computable as a power series in $\alpha^{'}$. $\bullet$ On a more conceptual level, the approximate duality of [@SeiWit] between a microscopic $SU(2)$ theory (at certain points in its moduli space) and a $U(1)$ monopole/dyon theory is promoted to an $\it exact$ duality, valid at all wavelengths, between heterotic and type II strings. $\bullet$ There is evidence that at strong heterotic coupling, new gauge bosons and charged matter fields appear, sometimes giving rise to new branches of the moduli space [@KMP; @KM]. $\bullet$ The $e^{-1/g}$ corrections to the hypermultiplet moduli space of type II strings are in principle exactly computable using duality (and may be of some mathematical interest). One might wonder what is special about the Calabi-Yau manifolds which are dual to weakly coupled heterotic strings. In fact it was soon realized that the examples of duality in [@KV] involve Calabi-Yau manifolds which are $K_3$ fibrations [@KLM]. That is, locally the Calabi-Yau looks like $CP^{1}\times K_{3}$. In fact, one can prove that if the type IIA string on a Calabi-Yau $M$ (at large radius) is dual to a weakly coupled heterotic string, then $M$ must be a $K_3$ fibration [@AL]. To make this more concrete, in the example of the previous section, we saw $M$ was defined by the vanishing of $$p = z_{1}^{12} + z_{2}^{12} + z_{3}^{6} + z_{4}^{6} + z_{5}^{2} + ...$$ in $WP^{4}_{1,1,2,2,6}$. Set $z_{1}=\lambda z_{2}$ and define $y=z_{1}^{2}$ (which is an allowed change of variables since an identification on the $WP^{4}$ takes $z_{1} \rightarrow -z_{1}$ without acting on $z_{3,4,5}$). Then the polynomial becomes (after suitably rescaling to absorb $\lambda$) $$p = y^{6} + z_{3}^{6} + z_{4}^{6} + z_{5}^{2} + ...$$ which defines a $K_{3}$ surfaces in $WCP^{3}_{1,1,1,3}$. The choice of $\lambda$ in $z_{1}=\lambda z_{2}$ is a point on $CP^{1}$, and the $K_{3}$ for fixed choice of $\lambda$ is the fiber. It is not surprising that $K_3$ fibrations play a special role in 4d N=2 heterotic/type II duality. Indeed the most famous example of heterotic/type II duality is the 6d duality between heterotic strings on $T^{4}$ and type IIA strings on $K_{3}$ [@HT; @Witten]. If one compactifies the type IIA string on a CY threefold which is a $K_3$ fibration, and simultaneously compactifies the heterotic string on a $K_{3}\times T^{2}$ where the $K_3$ is an elliptic fibration, then locally one can imagine taking the bases of both fibrations to be large and obtaining in six dimensions an example of the well-understood 6d string-string duality [@VW]. This picture is not quite precise because of the singularities in the $K_3$ fibration, but it does provide an intuitive understanding of the special role of $K_3$ fibrations. N=1 Duality and Gaugino Condensation ==================================== Starting with an $N=2$ dual pair of the sort discussed above, one can try to obtain an $N=1$ dual pair by orbifolding both sides by freely acting symmetries. This strategy was used in [@VW; @HLS] where several examples with trivial infrared dynamics were obtained. Here we will find that examples with highly nontrivial infrared dynamics can also be constructed [@KS]. Our starting point is an N=2 dual pair (IIA on a Calabi-Yau $M$ and heterotic on $K_{3}\times T^{2}$) where the heterotic gauge group takes the form $$G ~=~E_{8}^{H} \otimes E_{7}^{obs}\otimes ...$$ $H$ denotes the hidden sector and $obs$ the observable sector. We will first discuss the technical details of the $Z_2$ symmetry by which we can orbifold both sides to obtain an $N=1$ dual pair, and then we discuss the physics of the duality. Orbifold the heterotic side by the Enriques involution acting on $K_3$ and a total reflection on the $T^{2}$. This acts on the base of the elliptic fibration $(z_{1},z_{2})$ by $$(z_{1}, z_{2}) ~ \rightarrow ~ (\bar z_{2}, - \bar z_{1})$$ taking $CP^{1} \rightarrow RP^{2}$. In addition, we need to choose a lifting of the orbifold group to the gauge degrees of freedom. We do this as follows: $\bullet$ Put a modular invariant embedding into the “observable” part of the gauge group alone. $\bullet$ Embed the translations which generate the $T^2$ into $E_{8}^{H}$, constrained by maintaining level-matching and the relations of the space group. For example one could take Wilson lines $A_{1,2}$ along the $a$ and $b$ cycle of the $T^{2}$ given by $$A_{1} = {1\over 2} (0,0,0,0,1,1,1,1),~~~A_{2} = {1\over 2}(-2,0,0,0,0,0,0,0)$$ Here $A_{1,2} = {1\over 2}L_{1,2}$ where $L_{1,2}$ are vectors in the $E_8$ root lattice. These Wilson lines break the $E_{8}^{H}$ gauge symmetry to $SO(8)_{1} \otimes SO(8)_{2}$. How does the $Z_2$ map over to the type II side? &gt;From the action $$(z_{1},z_{2}) ~\rightarrow ~(\bar z_{2}, -\bar z_{1})$$ on the $CP^{1}$ base (which is common to both the heterotic and type II sides), we infer that the $Z_2$ must be an antiholomorphic, orientation-reversing symmetry of the Calabi-Yau manifold $M$. To make this a symmetry of the type IIA string theory, we must simultaneously flip the worldsheet orientation, giving us an “orientifold.” In such a string theory, one only includes maps $\Phi$ of the worldsheet $\Sigma$ to spacetime $M/Z_{2}$ if they satisfy $$\Phi^{*}(w_{1}(M/Z_{2})) = w_{1}(\Sigma)$$ where $w_1$ is the first Stieffel-Whitney class. We know from 6d string-string duality that the Narain lattice $\Gamma^{20,4}$ of heterotic string compactification on $T^4$ maps to the integral cohomology lattice of the dual $K_{3}$. This means that we can infer from the action of the $Z_2$ on the heterotic gauge degrees of freedom, what the action of the $Z_2$ must be on the integral cohomology of the $K_3$ fiber on the IIA side. Since we are frozen on the heterotic side at a point with $SO(8)^{2}$ gauge symmetry in the hidden sector, the dual $K_3$ must be frozen at its singular enhanced gauge symmetry locus. The $K_3$ dual to heterotic enhanced gauge symmetry $G$ has rational curves $C_i$, $i = 1,...,rank(G)$ shrinking to zero area (with the associated $\theta_{i}=0$ too). It is easy to see, e.g. from Witten’s gauged linear sigma model that in this situation the type II theory indeed exhibits an extra $Z_2$ symmetry. The bosonic potential of the relevant gauged linear sigma model (for the case of a single shrinking curve) is given by $$V= {1\over{2e^2}} \sum_i\biggl\{\biggl(\bigl[\sum_\alpha Q_i^\alpha(|\phi^i_\alpha|^2 -|\tilde\phi^i_\alpha|^2)\bigr]-r_i^0\biggr)^2$$ $$+\biggl(Re(\sum_\alpha\phi^i_\alpha\tilde\phi^i_\alpha)-r_i^1\biggr)^2 +\biggl(Im(\sum_\alpha\phi^i_\alpha\tilde\phi^i_\alpha)-r_i^2\biggr)^2 \biggr\}$$ $$+{1\over 2}\sum_i\bigl[\sum_\alpha Q^{\alpha~2}_i (|\phi^i_\alpha|^2+|\tilde\phi^i_\alpha|^2)\bigr]|\sigma_i|^2$$ Here the $\phi$s represent the $K_3$ coordinates while $r$ parametrizes the size of the curve and $\sigma$ is the Kahler modulus. Precisely when $\vec r \rightarrow 0$, the model has the $Z_2$ symmetry $\phi \rightarrow -\tilde \phi$, $\sigma \rightarrow - \sigma$. Orbifolding by this $Z_2$ then freezes the $K_3$ at its enhanced gauge symmetry locus, as expected. What is the physics of the dual pairs that one constructs in this manner? In the heterotic string, when there is a hidden sector pure gauge group $$G^{hidden} = \Pi ~G^{b}$$ one expects gaugino condensation to occur. This induces an effective superpotential $$W = \sum ~h_{b}~ \Lambda_{b}^{3}(S)$$ where $\Lambda_{b}(S) \sim e^{-\alpha_{b} S}$ and $\alpha_{b}$ is related to the beta function for the running $G_b$ coupling. It was realized early on [@Krasnikov; @DKLP] that in such models (with more than one hidden factor) one might expect both stabilization of the dilaton and supersymmetry breaking. It has remained a formidable problem to determine which (if any) such models actually do have a stable minimum at weak coupling with broken supersymmetry. For now, we will be content to simply understand how the $\it qualitative$ structure of the heterotic theory (e.g. the gaugino-condensation induced effective superpotential) is reproduced by the type II side. This is mysterious because the type II N=2 theory we orientifolded had only abelian gauge fields in its spectrum, so we need to reproduce the strongly coupled nonabelian dynamics of the heterotic string with an $\it abelian$ gauge theory on the type II side. The heterotic orbifold indicates the spectrum of the string theory as $g_{het} \rightarrow 0$. The heterotic dilaton $S$ maps to the radius $R$ of the $RP^{2}$ base of the type II orientifold (recall one obtains the $RP^2$ by orbifolding the base $P^1$ of the $K_3$ fibration) $$S_{het} \leftrightarrow R_{RP^{2}}$$ The purported stable vacuum of the heterotic theory should then be expected to lie at large radius for the base, and on the (orientifold of the) conifold locus dual to enhanced gauge symmetry. There are two crucial features of this locus: 1\) The $RP^2$ base has $\pi_{1}(RP^{2}) = Z_{2}$. So a state projected out in orientifolding the N=2 theory will have a massive version invariant under the $Z_2$. Say $\beta \in \pi_{1}(RP^{2})$ is the nontrivial element. Take $x$ a coordinate along an appropriate representative of $\beta$ – a representative can be obtained by taking the image of a great circle on the original base $P^1$ after orientifolding. Then if the original non-invariant vertex operator was $V$, a new invariant vertex operator is given adiabatically by $$V^{\prime} = e^{ix \over R} V$$ The $Z_2$ takes $x$ to $x + \pi R$ and therefore $V^{\prime}$ is invariant if $V$ was not. In particular this gives us massive versions of the scalars $a^{i}_{b,D}$ in the N=2 vector multiplets for $G^{b}$ with masses $$M_{a} \sim {1\over R^{2}}$$ Effectively, for very large $R$, one is restoring the original N=2 supersymmetry. 2\) The low energy theory for IIA at the conifold locus contains massless $\it solitonic$ states [@Strominger]. One can see that they survive the N=2 $\rightarrow$ N=1 orientifolding by examining the behavior of the gauge couplings [@VW]. These extra solitonic states play the role of the “monopole hypermultiplets” $M_{i}^{b}, \tilde M_{i}^{b}$ of the N=2 theory. These two facts taken together imply that as $R \rightarrow \infty$ there is an effective superpotential $$W_{II} = \sum_{b} \left( m_{b}u^{b}_{2}(a^{i}_{b,D},R) + \sum_{i=1}^{rank(b)} M_{i}^{b} a^{i}_{b,D}\tilde M^{b}_{i} \right)$$ where $u^{b}_{2}$ is the precise analogue of $u$ of §2 for $G^b$ and its functional dependence on $R$ can be found from the $N=2$ dual pair. As we’ll now discuss, this structure a\) Allows us to reproduce the gaugino-condensation induced effective superpotential of the heterotic side. b\) Implies $\langle M \rangle \neq 0$, suggesting a geometrical description of the type II side by analogy with N=2 conifold transitions. To see a), recall how the physics of N=1 $SU(2)$ gauge theory is recovered from the N=2 theory in [@SeiWit]. One can obtain the N=1 theory by giving a bare mass to the adjoint scalar in the N=2 vector multiplet and integrating it out. In the vicinity of the monopole points this means there is an effective superpotential $$W = m u(a_{D}) + \sqrt{2} a_{D} M \tilde M$$ Using the equations of motion and D-flatness, one finds $$\vert \langle M \rangle \vert = \vert \langle \tilde M \rangle \vert = ( -mu^{\prime}(0)/\sqrt{2})^{1/2},~~a_{D} = 0$$ The monopoles condense and given a mass to the (dual) $U(1)$ gauge field by the Higgs mechanism, leaving a mass gap. Two vacua arise in this way – one at each of the monopole/dyon points – in agreement with the Witten index computation for pure $SU(2)$ gauge theory. In our case, we expect that condensation of the massless solitons will lead to the gaugino condensation induced superpotential (and, perhaps, supersymmetry breaking). To see this we must expand $W_{II}$ in $a^{i}_{b,D}$ in anticipation of finding a minimum at small $a_{D}$ (since the minimum was at $a_{D} = 0$ in the global case). Using $$u_{2}(a^{i}_{b,D},R)) = e^{i\gamma^{b}}\lambda_{b}^{2}(R) + \cdots$$ (where $\gamma^{b}$ comes from the phase of the gaugino condensate) as well as the matching condition $$m_{b}\Lambda_{b,high}^{2} = \Lambda_{b,low}^{3}$$ one obtains from integrating out the massive adjoint scalar, one sees $$W_{II} = \sum_{b} e^{i\gamma^{b}}\Lambda_{b}^{3}(S) + \cdots$$ Simply minimizing the supergravity scalar potential $$V = e^{K}(D_{i}W G^{i\bar j}D_{\bar j}W - 3 \vert W \vert^{2}) + {1\over 2}g^{2}D^{2}$$ we also find that $$\langle M_{i}^{b} \tilde M_{i}^{b} \rangle = -h_{b}m_{b}u_{2,i}^{b}(S) - K_{i} W$$ That is, the “wrapped two-branes” which give us the massless monopoles have condensed, in accord with the global result. So integrating out the massive $M, \tilde M$ and adjoint scalar degrees of freedom yields the same form of bosonic potential that we expect from gaugino condensation on the heterotic side. In summary, we have argued that the type II dual description of the effects of gaugino condensation involves a mass perturbation breaking N=2 supersymmetry. One cannot add mass terms by hand in string theory: The type II orientifold produces the requisite massive mode as a Kaluza-Klein excitation of the original N=2 degrees of freedom that were projected out. One intriguing feature of the IIA vacuum is the nonzero VEVs for the wrapped two-branes $M, \tilde M$. In the N=2 context $\langle M \rangle \neq 0 \rightarrow$ conifold transition. There is a well known geometrical description of the conifold points. For example, in the IIB theory the conifold in vector multiplet moduli space is obtained by going to a point in ${\cal M}_{complex}$ where there is a cone over $S^{3}\times S^2$ in the Calabi-Yau. One can either “deform the complex structure” (return to the Coulomb phase, in physics language) by deforming the tip of the cone into an $S^3$, or one can do a “small resolution” and blow the tip of the cone into an $S^2$. The latter corresponds to moving to a new Higgs phase, in the N=2 examples [@GMS]. It was noted long ago by Candelas, De La Ossa, Green, and Parkes [@CDGP] that at a $\it generic$ conifold singularity such a small resolution does not produce a Kahler manifold. They speculated that such nonKahler resolutions might correspond to supersymmetry breaking directions. It is natural to suggest that we might be seeing a realization of that idea by duality. The analogy with N=2 conifold transitions suggests that $\langle M \rangle \neq 0 \rightarrow$ nonKahler resolution. One can hope that this will provide a useful dual view of supersymmetry breaking in string theory. [99]{} K. Kikkawa and M. Yamasaki, Phys. Lett. [**B149**]{} (1984) 357. J. Schwarz, talk presented at this conference and references therein, hep-th/9601077. N. Seiberg and E. Witten, Nucl. Phys. [**B426**]{} (1994) 19. S. Kachru and C. Vafa, Nucl. Phys. [**B450**]{} (1995) 69. S. Ferrara, J. Harvey, A. Strominger, and C. Vafa, Phys. Lett. [**B361**]{} (1995) 59. S. Kachru and E. Silverstein, hep-th/9511228. B. de Wit, P. Lauwers, and A. Van Proeyen, Nucl. Phys. [**B255**]{} (1985) 569. K. Becker, M. Becker, and A. Strominger, Nucl. Phys. [**B456**]{} (1995) 130. P. Candelas, X. De La Ossa, P. Green, and L. Parkes, Nucl. Phys. [**B359**]{} (1991) 21. A. Strominger, Nucl. Phys. [**B451**]{} (1995) 96. B. Greene, D. Morrison, and A. Strominger, Nucl. Phys. [**B451**]{} (1995) 109. T. Chiang, B. Greene, M. Gross, and Y. Kanter, hep-th/9511204. A. Avram, P. Candelas, D. Jancic, and M. Mandelberg, hep-th/9511230. G. Aldazabal, A. Font, L. Ibanez, and F. Quevedo, hep-th/9510093. S. Hosono, A. Klemm, S. Theisen, and S.T. Yau, Commun. Math. Phys. [**167**]{} (1995) 301. P. Candelas, X. De La Ossa, A. Font, S. Katz, and D. Morrison, Nucl. Phys. [**B416**]{} (1994) 481. B. Greene and M. Plesser, Nucl. Phys. [**B338**]{} (1990) 15. V. Kaplunovsky, J. Louis, and S. Theisen, Phys. Lett. [**B357**]{} (1995) 71. I. Antoniadis, E. Gava, K. Narain, and T. Taylor, Nucl. Phys. [**B455**]{} (1995) 109. S. Kachru, A. Klemm, W. Lerche, P. Mayr and C. Vafa, hep-th/9508155. S. Katz, D. Morrison, and R. Plesser, hep-th/9601108. A. Klemm and P. Mayr, hep-th/9601014. A. Klemm, W. Lerche, and P. Mayr, Phys. Lett. [**B357**]{} (1995) 322. P. Aspinwall and J. Louis, hep-th/9510234. C. Hull and P. Townsend, Nucl. Phys. [**B438**]{} (1995) 109. E. Witten, Nucl. Phys. [**B443**]{} (1995) 85. C. Vafa and E. Witten, hep-th/9507050. J. Harvey, D. Lowe, and A. Strominger, Phys. Lett. [**B362**]{} (1995) 65. N. Krasnikov, Phys. Lett. [**B193**]{} (1987) 37. L. Dixon, “Supersymmetry Breaking in String Theory,” invited talk presented at 15th APS DPF general meeting, Houston, TX, and references therein. [^1]: Research supported in part by the Harvard Society of Fellows. [^2]: Email: kachru@string.harvard.edu

ArXiv

--- author: - Konstantinos Sotiropoulos - 'John W. Byers' - Polyvios Pratikakis - 'Charalampos E. Tsourakakis' bibliography: - 'bibliography.bib' title: | TwitterMancer:\ Predicting Interactions on Twitter Accurately ---

ArXiv

--- abstract: 'In this work we define a relative version of the flux homomorphism, introduced by Calabi in 1969, for a symplectic manifold. We use it to study (the universal cover of) the group of symplectomorphisms of a symplectic manifold leaving a Lagrangian submanifold invariant. We also show that some quotients of the universal covering of the group of symplectomorphisms are stable under symplectic reduction.' address: 'Department of Mathematics, Middle East Technical University, 06531 Ankara, TURKEY' author: - 'Y[i]{}ld[i]{}ray Ozan' title: Relative Flux Homomorphism in Symplectic Geometry --- Introduction ============ In 1969 Calabi introduced the flux homomorphism for a symplectic manifold $(M,\omega)$ ([@Cal]). Let $\textmd{Symp}_0(M,\omega)$ denote the path component of the Lie group of symplectomorphisms of $(M,\omega)$ containing the identity. Also, let $\widetilde{\textmd{Symp}}_0(M,\omega)$ denote the universal covering space of $\textmd{Symp}_0(M,\omega)$. Note that an element of $\widetilde{\textmd{Symp}}_0(M,\omega)$ is represented by a homotopy class of a smooth path $\psi_t$ in $\textmd{Symp}_0(M,\omega)$ connecting the identity $\psi_0=id$ to $\psi=\psi_1$, where the homotopies fix the end points of the paths. We will denote the homotopy class of $\psi_t$ by $\{ \psi_t \} \in \widetilde{\textmd{Symp}}_0(M,\omega)$. Any such $\psi_t$ determines a smooth family of vector fields $X_t:M\rightarrow T_*M$ defined by the formula $$\frac{d}{dt}\psi_t=X_t\circ \psi_t.$$ Now the flux homomorphism $\textmd{Flux}:\widetilde{\textmd{Symp}}_0(M,\omega)\rightarrow H^1(M,{\mathbb R})$ is defined, for any $\{\psi_t\} \in \widetilde{\textmd{Symp}}_0(M,\omega)$, as $$\textmd{Flux}(\{ \psi_t\})=\int_0^1[\imath(X_t)\omega] \ dt.$$ For connected $M$ identifying $H^1(M,{\mathbb R})$ with $\textmd{Hom}(\pi_1(M),{\mathbb R})$ the cohomology class $\textmd{Flux}(\{ \psi_t\})$ corresponds to the homomorphism $\pi_1(M)\rightarrow {\mathbb R}$ defined by $$\gamma \mapsto \int_0^1 \int_0^1 \omega(X_t(\gamma(s), \dot{\gamma}(s))) \ ds \ dt$$ for any smooth loop $\gamma :S^1\rightarrow M$. In the next section, we will define a relative flux homomorphism in the presence of a Lagrangian submanifold $L\subseteq M$, mainly following both the treatment and the notation of Section $10$ of [@McS]. Most of the results and proofs of this section are analogous to those of [@McS] and we will omit the proofs of some of them unless the proof includes some new ingredients or makes the text more comprehensible. In the third section, we will show that some quotients of $\widetilde{\textmd{Symp}}_0(M,\omega)$ are stable under symplectic reduction. Relative Flux Homomorphism ========================== Throughout this section, unless stated otherwise, $(M,\omega)$ is a connected closed symplectic manifold and $L$ a closed Lagrangian submanifold. Let $\textmd{Symp}(M,L,\omega)$ denote the subgroup of $\textmd{Symp}(M,\omega)$ consisting of symplectomorphisms leaving the Lagrangian submanifold $L$ invariant and $\textmd{Symp}_0(M,L,\omega)$ the path component of $\textmd{Symp}(M,L,\omega)$ containing the identity. \[Rem-rel\] In general, $\textmd{Symp}(M,L,\omega)\cap\textmd{Symp}_0(M,\omega)$ is not path connected. Let $S^2$ be the unit sphere in ${\mathbb R}^3$ equipped with the standard symplectic form and $S^1$ be the intersection of $S^2$ with the $xy-$plane. Then $S^1$ is a lagrangian submanifold $S^2$. Let $\psi_t$, $t\in [0,1]$, denote the rotation of ${\mathbb R}^3$ about the $x-$axis $t\pi$ radians and $\psi=\psi_1$. Since $\psi_{|S^1}$ is orientation reversing $\psi$ is not in $\textmd{Symp}_0(M,L,\omega)$, even though it lies in $\textmd{Symp}(M,L,\omega)\cap\textmd{Symp}_0(M,\omega)$ trivially. Hence, $\textmd{Symp}(M,L,\omega)\cap\textmd{Symp}_0(M,\omega)$ is not path connected. Identifying the homotopy class of the constant loop at the identity element of the group ${\textmd{Symp}}_0(M,L,\omega)$ with that of ${\textmd{Symp}}_0(M,\omega)$, we obtain a canonical map from $\widetilde{\textmd{Symp}}_0(M,L,\omega)$ onto a subgroup of $\widetilde{\textmd{Symp}}_0(M,\omega)$, the subgroup of homotopy classes $\{ \psi_t \}$ in $\widetilde{\textmd{Symp}}_0(M,\omega)$, where each $\psi_t$ leaves $L$ invariant. Then we have the following result. \[lem-defn\] The flux homomorphism is well defined on $\widetilde{\textmd{\em Symp}}_0(M,L,\omega)$ and takes values in the kernel of the restriction map $H^1(M,{\mathbb R})\rightarrow H^1(L,{\mathbb R})$. If $L$ is connected the flux homomorphism takes values in $H^1(M,L,{\mathbb R})$. Moreover, in this case the flux homomorphism $$\textmd{\em Flux}:\widetilde{\textmd{\em Symp}}_0(M,L,\omega)\rightarrow H^1(M,L,{\mathbb R})$$ is onto. The Flux homomorphism is well defined on $\widetilde{\textmd{Symp}}_0(M,\omega)$ since homotopic loops in $\textmd{Symp}_0(M,\omega)$ have the same value under the flux homomorphism (cf. see Lemma $10.7$ of [@McS]). Therefore, homotopic loops in $\textmd{Symp}_0(M,L,\omega)$ have the same value and hence the flux homomorphism is well defined on $\widetilde{\textmd{Symp}}_0(M,L,\omega)$. Let $\{\psi_t\} \in \widetilde{\textmd{Symp}}_0(M,L,\omega)$ and $X_t$ the vector field defined by $$\displaystyle \frac{d}{dt}\psi_t=X_t\circ \psi_t.$$ Since $\psi_t$ leaves $L$ invariant, for any $p\in L$ we have $X_t(p)\in T_pL$. Note that to prove the first assertion it suffices to show that for any smooth loop $\gamma :S^1\rightarrow L$ the integral $$\displaystyle \int_0^1 \int_0^1 \omega(X_t(\gamma(s), \dot{\gamma}(s))) \ ds \ dt=0.$$ However, this trivially holds since both $X_t(\gamma(s))$ and $\dot{\gamma}(s)$ lie in $T_{\gamma(s)}L$, and $L$ is a Lagrangian submanifold. Now suppose that $L$ is connected. Then the exact sequence $$\rightarrow H^0(M,{\mathbb R})\stackrel{\simeq}{\rightarrow} H^0(L,{\mathbb R})\stackrel{0}{\rightarrow} H^1(M,L,{\mathbb R})\rightarrow H^1(M,{\mathbb R})\rightarrow H^1(L,{\mathbb R})\rightarrow$$ implies that the flux homomorphism takes values in the isomorphic image of the relative cohomology group $H^1(M,L,{\mathbb R})$ in $H^1(M,{\mathbb R})$. To show surjectivity let $u\in \Omega^1(M)$ be a closed $1$-form which is exact on $L$. Then $u_{|L}=dh$ for some smooth function $h:L\rightarrow {\mathbb R}$. We can extend $h$ first to a tubular neighborhood and then to whole $M$. Replacing $u$ by $u-dh$ we can assume that $u_{|L}=0$. The closed $1$-form $u$ gives a symplectic vector field $X_t$ on $M$, which is indeed constant in time. If $\psi_t$ denotes the smooth $1$-parameter family of symplectomorphisms generated by $X_t$, then clearly $$\textmd{Flux}(\{ \psi_t \})=\int_0^1[\imath(X_t)\omega] \ dt=\int_0^1 [u] \ dt =[u].$$ We need to show that $\psi_t$ leaves $L$ invariant. Let $p \in L$ and $Y\in T_pL$ be any vector. Then by the construction of $u$ we have $0=u(p)(Y)=\imath(X_t)\omega(p) (Y)=\omega (X_t(p),Y)$ for any $Y\in T_pL$, which implies that $X_t(p)\in T_pL$, because $L$ is a Lagrangian submanifold. This finishes the proof. The following theorem is the relative version of Theorem $10.12$ of [@McS], which describes the kernel of the flux homomorphism. First we define the Hamiltonian symplectomorphisms of $M$ leaving $L$ invariant, denoted $\textmd{Ham}(M,L)$, as the subgroup of $\textmd{Symp}(M,L)$ consisting of symplectomorphisms $\psi$ such that there is a Hamiltonian isotopy $\psi_t:(M,L)\rightarrow (M,L)$, $t\in [0,1]$, such that $\psi_0=id$ and $\psi_1=\psi$; i.e., $\psi_t$ is a Hamiltonian isotopy of $M$ such that $\psi_t(L)=L$, for any $t\in [0,1]$. Since $L$ is Lagrangian a Hamiltonian $H_t:M\rightarrow {\mathbb R}$ generating the above isotopy is locally constant on $L$. \[thm-kernel\] Let $\psi \in \textmd{\em Symp}_0(M,L,\omega)$. Then $\psi$ is a Hamiltonian symplectomorphism if and only if there exists a symplectic isotopy, $\psi_t$, $[0,1]\rightarrow \textmd{\em Symp}_0(M,L,\omega)$ such that $\psi_0=id$, $\psi_1=\psi$ and $\textmd{\em Flux}(\{ \psi_t\})=0$. Moreover, if $\textmd{\em Flux}(\{ \psi_t\})=0$ then $\{ \psi_t\}$ is isotopic with fixed end points to a Hamiltonian isotopy through paths in ${\textmd{\em Symp}}_0(M,L,\omega)$. The proof of the above theorem is almost the same as that of Theorem $10.12$ of [@McS], where one has to observe that the symplectomorphisms and isotopies of $M$ involved in the proof leave $L$ invariant. In [@McS] the authors define the subgroup $$\Gamma(M) =\textmd{Flux}(\pi_1(\textmd{Symp}_0(M)))\subseteq H^1(M,{\mathbb R}).$$ Similarly, we define its relative version $$\Gamma(M,L) =\textmd{Flux}(\pi_1(\textmd{Symp}_0(M,L,\omega)))\subseteq H^1(M,L,{\mathbb R})$$ as the image of the fundamental group of the identity component of the relative symplectic group $\textmd{Symp}_0(M,L,\omega)$ under the flux homomorphism. As $\Gamma(M)$ the group $\Gamma(M,L)$ is a subgroup of $H^1(M,P_{\omega})$, where $P_{\omega}$ is the additive subgroup $[ \omega ] \cdot H_2(M,{\mathbb Z})\subseteq {\mathbb R}$, and hence both $\Gamma(M)$ and $\Gamma(M,L)$ are countable. To state the next lemma we need to recall the correspondence between symplectomorphisms of $(M,\omega)$, which are $C^1-$close to the identity and closed $1$-forms on $M$, which are close to the zero form. Consider the symplectic manifold $(M\times M, (- \omega)\oplus \omega)$ whose diagonal $\Delta \subset M\times M$ is a Lagrangian submanifold, diffeomorphic to $M$. Then the Lagrangian neighborhood theorem implies that there is a symplectomorphism $\Psi :\mathcal{N}(\Delta)\rightarrow \mathcal{N}(M_0)$ between the neighborhoods of the Lagrangian submanifolds $\Delta \subset M\times M$ and $M_0\subseteq (T^*M,\omega_{can})$, where $M_0$ is the zero section of the $T^*M$ equipped with the canonical symplectic form $\omega_{can}$, satisfying $\Psi^*(\omega_{can})=(- \omega)\oplus \omega$ and $\Psi(q,q)=q$, for any $q\in M$. Then the correspondence $$(\psi:M\rightarrow M) \mapsto\sigma = \mathcal{C}(\psi)\in \Omega^1(M)$$ is defined by $$\Psi(\textmd{graph}(\psi))=\textmd{graph}(\sigma)$$ provided that $\psi \in \textmd{Symp}(M)$ is sufficiently $C^1-$close to the identity. Studying the proof of the existence of a such symplectomorphism $\Psi$, one realizes that $\Psi$ can be chosen so that symplectomorphisms of $M$, $C^1-$close to the identity, leaving a Lagrangian submanifold $L$ invariant correspond to closed $1$-forms which evaluates zero on $TL$ (cf. see Theorem 3.32 and Theorem 3.14 of [@McS]). Indeed, for the proof of the below relative version one has to note that a symplectomorphism, $C^1-$close to the identity, leaves $L$ invariant if and only if the various $1$-forms involved in the proofs of Theorem 3.32 and Theorem 3.14 of [@McS] vanish on $(L\times L)\cap\Delta$ and the vector fields are parallel to $L\times L$. \[lem-diagonal\] Assume the above notation. Then there is a symplectomorphism $\Psi :\mathcal{N}(\Delta)\rightarrow \mathcal{N}(M_0)$ satisfying $\Psi(q,q)=q$ for any $q\in M$, such that if $\psi:(M,\omega)\rightarrow (M,\omega)$ is a symplectomorphism, which is sufficiently $C^1-$close to the identity map, then $\psi$ leaves a Lagrangian submanifold $L$ invariant if and only if $\sigma_{|T_qL}=0$, for any $q \in L$. \[rem-diagonal\] The above lemma can be restated as follows: Let $L$ be a Lagrangian submanifold of a symplectic manifold $(M,\omega)$. Then there is a symplectomorphism $\Psi :\mathcal{N}(\Delta)\rightarrow \mathcal{N}(M_0)$ satisfying $\Psi(q,q)=q$ for any $q\in M$, which maps a neighborhood of $(L\times L)\cap \Delta$ in $L\times L$ onto a neighborhood of $L_0=L\cap M_0$ in the conormal bundle to $L_0\subseteq T^*M$, which is $\{v\in T^*_qM \ | \ q \in L, \ ~ v_{|T_qL}=0 \}$. An immediate consequence of the above lemma is the following corollary. \[cor-diagonal\] Let $\psi:(M,\omega)\rightarrow (M,\omega)$ and $\sigma = \mathcal{C}(\psi)\in \Omega^1(M)$ be as in the above lemma. Assume that $\psi$ leaves a Lagrangian submanifold $L$ invariant. Then for any $t\in [0,1]$, the symplectomorphism $\psi_t$, corresponding to the closed $1$-form $t\sigma$, leaves $L$ invariant. Now we can prove the following relative version of Lemma $10.16$ of [@McS]. \[lem-gamma\] If $\psi \in \textmd{\em Symp}_0(M,L,\omega)$ is sufficiently $C^1$-close to the identity and $\sigma= \mathcal{C}(\psi_t)\in \Omega^1(M)$, then $\psi \in \textmd{\em Ham}(M,L)$ if and only if $[\sigma]\in \Gamma (M,L)$. Let $\psi_t$, $t \in [0,1]$, be the symplectic isotopy with $\mathcal{C}(\psi_t)=t\sigma$. Note that by the above corollary $\psi_t \in \textmd{Symp}_0(M,L,\omega)$. Lemma $10.15$ of [@McS] implies that $\textmd{Flux}(\{\psi_t\})=-[\sigma]$. Since $\psi$ is a Hamiltonian symplectomorphism, the path $\psi_t$ extends to a loop $[0,2]\mapsto \textmd{Symp}_0(M,L,\omega)$, which is generated by Hamiltonian vector fields with end points $\psi_0=id=\psi_2$. Moreover, $$\textmd{Flux}(\{\psi_t\}_{0\leq t \leq 2})=\textmd{Flux}(\{\psi_t\}_{0\leq t \leq 1})=-[\sigma]$$ and hence $[\sigma]\in \Gamma (M,L)$. Conversely, let $[\sigma]\in \Gamma (M,L)$. Choose a loop $\psi_t \in \textmd{Symp}_0(M,L,\omega)$ such that $\textmd{Flux}(\{\psi_t\})=-[\sigma]$. This extends to the interval $[0,2]$ by $\mathcal{C}(\psi_t)=[(t-1)\sigma]$, for $1\leq t \leq 2$. Then the resulting path $\psi_t$, $t\in [0,2]$, has zero flux. Now by Theorem \[thm-kernel\] this path can be deformed to a Hamiltonian isotopy via a homotopy with fixed end points. Therefore $\psi=\psi_2$ is a Hamiltonian isotopy. \[lem-ham\] Every smooth path $\psi_t\in \textmd{\em Ham}(M,L)$ is generated by Hamiltonian vector fields. The proof of the above lemma is completely analogous to that of Proposition $10.17$ of [@McS] and thus will be omitted. The following corollary which is the relative version of Corollary $10.18$ of [@McS] is the main result of this section. \[cor-rel\] Let $(M,\omega)$ be a closed connected symplectic manifold and $L$ is a closed connected Lagrangian submanifold. [**i)**]{} There is an exact sequence of simply connected Lie groups $$1\rightarrow \widetilde{\textmd{Ham}}(M,L)\rightarrow \widetilde{\textmd{Symp}}_0(M,L,\omega)\rightarrow H^1(M,L,{\mathbb R})\rightarrow 0$$ where $\widetilde{\textmd{ Ham}}(M,L)$ is the universal cover of $\textmd{Ham}(M,L)$ and the third homomorphism is the flux homomorphism. [**ii)**]{} There is an exact sequence of Lie algebras $$0\rightarrow {\mathbb R}\rightarrow C^{\infty}(M,L)\rightarrow \chi(M,L,\omega) \rightarrow H^1(M,L,{\mathbb R})\rightarrow 0.$$ Here the third map is $H\mapsto X_H$, the fourth map is $X\mapsto [\imath(X)\omega]$, while $C^{\infty}(M,L)$ denotes the algebra of smooth functions on $M$, which are constant on $L$ and $\chi(M,L,\omega)$ is the algebra of symplectic vector fields $X$ on $M$ such that $X(q)\in T_qL$, for any $q \in L$. [**iii)**]{} The sequence of groups $$1\rightarrow \pi_1(\textmd{Ham}(M,L)) \rightarrow \pi_1(\textmd{Symp}_0(M,L,\omega))\rightarrow \Gamma (M,L)\rightarrow 0$$ is exact. [**iv)**]{} There is an exact sequence of groups $$1\rightarrow \textmd{Ham}(M,L)\rightarrow \textmd{Symp}_0(M,L,\omega)\rightarrow H^1(M,L,{\mathbb R})/ \Gamma (M,L) \rightarrow 0$$ where the third map is induced by the flux homomorphism. By Lemma \[lem-ham\] every smooth path $\psi_t \in \textmd{ Ham}(M,L)$ which starts at the identity is a Hamiltonian isotopy and therefore has zero flux. This implies that $\widetilde{\textmd{Ham}}(M,L)\subseteq \ker(\textmd{Flux})$. On the other hand, by Theorem \[thm-kernel\] if $\textmd{Flux}(\{\psi_t\})=0$ then the path $\psi_t$ is homotopic, with fixed end points, to a Hamiltonian isotopy and hence $\{ \psi_t\} \in \widetilde{\textmd{Ham}}(M,L)$. Now, the first statement follows from the surjectivity of the flux homomorphism (see Lemma \[lem-defn\]). The second statement is easy and indeed follows from the proof of Lemma \[lem-defn\]. The only nontrivial part in the third statement is the fact that the homomorphism $\pi_1(\textmd{Ham}(M,L)) \rightarrow \pi_1(\textmd{ Symp}_0(M,L,\omega))$ is injective. To see this, it suffices to show that any path $[0,1] \rightarrow \widetilde{\textmd{Symp}}_0(M,L,\omega)$ with end points in $\widetilde{\textmd{Ham}}(M,L)$ is isotopic with fixed end points to a path in $\widetilde{\textmd{Ham}}(M,L)=\ker(\textmd{Flux})$. However, this is just the parameterized version of the first statement. The last statement is also obvious. Applications and Stability under Symplectic Reduction ===================================================== Combining Part (i) of Corollary \[cor-rel\] with its absolute version $$0\rightarrow \widetilde{\textmd{Ham}}(M)\rightarrow \widetilde{\textmd{Symp}}_0(M,\omega)\rightarrow H^1(M,{\mathbb R})\rightarrow 0,$$ Corollary $10.18$ of [@McS], we derive the following corollaries. \[cor3\] Suppose that $(M,\omega)$ and $L\subseteq M$ are as in Corollary \[cor-rel\]. If the first cohomology group, $H^1(M,\mathbb R)$, is trivial then $$\widetilde{\textmd{Ham}}(M)=\widetilde{\textmd{Symp}}_0(M,\omega)$$ and $$\widetilde{\textmd{Ham}}(M,L)=\widetilde{\textmd{Symp}}_0(M,L,\omega).$$ \[cor4\] Let $(M,\omega)$ be closed connected symplectic manifold and $L\subseteq M$ a closed connected Lagrangian submanifold. If $ImH^1(L,{\mathbb R})$ denotes the image of the restriction map $H^1(M,{\mathbb R})\rightarrow H^1(L,{\mathbb R})$ then the flux homomorphism induces an exact sequence $$1\rightarrow N(M,L)\rightarrow\widetilde{\textmd{Symp}}_0(M,\omega) \rightarrow ImH^1(L,{\mathbb R})\rightarrow 0,$$ where $N(M,L)$ is the normal closure of the product of the image of the canonical map $$\widetilde{\textmd{Symp}}_0(M,L,\omega)\rightarrow \widetilde{\textmd{Symp}}_0(M,\omega)$$ and $\widetilde{\textmd{Ham}}(M)$  in $\widetilde{\textmd{Symp}}_0(M,\omega)$. \[Exm-rel\] [**i)**]{} Let $L=S^1\subseteq S^2=M$, where $S^2$ is equipped with any symplectic form. Then by the above corollaries $$\widetilde{\textmd{Ham}}(M)= \widetilde{\textmd{Symp}}_0(M,\omega),$$ $$\widetilde{\textmd{Ham}}(M,L)= \widetilde{\textmd{Symp}}_0(M,L,\omega)$$ and $$\widetilde{\textmd{Symp}}_0(M,\omega)=N(M,L).$$ [**ii)**]{} Let $L=S^1\times \{pt\} \subseteq S^1 \times S^1=M$, where $M$ is equipped with any symplectic form. Then $$\widetilde{\textmd{Symp}}_0(M,\omega)/\widetilde{\textmd{Ham}}(M) \simeq{\mathbb R}^2,$$ $$\widetilde{\textmd{Symp}}_0(M,L,\omega)/\widetilde{\textmd{Ham}}(M,L) \simeq{\mathbb R}$$ and $$\widetilde{\textmd{Symp}}_0(M,\omega)/N(M,L)\simeq{\mathbb R}.$$ [**iii)**]{} Again let $M=S^1\times S^1$ and $L$ any smoothly embedded nullhomotopic circle in $M$. Then $$\widetilde{\textmd{Symp}}_0(M,\omega)/\widetilde{\textmd{Ham}}(M) \simeq{\mathbb R}^2,$$ $$\widetilde{\textmd{Symp}}_0(M,L,\omega)/\widetilde{\textmd{Ham}}(M,L) \simeq{\mathbb R}^2$$ and $$\widetilde{\textmd{Symp}}_0(M,\omega)=N(M,L).$$ Stability under symplectic reductions ------------------------------------- The next result is about the stability of the quotients $$\widetilde{\textmd{Symp}}_0(M,\omega)/\widetilde{\textmd{Ham}}(M) \ \ \mbox{and} \ \ \widetilde{\textmd{Symp}}_0/N(M,L)$$ under symplectic reduction. Let $S^1$ act in a Hamiltonian fashion on the symplectic manifold $(M,\omega)$ and $L$ be a Lagrangian submanifold contained in a level set $\mu^{-1}(c)$ of the moment map, $\mu :M\rightarrow {\mathbb R}$, of the $S^1$-action. Assume that $L$ is $S^1$ invariant and the $S^1$-action on the level set $\mu^{-1}(c)$ is free. Let $M_{red}=\mu^{-1}(c)/S^1$ be the symplectic quotient with the symplectic structure $\omega_{red}$ and $L_{red}=L/S^1$, which is a Lagrangian submanifold of $(M_{red},\omega_{red})$. A theorem of Li ([@Li]) states that the fundamental groups and hence the first cohomology groups of $M$, $\mu^{-1}(c)$ and $M_{red}$ are all isomorphic under the canonical homomorphisms. On the other hand, if $p:L\rightarrow L_{red}$ denotes the quotient map then we have the following result (see [@Oz1]). \[cor-final\] Let $M$, $M_{red}$, $L$ and $L_{red}$ be as above. Then the map $$p^*:Im(H^i(M_{red},{\mathbb Q})\rightarrow H^i(L_{red},{\mathbb Q})) \rightarrow Im(H^i(M,{\mathbb Q})\rightarrow H^i(L,{\mathbb Q}))$$ is onto for any $i$, and is an isomorphism for $i=1$. The above arguments together with Corollary \[cor4\] imply the following stability result. \[cor-stable\] Let $M$, $M_{red}$, $L$ and $L_{red}$ be as above. Then we have $$\widetilde{\textmd{Symp}}_0(M,\omega)/ \widetilde{\textmd{Ham}}(M) \simeq \widetilde{\textmd{Symp}}_0 (M_{red},\omega_{red})/\widetilde{\textmd{Ham}}(M_{red})$$ and $$\widetilde{\textmd{Symp}}_0(M,\omega)/N(M,L)\simeq \widetilde{\textmd{Symp}}_0(M_{red},\omega_{red})/N(M_{red},L_{red}).$$ Real algebraic varieties and their complexifications ---------------------------------------------------- In this subsection we will mention a different kind of stability of the above groups for real algebraic varieties. Let $X$ be a nonsingular compact real algebraic variety with a nonsingular projective complexification $i:X \rightarrow X_{\mathbb C}$. Note that any two nonsingular projective complexifications of $X$ are always birationally isomorphic and thus have the same fundamental group and the same first cohomology group (cf. see p.494 of [@GH]). Clearly $X_{\mathbb C}$ carries a Kähler and hence a symplectic structure such that $X$ becomes a Lagrangian submanifold. So, by Corollary $10.18$ of [@McS] (or just let $L$ be the empty set in Corollary \[cor-rel\](i)) the quotient group $$\widetilde{\textmd{Symp}}_0(X_{\mathbb C},\omega)/ \widetilde{\textmd{Ham}}(X_{\mathbb C})$$ is determined only by $X$ and hence is independent of the projective complexification $i:X\rightarrow X_{\mathbb C}$. Define $KH_i(X,{\mathbb R})$ as the kernel of the induced homomorphism $$i_*:H_i(X,{\mathbb R}) \rightarrow H_i(X_{\mathbb C},{\mathbb R})$$ and $ImH^i(X,{\mathbb R})$ as the image of the induced homomorphism $$i^*:H^i(X_{\mathbb C},{\mathbb R}) \rightarrow H^i(X,{\mathbb R}).$$ In [@Oz2; @Oz3] it is shown that both $KH_i(X,{\mathbb R})$ and $ImH^i(X,{\mathbb R})$ are independent of the projective complexification $i:X \rightarrow X_{\mathbb C}$ and thus (entire rational) isomorphism invariants of $X$. Moreover, by Corollary \[cor4\], for any topological component $X_0$ of $X$ the quotient group $$\widetilde{\textmd{Symp}}_0(X_{\mathbb C},\omega)/N(X_{\mathbb C},X_0)$$ is independent of the smooth projective complexification $i:X\rightarrow X_{\mathbb C}$ and hence is determined only by $X$ or equally by $X_0$. [10]{} E. Calabi, *On the group of automorphisms of a symplectic manifold*, Problems in Analysis (ed. R. Gunning), Princeton University Press, New Jersey, 1970. P. Griffiths, J. Harris, *Principles of Algebraic Geometry*, John Wiles $\&$ Sons, Inc., New York, 1994. H. Li, *$\pi_1$ of Hamiltonian $S^1$ manifolds*, (2002), arXiv:math.SG/0203075v1. D. McDuff, D. Salamon, *Introduction to symplectic topology*, Oxford University Press, New York, 1997. Y. Ozan, *On cohomology of invariant submanifolds of Hamiltonian actions*, preprint. ——– *Homology of non orientable real algebraic varieties* preprint. ——– *On homology of real algebraic varieties*, Proc. Amer. Math. Soc. **129** (2001), 3167-3175.

ArXiv

--- abstract: 'We have serendipitously discovered a previously-unknown, bright source ($B_T$ = 11.75$\pm$0.07mag) with a very blue $V_T-K_{\rm s}$ color, to which we have named Albus 1. A photometric and astrometric study using Virtual Observatory tools has shown that it possesses an appreciable proper motion and magnitudes and colors very similar to those of the well known white dwarf G 191–B2B. We consider Albus 1 as a DA-type white dwarf located at about 40pc. If confirmed its nature, Albus 1 would be the sixth brightest isolated white dwarf in the sky, which would make it an excellent spectrophotometric standard.' author: - José Antonio Caballero - Enrique Solano title: 'Albus 1: A very bright white dwarf candidate' --- Introduction {#intro} ============ The three classical white dwarfs were, at the beginning of the twentieth century, , , and the . Although $o^2$ Eri B had been discovered by Herschel (1785) and Sirius B had been predicted by Bessel (1844), it was not until the 1920s when other great astronomers noticed their oddness (Luyten 1922) and popularized the term “white dwarf” (Eddington 1924). The extreme physical conditions by which the white dwarfs are supported against gravitational collapse could not be understood until the Quantum Mechanic was properly developed. Since the 1930s, the number of known white dwarfs has exponentially increased, from 18 in 1939 (Schatzman 1958) and over a hundred in 1950 (Luyten 1950) to a few thousands at present (McCook & Sion 1999; Eisenstein et al. 2006). See Liebert (1980), Koester (2002) and Hansen & Liebert (2003) for extensive reviews of the general properties of white dwarfs. The vast majority of the known white dwarfs are very faint, with typical magnitudes in the optical from $V$ = 15 to 20mag, or even fainter. Only a few very bright white dwarfs ($V <$ 12mag), including the three classical white dwarfs and , are known. Many of them, especially those that are not in double degenerate systems, are extensively used as spectrophotometric stars (see, e.g., the recent catalogue by Landolt & Uomoto 2007). Except for rare exceptions, as the very hot dwarfs and cataclysmic variables found in the [*ROSAT*]{} all-sky survey of extreme-ultraviolet sources by Pounds et al. (1993), all the very bright white dwarfs were discovered during photometric and astrometric surveys before the early 70s (e.g. Kuiper 1941; Luyten 1949; Thackeray 1961; Eggen & Greenstein 1965; Giclas, Burnham & Thomas 1965; Schwartz 1972). Afterwards, and in particular with the advent of the Sloan Digital Sky Survey, the detection of new white dwarfs and blue subdwarfs has been biased towards magnitudes fainter than $V$ = 12mag. Because of that, the photometry-based discovery of a very bright white dwarf candidate 35 years later would be outstanding. If confirmed, it would yield doubts on the real knowledge that we have of the solar neighbourhood and on the completeness of previous and current surveys for white dwarfs. In this work we present , a previously-unknown very bright ($V_T$ = 11.80$\pm$0.14mag) white dwarf candidate[^1]. Its finding chart is provided in Fig. \[findingchart\]. Albus 1 was serendipitously discovered during an optical-near infrared photometric study by Caballero & Solano (2007), devoted to characterize the young stars and brown dwarfs surrounding ($\epsilon$ Ori) and ($\delta$ Ori). As part of this study, they made a correlation between the Tycho-2 (H[ø]{}g et al. 2000) and the 2MASS (Cutri et al. 2003) catalogues in ten 45arcmin-radius comparison fields at the same galactic latitude of the brightest stars of the young Association (the Orion Belt; $b \sim$ –17.5deg). The total investigated area was only 17.7deg$^2$ ($\sim$0.04% of the whole sky). Albus 1 has a $V_T-K_{\rm s}$ color that clearly deviates from those of the other 1275 investigated sources (see Fig. \[colormagnitude\]). In particular, while the bluest remaining sources have $V_T-K_{\rm s} \gtrsim$ –0.3mag, Albus 1 has a color $V_T-K_{\rm s}$ = –0.95$\pm$0.14mag. The Tycho-2 $B_T V_T$ and 2MASS $JHK_{\rm s}$ photometry shows that the object is extremely blue at all the wavelenghts from 0.4 to 2.2$\mu$m. Given the extreme blueing of Albus 1, we decided to investigate it in detail. Analysis ======== In this work we have taken advantage of the tools offered by the Virtual Observatory (VO; [http://www.ivoa.net]{}), which is an international, community-based initiative to provide seamless access to the data available from astronomical archive and services. The VO also aims to provide state-of-the-art tools for the efficient analysis of this huge amount of information. In particular, we have used Aladin ([http://aladin.u-strasbg.fr/aladin.gml]{}), a VO-compliant interactive sky atlas developed by CDS that allows the user to visualize and analyze astronomical images, spectra and catalogues available from the VO services. Albus 1 has an appreciable Tycho-2 proper motion of 19masa$^{-1}$. Since the comparison fields are relatively close to the antapex (the point the Sun is moving away from), the foreground objects in this region have not very large proper motions except for some very nearby stars with large tangential velocities. Indeed, the proper motion of Albus 1 is in the percentile 23% of the investigated sources (i.e. 77% of the investigated Tycho-2 stars have $\mu <$ 19masa$^{-1}$). This makes Albus 1 to be a nearby Galactic object. Apart from the coordinates, proper motion and $B_T V_T$ magnitudes from Tycho-2 and the $JHK_{\rm s}$ magnitudes from 2MASS, taken from Caballero & Solano (2007), we have also collected additional photometric and astrometric data from other catalogues: SuperCOSMOS Science Archive (Hambly et al. 2001), USNO-B1 (Monet et al. 2003), NOMAD1 (Zacharias et al. 2005) and DENIS (DENIS Consortium 2005). They nicely match between them and Tycho-2 and 2MASS, except for the fact that for very blue objects the Tycho-2 $B_T$ and photographic $B_J$ photometry are not comparable. To avoid superfluous, repetitive information in Table \[albus1\], we provide only the Tycho-2 and 2MASS information and the $RI_N$ photometric data from USNO-B1. The search with Aladin concluded that there is no radio (NRAO VLA), mid-infrared ($IRAS$), ultraviolet ($EUVE$), X-ray ($ROSAT$) source, or object discussed in the literature (SIMBAD) at less than 4arcmin to Albus 1. Neither spectroscopic information exists nor photometry in the Johnson passbands has been obtained yet. Results ======= To ascertain the real nature of Albus 1, we must compare its photometry with that of other very blue objects. There is a limited number of Galactic objects with $V_T-K_{\rm s}$ colors as blue as those of Albus 1: white dwarfs, hot subdwarfs, and early-type main sequence, blue horizontal branch and Population II stars. Fig. \[colorcolor\] compares the optical-near infrared colors of Albus 1 with those of dwarf and giant stars in the direction to Alnilam and Mintaka. Our blue source is even bluer than the late O- and early B-type stars. Besides, such luminous stars are located at long heliocentric distances ($d \gtrsim$ 0.4kpc), which implies very low proper motions, in contrast with what we have measured for Albus 1. Population II stars are common in the bulge near the centre of the Galaxy and in the Galactic halo. Some of the latters cross the Solar neighbourhood, but only a few of them display extremely blue colors (see, e.g., the recent photometric study of horizontal-branch and metal-poor candidates by Beers et al. 2007). Therefore, Albus 1 is an early-type hot subdwarf or a white dwarf. The extreme blueing of Albus 1 prevents from transforming the $B_T V_T$ magnitudes to Johnson $BV$ magnitudes using standard relations and, therefore, to compare its colors with other white dwarfs tabulated in exhaustive works such as in Bergeron, Leggett & Ruiz (2001). A new comparison may come, instead, from the available Tycho-2 and 2MASS data. In Table \[WDs\] we have compiled the basic data of the brightest white dwarfs and blue subdwarfs identified in the Tycho-2 catalogue. We used the white dwarf lists by McCook & Sion (1999) and Holberg, Oswalt & Sion (2002), and looked for the Tycho-2 counterparts of the white dwarfs brighter than $V$ = 13.0mag. Fainter objects were not considered due to their poor photometric accuracy. This list surpasses the sample of white dwarf observed by the $Hipparcos$ satellite in Vauclair et al. (1997). There are three evident absences: Sirius B, $o^2$ Eri B, and Procyon B (the three brightest known white dwarfs), which are too close to other bright stars and were, thus, not identified by Tycho-2. Among the tabulated objects, there is only one hot subdwarf with blue colors, GJ 3435, which indicates its rarity. Data in Table \[WDs\] is represented in Fig. \[colorcolor\]. Albus 1 is located in the color-color diagram very close to the well known white dwarfs G 191–B2B (DA1) and GJ 433.1 (DA3), at $V_T-J \sim$ –0.8mag, $J-K_{\rm s} \sim$ –0.2mag. The resemblance between the spectral energy distributions of Albus 1 and G 191–B2B (“the best studied of all hot white dwarfs”; Barstow et al. 2003), shown in Fig. \[sed\], is evident. Both of them have the same $K_{\rm s}$ magnitude within the error bars (Albus 1: $K_{\rm s}$ = 12.76$\pm$0.03mag; G 191–B2B: $K_{\rm s}$ = 12.76$\pm$0.02mag), but G 191–B2B is 0.40$\pm$0.11mag brighter in $B_T$. It leads to tentatively classify Albus 1 as an early DA white dwarf slightly cooler than G 191–B2B and, therefore, slightly closer to the Sun. The stellar common proper-motion companion of G 191–B2B has an accurate parallax determination by [*Hipparcos*]{} at $d$ = 46$\pm$4pc. Hence, Albus 1 could be located at about 40pc, which would explain its appreciable proper-motion. The probability of Albus 1 being a more distant blue subdwarf is smaller (see Table \[WDs\]). From the blue $J-K_{\rm s}$ color in Fig. \[colorcolor\], it is deduced, besides, that Albus 1 has no main sequence close companion or forms part of a cataclysmic variable system. As shown by Salim & Gould (2002), an optical-infrared reduced proper motion digram (e.g. $V + 5 \log{\mu}$ vs. $V-J$) can be used to classify stars even if no parallax information is available. In particular, white dwarfs and subdwarf stars are easily distinguished from main sequence stars as they are several magnitudes dimmer at the same color. The position of Albus 1 in the reduced proper motion diagram in fig. 4 in Gould & Morgan (2003; $V_T + 5 \log{\mu} =$ 3.2$\pm$0.6mag, $V_T-J$ = –0.76$\pm$0.14mag) agrees with this requirement. Conclusions =========== Of the 30 white dwarfs and blue subdwarfs listed in Table \[WDs\], only thirteen have Tycho-2 $V_T$ magnitudes brighter than 12.0mag. Albus 1, with $V_T$ = 11.80$\pm$0.14mag, is included in this group. Accounting for the three brightest known white dwarfs not in the Table, and discarding the close binary systems BL Psc AB, V841 Ara AB, V3885 Sgr AB, and BD+28 4211 AB, whose spectral energy distributions are affected by the main sequence close companions, then Albus 1 is the 12th brightest white dwarf yet known. Since six of the white dwarfs brighter than it are in multiple systems (the binary status of GJ 127.1 AB claimed by Gill & Kaptein 1896 is, however, not confirmed), then Albus 1 would be the sixth brightest isolated white dwarf, after the long-time known Feige 34, L 145–141, BD–07 3632, and HD 340611 (Luyten 1949; Eggen & Greenstein 1965) and the very hot white dwarf and extreme ultraviolet source RE J2214–49 (Holberg et al. 1993). Albus 1, although located in the southern hemisphere, is visible from the most important northern observatories. This fact, together with its brightness, makes our blue source an appropiate candidate spectrophotometric standard provided that its white dwarf or hot subdwarf nature is spectroscopically confirmed. Our serendipitous detection has also shown that the $V_T-K_{\rm s}$ color is a good and simple discriminator to look for very blue, relatively bright objects. A search for new very bright white dwarf candidates using Tycho-2 and 2MASS is currently ongoing. This research has made use of the Spanish Virtual Observatory supported from the Spanish Ministerio de Educación y Ciencia through grants AyA2005–04286 and AyA2005–24102–E of the Plan Nacional de Astronomía y Astrofísica. Barstow M. A., Good S. A., Holberg J. B., Hubeny I., Bannister N. P., Bruhweiler F. C., Burleigh M. R. & Napiwotzki R. 2003, MNRAS, 341, 870 Beers T. C., Flynn C., Rossi S. et al. 2007, ApJS, 168, 128 Bergeron P., Leggett S. K. & Ruiz M. T. 2001, ApJS, 133, 413 Bessel F. W. 1844, MNRAS, 6R, 136 Caballero J. A. & Solano E. 2007, A&A, to be submitted Cutri R. M., Skrutskie M. F., van Dyk S. et al. 2003, VizieR On-line Data Catalog: II/246. Originally published in: University of Massachusetts and Infrared Processing and Analysis Center (IPAC/California Institute of Technology) DENIS Consortium 2005, VizieR On-line Data Catalog: B/denis. Originally published in: The DENIS consortium (2005) Desidera S. & Barbieri M. 2007, A&A, 462, 345 Eddington A. S. 1924, MNRAS, 84, 308 Eggen O. J. & Greenstein J. L. 1965, ApJ, 141, 83 Eisenstein D. J., Liebert J., Harris H. C. et al. 2006, ApJS, 167, 40 Giclas H. L., Burnham R. & Thomas N. G. 1965, Bulletin Lowell Observatory, 6, 155 Gill D. & Kapteyn J. C. 1896, The Cape photographic durchmusterung for the equinox 1875, London, printed for H. M. Stationery off., by Darling & son, ltd., 1896–1900 Gould A. & Morgan C. W. 2003, ApJ, 585, 1056 Hambly N. C., MacGillivray H. T., Read M. A. et al. 2001, MNRAS, 326, 1279 Hansen B. M. S. & Liebert J. 2003, ARA&A, 41, 465 Herschel W. 1785, Philosophical Transactions of the Royal Society of London, 75, 40 H[ø]{}g E., Fabricius C., Makarov V. V., Urban S., Corbin T., Wycoff G., Bastian U., Schwekendiek P. & Wicenec A. 2000, A&A, 355, L27 Holberg J. B., Barstow M. A. & Buckley D. A. H. 1993, ApJ, 416, 806 Holberg J. B., Oswalt T. D. & Sion E. M. 2002, ApJ, 571, 512 Koester D. 2002, A&ARv, 11, 33 Kuiper G. K. 1941, PASP, 53, 248 Landolt A. U. & Uomoto A. K. 2007, AJ, 133, 768 Liebert J. 1980, ARA&A, 18, 363 Luyten W. J. 1922, PASP, 34, 156 Luyten W. J. 1949, ApJ, 109, 528 Luyten W. J. 1950, AJ, 55, 86 McAlister H. A., Mason B. D., Hartkopf W. I., Roberts L. C. Jr. & Shara M. M. 1996, AJ, 112, 1169 McCook G. P. & Sion E. M. 1999, ApJS, 121, 1 Monet D. G., Levine S. E., Casian B. et al. 2003, AJ, 125, 984 Pounds K. A., Allan D. J., Barber C. et al. 1993, MNRAS, 260, 77 Ribeiro F. M. A. & Diaz M. P. 2007, AJ, 133, 2659 Salim S. & Gould A. 2002, ApJ, 585, L83 Schatzman E. 1958, White Dwarfs, Amsterdam, North-Holland Pub. Co.; New York, Interscience Publishers, 1958. Schwartz R. D. 1972, PASP, 84, 28 Thackeray A. D. 1961, MNSSA, 20, 40 Vauclair G., Schmidt H., Koester D. & Allard N. 1997, A&A, 325, 1055 Zacharias N., Monet D. G., Levine S. E., Urban S. E., Gaume R. & Wycoff G. L. 2004, American Astronomical Society Meeting 205, \#48.15; Bulletin of the American Astronomical Society, Vol. 36, p. 1418 [lccc]{} Name & Albus 1 & & 1\ WD number & WD 0604–203 & & 1\ $\alpha$ (J2000)& 06 06 13.39 & & 2\ $\delta$ (J2000)& –20 21 07.3 & & 2\ $\mu_\alpha \cos{\delta}$& +7$\pm$3 & masa$^{-1}$ & 2\ $\mu_\delta$ & –18$\pm$3 & masa$^{-1}$ & 2\ $B_T$ & 11.75$\pm$0.07 & mag & 2\ $V_T$ & 11.80$\pm$0.14 & mag & 2\ $R$ & 11.84 & mag & 3\ $I_N$ & 11.90 & mag & 3\ $J$ & 12.56$\pm$0.02 & mag & 4\ $H$ & 12.66$\pm$0.03 & mag & 4\ $K_{\rm s}$ & 12.76$\pm$0.03 & mag & 4\ Sp. type & DA? & & 1\ $d$ & $\sim$40? & pc & 1\ [llcccc]{} &  AB & DA2+K0IV & 10.18$\pm$0.04 & 8.45$\pm$0.03& 7.80$\pm$0.03\ & & DZ7 & 12.559$\pm$0.018 & 11.69$\pm$0.02& 11.50$\pm$0.02\ & & DA3.5 & 12.552$\pm$0.010 & 12.77$\pm$0.02& 12.85$\pm$0.03\ & , Feige 22 & DA2.5 & 12.848$\pm$0.019 & 13.28$\pm$0.03& 13.42$\pm$0.02\ &  AB, Feige 24 & DA+M1V & 12.6$\pm$0.2 & 11.26$\pm$0.03& 10.557$\pm$0.019\ & $^a$ & DA3 & 11.15$\pm$0.07 & 11.76$\pm$0.02& 11.86$\pm$0.02\ & $^b$ & DQ7 & 12.13$\pm$0.18 & 6.62$\pm$0.02& 5.72$\pm$0.02\ & & DA1 & 11.65$\pm$0.17 & 12.54$\pm$0.02& 12.76$\pm$0.02\ **[WD 0604–203]{} & **[Albus 1]{} & **[DA?]{} & **[11.80$\pm$0.14]{} & **[12.56$\pm$0.02]{}& **[12.76$\pm$0.03]{}\ & & DA1 & 12.3$\pm$0.2 & 12.85$\pm$0.03& 13.09$\pm$0.03\ & , He 3 & DA2.5 & 12.2$\pm$0.2 & 12.7$\pm$0.3 & 12.8$\pm$0.3\ & , GD 294 & sdB & 12.03$\pm$0.16 & 11.77$\pm$0.02& 11.721$\pm$0.018\ & , Feige 34 & DA0: & 11.10$\pm$0.07 & 11.64$\pm$0.02& 11.540$\pm$0.019\ & & DA3 & 12.14$\pm$0.19 & 12.99$\pm$0.02& 13.18$\pm$0.03\ & , L 145–141 & DQ6 & 11.34$\pm$0.09 & 11.19$\pm$0.02& 11.10$\pm$0.03\ &  AB$^c$ & DA1+M3.5Ve & 12.667$\pm$0.009 &10.373$\pm$0.019& 9.56$\pm$0.02\ & $^d$ & DA4 & 11.77$\pm$0.18 & 12.62$\pm$0.04& 12.74$\pm$0.05\ & & DA3 & 12.839$\pm$0.010 & 13.25$\pm$0.02& 13.45$\pm$0.04\ & $^e$ & DA2 & 11.00$\pm$0.08 & 11.58$\pm$0.02& 11.77$\pm$0.02\ &  AB$^f$ & DOZ1+K0V & 11.26$\pm$0.07 & 7.12$\pm$0.02& 6.57$\pm$0.03\ & , G 226–29$^g$& DAV4.7 & 12.00$\pm$0.16 & 12.42$\pm$0.02& 12.52$\pm$0.03\ & $^h$ & DBQA5 & 13.1$\pm$0.3 & 12.35$\pm$0.03& 12.42$\pm$0.03\ &  AB$^i$ & DB:p+M:V & 10.33$\pm$0.05 & 9.96$\pm$0.03& 9.62$\pm$0.02\ & & DA4 & 12.9$\pm$0.3 & 12.58$\pm$0.02& 12.70$\pm$0.03\ & & DA2.5 & 11.55$\pm$0.10 & 12.04$\pm$0.03& 12.19$\pm$0.03\ & & DA2.5 & 13.1$\pm$0.3 & 12.82$\pm$0.03& 13.00$\pm$0.03\ & & DA3.5 & 12.7$\pm$0.3 & 12.68$\pm$0.02& 12.85$\pm$0.04\ &  AB$^j$ & sdO:p+G: & 10.53$\pm$0.05 & 11.28$\pm$0.03& 11.56$\pm$0.03\ & & DA3 & 13.2$\pm$0.4 & 13.20$\pm$0.02& 13.39$\pm$0.04\ & & DA.76 & 11.52$\pm$0.08 & 12.44$\pm$0.03& 12.64$\pm$0.03\ ************ [^1]: [*Albus*]{} is the Latin term for “white”.

ArXiv

--- abstract: 'Programming with logic for sophisticated applications must deal with recursion and negation, which have created significant challenges in logic, leading to many different, conflicting semantics of rules. This paper describes a unified language, DA logic, for design and analysis logic, based on the unifying founded semantics and constraint semantics, that support the power and ease of programming with different intended semantics. The key idea is to provide meta constraints, support the use of uncertain information in the form of either undefined values or possible combinations of values, and promote the use of knowledge units that can be instantiated by any new predicates, including predicates with additional arguments.' author: - | Yanhong A. Liu Scott D. Stoller\ Computer Science Department, Stony Brook University\ liu@cs.stonybrook.edu   stoller@cs.stonybrook.edu bibliography: - 'strings.bib' - 'references.bib' title: | Knowledge of Uncertain Worlds:\ Programming with Logical Constraints[[^1]]{} --- Introduction {#sec:intro} ============ Programming with logic has allowed many design and analysis problems to be expressed more easily and clearly at a high level. Examples include problems in program analysis, network management, security frameworks, and decision support. However, when sophisticated problems require reasoning with negation and recursion, possibly causing contradiction in cyclic reasoning, programming with logic has been a challenge. Many languages and semantics have been proposed, but they have different, conflicting underlying assumptions that are subtle and do not work for all problems. This paper describes a unified language, DA logic, for design and analysis logic, for programming with logic using logical constraints. It supports logic rules with unrestricted negation in recursion, as well as unrestricted universal and existential quantification. It is based on the unifying founded semantics and constraint semantics, and it supports the power and ease of programming with different intended semantics without causing contradictions in cyclic reasoning. - The language provides meta-constraints on predicates. These meta-constraints capture the different underlying assumptions of different logic language semantics. - The language supports the use of uncertain information in the results of different semantics, in the form of either undefined values or possible combinations of values. - The language further supports the use of knowledge units that can be instantiated by any new predicates, including predicates with additional arguments. Together, the language allows complex problems to be expressed clearly and easily, where different assumptions can be easily used, combined, and compared for expressing and solving a problem modularly, unit by unit. We present examples from different games that show the power and ease of programming with DA logic. Need of easier programming with logic {#sec:motivation} ===================================== We discuss the challenges of understanding and programming with negation and recursion. We use a small well-known example, the win-not-win game, for illustration. Consider the following rule, called the win rule. It says that is a winning position if there is a move from to and is not a winning position. win(x) move(x,y) win(y). This seems to be a reasonable rule, because it captures the rule for winning for many games, including in chess for the King to not be captured, giving winning, losing, and draw positions. However, there could be potential problems. For example if there is a for some position , then the win rule would imply: if not , and thus the truth value of becomes unclear. #### Inductive definitions. Instead of the single win rule, one could use the following three rules to determine the winning, losing, and draw positions. win(x) y | move(x,y) lose(y). lose(x) y | move(x,y) win(y). draw(x) win(x) lose(x). The first two rules form inductive definitions [@hou2010fo; @dasseville2015semantics], avoiding the potential problems of the single win rule. The base case is the set of positions that have no moves to any other position and thus are losing positions. With winning and losing positions defined, the draw positions are those in cycles of moves that have no moves to losing positions. However, clearly, these rules are much more cumbersome than the single win rule. #### Well-founded semantics. Indeed, with well-founded semantics (WFS) [@van+91well], which computes a 3-valued model, the single win rule above gives being True, False, or Unknown for each , corresponding to being a winning, losing, or draw position, respectively. However, being 3-valued does not allow the three outcomes to be used as three predicates or sets for further computation; the three predicates defined by the three rules do allow this. For example, there is no way to use the Unknown positions explicitly, say to find all reachable nodes following another kind of moves from draw positions. One might try to do it by adding two additional rules to the single win rule: lose(x) win(x) draw(x) win(x) lose(x) However, the result is that is False for all positions that is True or False, and is Unknown for all draw positions. #### Stable model semantics. Stable model semantics (SMS) [@GelLif88] computes a set of 2-valued models, instead a single 3-valued model. It has been used for solving many constraint problems in answer set programming (ASP), because its set of 2-valued models can provide the set of satisfying solutions. For example, for the singl win rule, if besides winning and losing positions, there is a separate cycle of even length, say and , then instead of being Unknown for and as in WFS, SMS returns two models: one with being True for and other winning positions but not , and one with being True for and other winning positions but not . This is a very different interpretation of the win-not-win rule. However, for the single rule above, when there are draw positions, SMS may also return just an empty set, that is, a set with no models at all. For example, if besides winning and losing positions, there is a separate cycle of moves of odd length, say simply , then SMS returns simply the empty set. This is clearly undesired for the win-not-win game. #### Founded semantics and constraint semantics. Founded semantics and constraint semantics [@LiuSto18Founded-LFCS] unify different prior semantics. It allows different underlying assumptions to be specified for each predicate, and computes the desired semantics as a simple least fixed point to return a 3-valued model and, if there are undefined values, as constraint solving to return a set of 2-valued models. For the win-not-win game, one can write the single win rule, with the default assumption that is [*complete*]{}, that is, the win rule is the only rule that infers , which is an implicit assumption underlying WFS and SMS. - With founded semantics, the three rules that use inductive definitions can be automatically derived, and True, False, and Undefined positions for are inferred, corresponding to the three predicates from inductive definitions and the 3-valued results from WFS. - Then constraint semantics, if desired, computes all combinations of True and False values for the Undefined values for the draw positions, that satisfy all the rules as constraints. It equals SMS for the single win rule. Both WFS and SMS also assume that if nothing is said about some , then is false. When this is not desired, some programming tricks are used to get around it. For example, with SMS, to allow to be possibly true in some models, one can introduce some new and two new rules as below, to make it possible that, in some models, is true and is false. p q q p Founded semantics and constraint semantics allow to be simply declared as [*uncertain*]{}. Both WFS and SMS also assume that if all ways that can infer require using in the condition of some rule, then is false. Founded semantics and constraint semantics allow this reasoning to be used where desired, by applying it if is declared as [*closed*]{}. Founded semantics and constraint semantics also allow unrestricted universal and existential quantifications and unrestricted nesting of Boolean conditions; these are not supported in WFS and SMS. However, founded semantics and constraint semantics alone do not address how to use different semantics seamlessly in a single logic program. #### Programming with logical constraints. Because different assumptions and semantics help solve different problems or different parts of a problem, easier programming with logic requires supporting all assumptions and semantics in a simple and integrated design. This paper treats different assumptions as different meta-constraints for expressing a problem or parts of a problem, and support results from different semantics to be used easily and directly. For the win-not-win example: - We name the positions for which is true, false, and undefined in founded semantics using three predicates, , , and , corresponding exactly to the inductively defined , , and . These predicates can be used explicitly and directly for further reasoning, unlike with the truth values of WFS or founded semantics. - We let be the constraint semantics of a set of rules and facts. For , we use to denote the truth value of in model . Predicate means exactly and can be used directly for further reasoning, unlike the set of models in SMS or constraint semantics. Table \[tab:summary\] summarizes the meta-constraints that can be used to express different assumptions, corresponding declarations and resulting predicates in founded semantics and constraint semantics, and corresponding other prior semantics if all predicates use the same meta-constraint. Columns 2 and 4 are presented and proved in our prior work [@LiuSto18Founded-LFCS]. Columns 1 and 3 are introduced in DA logic. More fundamentally, we must enable easy specification of problems with reusable parts and where different parts may use different assumptions and semantics. To that end, we support instantiation and re-use of existing parts, and allow predicates in any existing parts to be bound to other given predicates, including predicates with additional arguments. Even with all this power, DA logic is decidable, because it does not include function symbols and is over finite domains. [@l|p[21ex]{}|p[22ex]{}|p[25ex]{}@]{}Meta-constraint & & Other Prior Semantics\ on predicate & Declarations on & Resulting predicates &\ certain() & certain & , & Stratified (Perfect,Inductive Definition)\ open() & uncertain,not complete & , , ----- for : Meta constraints and corresponding prior semantics.[]{data-label="tab:summary"} & [@p[25ex]{}@]{}\ -Order Logic \ complete() & uncertain,complete & as above & [@p[25ex]{}@]{}Fitting (Kripke-Kleene)$\!\!$\ \ closed() & uncertain,complete, closed & as above & [@p[25ex]{}@]{}WFS\ \ Additional examples {#sec:examples} =================== We present additional examples that show the power of our language. They are challenging or impossible to express and solve using prior languages and semantics. We use to prefix comments. #### Same different games. The same win-not-win game can be over different kinds of moves, forming different games, as introduced with kunit instantiation. However, the fundamental winning, losing, or draw situations stay the same, parameterized by the moves. The moves could also be defined easily using another kunit instantiation. A new game can use winning, losing, draw positions defined by in Section \[sec:motivation\], whose moves use paths defined by , whose edges use given links. kunit path\_unit: path(x,y) edge(x,y) path(x,y) edge(x,z) path(z,y) kunit win\_path\_unit: link = {(1,2), (1,3), ...} – shorthand for link(1,2), link(1,3), ... use path\_unit (edge = link) – instantiate path\_unit with edge replaced by link use win\_unit (move = path) – instantiate win\_unit with move replaced by path One could also define in place of above, and then can be used without rebinding the name , as follows. kunit win\_path\_unit: – as above edge = {(1,2), (1,3), ...} – as above but use edge in place of link use path\_unit () – as above but without replacing edge by link use win\_unit (move = path) – as above #### Defined from undefined positions. Sets and predicates can be defined using the set of values of arguments for which a given predicate is undefined. This is not possible in previous 3-valued logic like WFS, because anything depending on undefined can only be undefined. Using the win-not-win game, the predicates and below define the set of positions that have a move to a draw position, and the set of positions that have a special move from a draw position, respectively. kunit draw\_unit: move = {(1,1), (2,3), (3,1)} use win\_unit () move\_to\_draw(x) move(x,y) win.U(y) special\_move = {(1,4), (4,2)} use path\_unit (edge = special\_move) reach\_from\_draw(y) win.U(x) path(x,y) In , we have , that is, 1 is a draw position. Then we have , and we have and . Note that we could copy the single win rule here in place of and obtain an equivalent . We avoid copying when possible because this is a good principle, and in general, a kunit may contain many rules and facts. #### Unique undefined positions. Among the most critical information is information that is true in all possible ways of satisfying given constraints but cannot be determined to be true by just following founded reasoning. Having both founded semantics and constraint semantics at the same time allows one to find such information. Predicate in below finds positions in the game in that are in the founded model but, if a constraint model exists, are winning in all possible models in constraint semantics . kunit win\_unit1: prolog asp asp prolog move(1,0) prolog move(1,0) asp move(1,1) use win\_unit ()kunit cmp\_unit: use win\_unit1 () unique(x) win.U(x) m win\_unit1.CS m win\_unit1.CS | m.win(x) In , founded semantics gives , , , and . , that is, is true, and is false. So and are imprecise, and is true in . #### Multiple uncertain worlds. Given multiple worlds with different models, different uncertainties can arise from different worlds, yielding multiple uncertain worlds. It is simple to represent this using predicates that are possibly 3-valued and that are parameterized by a 2-valued model. The game in uses on a set of moves. The game in has its own moves, but the moves are valid if and only if they follow from a position that is a winning position in the constraint semantics of . kunit win\_unit2: move = {(1,4),(4,1)} use win\_unit ()kunit win\_set\_unit: move = {(1,2),(2,3),(3,1),(4,4),(5,6)} valid\_move(x,y,m) move(x,y), win\_unit2.CS(m), m.win(z), path(z,x) use win\_unit (move = valid\_move(m), win = valid\_win(m)) win\_some(x) valid\_win(x,m) win\_each(x) win\_some(x) m win\_unit2.CS | valid\_win(x,m) In , there is a 2-edge cycle of moves, so , where and . In , each in leads to a separately defined predicate under argument , which is then used to define a separate predicate under argument by instantiating with and parameterized by additional argument . Related work and conclusion =========================== Many logic languages and semantics have been proposed. Several overview articles [@apt1994negation; @prz94well; @ramUll95survey; @fitting2002fixpoint; @trusz18sem-wbook] give a good sense of the complications and challenges when there is unrestricted negation. Notable different semantics include Clark completion [@clark78] and similar additions, e.g., [@lloyd84making; @sato84transformational; @Jaffar+86some; @chan88constructive; @foo88deduced; @stuckey91constructive], Fitting semantics or Kripke-Kleene semantics [@fitting85], supported model semantics [@apt88], stratified semantics [@apt88], WFS [@van+91well], and SMS [@GelLif88]. Note that these semantics disagree, in contrast to different styles of semantics that agree [@ershov1987semantic]. There are also a variety of works on relating and unifying different semantics. These include Dung’s study of relationships [@dung1992relations], partial stable models, also called stationary models [@prz94well], Loop fomulas [@lin2004assat], FO(ID) [@denecker2008logic], and founded semantics and constraint semantics [@LiuSto18Founded-LFCS]. FO(ID) is more powerful than works prior to it, by supporting both first-order logic and inductive definitions while also being similar to SMS [@bruynooghe2016first]. However, it does not support any 3-valued semantics. Founded semantics and constraint semantics uniquely unify different semantics, by capturing their different assumptions using predicates declared to be certain, complete, and closed, or not. However, founded semantics and constraint semantics by themselves do not provide a way for different semantics to be used for solving different parts of a problem or even the same part of the problem. DA logic supports these, and supports everything completely declaratively, in a unified language. Specifically, DA logic allows different assumptions under different semantics to be specified easily as meta-constraints, and allows the results of different semantics to be built upon, including defining predicates using undefined values in a 3-valued model and using models in a set of 2-valued models, and parameterizing predicates by a set of 2-valued models. More fundamentally, DA logic allows different parts of a problem to be solved with different knowledge units, where every predicate is a parameter that can be instantiated with new predicates, including new predicates with additional arguments. These are not supported in prior languages. Among many directions for future work, one particularly important and intriguing problem is to study precise complexity guarantees for inference and queries for DA logic. [^1]: This work was supported in part by NSF under grants CCF-1414078, CNS-1421893, and IIS-1447549.

ArXiv

--- abstract: | Let $G$ be a finite simple graph on $n$ vertices. Let $J_G \subset K[x_1, \ldots, x_n]$ be the cover ideal of $G$. In this article, we obtain the graded minimal free resolution and the Castelnuovo-Mumford regularity of $J_G^s$ for all $s \geq 1$ for certain multipartite graphs $G$. address: - 'Department of Mathematics, Indian Institute of Technology Madras, Chennai, INDIA - 60036' - 'Department of Mathematics, Indian Institute of Technology Bombay, Powai, Mumbai, INDIA - 400076' author: - 'A. V. Jayanthan' - Neeraj Kumar bibliography: - 'refs\_reg.bib' title: Resolution and regularity of cover ideals of certain multipartite graphs --- Introduction ============ Recently there have been a lot of research on various properties of powers of homogeneous ideals. In particular, there have been a lot of interest on Castelnuovo-Mumford regularity of powers of ideals. It was shown by Kodiyalam [@vijay] and independently by Cutkosky, Herzog and Trung [@CHT99] that if $I$ is a homogeneous ideal in a polynomial ring, then there exist non-negative integers $d, e$ and $s_0$ such that ${\operatorname{reg}}(I^s) = ds + e$ for all $s \geq s_0$, where ${\operatorname{reg}}(-)$ denote the Castelnuovo-Mumford regularity. Kodiyalam proved that $d \leq \deg(I)$, where $\deg(I)$ denotes the largest degree of a homogeneous minimal generator of $I$. In general, the stability index $s_0$ and the constant term $e$ are hard to compute. There have been discrete attempts in identifying $s_0$ and $e$ for certain classes of ideals. Given a finite simple graph, one can identify the vertices with indeterminates and associate ideals in a polynomial ring, for example, edge ideals and cover ideals (see Section 2 for definition) corresponding to the given graph. For edge ideals of various classes of graphs, $e$ and $s_0$ have been computed, see for example [@banerjee; @huneke; @froberg; @ha_adam; @sean_thesis; @jns17; @khosh_moradi; @kummini; @mohammad; @nevo_peeva; @adam; @Villarreal; @russ; @Zheng]. In the case of edge ideals, $d = 2$ and in the known cases, $e$ is connected to combinatorial invariants associated with the graph $G$. Not much is known about the regularity of powers of cover ideals. Since cover ideal is the Alexander dual of the edge ideal, its regularity is equal to the projective dimension of the edge ideal, [@terai99]. In [@fakhari16], Seyed Fakhari studied certain homological properties of symbolic powers of cover ideals of very well-covered and bipartite graphs. For a finite simple graph $G$ on the vertex set $\{x_1, \ldots, x_n\}$, let $J_G \subset R = K[x_1, \ldots, x_n]$, where $K$ is a field, denote the cover ideal of $G$. It was shown that if $G$ is a very well-covered graph and $J_G$ has a linear resolution, then $J_G^{(s)}$ has a linear resolution for all $s \geq 1$. Furthermore, it was proved that if $G$ is a bipartite graph with $n$ vertices, then for $s \geq 1$, $${\operatorname{reg}}(J_G^s) \leq s\deg(J_G) + {\operatorname{reg}}(J_G) + 1.$$ Hang and Trung, in [@ht17], studied unimodular hypergraphs and proved that if $\mathcal{H}$ is a unimodular hypergraph on $n$ vertices and rank $r$ and $J_{\mathcal{H}}$ is the cover ideal of $\mathcal{H}$, then there exists a non-negative integer $e \leq \dim(R/J_{\mathcal{H}}) - \deg(J_{\mathcal{H}}) + 1$ such that $$reg J_{\mathcal{H}}^s = \deg(J_{\mathcal{H}}) s + e$$ for all $s \geq \frac{rn}{2} + 1$. Since bipartite graphs are unimodular, their results hold true in the case of bipartite graphs as well. While the first result gives an upper bound for the constant term, the later result gives the upper bound for both the stability index and the constant term. The Betti numbers and regularity are classical invariants associated to a module which can be computed from the resolution. However, one may be able to compute them without completely describing the resolution. To compute the syzygies explicitly is a challenging task. In this article, we obtain the complete description of the minimal free resolution, including the syzygies, of $J_G^s$ for some classes of multipartite graphs, thereby obtaining a precise expression for ${\operatorname{reg}}(J_G^s)$. The paper is organized as follows. In Section 2, we collect the preliminaries required for the rest of the paper. We study the resolution of powers of cover ideals of certain bipartite graphs in Section 3. If $G$ is a complete bipartite graph, then $J_G$ is a regular sequence and hence the minimal graded free resolution of $J_G^s$ can be obtained from [@guardo-vtuyl05 Theorem 2.1]. It can be seen that, in this case the index of stability, $s_0 = 1$ and the constant term is one less than the size of the minimum vertex cover. We then move on to study some classes of bipartite graphs which are not complete. We obtain the resolution and precise expressions for the regularity of powers of cover ideals of certain bipartite graphs. Section 4 is devoted to the study of resolution and regularity of powers of cover ideals of certain complete multipartite graphs. When $G$ is the cycle of length three or the complete graph on $4$ vertices, we describe the graded minimal free resolution of $J_G^s$ for all $s \geq 1$. This allow us to compute the Betti numbers, Hilbert series and the regularity of $J_G^s$ for all $s \geq 1$. As a consequence, for cover ideals of complete tripartite and $4$-partite graphs, we obtain precise expressions for the Betti numbers and the regularity of $J_G^s$. We conclude our article with a conjecture on the resolution of $J_G^s$ for all $s \geq 1$, where $G$ is a complete multipartite graph. **Acknowledgements:** Part of the work was done while the second author was visiting Indian Institute of Technology Madras. He would like to thank IIT Madras for their hospitality during the visit. All our computations were done using Macaulay 2, [@M2]. Preliminaries ============= In this section, we set the notation for the rest of the paper. All the graphs that we consider in this article are finite, simple and without isolated vertices. For a graph $G$, $V(G)$ denotes the set of all vertices of $G$ and $E(G)$ denotes the set of all edges of $G$. A graph $G$ is said to be a *complete multipartite* graph if $V(G)$ can be partitioned into sets $V_1, \ldots, V_k$ for some $k \geq 2$ such that $\{x,y\} \in E(G)$ if and only if $x \in V_i$ and $y \in V_j$ for $i \neq j$. When $k = 2$, the graph is called a *complete bipartite* graph. If $k = 2$ with $|V_1| = m$ and $|V_2| = n$, we denote the corresponding complete bipartite graph by $K_{m,n}$ or by $K_{V_1,V_2}$. If $G$ and $H$ are graphs, then $G \cup H$ denote the graph on the vertex set $V(G) \cup V(H)$ with $E(G\cup H) = E(G) \cup E(H)$. A graph $G$ is called a *bipartite* graph if $V(G) = V_1 \sqcup V_2$ such that $\{x,y\} \in E(G)$ only if $x \in V_1$ and $y \in V_2$. A subset $w = \{x_{i_1}, \ldots, x_{i_r}\}$ of $V(G)$ is said to be a *vertex cover* of $G$ if $w \cap e \neq \emptyset$ for every $e \in E(G)$. A vertex cover is said to be minimal if it is minimal with respect to inclusion. Let $G$ be a graph with $V(G) = \{x_1, \ldots, x_n\}$. Let $K$ be a field and $R = K[x_1,\ldots,x_n]$. The ideal $I(G) = \langle\{x_ix_j ~ : ~ \{x_i,x_j\} \in E(G)\} \rangle \subset S$ is called the *edge ideal* of $G$ and the ideal $J_G = \langle \{x_{i_1}\cdots x_{i_r} ~ : ~ \{x_{i_1}, \ldots, x_{i_r}\} \text{ is a vertex cover of } G \} \rangle$ is called the *cover ideal* of $G$. It can also be seen that $J_G$ is the Alexander dual of $I(G)$. Let $S=R/I$, where $R$ is a polynomial ring over $K$ and $I$ a homogeneous ideal of $R$. For a finitely generated graded $S$-module $M=\oplus M_i$, set $$t_i^{S}(M)=\max \{j : \operatorname{Tor}_{i}^{S}(M,K)_j \neq 0 \},$$ with $t_i^{S}(M)=-\infty $ if $\operatorname{Tor}_i^S(M,K) = 0.$ The *Castelnuovo-Mumford regularity*, denoted by ${\operatorname{reg}}_S(M)$, of an $S$-module $M$ is defined to be $${\operatorname{reg}}_SM= \max \{t_i^{S}(M)-i: i \geq 0\}.$$ Bipartite Graphs ================ In this section, we study the regularity of powers of cover ideals of certain bipartite graphs. We begin with a simple observation concerning the vertex covers of a bipartite graph. Let $G$ be a bipartite graph on $n+m$ vertices. Then $G$ is a complete bipartite graph if and only if $J_G$ is generated by a regular sequence. Let $V(G) = X \sqcup Y$ be the partition of the vertex set of $G$ with $X = \{x_1,\ldots,x_n\}$ and $Y = \{y_{n+1},\ldots,y_{n+m}\}$. First note that $J_G$ is generated by a regular sequence if and only if for any two minimal vertex covers $w, w'$, $w \cap w' = \emptyset$. If $G = K_{n,m}$, then $J_G = (x_1\cdots x_n, y_{n+1}\cdots y_{n+m})$ which is a regular sequence. Conversely, suppose $G$ is not a complete bipartite graph. Since $G$ is a bipartite graph, note that $\prod_{x_i \in X} x_i, \prod_{y_j \in Y}y_j \in J_G$ are minimal generators of $J_G$. Therefore, there exist $x_{i_0} \in X$ and $y_{i_0} \in Y$ such that $\{x_{i_0}, y_{i_0}\} \notin E(G)$. Then $w = \{x_i, y_j ~ : ~ i \neq i_0 \text{ and }y_j \in N_G(x_{i_0})\}$ is a minimal vertex cover of $G$ that intersects $X$ as well as $Y$ non-trivially. Therefore $J_G$ is not a complete intersection. First we discuss the regularity of powers of cover ideals of complete bipartite graphs. Since the cover ideal of a complete bipartite graph is a complete intersection, the result is a consequence of [@guardo-vtuyl05 Theorem 2.1]. \[com.bipar00\] Let $J = J_{K_{m,n}}$ be the cover ideal of the complete bipartite graph $K_{m,n}, ~ m \leq n$. Then ${\operatorname{reg}}(J^s) = sn + m -1$ for all $s \geq 1$. Consider the ideal $I = (T_1, T_2) \subset R = K[T_1, T_2]$ with $\deg T_1 = m$ and $\deg T_2 = n$. It follows from [@guardo-vtuyl05 Theorem 2.1] that the resolution of $I^s$ is $$\begin{aligned} \label{ci-res} 0 \to \underset{a_i \geq 1}{\bigoplus_{a_1+a_2 = s+1}}R(-a_1m-a_2n) \to \bigoplus_{a_1+a_2 = s} R(-a_1m-a_2n) \to I^s \to 0.\end{aligned}$$ Note that $J = (x_1\cdots x_m, y_{m+1}\cdots y_{m+n})$. Set $x_1\cdots x_m = T_1$ and $y_{m+1}\cdots y_{m+n} = T_2$. Then $J^s$ has the minimal free resolution as in (\[ci-res\]). If $m \leq n$, then ${\operatorname{reg}}(J^s) = sn + m - 1$. It follows from Theorem \[com.bipar00\] that in the case of the complete bipartite graph $K_{m,n}$, the stability index is $1$ and the constant term is $\tau-1$, where $\tau$ is the size of a minimum vertex cover. We now move on to study the cover ideals of bipartite graphs that are not complete. Since this is a huge class, we do not expect that a single expression may represent the regularity of powers of their cover ideals. Therefore, we restrict our attention to some of the structured subclasses of bipartite graphs. We begin our investigation by describing certain properties of the cover ideals of a certain class of Cohen-Macaulay bipartite graphs. \[thm:gen-ncb\] Let $G$ be the graph with $V(G) = \{x_1, \ldots, x_n, y_1, \ldots, y_n\}$ and $E(G) = \{\{x_1,y_j\}, \{x_i,y_i\} ~ : 1 \leq j \leq n, ~2 \leq i \leq n\}$. Let $J_G \subset R = K[x_1,\ldots,x_n,y_1,\ldots,y_n]$ be the cover ideal of $G$. Then 1. $\mu(J_G) = 2^{n-1}+1$; 2. $J_G$ is generated in degree $n$ and has a linear quotient; 3. $J_G$ has an $n$-linear resolution and hence ${\operatorname{reg}}J_G = n$; 4. $J_G^s$ has a $ns$-linear resolution and hence ${\operatorname{reg}}J_G^s = ns$; 5. $\operatorname{pdim}R/J_G = n$. $(1)$: Let $w$ denote a minimal vertex cover of $G$. If $x_1 \in w$, then one and only one of $\{x_i, y_i\}$ is in $w$. And, if $x_1 \notin w$, then $w = \{y_1, \ldots, y_n\}$. Therefore, $J_G$ is minimally generated by elements of the form $y_1\cdots y_n$ and $x_1 x_{i_1}\cdots x_{i_j} y_{i_{j+1}}\cdots y_{i_{n-1}},$ where $0 \leq j \leq n-1$ and $2 \leq i_k \leq n$ for all $k$. Therefore, $\mu(J_G) = 1+\sum_{j=0}^{n-1} {n-1 \choose j} = 2^{n-1}+1$. $(2)$: Clearly a minimal vertex cover has $n$ elements so that $J_G$ is generated in degree $n$. Write $J_G = (f_1, \ldots, f_{2^{n-1}+1})$, where $f_1 < f_2 < \cdots f_{2^{n-1}+1}$ written in the graded reverse lexicographic ordering induced by $x_1 > \cdots > x_n > y_1 > \cdots > y_n$. We prove that $J_G$ has linear quotients. Note that it is enough to prove that for an $i$ and $j < i$, there exists $k < i$ such that $f_k : f_i = u_t$ and $u_t | f_j$, where $u_t = x_t \text{ or } y_t$, [@jz10]. For $f_i$, let $A_i = \{i_j\in [n] ~ : ~ \deg_{x_{i_j}} f_i = 1\}$. Then $f_i = \prod_{r \in A}x_{r}\prod_{s \in A^c}y_s$. Let $f_j < f_i$ be given. Let $r \in A_i \setminus A_j$. Let $f_k = \frac{y_r}{x_r}f_i$. Then $f_k < f_i$ and $f_k : f_i = y_r$. Since $r \notin A_j$, $y_r | f_j$. Hence $J_G$ has linear quotients. $(3) \& (4)$: Since $J_G$ has linear quotients, $J_G^s$ has a linear resolution for all $s \geq 1$. Since $J_G^s$ is generated in degree $ns$, ${\operatorname{reg}}(J_G^s) = ns$. $(5)$: It follows from [@hh05a Theorem 3.4] that $G$ is a Cohen-Macaulay bipartite graph. Therefore, ${\operatorname{reg}}(I(G)) = \nu(G) + 1$, where $\nu(G)$ denotes the independence matching number of $G$, [@kummini]. It is easy to see that $\nu(G) = n-1$. Therefore, ${\operatorname{reg}}(I(G)) = 2n - (n-1) - 1 = n$. Since $J_G = I(G)^\vee$, by [@terai99] $\operatorname{pdim}(R/J_G) = n$. In a personal communication, we have been informed that Seyed Fakhari has proved that if $G$ is a bipartite graph, then for all $s \geq 1$, $$s \deg(J_G) \leq {\operatorname{reg}}(J_G^s) \leq (s-1)\deg(J_G) + |V(G)| - 1.$$ We note that the class of bipartite graphs discussed in Theorem \[thm:gen-ncb\] attain the lower bound. Let $U_1, \ldots, U_n$ and $V_1, \ldots, V_n$ be sets of vertices. Set $U = U_1 \cup \cdots \cup U_n$ and $V = V_1 \cup \cdots \cup V_n$. Let $G = K_{U_1,V} \cup K_{U_2,V_2} \cup \cdots \cup K_{U_n,V_n}$. For $1 \leq i, j \leq n$, set $U_i = \{x_{i_1},\ldots, x_{i_{n_i}}\}$ and $V_j = \{y_{j_1},\ldots, y_{j_{m_j}}\}$. Let $J_G$ denote the cover ideal of $G$. Then $J_G$ can be identified with the cover ideal described in Theorem \[thm:gen-ncb\] by taking $x_i = \prod_{x_{i_r} \in U_i} x_{i_r}$ and $y_j = \prod_{y_{j_s} \in V_j} y_{j_s}$. But in this case, since $\deg x_i$ is not necessarily one, the resolution need not be linear and hence the computation of regularity is non-trivial. We now study the resolution and the regularity for $n = 2$. \[thm:bipar02\] Let $U = U_1 \sqcup U_2$ and $V = V_1 \sqcup V_2$ be a collection of vertices with $|U| = n, ~|U_i| = n_i, ~ |V| = m, ~|V_i| = m_i$ and $1 \leq n_i, m_i$ for $i = 1, 2$. Let $G$ be the bipartite graph $K_{U_1, V} \cup K_{U_2, V_2}$. Let $R = K[x_1, \ldots, x_n, y_1, \ldots, y_m]$. Let $J_G \subset R$ denote the cover ideal of $G$. Then the graded minimal free resolution of $R/J_G$ is of the form: $$0 \longrightarrow R(-n-m_2) \oplus R(-m-n_1) \longrightarrow R(- (n_1+m_2)) \oplus R(-m) \oplus R(-n) \longrightarrow R \longrightarrow 0.$$ In particular, $${\operatorname{reg}}(J_G) = \max \{ n+ m_2-1, m+ n_1 -1\}$$ It can easily be seen that the cover ideal $J_G$ is generated by $g_1 = x_1 \cdots x_n,\; g_2 = y_1 \cdots y_m,$ and $g_3 = x_1\cdots x_{n_1} y_{m_1+1}\cdots y_m$. Set $X_1 = x_1 \cdots x_{n_1}; ~X_2 = x_{n_1+1}\cdots x_n; ~ Y_1 = y_1 \cdots y_{m_1}$ and $Y_2 = y_{m_1+1}\cdots y_m$. Then we can write $g_1 = X_1X_2, ~g_2 = Y_1Y_2$ and $g_3 = X_1Y_2$. Consider the minimal graded free resolution of $R/J_G$ over $R$: $$\cdots \longrightarrow F \overset{\partial_2}{\longrightarrow} R(-n) \oplus R(-m) \oplus R(- (n_1+m_2)) \overset{\partial_1}{\longrightarrow} R \longrightarrow 0,$$ where $ \partial_1(e_1)= g_1$, $ \partial_1(e_2)= g_2$, and $ \partial_1(e_3)= g_3$. Let $ae_1+be_2+ce_3 \in \ker \partial_1$. Then $aX_1X_2+bY_1Y_2+cX_1Y_2 = 0$. Solving the above equation, it can be seen that, $$\ker \partial_1 = {\operatorname{Span}}_R\{Y_2e_1 - X_2e_3, X_1e_2 - Y_1e_3\}.$$ Also, it is easily verified that these two generators are $R$-linearly independent. Hence $\ker \partial_1 \cong R^2$. Note that $\deg (Y_1e_1 - X_2e_3) = n+m_1$ and $\deg(X_1e_2 - Y_1e_3) = n_1 + m$. Therefore, we get the minimal free resolution of $R/J_G$ as: $$0 \longrightarrow R(-n-m_2) \oplus R(-m-n_1) \overset{\partial_2}{\longrightarrow} R(- (n_1+m_2)) \oplus R(-m) \oplus R(-n) \overset{\partial_1}{\longrightarrow} R \longrightarrow 0,$$ where $\partial_2(a,b) = a(Y_1e_1 - X_2e_3) + b(X_1e_2 - Y_1e_3)$. The regularity assertion follows immediately from the resolution. Now, our aim is to compute the resolution and regularity of $J_G^s$, where $J_G$ is the cover ideal discussed in Theorem \[thm:bipar02\]. For this, we first study the resolution of powers of the ideal $(X_1X_2,X_1Y_2, Y_1Y_2)$ and obtain the resolution and regularity of the cover ideal as a consequence. \[thm:two-bipartite-graph\] Let $R=K[X_1,X_2,Y_1,Y_2]$ and $J=(X_1X_2,X_1Y_2,Y_1Y_2)$ be an ideal of $R$. Then, for $s \geq 2$, the minimal free resolution of $R/J^s$ is of the form $$0 \longrightarrow R^{s\choose 2} \longrightarrow R^{2{s+1\choose 2}} \longrightarrow R^{s+2\choose 2} \longrightarrow R \longrightarrow 0,$$ and ${\operatorname{reg}}(J^s) = 2s$. Denote the generators of $J$ by $g_1=X_1X_2$, $g_2=X_1Y_2$, and $g_3=Y_1Y_2$. Note that $J$ is the edge ideal of $P_4$, the path graph on the vertices $\{X_1,X_2,Y_1,Y_2\}$. Since $P_4$ is chordal, it follows that $R/J^s$ has a linear resolution for all $s \geq 1$, [@hhz04]. Now we compute the Betti number of the $R/J^s$. Write $$(g_1, g_2, g_3)^s= ( g_1^s, g_1^{s-1}(g_2,g_3), g_1^{s-2}(g_2,g_3)^{2}, \dots , g_1(g_2,g_3)^{s-1}, (g_2,g_3)^{s})$$ where $$(g_2,g_3)^{t}= (g_2^t, g_2^{t-1}g_3 , g_2^{t-2}g_3^2, \dots, g_2 g_3^{t-1}, g_3^t).$$ For $i \geq j$, set $M_{i,j} = g_1^{s-i}g_2^{i-j}g_3^j = (X_1X_2)^{s-i}Y_2^iY_1^jX_1^{i-j}.$ It follows that $\mu(J^s) = \frac{(s+1)(s+2)}{2}$. Set $\beta_1 = \frac{(s+1)(s+2)}{2}$. Let $\{e_{p,q} \mid 0 \leq p \leq s; 0 \leq q \leq p\}$ denote the standard basis for $R^{\beta_1}$. Let $\partial_1 : R^{\beta_1} \longrightarrow R$ be the map $\partial_1(e_{p,q}) = M_{p,q}$. Since $g_i$’s are monomials, the kernel is generated by binomials of the form $m_{i,j}M_{i,j} - m_{k,l}M_{k,l}$, where $m_{p,q}$’s are monomials in $R$. Since the resolution of $R/J^s$ is linear, it is enough to find the linear syzygy relations among the generators of $\ker(\partial_1)$. To find these linear syzygies, we need to find conditions on $i,j,k,l$ such that $\frac{M_{i,j}}{M_{k,l}}$ is equal to $\frac{X_p}{Y_q}$ or $\frac{Y_q}{X_p}$ for some $p, q$. First of all, note that for such linear syzygies, $|i-k|, |j-l| \leq 1$. If $i = k$ and $j = l+1$, then $\displaystyle{\frac{M_{i,j}}{M_{i,j+1}} = \frac{g_3}{g_2} = \frac{Y_1}{X_1}}$. We get the same relation if $i = k$ and $j = l-1$. If $i = k+1$ and $j = l$, then $\displaystyle{\frac{M_{i,j}}{M_{i-1,j}} = \frac{g_2}{g_1} = \frac{Y_2}{X_2}}$. As before, $i = k-1$ yields the same relation. Therefore, the kernel is minimally generated by $$\left\{Y_1 e_{i,j} - X_1 e_{i,j+1}, X_2 e_{i,j} - Y_2 e_{i-1,j} \mid 0 \leq j < i \leq s \right\}.$$ Hence, $\mu(\ker \partial_1) = 2{s+1 \choose 2}$. Write the basis elements of $R^{2{s+1\choose 2}}$ as $$\{e_{1,i,p}, e_{2,i,q} \mid 1 \leq i \leq s, 0 \leq p < i \text{ and } 0 \leq q < i\}$$ and define $\partial_2 : R^{2{s+1\choose 2}} \longrightarrow R^{\beta_1}$ by $$\begin{aligned} \partial_2(e_{1,i,p}) & = & Y_1e_{i,p} - X_1e_{i,p+1} \\ \partial_2(e_{2,i,q}) & = & X_2e_{i,q} - Y_2e_{i-1,q}.\end{aligned}$$ By [@morey10 Proposition 3.2], $\operatorname{pdim}(R/J^s) = 3$ for all $s \geq 2$. Hence we conclude that the minimal graded free resolution of $R/J^s$ is of the form $$0 \longrightarrow R^{\beta_3} \longrightarrow R^{2{s+1 \choose 2}} \overset{\partial_2}{\longrightarrow} R^{{s+2 \choose 2}} \overset{\partial_1}{\longrightarrow} R \longrightarrow 0.$$ Therefore, $$\beta_3 - 2{s+1 \choose 2} + {s+2\choose2} - 1=0,$$ so that $\beta_3={s \choose 2}$. Now we compute the generators of the second syzygy. Again, since the resolution is linear, it is enough to compute linear generators. First, note that $$B = \{Y_2e_{1,i,j} - X_2 e_{1,i+1,j} + Y_1 e_{2,i+1,j} - X_1 e_{2,i+1,j+1} \mid 0 \leq j < i < s \} \subseteq \ker \partial_2.$$ It can easily be verified that $B$ is $R$-linearly independent. Since $\mu(\ker \partial_2) = {s \choose 2}$, $B$ generates $\ker \partial_2$. Let $\{E_{i,j} \mid 0 \leq j < i < s\}$ denote the standard basis for $R^{s \choose 2}$. Define $\partial_3 : R^{s \choose 2} \to R^{s(s+1)}$ by $$\partial_3(E_{i,j}) = Y_2e_{1,i,j} - X_2 e_{1,i+1,j} + X_1 e_{2,i+1,j} - Y_1 e_{2,i+1,j+1}.$$ Therefore, we get the minimal free resolution of $R/J^s$ as $$0 \rightarrow R^{s \choose 2} \xrightarrow{\partial_3} R^{2{s+1\choose 2}} \xrightarrow{\partial_2} R^{s+2\choose 2} \xrightarrow{\partial_1} R \longrightarrow 0.$$ Since the resolution is linear, ${\operatorname{reg}}J^s = 2s$. It is to be noted that the resolution and the projective dimension of the powers of the edge ideals of path graphs are known in the literature. In Theorem \[thm:two-bipartite-graph\], we obtain the syzygies and the Betti numbers explicitly. The knowledge of syzygies is crucial in obtaining the regularity of powers of edge ideals of certain bipartite graphs as seen below: \[cor:bipar02\] Let $U = U_1 \sqcup U_2$ and $V = V_1 \sqcup V_2$ be a collection of vertices with $|U| = n, ~|U_i| = n_i, ~ |V| = m, ~|V_i| = m_i$ and $1 \leq n_i, m_i$ for $i = 1, 2$. Let $G$ be the bipartite graph $K_{U_1, V} \cup K_{U_2, V_2}$. Let $R = K[x_1, \ldots, x_n, y_1, \ldots, y_m]$. Let $J_G \subset R$ denote the cover ideal of $G$. Then the minimal free resolution of $R/J_G^s$ is of the form: $$0 \longrightarrow R^{s\choose 2} \longrightarrow R^{2{s+1\choose 2}} \longrightarrow R^{s+2\choose 2} \longrightarrow R \longrightarrow 0.$$ Moreover, $${\operatorname{reg}}J_G^s = \max \left\{ \begin{array}{ll} (s-j)n_1 + (s-i)n_2 + jm_1 + im_2 & \text{ for } 0 \leq j \leq i \leq s\\ (s-j)n_1 + (s-i)n_2 + (j+1)m_1 + im_2 - 1& \text{ for } 0 \leq j < i \leq s\\ (s-j)n_1 + (s-i+1)n_2 + jm_1 + im_2 - 1& \text{ for } 0 \leq j < i \leq s\\ (s-j)n_1 + (s-i)n_2 + (j+1)m_1 + (i+1)m_2 - 2& \text{ for } 0 \leq j < i < s \end{array} \right. .$$ Let $R = K[x_1,\ldots,x_n,y_1,\ldots,y_m]$. Then $J_G = (x_1\cdots x_n, y_1\cdots y_m, x_1\cdots x_{n_1}y_{m_1+1}\cdots y_m)$. Set $X_1 = x_1\cdots x_{n_1}, X_2 = x_{n_1+1}\cdots x_n, Y_1 = y_1\cdots y_{m_1}$ and $Y_2 = y_{m_1+1}\cdots y_m$. Then $J_G = (X_1X_2, Y_1Y_2, X_1Y_2)$. Therefore, it follows from Theorem \[thm:bipar02\] that $J_G^s$ has the given minimal free resolution. To compute the regularity, we need to obtain the degrees of the syzygies. These shifts are given by the degrees of the minimal generators of the syzygies. Following the notation in the proof of Theorem \[thm:bipar02\], we can see that $$\begin{aligned} \deg e_{i,j} & = & (s-j)n_1 + (s-i)n_2 + jm_1 + im_2, \\ \deg e_{1,i,j} & = & (s-j)n_1 + (s-i)n_2 + (j+1)m_1 + im_2, \\ \deg e_{2,i,j} & = & (s-j)n_1 + (s-i+1)n_2 + jm_1 + im_2, \\ \deg E_{i,j} & = & (s-j)n_1 + (s-i)n_2 + (j+1)m_1 + (i+1)m_2.\end{aligned}$$ Therefore, the assertion on the regularity follows. Discussion {#disc} ---------- It has been proved by Hang and Trung, [@ht2017], that if $G$ is a bipartite graph on $n$ vertices and $J_G$ is the cover ideal of $G$, then there exists a non-negative integer $e$ such that for $s \geq n+2 $, ${\operatorname{reg}}(J_G^s) = \deg(J_G) s + e$, where $\deg(J_G)$ denote the maximal degree of minimal monomial generators of $J_G$. It follows from Theorem \[com.bipar00\] that if $G = K_{n,m}$ with $n \geq m$, then $e = m-1$ and the index of stability is $1$. If the graph is not a complete bipartite graph, then $e$ does not uniformly represent a combinatorial invariant associated to the graph as can be seen in the computations below. We compute the polynomial ${\operatorname{reg}}(J_G^s)$ for some classes of bipartite graphs that are considered in Corollary \[cor:bipar02\]. We see that $e$ depends on the relation between the integers $n_1, n_2, m_1$ and $m_2$. To find the polynomial ${\operatorname{reg}}(J_G^s)$, we need to identify which of those relations give rise to the regularity. If we assume that all those integers are at least one, then $\deg e_{i,j}$ will not play a role in deciding the regularity. Note that $\deg e_{1,i,j} - \deg e_{2,i,j} = m_1 - n_2$ and $\deg e_{2,i,j} - \deg E_{i,j} = n_2 - (m_1+m_2)$. Therefore, depending on the relation between these integers, one can decide the expression for the regularity of $J_G^s$. We first deal with two simple cases. The first case shows that the constant term can as well be zero. 1. If $m_1 = m_2 = 1$, then it can be seen that for $s \geq 2$, $${\operatorname{reg}}(J_G^s) = \max\{(s-j)n_1 + (s-i+1)n_2 + jm_1 + i m_2 - 1 ~ : ~ 0 \leq j < i \leq s\}.$$ Since $(s-j)n_1 + (s-i+1)n_2 + jm_1 + i m_2 - 1 = s(n_1+n_2) + j(1-n_1) + i(1-n_2) + n_2-1$ and $n_1 \geq 1, ~ n_2 \geq 1$, this expression attains maximum when $i$ and $j$ attain minimum, i.e., if $j = 0$ and $i = 1$. Therefore, ${\operatorname{reg}}(J_G^s) = n s$. Thus, in this case, $e = 0$. It can also be noted that, since $n \geq 2$, it follows from Theorem \[thm:bipar02\] that ${\operatorname{reg}}(J_G) = n$. Therefore, in this case, the stability index is also equal to $1$. 2. If $n_1 = n_2 = m_1 = m_2 = \ell > 1$, then for $s \geq 2$, $${\operatorname{reg}}(J_G^s) = \max\{(s-j)n_1 + (s-i)n_2 + (j+1)m_1 + (i+1) m_2 - 2 ~ : ~ 0 \leq j < i \leq s\}.$$ Therefore, ${\operatorname{reg}}(J_G^s) = 2\ell s + (2\ell-2)$ and hence $e = 2\ell - 2$. Note that in this case, the stability index is $2$. Now we discuss some more cases and compute the polynomial ${\operatorname{reg}}(J_G^s)$, $s \geq 2$. Assume that $n_2 = m_1$. 1. $n_1 \geq m_2 \geq n_2=m_1$: Note that, in this case, $\deg(J_G) = n_1+m_2$. We have $${\operatorname{reg}}(J_G^s) = \max\{ (s-j)n_1+(s-i)n_2 + (j+1)m_1 + (i+1)m_2 -2~ : ~ 0 \leq j < i < s\}.$$ Since $(s-j)n_1 + (s-i)n_2 + (j+1)m_1 + (i+1)m_2 -2 = s(n_1+n_2) + j(m_1-n_1) + i(m_2-n_2) + m_1+m_2-2$. Since $n_1 \geq m_1$ and $m_2 \geq n_2$, the above expression attains the maximum when $i$ attains the maximum and $j$ attains the minimum, i.e., if $i = s-1$ and $j = 0$. Therefore, ${\operatorname{reg}}(J_G^s) = (n_1+m_2)s + (n_2+m_1-2)$ and hence $e = n_2+m_1-2$. 2. $n_1 \geq n_2 = m_1 \geq m_2$: In this case, $\deg(J_G) = n$ and $${\operatorname{reg}}(J_G^s) = \max\{ (s-j)n_1+(s-i)n_2 + (j+1)m_1 + (i+1)m_2 -2~ : ~ 0 \leq j < i < s\}.$$ As in the previous case, one can conclude that the maximum is attained when $i = 1$ and $j = 0$. Therefore, ${\operatorname{reg}}(J_G^s) = n s + (2m_2 - 2)$. Thus $e = 2m_2-2$. 3. $n_2 = m_1 \geq n_1 \geq m_2$: As done earlier, one can conclude that ${\operatorname{reg}}(J_G)^s = n s + (2m_2 - 2)$ and hence $e = 2m_2-2$. 4. Similarly analyzing the relation between these integers, one can obtain the regularity expressions as follows: $\mathbf{n_2 = m_1}$ $\mathbf{{\operatorname{reg}}(J_G^s)}, ~ ~ s \geq 2$ ------------------------------- ------------------------------------------------------ $n_2 = m_1 \leq n_1 \leq m_2$ $(n_1+m_2) s + (m_1+n_2-2)$ $n_1 \leq n_2=m_1 \leq m_2$ $(m_1+m_2)s + (n_1+m_2-2)$ $n_1 \leq m_2\leq n_2 = m_1$ $(m_1+m_2)s + (n_1+m_2-2)$ It can be noted that in the case $n_2 = m_1$, the stability index is $2$. If $G = K_{U_1,V} \cup K_{U_2,V_2} \cup K_{U_3,V_3}$, for some set of vertices $U_i, V_i$, then one can still describe the complete resolution and the regularity of $J_G^s$ using a similar approach. However, the resulting syzygies are not so easy to describe though the generating sets are similar. Therefore, we restrict ourselves to the above discussion. Complete Multipartite Graphs ============================ In this section our goal is to understand the resolution and regularity of the powers of cover ideals of complete $m$-partite graphs. Let $G$ be a complete $m$-partite graph and let $J_G$ be the cover ideal of $G$. The main idea in constructing the resolution of $J_G$ is to identify $J_G$ with the cover ideal of the complete graph $K_m$ on $m$ vertices. Then one replaces the variables by the product of a set of variables from each partitions of $m$-partite graphs to get the cover ideal $J_G$. Let $R=K[x_1,\ldots,x_m]$. It is known that the cover ideal $J_G$ of complete graph $G=K_m$ is generated by all squarefree monomials $x_1x_2\cdots \hat{x_i} \cdots x_m$ of degree $m-1$. Moreover one can also identify this cover ideal with the squarefree Veronese ideal $I=I_{m,m-1}$, and hence it is a polymatroidal ideal, [@hh05]. Let $I \subset R$ be an ideal of $R$. The ideal $I$ is said to have linear quotients if there exists an ordered system of homogeneous generators $f_1,\ldots,f_{\ell}$ of $I$ such that for all $j=1,\ldots,\ell$ the colon ideals $(f_1,\ldots,f_{j-1}):f_j$ are generated by a subset of $\{ x_1,\ldots,x_n\}$. It was proved by Conca and Herzog that a polymatroidal ideal has linear quotients with respect to the reverse lexicographical order of the generators [@ch03 Proposition 5.2]. Herzog and Hibi proved that if $I$ is generated in degree $d$ and has linear quotients, then $I$ has a $d$-linear resolution, [@Herzog'sBook Proposition 8.2.1]. Furthermore product of polymatroidal monomial ideals are polymatroidal [@ch03 Theorem 5.3]. Hence powers of polymatroidal monomial ideals are polymatroidal. Therefore we have the following: \[rem:lin res powers\] The cover ideal $J_G$ of complete graph $G=K_m$ has linear quotients and hence has linear resolution. Moreover $J_G^s$ has linear resolution for all $s \geq 1$. If $I$ is an ideal of $R$ all of whose powers have linear resolution, then depth $R/I^k$ is a non-increasing function of $k$ and $\operatorname{depth}R/I^k$ is constant for all $k \gg 0$, [@hh05 Proposition 2.1]. Further, we have: [@hh05 Corollary 3.4] \[rem:depth powers\] Let $R=K[x_1,\cdots,x_m]$ and $J_G$ be the cover ideal of $G = K_m$. Then $$\operatorname{depth}R/{J_G^s} = \max \{ 0, m-s-1 \}.$$ In particular, $\operatorname{depth}R/{J_G^s} =0$ for all $s \geq m-1$. We begin our investigation by studying the resolution of powers of cover ideal of $K_3$. Complete tripartite: -------------------- It has been shown by Beyarslan et al., [@selvi_ha Theorem 5.2], that ${\operatorname{reg}}(I(C_3)^s) = 2s$ for all $s \geq 2$. We first describe the graded minimal free resolution of $I(C_3)^s$ for all $s \geq 1$. We also obtain the Hilbert series of the powers. \[C3\] Let $R = K[x_1, x_2, x_3]$ and $I = (x_1x_2, x_1x_3, x_2x_3)$. Then the graded minimal free resolution of $R/I$ is of the form: $$0 \to R(-3)^2 \to R(-2)^3 \to R \to 0.$$ For $s \geq 2,$ the graded minimal free resolution of $R/I^s$ is of the form: $$0 \to R(-2s-2)^{s \choose 2} \to R(-2s-1)^{2{s+1\choose 2}} \to R(-2s)^{s+2\choose 2} \to R \to 0$$ so that ${\operatorname{reg}}_R (I^s)=2s$. Moreover, the Hilbert series of $R/I^s$ is given by $$\begin{aligned} H(R/I^s,t) = \frac{1+2t+3t^2+ \cdots +2st^{2s-1} - \left({s +2 \choose 2}-2s-1 \right)t^{2s} }{(1-t)}.\end{aligned}$$ It is clear that the resolution of $I$ is as given in the assertion of the theorem. Therefore ${\operatorname{reg}}(R/I) = 1$. It follows from [@hhz04], that $I^s$ has a linear minimal free resolution for all $s \geq 1$. Note that by [@cms02 Lemma 3.1], $\operatorname{depth}R/I^s = 0$ for all $s \geq 2$ so that $\operatorname{pdim}R/I^s = 3$ for all $s \geq 2$. Hence, the minimal free resolution of $R/I^s$ is of the form: $$0 \to R(-2s-2)^{\beta_3} \overset{\partial_3}{\longrightarrow} R(-2s-1)^{\beta_2} \overset{\partial_2}{\longrightarrow} R(-2s)^{\beta_1} \overset{\partial_1}{\longrightarrow} R \to 0$$ Now we describe completely the minimal free resolution of $R/I^s$ for $s \geq 2$. Let $g_1 = x_1x_2, g_2 = x_1x_3$ and $g_3 = x_2x_3$. Write $I^s = (g_1^s, g_1^{s-1}(g_2, g_3), \ldots, g_1(g_2,g_3)^{s-1}, (g_2,g_3)^s).$ The generators of $I^s$ are of the form $g_1^{\ell_1}g_2^{\ell_2}g_3^{\ell_3}$, where $0 \leq \ell_1,\ell_2,\ell_3 \leq s$ and $\ell_1 + \ell_2 +\ell_3 =s$. Denote by $f_{\ell_1,\ell_2,\ell_3}=g_1^{\ell_1}g_2^{\ell_2}g_3^{\ell_3}$. Then $f_{\ell_1,\ell_2,\ell_3}=x_1^{s-\ell_3}x_2^{s- \ell_2}x_3^{s- \ell_1}$. It is easy to see that $\mu(I^s)= {s+2 \choose 2}$, since the minimal number of generating elements of the monomial ideal $I^s$ is same as the total number of non-negative integral solution of $\ell_1 + \ell_2 +\ell_3 =s$, which is ${s+2 \choose 2}$. Let $\{ e_{\ell_1,\ell_2,\ell_3} ~ : ~ 0 \leq \ell_1,\ell_2,\ell_3 \leq s; \ell_1 + \ell_2 + \ell_3 = s \}$ denote the standard basis for $R^{s+2 \choose 2}$ and consider the map $\partial_1 : R^{s+2 \choose 2} \to R$ defined by $\partial_1(e_{\ell_1,\ell_2,\ell_3}) =f_{\ell_1,\ell_2,\ell_3}$. We find the minimal generators for $\ker \partial_1$. Since $f_{\ell_1,\ell_2,\ell_3}$’s are monomials, the kernel is generated by binomials of the form $m_{\ell_1,\ell_2,\ell_3}f_{\ell_1,\ell_2,\ell_3} - m_{t_1,t_2,t_3}f_{t_1,t_2,t_3}$, where $m_{i,j,k}$’s are monomials in $R$. Also, since the minimal free resolution is linear, the kernel is generated in degree $1$. Note that $\displaystyle{\frac{f_{\ell_1,\ell_2,\ell_3}}{f_{t_1,t_2,t_3}} = x_1^{t_3-\ell_3}x_2^{t_2-\ell_2}x_3^{t_1-\ell_1}}$. Hence, for $\displaystyle{\frac{f_{\ell_1,\ell_2,\ell_3}}{f_{t_1,t_2,t_3}}}$ to be a linear fraction, $|t_i - \ell_i| \leq 1$ for $i = 1,2,3$. Let $t_3=\ell_3$, $t_2=\ell_2+1$ and $t_1=\ell_1-1$. The corresponding linear syzygy relation is $$\begin{aligned} \label{thm:3-partite-2ndsyz1} x_3 \cdot f_{\ell_1,\ell_2,\ell_3} - x_2 \cdot f_{\ell_1-1,\ell_2+1,\ell_3} =0,\end{aligned}$$ where $ 1 \leq \ell_1 \leq s$, and $0 \leq \ell_2,\ell_3 \leq s-1$. Note that the number of such relations is equal to the number of integral solution to $(\ell_1-1) + \ell_2 + \ell_3 = s$, i.e., ${s+1 \choose 2}$. Similarly, if $t_3=\ell_3+1$, $t_2=\ell_2-1$ and $t_1=\ell_1$, then we get the corresponding linear syzygy relation as $$\begin{aligned} \label{thm:3-partite-2ndsyz2} x_2 \cdot f_{\ell_1,\ell_2,\ell_3} - x_1 \cdot f_{\ell_1,\ell_2-1,\ell_3+1} =0,\end{aligned}$$ where $ 0 \leq \ell_1, \ell_3 \leq s-1$, and $1 \leq \ell_2 \leq s$. Note that the linear syzygy relation obtained by fixing $\ell_2$ and taking $|\ell_i - t_i| = 1$ for $i = 1, 2$ $$x_3 \cdot f_{\ell_1,\ell_2,\ell_3} -x_1 \cdot f_{\ell_1-1,\ell_2,\ell_3+1} =0$$ can be obtained from Equations $(\ref{thm:3-partite-2ndsyz1}),\; (\ref{thm:3-partite-2ndsyz2})$ by setting the $\ell_i$’s appropriately. Therefore, $$\ker \partial_1 = \langle x_3 \cdot e_{\ell_1,\ell_2,\ell_3} - x_2 \cdot e_{\ell_1-1,\ell_2+1,\ell_3}, ~ x_2 \cdot e_{\ell_1-1,\ell_2+1,\ell_3} - x_1 \cdot e_{\ell_1-1,\ell_2,\ell_3+1} ~ : ~ 1 \leq \ell_1 \leq s, \;0 \leq \ell_2,\ell_3 \leq s-1 \rangle.$$ Since there are ${s+1 \choose 2}$ minimal generators of each type in the list above, $\mu(\ker\partial_1) = \beta_2 = 2{s+1 \choose 2}$. Write the standard basis of $R^{2{s+1\choose 2}}$ as $$B_2 = \left\{ \begin{array}{ll} e_{1,(\ell_2+1,\ell_3),\ell_1-1 }\\ e_{2,(\ell_1,\ell_2),\ell_3 } \end{array} : \; 1 \leq \ell_1 \leq s,\; 0 \leq \ell_2,\ell_3 \leq s-1 \right\}$$ and define $\partial_2 : R^{2{s+1\choose 2}} \to R^{s+2 \choose 2}$ by $$\begin{aligned} \partial_2(e_{1,(\ell_2+1,\ell_3),\ell_1-1 }) & = & x_2 \cdot e_{\ell_1-1,\ell_2+1,\ell_3} - x_1 \cdot e_{\ell_1-1,\ell_2,\ell_3+1},\\ \partial_2(e_{2,(\ell_1,\ell_2),\ell_3 }) & = & x_3 \cdot e_{\ell_1,\ell_2,\ell_3} - x_2 \cdot e_{\ell_1-1,\ell_2+1,\ell_3}.\end{aligned}$$ From the equation, $\beta_3 - {2(s+1) \choose 2} + {s+2 \choose 2} -1 = 0$, it follows that $\beta_3 = {s \choose 2}$. Now, we describe $\ker \partial_2$. For $ 2 \leq \ell_1 \leq s \text{ and } 0 \leq \ell_2, \ell_3 \leq s-2$, set $$E_{\ell_1,\ell_2,\ell_3}= -x_3 \cdot e_{2,(\ell_1-1,\ell_2),\ell_3 +1 } + x_2 \cdot e_{2,(\ell_1-1,\ell_2+1),\ell_3 } +x_2 \cdot e_{1,(\ell_2+2,\ell_3),\ell_1-2 } -x_3 \cdot e_{1,(\ell_2+1,\ell_3),\ell_1-1 }.$$ Note that $$B_3 = \{ E_{\ell_1,\ell_2,\ell_3} ~ : ~ 2 \leq \ell_1 \leq s,\; 0 \leq \ell_2,\ell_3 \leq s-2\} \subset \ker \partial_2.$$ Since $B_3$ is $R$-linearly independent and $|B_3| = {s \choose 2}$, $\ker \partial_2 = \langle B_3 \rangle$. Let $\{ H_{\ell_1,\ell_2,\ell_3} ~ : ~ 2 \leq \ell_1 \leq s,\; 0 \leq \ell_2,\ell_3 \leq s-2\}$ denote the standard basis of $R^{s \choose 2}$. Define $\partial_3 : R^{s \choose 2} \to R^{2(s+1) \choose 2}$ by $ \partial_3(H_{\ell_1,\ell_2,\ell_3}) = E_{\ell_1,\ell_2,\ell_3}$ which completes the description of the resolution. It can also be seen that $\deg e_{\ell_1,\ell_2,\ell_3} = 2s, ~ \deg e_{1,(\ell_2+1,\ell_3),\ell_1-1 } = \deg e_{2,(\ell_1,\ell_2),\ell_3 } = 2s+1$ and $\deg H_{\ell_1,\ell_2,\ell_3} = 2s+2$. Therefore we get the minimal graded free resolution of $R/I^s$ as $$0 \to R(-2s-2)^{s \choose 2} \overset{\partial_3}{\longrightarrow} R(-2s-1)^{2(s+1) \choose 2} \overset{\partial_2}{\longrightarrow} R(-2s)^{s+2 \choose 2} \overset{\partial_1}{\longrightarrow} R \to 0.$$ Therefore, the Hilbert series of $R/I^s$ is given by $$\begin{aligned} H(R/I^s, t) & = & \frac{1-{s+2 \choose 2}t^{2s} + 2{s+1 \choose 2}t^{2s+1} - {s \choose 2}t^{2s+2}}{(1-t)^3}\\ & = & (1-t)^{-2}\frac{(1-{s+2 \choose 2}t^{2s} + 2{s+1 \choose 2}t^{2s+1} - {s \choose 2}t^{2s+2})}{(1-t)}.\end{aligned}$$ By expanding $(1-t)^{-2}$ in the power series form and multiplying with the numerator, we get the required expression. We now proceed to compute the minimal graded free resolution of powers of complete tripartite graphs. \[3-partite-notn\] Let $G$ denote a complete tripartite graph with $V(G) = V_1 \sqcup V_2 \sqcup V_3$ and $E(G) = \{\{a,b\} ~ : ~ b \in V_i, b \in V_j, i \neq j \}$. Set $V_1 = \{x_1, \ldots, x_\ell\}, ~ V_2 = \{y_1, \ldots, y_m \}$ and $V_3 = \{z_1, \ldots, z_n\}$. Let $J_G$ denote the vertex cover ideal of $G$. Let $X = \prod_{i=1}^\ell x_i, ~ Y = \prod_{j=1}^m y_i$ and $Z = \prod_{k=1}^n z_i$. It can be seen that $J_G =(XY, XZ, YZ)$. \[3-partite\] Let $R = K[x_1, \ldots x_{m_1}, y_1,\ldots,y_{m_2}, z_1, \ldots, z_{m_3}]$. Let $G$ be a complete tripartite graph as in Notation \[3-partite-notn\]. Let $J_G \subset R$ denote the cover ideal of $G$. Then for all $s \geq 2$, the minimal free resolution of $R/J_G^s$ is of the form: $$0 \to R^{s \choose 2} \overset{\partial_3}{\longrightarrow} R^{2(s+1) \choose 2} \overset{\partial_2}{\longrightarrow} R^{s+2 \choose 2} \overset{\partial_1}{\longrightarrow} R \to 0.$$ Set $\alpha = (s-\ell_3)m_1 +(s-\ell_2)m_2+ (s-\ell_1)m_3$. Then for all $s \geq 2$, $${\operatorname{reg}}(J_G^s) = \max \left\{ \begin{array}{ll} \alpha, & \text{ for } 0 \leq \ell_1, \ell_2,\ell_3 \leq s, \\ \alpha + m_3 -1, & \text{ for } 1 \leq \ell_1 \leq s,\; 0 \leq \ell_2,\ell_3 \leq s-1,\\ \alpha + 2m_3 -2, & \text{ for } 2 \leq \ell_1 \leq s,\; 0 \leq \ell_2,\ell_3 \leq s-2. \end{array} \right.$$ Taking $X = x_1, ~Y = x_2$ and $Z = x_3$ in Theorem \[C3\], it follows that the minimal free resolution of $S/J_G^s$ is of the given form: $$0 \to R^{s \choose 2} \overset{\partial_3}{\longrightarrow} R^{2(s+1) \choose 2} \overset{\partial_2}{\longrightarrow} R^{s+2 \choose 2} \overset{\partial_1}{\longrightarrow} R \to 0.$$ We now compute the degrees of the generators and hence obtain the regularity. Let $ \deg X_1 = m_1, \deg X_2=m_2,$ and $\deg X_3=m_3$. Then it follows that $$\deg e_{\ell_1,\ell_2,\ell_3} = \deg \left( X_1^{s-\ell_3}X_2^{s- \ell_3}X_3^{s- \ell_1} \right) = (s-\ell_3)m_1 +(s-\ell_2)m_2+ (s-\ell_1)m_3.$$ We observe that $$\begin{aligned} & \deg e_{1,(\ell_2+1,\ell_3),\ell_1-1 } = \deg e_{2,(\ell_1,\ell_2),\ell_3 } = \deg e_{\ell_1,\ell_2,\ell_3} + \deg X_3, \\ & \deg H_{\ell_1,\ell_2,\ell_3} = \deg e_{\ell_1,\ell_2,\ell_3} + 2 \deg X_3.\end{aligned}$$ Therefore, $${\operatorname{reg}}(J_G^s) = \max \left\{ \begin{array}{ll} \deg e_{\ell_1,\ell_2,\ell_3}, & \text{ for } 0 \leq \ell_1, \ell_2, \ell_3 \leq s, \\ \deg e_{\ell_1,\ell_2,\ell_3} + \deg(X_3)-1, & \text{ for } 1 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3 \leq s-1, \\ \deg e_{\ell_1,\ell_2,\ell_3} + 2\deg(X_3)-2, & \text{ for } 2 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3 \leq s-2, \end{array} \right.$$ where $ \ell_1 + \ell_2 +\ell_3 =s$. Let $\alpha = \deg e_{\ell_1,\ell_2,\ell_3}$. Then the regularity assertion follows. Observe that the above expression for ${\operatorname{reg}}(J_G^s)$ is not really in the form $as + b$. Given a graph, one can derive this form from the above expression. Depending on the size of partitions, the leading coefficient and the constant term will change. For example, suppose the graph is unmixed, i.e., all the partitions are of same cardinality. Then ${\operatorname{reg}}(J_G^s) = 2\ell s + (2\ell-2)$ for all $s \geq 2$, where $\ell = m_1 = m_2 = m_3$. Note also that the stabilization index in this case is $2$. As done in Discussion (\[disc\]), one can derive various expressions for ${\operatorname{reg}}(J_G^s)$ for different cases as well. Consider the arithmetic progression $m_1 = m+2r$, $m_2 =m$, and $ m_3 =m+r$: Let $m,r$ be any two positive integers. Let $m_1 = m+2r$, $m_2 =m$, and $ m_3 =m+r$ in Theorem \[3-partite\]. Then for all $s \geq 2$, we have $${\operatorname{reg}}(J_G^s) = s(2m+3r)+2m-2.$$ By Theorem \[3-partite\], $ \alpha = (s-\ell_3)m_1 +(s-\ell_2)m_2+ (s-\ell_1)m_3$. Hence we get $$\alpha = s(3m+3r) -\ell_3 (m+2r) -\ell_2 (m) -\ell_1 (m+r).$$ Using Theorem \[3-partite\], we have for all $s \geq 2$, $${\operatorname{reg}}(J_G^s) = \max \left\{ \begin{array}{ll} \alpha, & \text{ for } 0 \leq \ell_1, \ell_2, \ell_3 \leq s, \\ \alpha + (m+r)-1, & \text{ for } 1 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3 \leq s-1, \\ \alpha + 2(m+r)-2, & \text{ for } 2 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3 \leq s-2, \end{array} \right.$$ where $ \ell_1 + \ell_2 +\ell_3=s$. Since regularity ${\operatorname{reg}}(J_G^s)$ is maximum of all the numbers, we need to maximize the value of $\alpha$. For this to happen, negative terms in $\alpha$ should be minimum. The coefficient of $\ell_3$ is largest among the negative terms in $\alpha$, so $\ell_3$ should be assigned the least value. After $\ell_3$, assign the minimum value to $\ell_1$, and finally take $ \ell_2 = s-\ell_1 -\ell_3 $. For example, to get the maximum of $\alpha$ when $2 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3 \leq s-2$, put $\ell_3=0,\ell_1=2$, and $\ell_2=s-2$. We get for all $s \geq 2$ $${\operatorname{reg}}(J_G^s) = \max \left\{ \begin{array}{ll} s (2m+3r), & \text{ for } 0 \leq \ell_1, \ell_2, \ell_3, \ell_4 \leq s, \\ s (2m+3r)+m-1, & \text{ for } 1 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-1, \\ s (2m+3r)+2m-2, & \text{ for } 2 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-2. \end{array} \right.$$ For all $m \geq 1$, and for all $s \geq 2$, we get $${\operatorname{reg}}(J_G^s) = s (2m+3r)+2m-2.$$ Complete 4-partite graphs ------------------------- We now describe the resolution and regularity of powers of cover ideals of $4$-partite graphs. For this purpose, we first study the resolution of powers of cover ideal of the complete graph $K_4$. \[C4\] Let $R = K[x_1, x_2, x_3, x_4]$ and $I = (x_1x_2x_3, x_1x_2x_4, x_1x_3x_4, x_2x_3x_4)$. Then, for $s \geq 3$, the minimal graded free resolution of $R/I^s$ is of the form $$0 \to R(-3s-3)^{\beta_4} {\longrightarrow} R(-3s-2)^{\beta_3} {\longrightarrow} R(-3s-1)^{\beta_2} {\longrightarrow} R(-3s)^{\beta_1} {\longrightarrow} R \to 0$$ where $$\beta_1 = { s+3 \choose 3}, ~ \beta_2 = 3{ s+2 \choose 3},~\beta_3 = 3{ s+1 \choose 3}, \text{ and } \beta_4 = { s \choose 3}.$$ In particular, ${\operatorname{reg}}_R (R/I^s)=3s$. Moreover the Hilbert series of $R/I^s$ is given by $$\begin{aligned} H(R/I^s,t) & = & \frac{1+2t+3t^2+4t^3+ \cdots +3st^{3s-1} - \left({s +3 \choose 3}-3s-1 \right)t^{3s} + {s \choose 3} t^{3s+1}}{(1-t)^2} \\ & = & \frac{ \sum_{i=0}^{3s-1} (i+1)t^i - (3s+1 - {s +3 \choose 3} )t^{3s} + {s \choose 3} t^{3s+1}}{(1-t)^2}.\end{aligned}$$ By Remark \[rem:lin res powers\], $I^s$ has linear resolution for all $s \geq 1$. Moreover, by Remark \[rem:depth powers\], $\operatorname{depth}R/I^s = 0$ for all $s \geq 3$ and hence $\operatorname{pdim}R/I^s = 4$ for all $s \geq 3$. Therefore, the minimal graded free resolution of $R/I^s$, for $s \geq 3$, is of the form $$0 \to R(-3s-3)^{\beta_4} \overset{\partial_4}{\longrightarrow} R(-3s-2)^{\beta_3} \overset{\partial_3}{\longrightarrow} R(-3s-1)^{\beta_2} \overset{\partial_2}{\longrightarrow} R(-3s)^{\beta_1} \overset{\partial_1}{\longrightarrow} R \to 0.$$ Let $g_1 = x_1x_2x_3, g_2 = x_1x_2x_4, g_3=x_1x_3x_4$ and $g_4 = x_2x_3x_4$. The minimal generators of $I^s$ are of the form $g_1^{\ell_1}g_2^{\ell_2}g_3^{\ell_3}g_4^{\ell_4}$ where $0 \leq \ell_i \leq s$ for every $i$ and $\ell_1 + \ell_2 +\ell_3 + \ell_4=s$. The number of non-negative integral solution to the linear equation $\ell_1 + \ell_2 +\ell_3 + \ell_4=s$ is ${s+3 \choose 3}$. Hence we have $\mu(J^s)= \beta_1 = {s+3 \choose 3}$. Note that $g_1^{\ell_1}g_2^{\ell_2}g_3^{\ell_3}g_4^{\ell_4}=x_1^{s-\ell_4}x_2^{s- \ell_3}x_3^{s- \ell_2}x_4^{s- \ell_1} $ with $0 \leq \ell_1,\ell_2,\ell_3,\ell_4 \leq s$. Set $f_{\ell_1,\ell_2,\ell_3,\ell_4}=x_1^{s-\ell_4}x_2^{s- \ell_3}x_3^{s- \ell_2}x_4^{s- \ell_1} $. Let $$\{e_{\ell_1,\ell_2,\ell_3,\ell_4} ~ : ~ 0 \leq \ell_1,\ell_2,\ell_3,\ell_4 \leq s \text{ and } \ell_1 + \ell_2 +\ell_3 + \ell_4=s \}$$ denote the standard basis of $R^{s+3 \choose 3}$ and consider the map $\partial_1 : R^{s+3 \choose 3} \to R$ defined by $\partial_1(e_{\ell_1,\ell_2,\ell_3,\ell_4}) = f_{\ell_1,\ell_2,\ell_3,\ell_4}$. Now we find the minimal generators for $\ker \partial_1$. Let $f_{\ell_1,\ell_2,\ell_3,\ell_4}, f_{t_1,t_2,t_3,t_4}$ be any two minimal monomial generators of $I^s$. It is known that the kernel of $\partial_1$ is generated by binomials of the form $m_{\ell_1,\ell_2,\ell_3,\ell_4}f_{\ell_1,\ell_2,\ell_3,\ell_4} - m_{t_1,t_2,t_3,t_4}f_{t_1,t_2,t_3,t_4}$, where $m_{i,j,k,l}$’s are monomials in $R$. Since the syzygy is generated by linear binomials, we need to find conditions on $\ell_1,\ell_2,\ell_3, \ell_4$ such that $\frac{f_{\ell_1,\ell_2,\ell_3,\ell_4}}{f_{t_1,t_2,t_3,t_4}}$ is equal to $\frac{x_i}{x_j}$ for some $i,j$. Observe that $$\frac{f_{\ell_1,\ell_2,\ell_3,\ell_4}}{f_{t_1,t_2,t_3,t_4}} = x_1^{t_4-\ell_4} x_2^{t_3-\ell_3} x_3^{t_2-\ell_2} x_4^{t_1-\ell_1}.$$ Suppose $t_4=\ell_4 +1$, $t_{3}=\ell_3-1$ and $t_j=\ell_j$ for $j=1,2$. Then we get the linear syzygy relation $$\begin{aligned} \label{eqn:beta_2 relations00} x_{2} f_{\ell_1,\ell_2,\ell_3,\ell_4} - x_1 f_{\ell_1,\ell_{2}, \ell_3-1, \ell_4 +1}=0.\end{aligned}$$ Fixing $t_3$ and $t_4$, with $1 \leq t_3 + t_4 \leq s$, there are as many linear syzygies are there as the number of solutions of $t_1+t_2 = s-(t_3+t_4)$. Therefore, for the pair $(t_3,t_4)$, there are ${s+1 \choose 2} + {s \choose 2} + \cdots + {2 \choose 2} = {s+2 \choose 3}$ number of solutions. Similarly for each pair $(t_4, t_2), (t_4, t_1), (t_3, t_2), (t_3, t_1)$ and $(t_2,t_1)$, we get ${s+2 \choose 3}$ linear syzygies. Note that the syzygies $x_4f_{t_1+1,t_2-1,t_3,t_4} - x_3f_{t_1,t_2,t_3,t_4}$ and $x_3f_{t_1,t_2,t_3,t_4} - x_2f_{t_1,t_2-1,t_3+1,t_4}$ give rise to another linear syzygy $x_4f_{t_1+1,t_2-1,t_3,t_4} -x_2f_{t_1,t_2-1,t_3+1,t_4}$. The same linear syzygy can also be obtained from a combination of linear syzygies that arise out of the pairs $(t_1,t_4)$ and $(t_3,t_4)$. Therefore, to get a minimal generating set, we only need to consider linear syzygies corresponding to the pairs $(t_1,t_2), ~(t_2,t_3)$ and $(t_3,t_4)$. For each such pair, we have ${s+2 \choose 3}$ number of linear syzygies. Hence $\beta_2 = 3 {s+2 \choose 3}$. Write the basis elements of $R^{3{s+2\choose 3}}$ as $$B_2 = \left\{ \begin{array}{ll} e_{(1,\ell_1-1,\ell_2),\ell_3+1,\ell_4}\\ e_{(2,\ell_1-1,\ell_4),\ell_2+1,\ell_3}\\ e_{(3,\ell_3,\ell_4),\ell_1,\ell_2} \end{array} : \; 1 \leq \ell_1 \leq s,\; 0 \leq \ell_2,\ell_3,\ell_4 \leq s-1 \right\}$$ and define $\partial_2 : R^{3{s+1\choose 2}} \longrightarrow R^{\beta_1}$ by $$\begin{aligned} \partial_2(e_{(1,\ell_1-1,\ell_2),\ell_3+1,\ell_4}) & = & x_{2} e_{\ell_1-1,\ell_2,\ell_3+1,\ell_4} - x_1 e_{\ell_1-1,\ell_{2}, \ell_3, \ell_4 +1}; \\ \partial_2(e_{(2,\ell_1-1,\ell_4),\ell_2+1,\ell_3}) & = & x_{3} e_{\ell_1-1,\ell_2+1,\ell_3,\ell_4} - x_2 e_{\ell_1-1,\ell_{2}, \ell_3+1, \ell_4};\\ \partial_2(e_{(3,\ell_3,\ell_4),\ell_1,\ell_2}) & = & x_{4} e_{\ell_1,\ell_2,\ell_3,\ell_4} - x_3 e_{\ell_1-1,\ell_{2}+1, \ell_3, \ell_4}.\end{aligned}$$ Now we decipher the Betti numbers $\beta_3$ and $\beta_4$ to complete the resolution. The Hilbert series of $R/I^s$ is $$H(R/I^s,t) = \frac{{ \left( 1- \beta_1 t^{3s} + \beta_2 t^{3s+1} - \beta_3t^{3s+2} +\beta_4t^{3s+3} \right)}}{(1-t)^4}=\frac{p(t)}{(1-t)^4}.$$ Since $\dim R/I^s=2$, the polynomial $p(t)$ has a factor $(1-t)^2$. Note that $(1-t)^2$ is a monic polynomial of degree $2$, hence we can write $p(t)= (1-t)^2 \cdot q(t)$ where $$\begin{aligned} \label{eqn:q(t) polynomial} q(t)=\beta_4 t^{3s+1}+ (2\beta_4-\beta_3)t^{3s}+a_{3s-1}t^{3s-1}+\cdots + a_1 t + 1.\end{aligned}$$ On the other hand, we also have $$\begin{aligned} \label{eqn:q(t) polynomial from HS} \frac{\left( 1- \beta_1t^{3s} + \beta_2t^{3s+1} - \beta_3t^{3s+2} +\beta_4t^{3s+3} \right)}{(1-t)^{2}} = \end{aligned}$$ $$(1+2t+3t^2+4t^3+\cdots+(n+1)t^n+\cdots) \left( 1- \beta_1t^{3s} + \beta_2t^{3s+1} - \beta_3t^{3s+2} +\beta_4t^{3s+3} \right)$$ $$=(1+2t+3t^2+\cdots+(3s)t^{3s-1}+ (3s+1 - \beta_1 ) t^{3s}+ (3s+2 -2\beta_1 +\beta_2)t^{3s+1} + a_{j}t^{j}+\cdots.$$ For the expressions in Equations $(\ref{eqn:q(t) polynomial})$ and $(\ref{eqn:q(t) polynomial from HS})$ to be equal, they should agree coefficient-wise. In particular, we should have that $$\beta_4= 3s+2 -2\beta_1 +\beta_2 \text{ and } 2\beta_4-\beta_3 = 3s+1 - \beta_1.$$ On substituting $\beta_1 = { s+3 \choose 3}$ and $\beta_2 = 3 {s+2 \choose 3}$, we get $\beta_4= {s \choose 3}$ from the first equation and substituting the value in the second equation, we get $\beta_3=3{s+1 \choose 3} $. Now we verify that $a_j=0$ for all $j \geq 3s+2$. Note that for $r \geq 2$, the coefficient of $t^{3s+r}$ in Equation (\[eqn:q(t) polynomial from HS\]) is $$a_{3s+r}= (3s+r+1) + (r-2) \beta_4 -(r-1) \beta_3 + r \beta_2 - ( r+1) \beta_1.$$ Since, $1 - \beta_1 + \beta_2 - \beta_3 + \beta_4 = 0$, this equation is reduced to $a_{3s+r}= (3s+3) - \beta_3 + 2 \beta_2 - 3 \beta_1.$ Applying the binomial identify $ {n+1 \choose r+1} = {n \choose r+1}+{n \choose r}$ repeatedly, we get $a_{3s+r}=0$ for all $r \geq 2$. Hence the Hilbert series of $R/I^s$ is $$H(R/I^s,t) = \frac{1+2t+3t^2+4t^3+ \cdots +3st^{3s-1} - \left({s +3 \choose 3}-3s-1 \right)t^{3s} + {s \choose 3} t^{3s+1}}{(1-t)^2}.$$ We now complete the description of the resolution. Write the basis elements of $R^{3{s+1 \choose 3}}$ as $$B_3 = \left\{ \begin{array}{ll} E_{1,\ell_1,\ell_2,\ell_3,\ell_4};\\ E_{2,\ell_1,\ell_2,\ell_3,\ell_4},\\ E_{3,\ell_1,\ell_2,\ell_3,\ell_4}, \end{array} : \; 2 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-2 \right\}.$$ Note that $|B_3|=3{s+1 \choose 3}$. Now define the map $\partial_3 : R^{3{s+1 \choose 3}} \longrightarrow R^{3 {s+2 \choose 3}}$ by $$\begin{aligned} \partial_3(E_{1,\ell_1,\ell_2,\ell_3,\ell_4}) & = & x_{4} e_{(2,\ell_1-1,\ell_4),\ell_2+1,\ell_3 } + x_4 e_{(1,\ell_1-1,\ell_2),\ell_3+1,\ell_4}-x_3 e_{(3,\ell_3,\ell_4),\ell_1-1,\ell_2+1} \\ & & -x_3 e_{(2,\ell_1-2,\ell_4),\ell_2+2,\ell_3 } -x_3 e_{(1,\ell_1-2,\ell_2+1),\ell_3+1,\ell_4} +x_1 e_{(3,\ell_3,\ell_4 +1),\ell_1-1,\ell_2};\\ \partial_3(E_{2,\ell_1,\ell_2,\ell_3,\ell_4}) & = & x_{4} e_{(1,\ell_1-1,\ell_2),\ell_3+1,\ell_4} - x_3 e_{(1,\ell_1-2,\ell_2+1),\ell_3+1,\ell_4} - x_2 e_{(3,\ell_3+1,\ell_4),\ell_1-1,\ell_2} \\ & & +x_1 e_{(3,\ell_3,\ell_4+1),\ell_1-1,\ell_2};\\ \partial_3(E_{3,\ell_1,\ell_2,\ell_3,\ell_4}) & = & x_{3} e_{(1,\ell_1-2,\ell_2+1),\ell_3+1,\ell_4} - x_2 e_{(2,\ell_1-2,\ell_4),\ell_2+1,\ell_3+1 } -x_2 e_{(1,\ell_1-2,\ell_2),\ell_3+2,\ell_4} \\ & & + x_1 e_{(2,\ell_1-2,\ell_4+1),\ell_2+1,\ell_3 }.\end{aligned}$$ We now compute the kernel of $\partial_3$. Consider the set We know that $\beta_4= {s \choose 3}$. Since $\operatorname{pdim}R/I^s = 4$, the image of the map $\partial_4 : R^{{s \choose 3}} \longrightarrow R^{3 {s+1 \choose 3}}$ is the kernel of the map $\partial_3 : R^{3{s+1 \choose 3}} \longrightarrow R^{3 {s+2 \choose 3}}$. Hence to describe the map $\partial_4$, it is enough to compute the kernel of the map $\partial_3$. The kernel of the map $\partial_3$ is given by $$A = \left\{ H_{\ell_1,\ell_2,\ell_3,\ell_4} : 3 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-3 \right\},$$ where $$\begin{aligned} H_{\ell_1,\ell_2,\ell_3,\ell_4} & = & x_4 E_{3,\ell_1,\ell_2,\ell_3,\ell_4} -x_3 E_{2,\ell_1-1,\ell_2+1,\ell_3,\ell_4} - x_3 E_{3,\ell_1-1,\ell_2+1,\ell_3,\ell_4} + x_2 E_{1,\ell_1-1,\ell_2,\ell_3+1,\ell_4} \\ & & - x_1 E_{1,\ell_1-1,\ell_2,\ell_3,\ell_4+1} + x_1 E_{2,\ell_1-1,\ell_2,\ell_3,\ell_4+1}.\end{aligned}$$ One can verify that $\partial_3 (H_{\ell_1,\ell_2,\ell_3,\ell_4})=0$. Hence the set $A$ lies in the kernel of the map $\partial_3$. To say that the set $A$ is precisely the kernel of the map $\partial_3$, one needs to show that $\mu(A)={s \choose 3}$. Let $\ell_1^{\prime}=\ell_1-3$, then one has $ \ell_1^{\prime} + \ell_2 +\ell_3 +\ell_4 =s-3$. The total number of non-negative integral solution of this linear equation is precisely ${s \choose 3}$, hence $\mu(A)={s \choose 3}$. Write the basis elements of $R^{{s \choose 3}}$ as $B_4= \left\{ G_{\ell_1,\ell_2,\ell_3,\ell_4} : \; 3 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-3 \right\} $ and define the map $\partial_4 : R^{{s \choose 3}} \longrightarrow R^{3 {s+1 \choose 3}}$ by $$\begin{aligned} \partial_4(G_{\ell_1,\ell_2,\ell_3,\ell_4}) = & H_{\ell_1,\ell_2,\ell_3,\ell_4}.\end{aligned}$$ This is an injective map and hence we get the complete resolution: $$0 \to R(-3s-3)^{\beta_4} {\longrightarrow} R(-3s-2)^{\beta_3} {\longrightarrow} R(-3s-1)^{\beta_2} {\longrightarrow} R(-3s)^{\beta_1} {\longrightarrow} R \to 0.$$ Note that in the above proof, we used $s \geq 3$ only to conclude that $\operatorname{pdim}(R/I^s) = 4$. By Remark \[rem:depth powers\], $\operatorname{depth}(R/I) = 2$ and hence $\operatorname{pdim}(R/I) = 2$. Similarly, $\operatorname{depth}(R/I^2) = 1$ and hence $\operatorname{pdim}(R/I^2) = 3$. This forces $\partial_2$ to be injective when $s = 1$ and $\partial_3$ to be injective when $s = 2$. The computations of syzygies in the cases of resolution of $R/I$ and $R/I^2$ remain the same as given in the above proof. Therefore, we get resolutions truncated at $R^{\beta_2}$ in the case of $R/I$ and truncated at $R^{\beta_3}$ in the case of $R/I^2$, with the expressions for $\beta_2$ and $\beta_3$ coinciding with the ones given in the proof. Therefore, we can conclude that in this case, ${\operatorname{reg}}(I^s) = 3s$ for all $s \geq 1$. As an immediate consequence, we obtain an expression for the asymptotic regularity of cover ideals of complete $4$-partite graphs. \[4-partite\] Let $G$ denote a complete $4$-partite graph with $V(G) = \sqcup_{i=1}^4 V_i$ and $E(G) = \{\{a,b\} ~ : ~ a \in V_i, b \in V_j, i \neq j \}$. Set $V_i = \{x_{i1}, \ldots, x_{im_i}\}$ for $i = 1, \ldots, 4$. Let $J_G \subset R = K[x_{ij} : 1 \leq i \leq 4; ~1\leq j \leq m_i]$ denote the cover ideal of $G$. Then the minimal free resolution of $R/J_G^s$ is of the form: $$0 \to R^{s \choose 3} {\longrightarrow} R^{3{s+1\choose 3}} {\longrightarrow} R^{3{s+2 \choose 3}} {\longrightarrow} R^{s+3 \choose 3} {\longrightarrow} R \to 0.$$ Set $ \alpha = (s-\ell_4)m_1 +(s-\ell_3)m_2+ (s-\ell_2)m_3 + (s-\ell_1)m_4$. Furthermore, we have $${\operatorname{reg}}(J_G^s) = \max \left\{ \begin{array}{ll} \alpha, & \text{ for } 0 \leq \ell_1, \ell_2, \ell_3, \ell_4 \leq s, \\ \alpha + m_4-1, & \text{ for } 1 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-1, \\ \alpha + 2m_4-2, & \text{ for } 2 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-2, \\ \alpha + 3m_4-3, & \text{ for } 3 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-3, \end{array} \right.$$ where $ \ell_1 + \ell_2 +\ell_3 +\ell_4 =s$. Let $X_i = \prod_{j=1}^{m_i} x_{ij}$. Then $J_G = (X_1X_2X_3, X_1X_2X_4, X_1X_3X_4, X_2X_3X_4)$. Set $X_1= x_1, ~X_2 = x_2,~ X_3=x_3$ and $X_4 = x_4$ in the previous theorem. Then it follows that the minimal free resolution of $R/J_G^s$ is of the given form $$0 \to R^{s \choose 3} {\longrightarrow} R^{3{s+1\choose 3}} {\longrightarrow} R^{3{s+2 \choose 3}} {\longrightarrow} R^{s+3 \choose 3} {\longrightarrow} R \to 0.$$ To compute the regularity of $R/J_G^s$, we first need to find the degree’s of the generators of the syzygies. Following the notation of the previous theorem, we have $$\deg e_{\ell_1,\ell_2,\ell_3,\ell_4} = \deg \left( X_1^{s-\ell_4}X_2^{s- \ell_3}X_3^{s- \ell_2}X_4^{s- \ell_1} \right) = (s-\ell_4)m_1 +(s-\ell_3)m_2+ (s-\ell_2)m_3 + (s-\ell_1)m_4$$ and we have $$\begin{aligned} \deg e_{(1,\ell_1-1,\ell_2),\ell_3,\ell_4} & = & \deg e_{(2,\ell_1-1,\ell_4),\ell_2+1,\ell_3} = \deg e_{(3,\ell_3,\ell_4),\ell_1,\ell_2} = \deg e_{\ell_1,\ell_2,\ell_3,\ell_4} + \deg(X_4);\\ \deg E_{1,\ell_1,\ell_2,\ell_3,\ell_4} & = & \deg E_{2,\ell_1,\ell_2,\ell_3,\ell_4} = \deg E_{3,\ell_1,\ell_2,\ell_3,\ell_4} = \deg e_{\ell_1,\ell_2,\ell_3,\ell_4} + 2\deg(X_4); \\ \deg G_{\ell_1,\ell_2,\ell_3,\ell_4} & = & \deg e_{\ell_1,\ell_2,\ell_3,\ell_4} + 3\deg(X_4).\end{aligned}$$ Therefore, by setting $\alpha = \deg e_{\ell_1, \ell_2,\ell_3,\ell_4}$, we get $${\operatorname{reg}}(J_G^s) = \max \left\{ \begin{array}{ll} \alpha, & \text{ for } 0 \leq \ell_1, \ell_2, \ell_3, \ell_4 \leq s, \\ \alpha + \deg(X_4)-1, & \text{ for } 1 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-1, \\ \alpha + 2\deg(X_4)-2, & \text{ for } 2 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-2, \\ \alpha + 3\deg(X_4)-3, & \text{ for } 3 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-3, \end{array} \right.,$$ where $ \ell_1 + \ell_2 +\ell_3 +\ell_4 =s$. Here also, we have obtained an expression for ${\operatorname{reg}}(J_G^s)$ not in the form of a linear polynomial. But, as we have demonstrated in the previous cases, this can always be derived for a given graph. Analyzing the interplay between the cardinalities of the partitions, one can obtain the polynomial expression. We discuss the unmixed case as an example. Let $m_1 = m_2 = m_3 = m_4 = m$. Then $$\begin{aligned} \alpha & = & (s-\ell_4)m_1 +(s-\ell_3)m_2+ (s-\ell_2)m_3 + (s-\ell_1)m_4 \\ & = & (4s - (\ell_1+\ell_2+\ell_3+\ell_4))m = 3ms.\end{aligned}$$ Therefore ${\operatorname{reg}}(J_G^s) = 3ms + (3m-3)$ for all $s \geq 3$. Let the partitions be in an arithmetic progression. Let $m,r$ be any two positive integers. Consider the arithmetic progression $m_1 = m$, $m_2 =m+r$, $ m_3 =m+2r$, and $m_4 = m+3r$ in Theorem \[4-partite\]. Then, for all $s \geq 3$, we have $${\operatorname{reg}}(J_G^s) = s (3m+6r)+3m-3.$$ We have from Theorem \[4-partite\], $ \alpha = (s-\ell_4)m_1 +(s-\ell_3)m_2+ (s-\ell_2)m_3 + (s-\ell_1)m_4$. On substituting the values of $m_i$’s in $ \alpha$, we get $$\alpha = s(4m+6r) -m\ell_4 -(m+r)\ell_3 -(m+2r)\ell_2 - (m+3r)\ell_1$$ By Theorem \[4-partite\], we have for all $s \geq 3$, $${\operatorname{reg}}(J_G^s) = \max \left\{ \begin{array}{ll} \alpha, & \text{ for } 0 \leq \ell_1, \ell_2, \ell_3, \ell_4 \leq s, \\ \alpha + (m+3r)-1, & \text{ for } 1 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-1, \\ \alpha + 2(m+3r)-2, & \text{ for } 2 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-2, \\ \alpha + 3(m+3r)-3, & \text{ for } 3 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-3, \end{array} \right.$$ where $ \ell_1 + \ell_2 +\ell_3 +\ell_4 =s$. To achieve the maximum value of $\alpha$, negative terms in $\alpha$ should be minimum. The coefficient of $\ell_1$ in negative terms in $\alpha$ is largest, so $\ell_1$ should be assigned the minimum value. After assigning the minimum value to $\ell_1$, assign the minimum value to $\ell_2$, and similarly minimum value to $\ell_3$. Then assign $ \ell_4 = s-\ell_2 -\ell_3 -\ell_4$. For instance, to get the maximum of $\alpha$ when $1 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-1$, put $\ell_1=1,\ell_2=0,\ell_3=0$, and $\ell_4=s-1$. With appropriate substitution, we get for all $s \geq 3$ $${\operatorname{reg}}(J_G^s) = \max \left\{ \begin{array}{ll} s (3m+6r), & \text{ for } 0 \leq \ell_1, \ell_2, \ell_3, \ell_4 \leq s, \\ s (3m+6r)+m-1, & \text{ for } 1 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-1, \\ s (3m+6r)+2m-2, & \text{ for } 2 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-2, \\ s (3m+6r)+3m-3, & \text{ for } 3 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \ell_3, \ell_4 \leq s-3. \end{array} \right.$$ Clearly for all $m \geq 1$, and for all $s \geq 3$, we get $${\operatorname{reg}}(J_G^s) = s (3m+6r)+3m-3.$$ Complete $m$-partite graphs --------------------------- Let $G$ be a complete graph on $m$-vertices. Then the cover ideal $J_G$ of $G$ is generated by $\{x_1\cdots \hat{x}_i \cdots x_m ~ : ~ 1\leq i \leq m\}$. It follows from Remark \[rem:depth powers\] that $\operatorname{depth}R/J_G^s = 0$ for all $s \geq m-1$. Moreover, by Remark \[rem:lin res powers\], we know that $R/J_G^s$ has linear resolution for all $s \geq 1$. Therefore, the minimal graded free resolution of $R/J_G^s$ for all $s \geq m-1$ is of the form $$0 \to R(-s(m-1)-m+1)^{\beta_m} {\longrightarrow} \to \cdots \to R(-s(m-1)-1)^{\beta_2} {\longrightarrow} R(-s(m-1))^{\beta_1} {\longrightarrow} R \to 0.$$ Let $g_1,g_2,\ldots,g_m$ be the minimal generators $J_G$. Then the elements in $J_G^s$ consists of elements $T_{\ell_1,\ell_2,\dots,\ell_m}$, where $T_{\ell_1,\ell_2,\dots,\ell_m}=g_1^{\ell_1}g_{2}^{\ell_2}\dots g_{m}^{\ell_m}$ such that $ \ell_1 + \ell_2 + \cdots + \ell_m = s$ and $0 \leq \ell_i \leq s$. Therefore the total number of elements in $J_G^s$ is same as the total number of non-negative integral solution to the linear equation $ \ell_1 + \ell_2 + \cdots + \ell_m = s$ which is ${s+m-1 \choose m-1 }$. Hence $\mu (J_G^s)= { s+m-1 \choose m-1}$. Therefore $\beta_1 = {s+m-1 \choose m-1}$. Let $\{e_{\ell_1,\ell_2,\dots,\ell_m} \mid 0 \leq \ell_i \leq s; \text{ and } \ell_1 + \ell_2 + \cdots + \ell_m = s\}$ denote the standard basis for $R^{\beta_1}$. Let $\partial_1 : R^{\beta_1} \longrightarrow R$ be the map $\partial_1(e_{\ell_1,\ell_2,\dots,\ell_m}) = T_{\ell_1,\ell_2,\dots,\ell_m}$. As done in the proofs of Theorems \[C3\] and \[C4\], we can see that the first syzygy is given by the relations of the form $$x_i \cdot T_{\ell_1,\ell_2, \cdots, \ell_{i-1}, {\ell_{i}-1 },{\ell_{i+1}+1},\cdots,\ell_m} - x_{i+1}\cdot T_{\ell_1,\ell_2, \cdots,\ell_{i-1},\ell_{i},\ell_{i+1}, \cdots, \ell_{m}}=0$$ for each $1 \leq i \leq m-1$. Set $\ell_i - 1 = \ell_i'$ and $\ell_{i+1}+1 = \ell_{i+1}'$. Then it can be seen that, for each $1 \leq i \leq m-1$, there exist as many such relations as the number of non-negative integer solutions of $\ell_1 + \cdots + \ell_i' + \cdots \ell_m = s-1$. Therefore, the total number of such linear relations is $(m-1){s+m-2 \choose m-1}$. Therefore $\beta_2 = {m-1 \choose 1}{s+m-2 \choose m-1}$. However it is not very difficult to realize that writing down the higher syzygy relations are quite challenging. Based on Theorems \[C3\] and \[C4\] and some of the experimental results using the computational commutative algebra package Macaulay 2 [@M2], we propose the following conjecture: Let $R = K[x_1, x_2, \dots,x_m]$ and let $J$ be the cover ideal of the complete graph $K_m$. The minimal graded free resolution of $R/I^s$ for all $s \geq m-1$ is of the form $$0 \to R(-s(m-1)-m+1)^{\beta_m} {\longrightarrow} \cdots \longrightarrow R(-s(m-1)-1)^{\beta_2} {\longrightarrow} R(-s(m-1))^{\beta_1} {\longrightarrow} R \to 0,$$ where $$\beta_i = {m-1 \choose i-1}{ s+m-i \choose m-1}.$$ Notice that proving the above conjecture will give the Betti numbers of powers of cover ideals of complete $m$-partite graphs. We conclude our article by proposing an expression for the regularity of powers of the cover ideals of complete $m$-partite graphs: Let $G$ denote a complete $m$-partite graph with $V(G) = \sqcup_{i=1}^m V_i$ and $E(G) = \{\{a,b\} ~ : ~ a \in V_i, b \in V_j, i \neq j \}$. Set $V_i = \{x_{i1}, \ldots, x_{in_i}\}$ for $i = 1, \ldots, m$. Let $J_G \subset R = K[x_{ij} ~ : ~ 1 \leq i \leq m; ~1\leq j \leq n_i]$ denote the cover ideal of $G$. Let $ 0 \leq \ell_1,\ell_2,\ldots,\ell_m \leq s$ be integers such that $ \ell_1 + \ell_2 + \cdots +\ell_m =s$. Set $$\alpha = s \cdot ( \sum_{i=1}^{m}n_i) - \sum_{i=1}^{m}n_i \ell_{m+1-i}.$$ Then for all $s \geq m-1$, one has $${\operatorname{reg}}(J_G^s) = \max \left\{ \begin{array}{ll} \alpha, & \text{ for } 0 \leq \ell_1, \ell_2, \cdots, \ell_m \leq s, \\ \alpha + n_m-1, & \text{ for } 1 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \cdots, \ell_m \leq s-1, \\ \alpha + 2(n_m-1), & \text{ for } 2 \leq \ell_1 \leq s, \; 0 \leq \ell_2,\cdots, \ell_m \leq s-2, \\ & \vdots\\ \alpha + (m-1)(n_m-1), & \text{ for } m-1 \leq \ell_1 \leq s, \; 0 \leq \ell_2, \cdots, \ell_m \leq s-(m-1). \end{array} \right.$$

ArXiv

--- abstract: 'We introduce an exactly solvable SU(2)-invariant spin-1/2 model with exotic spin excitations. With time reversal symmetry (TRS), the ground state is a spin liquid with gapless or gapped spin-1 but fermionic excitations. When TRS is broken, the resulting spin liquid exhibits deconfined vortex excitations which carry spin-1/2 and obey non-Abelian statistics. We show that this SU(2) invariant non-Abelian spin liquid exhibits spin quantum Hall effect with quantized spin Hall conductivity $\sigma^s_{xy}=\hbar/2\pi$, and that the spin response is effectively described by the SO(3) level-1 Chern-Simons theory at low energy. We further propose that a SU(2) level-2 Chern-Simons theory is the effective field theory describing the topological structure of the non-Abelian SU(2) invariant spin liquid.' author: - 'Hong Yao and Dung-Hai Lee' title: 'Fermionic magnons, non-Abelian spinons, and spin quantum Hall effect from an exactly solvable spin-1/2 Kitaev model with SU(2) symmetry' --- Quantum spin liquids are zero temperature exotic states of magnets which exhibit no classical order. Their existence is often due to strong quantum fluctuations and/or geometric frustrations. In two dimensions (2D), the notion of quantum spin liquids were first proposed by Anderson in 1973 as the candidate ground state of the spin-1/2 AFM Heisenberg model on the triangular lattice [@anderson1973]. In addition, soon after the discovery of the cuprate high temperature superconductors quantum spin liquids are proposed as the root states of the high $T_c$ superconducting states [@anderson1987; @kivelson1987; @lee2006]. However, up to now there seems no definitive proof of the existence of real spin liquid materials for $D\ge 2$ even though a few promising candidates have emerged recently [@lee2008b; @balents2010]. Consequently, further experimental scrutinies are needed to firmly establish the existence of real spin liquids in $D\ge 2$. Of course, microscopic theoretical models with reliable solution of spin liquid ground states are also desired. In order for the elementary excitations to carry well-defined spin quantum number, SU(2) invariance is important. Recently, a number of exactly solvable SU(2) invariant models [@moessner2001; @lee2003; @fujimoto2005; @raman2005_sondhi; @seidel2009; @cano2010] with spin liquid ground states have been proposed. They have the common trait that the low energy/long wave length physics is captured by the short-range resonating valence bond (RVB) [@anderson1987; @kivelson1987]. In this paper, we take a different route to construct a SU(2) invariant model with spin liquid ground states by generalizing the recently discovered Kitaev model [@kitaev2006] in a SU(2) symmetric way. Our model \[see \] is a spin-1/2 system on the decorated honeycomb lattice (also known as the star or 3-12 lattice, see ). Our spin model can be reduced to three species of free Majorana fermions coupled to background $Z_2$ gauge field such that the model is exactly solvable, in similar spirit of the original Kitaev model that have only one species of free Majorana fermions. In the presence of time reversal symmetry (TRS), it exhibits a spin liquid ground state with gapless or gapped spin-1 excitations, depending on parameters in the model. Strikingly, these spin-1 excitations obey fermionic statistics. When TRS is broken either spontaneously or explicitly, each flavor of Majorana fermions behave like the Bogoliubov quasiparticle of a chiral $p+ip$ superconductor. As a result, a $Z_2$ vortex binds three Majorana zero modes, protected by the SU(2) symmetry. Due to the existence of the odd number of Majarana zero modes, the vortex excitations obey non-Abelian statistics, which might find potential application in topological quantum computing [@nayak2008]. We further show that the non-Abelian vortex excitations carry spin-1/2 quantum number. As far as we know, this is the first realization of non-Abelian spinon in an exactly solvable SU(2) symmetric spin model. Note that a SU(2) symmetric wave function with non-Abelian spinon excitations was proposed recently in a spin-1 system[@greiter:prl09] but it is not clear which Hamiltonian can realize it as a ground state. ![The schematic representation of the decorated honeycomb lattice. The triangles are labeled by $i,j$, the sites within each triangle are labeled by $1,2,3$, and the type of inter-triangle bonds is $x$, $y$, or $z$. []{data-label="fig:star"}](decohoney) In the SU(2) invariant non-Abelian phase, the system exhibits a spin quantum Hall effect [@footnote1] with quantized spin Hall conductivity $\sigma^s_{xy}=\hbar/2\pi$ (which is twice of that in a $d+id$ superconductor [@senthil1999].) The quantized spin Hall response is shown to be described by the SO(3) level-1 Chern-Simons gauge theory at low energy. Furthermore, according to the threefold ground state degeneracy, non-Abelian statistics of vortices, and the SU(2)-invariance of the ground states, we propose that the low energy topological field theory for the non-Abelian phase is the SU(2) level-2 Chern-Simons theory [@wen1991; @fradkin1998]. [**The model:**]{} We consider the following SU(2) invariant Hamiltonian on the lattice shown in (a): H&=&J\_i \^2\_i + \_[ ]{} J\_, \[eq:ham\] where $i$, $j$ label the triangles and $\mathbf S_i=\mathbf S_{i,1}+\mathbf S_{i,2}+\mathbf S_{i,3}$ is the total spin of the $i$th triangle ($\mathbf S_{i,\alpha}$ is the spin-1/2 operator on site $\alpha=1,2,3$ of the $i$th triangle.) The operators $\tau^\lambda_i$ are defined as follows: $\tau^x_i=2(\mathbf S_{i,1}\cdot \mathbf S_{i,2}+1/4)$, $ \tau^y_i=2(\mathbf S_{i,1}\cdot \mathbf S_{i,3}-\mathbf S_{i,2}\cdot \mathbf S_{i,3})/\sqrt{3}$, and $\tau^z_i=4\mathbf S_{i,1}\cdot (\mathbf S_{i,2}\times \mathbf S_{i,3})/\sqrt{3}$. The parameter $J$ is the strength of the intra-triangle spin exchange coupling while $J_\lambda~(\lambda=x,y,z)$ describes the inter-triangle couplings on the type-$\lambda$ links. Because $[\mathbf S^2_i, \mathbf S_j]=0$ and $[\mathbf S^2_i, \tau^{\lambda}_j]=0$, the operator $\mathbf S_i^2$ commutes with the Hamiltonian for all $i$ [@wang2010]. As a result, the total spin of each triangle is a good quantum number, so we can use them to subdivide the Hilbert space. For each triangle, three spin-1/2’s can be decomposed in terms of their total spins: ${\frac12}\otimes{\frac12} \otimes{\frac12}={\frac32}\oplus{\frac12}\oplus{ \frac12}$. Within the space spanned by ${\frac12}\oplus{\frac12}$ it is straightforward to check that $\vec \tau_i$ satisfy the SU(2) algebra $[\tau^\alpha_i,\tau^\beta_i]= 2i\epsilon^{\alpha\beta\gamma}\tau^\gamma_i$ as well as the Clifford algebra $\{\tau^\alpha_i,\tau^\beta_i\}=2\delta^{\alpha\beta}$, which implies that $\tau^\lambda_i$ are the Pauli matrices. In fact the four states in ${\frac12}\oplus{\frac12}$ can be labeled by $\ket{\sigma^z_i=\pm 1,\tau^z_i=\pm 1}$ with $\vec \sigma_i=2 \vec S_i$ since $[\tau^\alpha_i,S^\beta_i]=0$. Consequently we may view $\tau^\lambda_i$ as pesudo-spin 1/2 which distinguishes the “orbital" degree of freedom within the two degenerate spin-1/2 multiplets. In the remaining spin-3/2 space, $\tau^z_i=\tau^y_i=0$ because the three spins are totally polarized for $S_i=3/2$; $\tau^\lambda_i$ do not satisfy SU(2) algebra. When $J\gg J_\lambda$, the ground state and the low lying excited states all lie in the sub-Hilbert space ${\cal L}_0$ where $S_i=1/2$ for all triangles. \[It can be shown that the ground state must lie in ${\cal L}_0$ as long as $J>\frac{3}{2}(J_x+J_y+J_z)$.\] For simplicity, we will assume $J\gg J_\lambda$ hereafter which allows us to focus on ${\cal L}_0$. In this sub-Hilbert space, apart from a constant, the Hamiltonian in reduces to H=14\_[ ]{}J\_, \[hl0\] which seems complicated but is actually exactly solvable as shown below. To solve the model, we introduce Majorana fermions representations [@kitaev2006] for the Pauli matrices $\sigma^\alpha_i$ and $\tau^\beta_i$ as follows: \^\_i \^\_i =ic\^\_i d\^\_i,  \^\_i =-ic\^\_i c\^\_i,  \^\_i=- id\^\_i d\^\_i. where $\alpha,\beta=x,y,z$ and $c^\alpha_i,~d^\alpha_i$ are Majorana fermions. As usual, the Majorana fermion representation is over-complete and the following constraint is needed for a physical wave function: $\ket{\Psi}_\text{phys}$ is in physical Hilbert space iff D\_i \_=\_,  i, \[eq:cons\] where $D_i=-i c^x_ic^y_ic^z_i d^x_i d^y_i d^z_i$. In other words, any state acted by the projection operator $P=\prod_i \left[\frac{1+D_i}{2}\right]$ is a physical state. In terms of Majorana fermion operators, it is straightforward to rewrite the spin Hamiltonian as follows =\_J\_[ij]{} u\_[ij]{} , \[eq:fer\_Ham\] where $ u_{ij}=-id^\lambda_i d^\lambda_j$ and $J_{ij}=J_\lambda/4$ on the type-$\lambda$ ($\lambda=x,y,z$) link $\avg{ij}$. It is clear that $H=P{\cal H} P$. Because $[ u_{ij}, {\cal H}]=0$, and $[ u_{ij}, u_{i'j'}]=0$, $ u_{ij}$ are good quantum numbers with eigenvalues $\pm 1$. It is obvious that is invariant under the following local $Z_2$ gauge transformation $c^\alpha_i\to \Lambda_ic^\alpha_i$ and $u_{ij}\to \Lambda_i u_{ij}\Lambda_j$, $\Lambda_i=\pm 1$. describes three species of free Majorana fermions coupled with background $Z_2$ gauge field $u_{ij}$. In addition to the $Z_2$ gauge symmetry has a global SO(3) symmetry which rotate among the three species of Majorana fermions, which is the consequence of the SU(2) symmetry of the original spin model. Lieb’s theorem [@lieb1994] requires that the ground state lies in the zero flux sector, namely $\phi_p=0$ for every hexagon plaquette $p$ where $\exp(i\phi_p)\equiv\prod_{\avg{jk}\in p}iu_{jk}$. The zero flux sector is realized by choosing $u_{ij}=1$ with $i(j)\in A(B)$ sublattice. In the zero flux sector, it is straightforward to show that there are fermionic excitations [@footnote2] that are gapped or gapless with Dirac-like dispersion, as discussed below. Besides these fermionic excitations, vortex excitation on plaquette $p$ is created when $\phi_p=\pi$. [**Gapless/gapped spin liquid:**]{} The spectrum of each species of Majorana fermions in can be obtained by a Fourier transform $c^{\alpha}_i=\sqrt{\frac2N}\sum_{\mathbf{k}\in \frac{\textrm{BZ}}{2}} \left[{\mathrm{e}}^{i\mathbf{k}\cdot \mathbf{r}_i} \psi_{\alpha,\mathbf{k},a}+{\mathrm{e}}^{-i\mathbf{k}\cdot \mathbf{r}_i} \psi^\dag_{\alpha,\mathbf{k},a}\right]$, where $a=A/B$ is the sublattice index of site $i$, $\{\psi^\dag_{\alpha,\mathbf{k},a}, \psi_{\beta,\mathbf{k}',b}\}=\delta_{\alpha\beta} \delta_{\mathbf{k,k'}}\delta_{ab}$, and the summation is over one-half of the Brillouin zone. The three species have identical dispersion $E_\mathbf{k}=\pm |\sum_\alpha J_\alpha {\mathrm{e}}^{i \mathbf{k}\cdot \mathbf{e}_\alpha}|$ , which is gapless with Dirac-like dispersion for $J_x+J_y>J_z$ and is gapped otherwise. (Here $\mathbf{e}_\alpha$ label nearest-neighbor vectors.) It is straightforward to calculate the spin-spin correlation with $\avg{S^z_i S^z_j}\sim |\mathbf{r}_i-\mathbf{r}_j|^{-4}$ for $|\mathbf{r}_i-\mathbf{r}_j|\gg 1$ in the gapless phase but in exponential decay for the gapped phase. This is in contrast with the original Kitaev model and its variants studied previously where the spins are uncorrelated beyond nearest neighbors [@baskaran2007; @tikhonov:arxiv10]. [**Fermionic spin-1 excitations**]{}: In the zero flux sector, what is the nature of fermionic excitations? We show below that these fermionic excitations carry spin-1 quantum number. For instance, by introducing the complex fermion operators $f_{i,z}=(c^x_i-ic^y_i)/2$, the Hamiltonian can be rewritten as =\_J\_[ij]{}u\_[ij]{}, which indicates that $f_{i,z}$ are free complex fermions and their fermion number is conserved. The dispersion of $f_z$ complex fermions and $c^z$ Majorana fermions is again $E_\mathbf{k}=\pm |\sum_\alpha J_\alpha {\mathrm{e}}^{i \mathbf{k}\cdot \mathbf{e}_\alpha}| $ but $\mathbf{k}$ is over the full (half) Brillouin zone for $f_z$ ($c^z$). Since $S^{z}_i=\sigma^z_i/2 =f^\dag_{i,z}f_{i,z}-1/2$, it is clear that the fermions created by $f^\dag_{i,z}$ carry $S^z=1$, which can be loosely called fermionic magnons with $S^z=1$. The $c^z$ quasi-particles carry $S^z=0$. In the gapped phase, due to total $f_z$ fermion number conservation (a result of total $S^z$ conservation), the Chern number of $f_z$ can be defined and it is zero. Similarly, a spectral Chern number can also be defined for $c^z$ Majorana fermions and it is also zero. Due to the zero Chern number, a vortex excitation has no fermion zero mode for $f_z$ and no Majorana zero mode for $c^z$; its spin quantum number is zero. ![The energy spectrum as a function of $k_y$ for $f_z$ fermions on a cylinder periodic in $y$-direction. There is exactly one gapless chiral edge state. The parameters used for both Fig. 2 and Fig. 3 are $J_x=J_y=J_z=1$ and $h=0.3$.[]{data-label="fig:edge"}](edge) [**Non-Abelian spinons**]{}: A spinon carries spin-1/2 quantum number and is essentially half of a spin-1 excitation. Deconfined spinon is commonly regarded as a hallmark of a quantum spin liquid. A spinon excitation with $S^z=1/2$ would carry half fermion number of $f_z$. Excitations with half fermion quantum number are known to exist as fermion zero modes in some sort of defect such as domain walls and vortices. Thus, we study vortex excitations. Since a fermion zero mode with half fermion number is expected in a vortex with flux $\pi$ when the insulator carries Chern number $\pm 1$. To obtain a finite Chern number for $f_z$, following Ref. [@lee2007], we consider additional interactions which breaks TRS externally (TRS can be spontaneously broken by “decorating” the star lattice [@yao2007]): H’=h\_[\_\_]{} \^ , \[eq:h\] where $\avg{ij}_\alpha \avg{jk}_{\beta}$ labels three neighboring sites, which are ordered in a clockwise way within the corresponding hexagon plaquette, forming a triad whose links are $\alpha$ and $\beta$ respectively. In term of fermions, $H'=\frac h4\sum_{\avg{ij}\avg{jk}} \hat u_{ij}\hat u_{jk} \left(i c^x_i c^x_k +i c^y_i c^y_k +i c^z_i c^z_k\right)=\frac h4\sum_{\avg{ij}\avg{jk}} \hat u_{ij}\hat u_{jk}[ 2(i f^\dag_{i,z} f_{k,z} -i f^\dag_{k,z}f_{i,z}) +i c^z_i c^z_k]$. This next-nearest-neighbor hopping term generates a mass term that gaps out the Dirac point of $f_z$ fermions and the resulting insulator of $f_z$ has Chern number $\nu=1$, which can be computed directly and is also indicated by the one gapless chiral edge mode of $f_z$ on a system with boundary, as shown in Fig. \[fig:edge\]. Since $f_z$ fermions carry $S^z=1$, the finite Chern number $\nu=1$ implies the spin quantum Hall effect $J^{z,i}=\sigma^s_{ij}\pa_j B^z$, where $J^{z,i}$ is the current of spin polarized along the $z$-direction flowing along the $i=x,y$-direction, with quantized spin Hall conductivity $\sigma^s_{xy}=\nu \frac{\hbar^2}{2\pi\hbar}=\hbar/2\pi$. Because of the finite spin Hall conductivity, by inserting a $\pi$ (or equivalently $-\pi$) flux locally on plaquette $p$, namely creating a vortex excitation, the spin is cumulated around the vortex by the amount $S^z=\sigma^s_{xy}\cdot (\pm \pi)=\pm \hbar/2$, implying that the vortex carries spin-1/2 quantum number. The sign ambiguity reflects the facts that there is one zero mode associated with the vortex excitation and that occupying it or not gives rise to $S^z=\pm 1/2$. The spin-1/2 nature of a vortex excitation is further verified numerically. For instance, by creating two well separated vortices in a finite lattice and occupying the zero mode associated with each vortex, we obtain $S^z=1/2$ for each vortex, as shown in Fig. \[fig:spinon\](a). Similarly, the $c^z$ Majorona fermions have spectral Chern number $\pm 1$ and the edge has a gapless chiral Majorana mode. Moreover, there is one $c^z$ Majorana zero mode associated with each vortex excitation [@kitaev2006]. Due to this unpaired Majorana zero mode from $c^z$, the vortex excitations obey non-Abelian statistics. In the following we refer to them as “non-Abelian spinons” since they also carry spin-1/2 quantum number. It is now clear that a local vortex excitation actually binds three Majorana zero modes, which are treated “on bias” as a complex fermion zero mode plus a Majorana zero mode above. The three Majorana zero modes will not mix and split because of the SU(2) symmetry. We further computed the total energy of a pair of vortices as a function of the inter-vortex distance and the result is shown in (b). The fact that the energy decreases with the increasing distance indicates that the vortex-vortex interaction is repulsive and creating two far separated spinons only cost a finite energy. The non-Abelian spinons are deconfined. [**Topological field theory**]{}: A topological phase is generically described by a topological field theory. In the non-Abelian spin liquid discussed above, we expect that the finite spin quantum Hall effect have a topological field theory description which we will derive below. For simplicity, we assume $J_x=J_y=J_z$ hereafter. Without the $h$ term, the dispersion is gapless with the Dirac point at $\mathbf{K}$. (Note that each species of Majorana fermions has only one Dirac cone due to their Majorana nature.) A finite $h$ term acts like a mass term for the Dirac fermions at $\mathbf{K}$. In the continuum limit, the low energy physics is then described by the following Euclidean action S=dt d\^2x |, where $\psi=(\psi_{xA}, \psi_{xB}, \psi_{yA}, \psi_{yB}, \psi_{zA}, \psi_{zB})^T$ and $\bar \psi=i\psi^\dag\gamma^0$. Here $m$ is the mass and $\gamma^{0,x,y}=\mathbb{I}\otimes\sigma^{z,y,x}$ where $\mathbb{I}$ is a 3 by 3 identity matrix with vector indices $x,y,z$ and $\vec \sigma$ Pauli matrices with sublattice indices. It is clear that the three species of massive Dirac fermions possess a global SO(3) symmetry, which is inherited from the SU(2) symmetry of the original spin model. The continuous SO(3) symmetry allows us to introduce external spin gauge fields $A^a_\mu$, which couple with the spin current $J^{a\mu}=\psi^\dag t^a\gamma^\mu\psi$, where $[t^a]_{bc}=i\epsilon^{abc}$, in the following way: S=dtd\^2x |(t,x)(t,x). \[eq:gaugecoup\] By integrating out fermions, we obtain an induced action for $A^a_\mu$, whose lowest-order imaginary part is a topological term: S\_\[A\]= dtd\^2x \^ ,       \[tpa\] which is a SO(3) level-1 Chern-Simons action. This topological term describes the spin responses of the system to the external spin gauge fields. Physically, $A^a_0=B^a$ is the external magnetic field. From J\^[a]{}=-i=\^\[\_A\^a\_+\^[abc]{}A\^b\_A\^c\_\],     we obtain $J^{a i}=\frac1{2\pi}\epsilon^{ij}\pa_j A^a_0=\frac1{2\pi}\epsilon^{ij}\pa_j B^a$ ($i,j=x,y$) which implies a quantized spin Hall response to the gradient of external magnetic field $A^a_0=B^a$ with quantized spin Hall conductance $\sigma^s_{xy}=\frac{1}{2\pi}$, as expected. To better understand the topological properties of the non-Abelian spin liquid, an effective topological field theory describing its long-distance and low-energy physics [@lee1989; @zhang1989; @wen1992] would be desired. To do so, we write $J^{a\mu}=\frac{1}{2\pi}\epsilon^{\mu\nu\lambda}\pa_\nu a^a_\lambda$ so that it satisfies the continuity condition $\pa_\mu J^{a\mu}=0$ automatically. Without derivations, we propose that the following SU(2) level-2 \[SU(2)$_2$\] Chern-Simons theory [@wen1991; @fradkin1998] S\_\[a\]=idt d\^2x \^,       \[eq:tft\] is the low-energy effective theory for the non-Abelian spin liquid phase. There are several reasons for such a proposal. First it is natural to use SU(2) as the gauge group; the Majorana fermions has both SO(3) and $Z_2$ symmetries, which combine leading to an effective SU(2) as expected from the spin SU(2) symmetry of the model. Secondly, as discussed in Ref. [@fradkin1998] the edge theory corresponds to the SU(2)$_2$ Chern-Simons theory \[\] is the chiral sector of SU(2)$_2$ Wess-Zumino-Witten model with central charge $c=3/2$, which is consistent with the fact that the edge theory here has three copies of chiral Majorana modes. Thirdly, the SU(2)$_2$ Chern-Simons theory on a torus is threefold degenerate, which is identical with the degeneracy computed from our lattice model. [**Concluding remarks:**]{} We have shown that the exactly solvable SU(2)-invariant spin-1/2 model on the decorated honeycomb lattice exhibits quantum spin liquid ground states with fermionic magnons or non-Abelian spinons. Interestingly, a recently discovered material, called Iron Acetate [@ironacetate], realizes a spin model on the decorated honeycomb lattice, which adds some hope that our model may be realized in similar family of materials. Moreover, we believe that the model can be potentially realized by loading cold atoms in specially designed optical lattices under appropriate circumstances [@coldatom]. We sincerely thank Joseph Maciejko, Xiao-Liang Qi, Shinsei Ryu, Ashvin Vishwanath, Zheng-Yu Weng, Shou-Cheng Zhang, and especially Steve Kivelson for helpful discussions. This work is partly supported by DOE grant DE-AC02-05CH11231. [29]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\ 12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} [****,  ()](\doibase 10.1103/PhysRevLett.102.207203) Note that spin quantum Hall effect is qualitatively different from the quantum spin Hall effect. The former is the quantized spin Hall response to the gradient of magnetic fields while the latter concerns to electric fields and the spin Hall response is not quantized. @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} As in the original Kitaev model, in each flux sector, including the zero flux sector where the ground state lies, fermionic excitations with a certain but fixed fermion parity are in the physical Hilbert space. @noop [****,  ()]{} @noop @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} L. Santos [*et. al.*]{}, Phys. Rev. Lett. [**93**]{}, 030601 (2004). Y.Z. Zheng [*et. al.*]{}, Angew. Chem. Int. Ed. [**46**]{}, 6076 (2007).

ArXiv

--- abstract: 'We develop a mathematical approach to the nonequilibrium work theorem which is traditionally referred to in statistical mechanics as Jarzynski’s identity. We suggest a mathematically rigorous formulation and proof of the identity.' author: - Evelina title: A mathematical approach to the nonequilibrium work theorem --- Introduction ============ The nonequilibrium work theorem is an equation in statistical mechanics that relates the free energy difference ${\Delta}F$ to the work $W$ carried out on a system during a nonequilibrium transformation. The identity appeared in different, but as we show below equivalent, formulations in [@Jar97_1] and [@Jar97] in 1997, and in the series of papers [@bochkov1]–[@bochkov4] in 1977–1981. In physics literature, the identity is usually written in the form where the average is taken over all possible system trajectories in the phase space, and $\beta$ is an inverse temperature. The identity first appeared in this form in [@Jar97_1] and [@Jar97]. Traditional equilibrium thermodynamics tells us that ${\langle}W {\rangle}{\geqslant}{\Delta}F$ while the transformation of the system is infinitely slow. The identity (\[jarzynski\]) is a stronger statement, and in addition to this, it is valid for arbitrary transformations of the system. The identity is used effectively in computer simulations, as well as in experimental physics, to calculate the free energy difference between two states of the system by running many trajectories and taking the average value of $e^{-\beta W}$ (see [@Crooks2000]–[@Dellago2], [@Hummer_Szabo]– [@liphardt] and references therein). The paper [@Jar07] discusses the connection between two different versions of the identity, and shows that the papers [@bochkov1]–[@bochkov4] use a different definition of work. The identity obtained in [@bochkov1]–[@bochkov4] (referred to below as Bochkov–Kuzovlev’s identity) reads: $${\langle}e^{-\beta W_0} {\rangle}= 1,$$ where $W_0$ is the work (in Bochkov–Kuzovlev’s sense) performed on the system, and the angle brackets have the same meaning as in (\[jarzynski\]). The present paper shows that Jarzynski’s and Bochkov–Kuzovlev’s identities easily follow from each other. Since the identities involve taking an “average over trajectories”, it is natural to interpret this average as the expectation relative to a probability measure on trajectories, while assuming that the system evolves stochastically. In terms of expectations the identities can be represented by the formulas $$\mathbb E[e^{-\beta W}] = e^{-\beta{\Delta}F} \quad {\textrm}{and} \quad \mathbb E[e^{-\beta W_0}] = 1,$$ where $\mathbb E$ is the expectation relative to a probability measure on phase space paths. For this probability measure, some analytical assumptions under which the identities hold are found. Notation and assumptions ======================== Let us assume that the evolution of our system is described by a Markov process $\Gamma_t({\omega})$, $t\in[0,T]$, given through its transition density function. Let $X=\Rnu^{2d}$ be the phase space for our system, i.e. the set of values of $\Gamma_t$. We assume that at time $t=0$ the distribution on the phase space $X$ is given by the following density function: $$q_{{\lambda}_0}(x)=\frac{e^{-\beta {\mathcal}H(x,{\lambda}_0)}}{\int_X e^{-\beta {\mathcal}H(x',{\lambda}_0)}\, dx'} =\frac1{Z_{{\lambda}_0}}\: e^{-\beta {\mathcal}H(x,{\lambda}_0)},$$ where ${\mathcal}H({\,\cdot\,}, {\lambda}) : X \to \Rnu$ is a Hamiltonian parametrized by an externally controlled parameter ${\lambda}\in {\Lambda}$, $\Lambda {\subset}\Rnu^l$ is an open set; $\beta = 1/(k_BT)$, $k_B$ is the Boltzmann constant, $T$ is the temperature of the system, $Z_{{\lambda}_0}=\int_X e^{-\beta {\mathcal}H(x,{\lambda}_0)}\, dx$ is the partition function. We assume that for all ${\lambda}_0 \in \Lambda$, $\int_X e^{-\beta {\mathcal}H(x,{\lambda}_0)}\, dx < \infty$. We consider the situation when the external parameter ${\lambda}$ is a function of time $[0,T]\to \Lambda$, i.e. we actually consider a time-dependent Hamiltonian ${\mathcal}H(x,{\lambda}(t))$. Let $\mathbb E_{{\lambda}_0}$ denote the expectation relative to the measure $q_{{\lambda}_0}(x)dx$. Below we assume that the changing in time external parameter ${\lambda}$ belongs to the space $$V[0,T]=\{{\lambda}: [0,T]\to \Lambda, \, {\lambda}= {\lambda_{\mathrm{c}}}+{\lambda_{\mathrm{step}}}; \; {\lambda_{\mathrm{c}}}\in {{\rm C}}^V, \, {\lambda_{\mathrm{step}}}\in {\mathcal}L_{\mathrm{step}}\}$$ where ${{\rm C}}^V={{\rm C}}^V[0,T]$ is the space continuous function of bounded variation on $[0,T]$, and $${\mathcal}L_{\mathrm{step}} = {\mathcal}L_{\mathrm{step}}[0,T] = \{ {\lambda}(\,\cdot\,)= \sum\nolimits_{i=0}^{n-1}{\lambda}_i{\mathbb I}_{[t_{i},t_{i+1})}(\,\cdot\,) + {\lambda}_n \}$$ is the space of right continuous step functions corresponding to different partitions ${\mathcal}P = \{0=t_0 < \cdots < t_n=T\}$ of $[0,T]$ and different finite sets of values $\{{\lambda}_i\}$. To emphasize the fact that the function ${\lambda}\in V[0,T]$ gives rise to the process ${\Gamma}_t$, we will use the notation ${\Gamma}_t^{\lambda}$. Let $p_{\lambda}(s,x,t,y)$, $s,\,t\in [0,T]$, $s<t$, $x,\,y\in X$, be the transition density function for $\Gamma_t^{\lambda}$. Let $X^{[0,T]}$ denote the space of all paths $[0,T]\to X$. In terms of $p_{\lambda}$ we can construct a probability measure $\mathbb L_{{\lambda}}$ on $X^{[0,T]}$ by means of the finite dimensional distributions where $\{0 = t_0 < \cdots < t_n =T\}$ is a partition of the interval $[0,T]$, and $f : X^{n+1} \to \Rnu$  is a bounded and measurable function. By Kolmogorov’s extension theorem, the right hand side of this equality defines a probability measure on the minimal ${\sigma}$-algebra of $X^{[0,T]}$ generated by all cylindrical sets. We denote this ${\sigma}$-algebra by ${\sigma_{\mathrm{c}}}(X^{[0,T]})$. We assume that $p_{\lambda}(s,x,t,y)$ satisfies one of the assumptions (see [@skorokhod], chapter 2, paragraph 1) that guaranties that the measure $\mathbb L_{{\lambda}}$ is concentrated on the right continuous trajectories without discontinuities of the second kind. Also, we assume that the ${\sigma}$-algebra ${\sigma_{\mathrm{c}}}(X^{[0,T]})$ is augmented with all subsets of $\mathbb L_{{\lambda}}$-null sets. We take $X^{[0,T]}$ as the probability space, i.e. we set ${\Omega}= X^{[0,T]}$. Then, $\mathbb L_{{\lambda}}$ is the distribution of ${\Gamma}_{t}^{\lambda}$, and $\mathbb L_{\lambda}$-a.s., ${\Gamma}^{\lambda}_t({\omega}) = {\omega}(t)$. For each fixed $\bar {\lambda}\in {\Lambda}$, we introduce another transition density function $p(s,x,t,y,\bar{\lambda})$ which represents the situation when the system evolves being controlled by a constant in time parameter. The following below Assumption \[asn1\] is the key assumption under which the nonequilibrium work theorem holds. \[asn1\] If ${\lambda}|_{[s,t]}\equiv \bar{\lambda}$, then $p_{\lambda}(s,x,t,y)= p(s,x,t,y,\bar{\lambda})$, where $p(s,x,t,y,\bar{\lambda})$ conserves the canonical distribution on the phase space $X$. Specifically, it satisfies the identity We make two further assumptions: \[asn2\] If ${\lambda}$ is constant on $[s,t)$, and discontinuous at the point $t$, then for Lebesgue almost all $x\in X$, å[ &&\_[0+]{} \_X dy p(s,x, t -, y, (s)) \_X dz p\_(t- , y, t, z) f(z)\ &&= \_X dy p(s,x, t, y, (s)) f(y), ]{} where $f: X\to \Rnu$ is bounded and continuous. \[asn2’\] For all $t\in [0,T)$, and for all compacts $K{\subset}X$, $$\lim_{{\delta}\to 0+} \sup_{x\in K} \Bigl[\int_X p(t,x,t+{\delta},y,{\lambda}(t))f(y) dy - f(x)\Bigr] = 0,$$ where $f: X\to \Rnu$ is bounded and continuous. \[lem1\] Let ${\mathcal}P =\{0=t_0<t_1< \cdots < t_n=T\}$ be a partition of $[0,T]$, and let $ {\lambda}(\,\cdot\,)= \sum_{i=0}^{n-1}{\lambda}_i{\mathbb I}_{[t_{i},t_{i+1})}(\,\cdot\,) + {\lambda}_n $. We assume that if ${\lambda}|_{[s,t]}\equiv \bar {\lambda}$, where $\bar {\lambda}\in {\Lambda}$ is constant, then $p_{\lambda}(s,x,t,y)= p(s,x,t,y,\bar{\lambda})$. Further we assume that the function $p_{\lambda}$ satisfies Assumption \[asn2\]. Then, for all $s,\,t$, $0{\leqslant}s < t {\leqslant}T$, for all $x\in X$, and for all bounded and continuous functions $f: X \to \Rnu$, where $\{s< \tau_1 < \cdots < \tau_k < t\} = ({\mathcal}P \cup \{s, t\}) \cap [s,t]$. Moreover, the right hand side of (\[9\]) defines a transition density function. First we prove that the right hand side of (\[9\]) defines a transition density function. We have to verify the Chapman–Kolmogorov equation. Let $r\in (s,t)$, and ${\tilde}{\mathcal}P = {\mathcal}P \cup \{s,r,t\}$. Further let ${\mathcal}P_1 =\{s<\tau_1 < \tau_2 < \cdots < \tau_l\} = {\tilde}{\mathcal}P \cap [s,r]$, and ${\mathcal}P_2 = \{r < \xi_1 < \xi_2 < \cdots < \xi_m < t \}= {\tilde}{\mathcal}P \cap [r,t]$. Then, ${\mathcal}P_1 \cup {\mathcal}P_2 = {\tilde}{\mathcal}P$, and $[\tau_l,\xi_1)$ is the interval of the partition ${\mathcal}P$ such that $r\in [\tau_l,\xi_1)$. We have: Note that ${\lambda}(r)={\lambda}(\tau_l)$, and hence, by the Chapman–Kolmogorov equation for the transition density function $p({\,\cdot\,},{\,\cdot\,},{\,\cdot\,},{\,\cdot\,},{\lambda}(\tau_l))$, we have: $$\int_X dy \, p(\tau_l,x_l,r,y,{\lambda}(\tau_l))\, p(r,y,\xi_1,z_1,{\lambda}(r))= p(\tau_l,x_l,\xi_1,z_1,{\lambda}(\tau_l)).$$ Let us take a bounded and continuous function $f$, and show (\[9\]). which holds for all ${\delta}_1,\ldots, {\delta}_{k}$ smaller than the mesh of ${\mathcal}P$. Taking the repeated limit $\lim_{{\delta}_1\to 0} \lim_{{\delta}_2\to 0} \cdots \lim_{{\delta}_{k}\to 0}$ of the right hand side of (\[long\]), applying Lebesgue’s theorem, and taking into consideration Assumption \[asn2\], we obtain the right hand side of (\[9\]). Jarzynski’s identity ==================== Let ${\lambda}= {\lambda_{\mathrm{c}}}+ {\lambda_{\mathrm{step}}}$ be the decomposition of ${\lambda}\in V[0,T]$ into a sum of a ${\lambda_{\mathrm{c}}}\in {{\rm C}}^V[0,T]$ and a ${\lambda_{\mathrm{step}}}\in {\mathcal}L_{\mathrm{step}}[0,T]$. Further let ${\partial}_{\lambda}{\mathcal}H: X{\times}{\Lambda}\to \Rnu^l$ denote the partial derivative with respect to the second argument (i.e. with respect to the control parameter). We assume that the restriction of ${\partial}_{\lambda}{\mathcal}H$ to $X{\times}{\lambda}([0,T])$ is bounded and measurable. Also, we assume that for each fixed $\bar {\lambda}\in {\lambda}([0,T])$, ${\mathcal}H({\,\cdot\,},\bar {\lambda})$ is bounded and measurable. Everywhere below, the probability space ${\Omega}$ is the space $X^{[0,T]}$. We define the work $W_{\lambda}: {\Omega}\to \Rnu$ performed on the system by the formula where $\{0=t_0 < t_1 < \cdots < t_n = T\}$ are discontinuity points of ${\lambda}$, $\langle \; , \, \rangle_{\Rnu^l}$ is the scalar product in $\Rnu^l \supset\Lambda$, and the integral on the right hand side is the Lebesgue–Stieltjes integral, i.e. the sum of the Lebesgue–Stieltjes integrals with respect to the components of ${\lambda_{\mathrm{c}}}$. In the following, we skip the sign of the scalar product in the first term of (\[8\]), and simply write å[ W\_() & = & \_0\^T \_H(\_t\^(),(t)) d[\_]{}(t)\ & + & \_[i=1]{}\^[n]{} (H(\^\_[t\_i]{}(), (t\_i)) - H(\^\_[t\_i]{}(),(t\_i-0))). ]{} Let $F_{\lambda}= -\frac1{\beta}\, \ln Z_{\lambda}$ (free energy of the system), and let ${\Delta}F = F_{{\lambda}(T)} - F_{{\lambda}(0)}$ (free energy difference). Let $\mathbb E_{\mathbb L_{{\lambda}}}$ denote the expectation relative to the measure $\mathbb L_{\lambda}$. Jarzynski’s Identity, case ${\lambda}\in {\mathcal}L_{\mathrm{step}}[0,T]$ -------------------------------------------------------------------------- Clearly, if ${\lambda}= \sum_{i=0}^{n-1} {\lambda}_i {\mathbb I}_{[t_{i},t_{i+1})} + {\lambda}_n \in {\mathcal}L_{\mathrm{step}}[0,T]$, $$W_{\lambda}({\omega})= \sum_{i=1}^{n} \bigl({\mathcal}H({\Gamma}^{\lambda}_{t_i}({\omega}), {\lambda}_{i}) - {\mathcal}H({\Gamma}^{\lambda}_{t_i}({\omega}),{\lambda}_{i-1})\bigr).$$ \[Jarzynski\_theorem\] Let ${\lambda}\in {\mathcal}L_{\mathrm{step}}[0,T]$, and let the transition density function $p_{\lambda}$ satisfy Assumptions \[asn1\] and \[asn2\]. Further let ${\mathcal}H({\,\cdot\,}, \bar{\lambda})$ be bounded and measurable on $X$ for each fixed $\bar {\lambda}\in {\lambda}([0,T])$. Then the function $e^{-\beta W_{\lambda}}$ is $\mathbb L_{\lambda}$-integrable, and $$\mathbb E_{\mathbb L_{\lambda}}[e^{-\beta W_{\lambda}}] = e^{-\beta{\Delta}F}.$$ $\mathbb L_{\lambda}$-a.s., $$W_{\lambda}({\omega}) = \sum_{i=1}^{n} \Bigl({\mathcal}H({\omega}(t_i),{\lambda}_{i}) - {\mathcal}H({\omega}(t_i),{\lambda}_{i-1})\Bigr).$$ Without loss of generality, we can assume that $t_0=0$, $t_n=T$, adding these points with zero jumps if necessary. Note that $W_{\lambda}({\omega})$ is a cylinder function. By Lemma \[lem1\], Note that Assumption \[asn1\] implies that for all $s<t$, for all $y\in X$, and $\bar{\lambda}\in \Lambda$, Taking into account this, and changing the order of integration in (\[long000\]), we obtain Starting from the end, we replace each integral in (\[long1\]) with $1$, which is valid by the relation (\[chaing\_simplify\_argument\]), until we reach the very first integral (taken with respect to $x_n$), which we replace with $Z_{{\lambda}_n}$. Noticing that ${\lambda}_0 = {\lambda}(0) $ and ${\lambda}_n = {\lambda}(T)$, we obtain $$\mathbb E_{\mathbb L_{\lambda}}[e^{-\beta W_{\lambda}}] = \frac{Z_{{\lambda}(T)}}{Z_{{\lambda}(0)}}.$$ The theorem is proved. Jarzynski’s Identity, case: [${\lambda}\in {{\rm C}}^V[0,T]$]{} --------------------------------------------------------------- \[Jarzynski\_theorem\_gen\] Let ${\lambda}\in {{\rm C}}^V[0,T]$, and let the transition density function $p_{\lambda}$ satisfy Assumptions \[asn1\], \[asn2\], and \[asn2’\], and the probability distribution $\mathbb L_{\lambda}$ of ${\Gamma}^{\lambda}_t$ be given by (\[prob\_dist\]). In addition, let the following assumptions be fulfilled: 4. \[asn3\] If ${\lambda}^n\in {\mathcal}L_{\mathrm{step}}[0,T]$, and as $n\to\infty$, ${\lambda}^n \to {\lambda}$ uniformly on $[0,T]$, then $\mathbb L_{{\lambda}^n} \to \mathbb L_{{\lambda}}$ weakly relative to the family of bounded continuous cylinder functions; 5. \[asn4’\] The function ${\mathcal}H({\,\cdot\,}, \bar{\lambda})$ is bounded and measurable on $X$ for each fixed $\bar {\lambda}\in {\lambda}([0,T])$; 6. \[asn4\] The function ${\partial}_{\lambda}{\mathcal}H$ is bounded on $X{\times}{\lambda}\bigl([0,T]\bigr)$; 7. \[asn44\] The functions ${\partial}_{\lambda}{\mathcal}H(x,{\,\cdot\,})$ are equicontinuous as a family of functions parameterized by $x\in X$; 8. \[asn5\] The function ${\partial}_{\lambda}{\mathcal}H({\,\cdot\,},\bar {\lambda})$ is continuous for each fixed $\bar{\lambda}\in \Lambda$. Then, the function $e^{-\beta W_{\lambda}}$ is $\mathbb L_{\lambda}$-integrable, and Let ${\lambda}\in {{\rm C}}^V[0,T]$. Then, under Assumptions \[asn4\]–\[asn5\], the function $W_{\lambda}({\omega})$ is ${\sigma}_c$-measurable. Below, for an arbitrary small ${\varepsilon}$ we construct a ${\sigma}_c$-measurable function $F: {\Omega}\to \Rnu$ such that We find a function ${\lambda}_{\mathrm{step}}= \sum_{i=0}^{n-1} {\lambda}(t_i) {\mathbb I}_{[t_i,t_{i+1})}+ {\lambda}(t_n) \in {\mathcal}L_{\mathrm{step}}[0,T]$ such that for all ${\omega}\in {\Omega}$, for all $t\in [0,T]$, $$|{\partial}_{\lambda}{\mathcal}H(\Gamma_t^{\lambda}({\omega}),{\lambda}(t))- {\partial}_{\lambda}{\mathcal}H(\Gamma_t^{\lambda}({\omega}),{\lambda}_{\mathrm{step}}(t))|<{\varepsilon}.$$ This is possible by Assumption \[asn44\]. Thus, it suffices to find a measurable function $F$ verifying (\[verify\]) for $W_{\lambda}$ of the form $\int_s^r {\partial}_{\lambda}{\mathcal}H(\Gamma_t^{\lambda}({\omega}),\bar {\lambda})\, d{\lambda_{\mathrm{c}}}(t)$, where $\bar {\lambda}\in {\lambda}([0,T])$ is fixed, and ${\lambda}_c$ is not a constant on $[s,r]$. By assumption, $\mathbb L_{\lambda}$-a.s., the paths of ${\Gamma}^{\lambda}_t$ are right continuous and without discontinuities of the second kind. This implies that the map ${\Gamma}^{\lambda}: [0,T] {\times}{\Omega}\to X$ is $({\mathscr}B([0,T])\otimes {\sigma}_c, {\mathscr}B(X))$-measurable (where ${\mathscr}B([0,T])$ and ${\mathscr}B(X)$ are Borel ${\sigma}$-algebras) which follows from [@skorokhod] (Chapter 2, Theorem 11). By Assumption \[asn4\], ${\partial}_{\lambda}{\mathcal}H({\,\cdot\,},{\,\cdot\,})$ is bounded, say by a constant $M$. We divide the ball of radius $M$ in $\Rnu^l$ with the center in the origin, by a finite number of sets $O_i$ whose diameter is smaller than ${\varepsilon}/V_0^T[{\lambda}_c]$, where $V_0^T[{\lambda}_c]$ denotes the variation of ${\lambda}_c$ on $[0,T]$. Further let $A_i = {\partial}_{\lambda}{\mathcal}H({\,\cdot\,},\bar {\lambda})^{-1}(O_i)$, and $C_i = ({\Gamma}^{\lambda})^{-1}(A_i){\subset}[0,T]{\times}{\Omega}$. The sets $A_i$ are open by Assumption \[asn5\]. The sets $C_i$ are ${\mathscr}B([0,T])\otimes {\sigma}_c$-measurable by the $({\mathscr}B([0,T])\otimes {\sigma}_c, {\mathscr}B(X))$-measurability of the map ${\Gamma}^{\lambda}$. Fix $x_i\in A_i$, and consider the function $\Phi({\omega},t) = \sum_i {\partial}_{\lambda}{\mathcal}H(x_i,\bar {\lambda}){\mathbb I}_{C_i}({\omega},t)$. Clearly, $ \sup_{t\in [0,T],{\omega}\in {\Omega}} |{\partial}_{\lambda}{\mathcal}H({\omega}(t),\bar {\lambda}) - \Phi({\omega},t)|<{\varepsilon}/ V_0^T[{\lambda}_c]. $ We define $$F({\omega}) = \int_s^r \Phi({\omega},t) {\lambda}(dt)= \sum_i {\partial}_{\lambda}{\mathcal}H(x_i,\bar {\lambda}) \mu_{\lambda}(\{t: ({\omega},t) \in C_i\}),$$ where $\mu_{\lambda}$ is the Lebesgue-Stieltjes measure on $[0,T]$ corresponding to the function ${\lambda}$. Fubini’s theorem implies that the function ${\omega}{\mapsto}\mu_{\lambda}(\{t: ({\omega},t) \in C_i\})$ is ${\sigma}_c$-measurable. The inequality (\[verify\]) is obviously satisfied. We take a sequence of partitions ${\mathcal}P_n = \{0=t^n_0 < t^n_1 < \cdots < t^n_n = T\}$, and consider the functions $${\lambda}^n(t) = \sum_{i=0}^{n-1}{\lambda}(t^n_i){\mathbb I}_{[t^n_{i},t^n_{i+1})}(t) + {\lambda}(T).$$ Clearly, ${\lambda}_n \rightrightarrows {\lambda}$ on $[0,T]$, and $V_0^T[{\lambda}^n] < V_0^T[{\lambda}]$, where $V_0^T$ is the variation on $[0,T]$. We set ${\lambda}^n_i={\lambda}(t^n_i)$, and define the functions $$\ffi_n({\omega}) = e^{-\beta \sum_{i=1}^{n} ({\mathcal}H({\omega}(t^n_i),{\lambda}^n_{i}) - {\mathcal}H({\omega}(t^n_i),{\lambda}^n_{i-1}))}.$$ By Theorem \[Jarzynski\_theorem\], $$ \int_{{\Omega}} \ffi_n({\omega})\, \mathbb L_{{\lambda}^n} (d{\omega}) = \frac{Z_{{\lambda}(T)}}{Z_{{\lambda}(0)}}$$ for all $n$. We denote $\ffi({\omega}) = e^{-\beta W_{\lambda}({\omega})}$, and prove that For this, we first replace the functions $\ffi_n$ in (\[term1\])–(\[term3\]) with more suitable functions $\hat \ffi_n$ such that $\bigl(\ffi_n({\omega}) - \hat\ffi_n({\omega})\bigr)\to 0$ as $n\to\infty$, uniformly on ${\Omega}$. We have: where in the first term in (\[2t\]) we applied the mean value theorem to each summand, and chose ${\tilde}{\lambda}_i^n \in [{\lambda}_i^n,{\lambda}_{i+1}^n]$. Since by assumption ${\mathcal}{\partial}_{\lambda}H(x,{\,\cdot\,})$ is equicontinuous, the absolute value of the difference in (\[2t\]) does not exceed ${\varepsilon}V_0^T[{\lambda}]$ where ${\varepsilon}$ is chosen so that $|{\partial}_{\lambda}{\mathcal}H(x,\bar{\lambda}_1) - {\partial}_{\lambda}{\mathcal}H(x,\bar{\lambda}_2)|<{\varepsilon}$ whenever $|\bar {\lambda}_1 - \bar {\lambda}_2| < {\delta}$, and ${\delta}$ is chosen by the equicontinuity argument. The relation (\[2t\]) shows that if we prove (\[term1\])–(\[term3\]) with ${\tilde}\ffi_n({\omega}) = e^{-\beta \int_0^T {\partial}_{\lambda}{\mathcal}H({\omega}(t),{\lambda}^n(t)) \, d{\lambda}^n(t)}$ substituted for $\ffi_n$, then we prove (\[term1\])–(\[term3\]). We define the functions: With the help of these functions, the second term in (\[2t\]) can be represented as By Assumption \[asn44\], we choose an ${\varepsilon}> 0$ so that $|{\partial}_{\lambda}{\mathcal}H({\tilde}{\omega}^n(t,{\omega}),{\tilde}{\lambda}^n(t)) - {\partial}_{\lambda}{\mathcal}H({\tilde}{\omega}^n(t,{\omega}),{\lambda}(t))|<{\varepsilon}$ whenever $\sup_{t\in [0,T]}|{\tilde}{\lambda}^n(t) - {\lambda}(t)| < {\delta}$, and ${\delta}$ is chosen by the equicontinuity argument. This means that the relations (\[term1\]), (\[term2\]), and (\[term3\]) are equivalent to where $\hat \ffi_n({\omega}) = e^{-\beta \int_0^T {\partial}_{\lambda}{\mathcal}H({\tilde}{\omega}^n(t,{\omega}),{\lambda}(t))d{\lambda}(t)}$, and ${\tilde}{\omega}^n(t,{\omega})$ is given by (\[td-om\_n\]). We show that the relations (\[term1aaa\]) hold. Since ${\partial}_{\lambda}{\mathcal}H({\,\cdot\,},{\,\cdot\,})$ is bounded by Assumption \[asn4\], ${\lambda}$ is a function of bounded variation on $[0,T]$, and the exponent is Lipschitz on bounded domains, for all $m$ we obtain the estimate where $K_L$ is the Lipschitz constant for the exponent, $|{\lambda}|(dt)$ is the Lebesgue-Stieltjes measure corresponding to the total variation function $|{\lambda}|(t)$, and Fubini’s theorem has been applied to pass to the last integral. The same estimate holds for $\mathbb L_{\lambda}$. Namely, We would like to show that Let $t\in (t^n_{i-1},t^n_{i}]$, and let ${\delta}_n = t_i^n -t$. Then ${\tilde}{\omega}^n(t,{\omega}) = {\omega}(t + {\delta}_n)$. Let us show (\[repeated\]). We have: å[ && \_ L\_[\^n]{} (d) |\_H((t+\_n),(t))-\_H((t),(t))|\ = & & \_X q\_[\_0]{}(x\_0) dx\_0 \_X dx p\_[\^n]{}(t\_0,x\_0,t,x)\ & & \_X dx\_i p(t,x,t+\_n,x\_i,(t)) |\_H(x,(t))-\_H(x\_i,(t))|. ]{} Note that å[ F\_[n]{} (x)= \_X dx\_i p(t,x,t + \_n,x\_i, (t)) |\_H(x\_i,(t)) -\_H(x,(t))| ]{} converges to zero uniformly in $x$ running over compacts in $X$ which follows from Assumption \[asn2’\]. Indeed, å[ |F\_n(x)|\^2 \_X dx\_i p(t,x,t + \_n,x\_i, (t)) |\_H(x\_i,(t)) -\_H(x,(t))|\^2. ]{} The right hand side of this inequality converges to zero uniformly in $x\in K {\subset}X$, where $K$ is an arbitrary compact. This easily follows from Assumption \[asn2’\] after we separate in the right hand side the terms depending on $x$ and on $x_i$. Define the measures: å[ &\_[\^n]{}(A) = \_X q\_[\_0]{}(x\_0) dx\_0 \_A p\_[\^n]{}(t\_0,x\_0,t,x) dx,\ &\_(A) = \_X q\_[\_0]{}(x\_0) dx\_0 \_A p\_(t\_0,x\_0,t,x) dx. ]{} By Assumption \[asn3\], $\mathbb L_{{\lambda}^n}$ converges weakly to $\mathbb L_{{\lambda}}$ relative to the family of bounded continuous cylinder function, whenever ${\lambda}^n \rightrightarrows {\lambda}$. This implies that $\mu_{{\lambda}^n} \to \mu_{\lambda}$ weakly (relative to the family of bounded continuous functions), as ${\lambda}^n \rightrightarrows {\lambda}$. By Prokhorov’s theorem, the family $\{\mu_{{\lambda}^n}, \mu_{\lambda}\}$ of probability measures on $X$ is tight. We fix an arbitrary ${\varepsilon}> 0$ and find a compact $K_{\varepsilon}$ such that $\mu_{{\lambda}^n}(X {\diagdown}K_{\varepsilon}) < {\varepsilon}$ for all $n$. Since $F_n \rightrightarrows 0$ on $K_{\varepsilon}$, and all $F_n$ are bounded on $X$ by a constant, say $M$, not depending on $n$, we can find an $N \in \Nnu$ such that $|F_n(x)|<{\varepsilon}$ on $K_{\varepsilon}$ for all $n>N$. We obtain the estimate: $$\Bigl| \int_X \mu_{{\lambda}^n}(dx) F_n(x) \Bigr| {\leqslant}\sup_{x\in K_{\varepsilon}} |F_n(x)| + M \mu_{{\lambda}^n}(X{\diagdown}K_{\varepsilon}) {\leqslant}{\varepsilon}+ M {\varepsilon}.$$ This proves (\[repeated\]). The relation (\[single\]) follows from the right continuity of ${\omega}$, Assumption \[asn5\], and Lebesgue’s theorem. The relation (\[repeated1\]) follows from Assumption \[asn3\], and from (\[single\]). Application of Lebesgue’s theorem to the integral (\[show\]) (taken over $[0,T]$ with respect to $|{\lambda}|(dt)$) implies that as $n\to\infty $, the integral (\[show\]) converges to zero. This implies (\[term1aaa\]), and thus (\[term1\]), (\[term2\]), and (\[term3\]) are proved. By (\[term1\]), we obtain: å[ &0& = \_[n]{} (\_n L\_[\^n]{} - L\_[\^n]{}) = \_[n]{} \_n L\_[\^n]{} - \_[n]{} L\_[\^n]{}\ & & = - \_[n]{} L\_[\^n]{}. ]{} On the other hand, (\[term2\]) and (\[term3\]) imply: å[ &0& = \_[n]{}\_[m]{} (\_[\^n]{} L\_[\^m]{} - L\_[\^m]{})\ & & = \_[n]{}\_[m]{} \_n L\_[\^m]{} - \_[m]{} L\_[\^m]{}\ & & = \_[n]{} \_n L\_- \_[m]{} L\_[\^m]{} = L\_- \_[m]{} L\_[\^m]{}. ]{} Comparing the last two relations gives: $$\int_{\Omega}\ffi({\omega}) \mathbb L_{\lambda}(d{\omega}) = \frac{Z_{{\lambda}(T)}}{Z_{{\lambda}(0)}}. $$ The theorem is proved for the case ${\lambda}\in {{\rm C}}^V[0,T]$. Jarzynski’s identity for ${\lambda}\in V[0,T]$ and its corollaries ------------------------------------------------------------------ \[corollary1\] Let the assumptions of Theorem \[Jarzynski\_theorem\_gen\] be fulfilled, and let $f:~X\to \Rnu$ be bounded and continuous. Then, where $\pi_t: X^{[0,T]}\to X, {\omega}{\mapsto}{\omega}(t)$ is the evaluation mapping, $\mathbb E_{{\lambda}(T)}$ is the expectation relative to the measure $\frac1{Z_{{\lambda}(T)}}\, e^{-\beta \,{\mathcal}H(x,{\lambda}(T))}dx$. Assuming that ${\lambda}\in {\mathcal}L_{\mathrm{step}}[0,T]$, we repeat the argument of (\[long\]) and (\[long1\]), while using the relation (\[chaing\_simplify\_argument\]). Specifically, we obtain: å[ & &E\_[L\_]{}\[(f\_T) e\^[- W\_]{}\]= 1 [Z\_[(t\_0)]{}]{} \_X [e\^[-H(x\_[n]{}, (t\_[n]{}))]{}]{} f(x\_n) dx\_n\ & & \_X p(t\_[n-1]{},x\_[n-1]{},t\_[n]{},x\_[n]{},(t\_[n-1]{})) dx\_[n-1]{} …\ & &\ & & \_X p(t\_0,x\_[0]{},t\_1,x\_[1]{},(t\_[0]{})) dx\_[0]{}\ & =& E\_[(T)]{}\[f\] = E\_[(T)]{}\[f\] E\_[L\_]{}\[e\^[- W\_]{}\]. ]{} Now let ${\lambda}\in {{\rm C}}^V[0,T]$, and let the constant $M_f$ be such that $\sup_{x\in X}|f(x)| < M_f$. As in the proof of Theorem \[Jarzynski\_theorem\_gen\], we find a sequence of step functions ${\lambda}^n$ converging to ${\lambda}$ uniformly on $[0,T]$. The identity (\[corollary\_relation\]) holds for each ${\lambda}^n$. The argument of passing to the limit as $n\to \infty$ is similar to the argument used in the proof of Theorem \[Jarzynski\_theorem\_gen\]. We define the function $F: {\Omega}\to \Rnu$, $F({\omega}) = f({\omega}(T))$, and repeat the argument of Theorem \[Jarzynski\_theorem\_gen\] until the inequality (\[show1\]) with the following replacements: $\ffi \leftrightarrow F \ffi$, $\ffi_n \leftrightarrow F \ffi_n$, ${\tilde}\ffi_n \leftrightarrow F {\tilde}\ffi_n$, $\hat\ffi_n \leftrightarrow F \hat\ffi_n$. The right hand sides of (\[show\]) and (\[show1\]) will be the same as in the proof of Theorem \[Jarzynski\_theorem\_gen\] but the Lipschitz constant $K_L$ will be replaced with $K_L M_f$. The part of the proof of Theorem \[Jarzynski\_theorem\_gen\] following after the inequality (\[show1\]) remains unchanged until the last two arguments of passing to the limit. Those arguments now will be: å[ &0& = \_[n]{} (F\_n L\_[\^n]{} - FL\_[\^n]{}) = \_[n]{} F \_n L\_[\^n]{} - \_[n]{} F L\_[\^n]{}\ & & = E\_[(T)]{}\[f\] E\_[L\_]{}\[e\^[- W\_]{}\] - \_[n]{}F L\_[\^n]{}. ]{} On the other hand, This implies the identity (\[corollary\_relation\]). \[corollary22\] Let the assumptions of Corollary \[corollary1\] except Assumption \[asn3\] of Theorem \[Jarzynski\_theorem\_gen\] be fulfilled with respect to the function ${\lambda}\in {{\rm C}}^V[0,T]$. Let $a\in \Lambda$, and let ${\lambda}^a = {\lambda}\,{\mathbb I}_{[0,T)} + a \,{\mathbb I}_{\{T\}}$. Let us assume that - If ${\lambda}^n\in {\mathcal}L_{\mathrm{step}}[0,T]$, and ${\lambda}^n \rightrightarrows {\lambda}^a$ on $[0,T]$, then $\mathbb L_{{\lambda}^n} \to \mathbb L_{{\lambda}^a}$ weakly relative to the family of bounded continuous cylinder functions; Then, If ${\lambda}^a \in {\mathcal}L_{\mathrm{step}}[0,T]$, i.e. ${\lambda}^a = \sum_{i=1}^{n-1} {\lambda}_i \,{\mathbb I}_{[t_i,t_{i+1})} + a\,{\mathbb I}_{\{T\}}$, then the representation (\[9\]) shows that $\mathbb E_{\mathbb L_{{\lambda}^a}}[(f\circ\pi_T) \, e^{-\beta \, W_{{\lambda}}}]$ is the same for the functions ${\lambda}^a$ with different $a$. Hence, $$\mathbb E_{\mathbb L_{{\lambda}^a}}[(f\circ\pi_T) \, e^{-\beta \, W_{{\lambda}}}] = \mathbb E_{{\lambda}(T)}[f]\; \frac{Z_{{\lambda}(T)}}{Z_{{\lambda}(0)}}\,.$$ Note that in the relations (\[show\]) and (\[show1\]) we integrate with respect to $|{\lambda}|(dt)$ where $|{\lambda}|$ is continuous. Hence, by Lebesgue’s theorem, the investigation of convergence at the point $T$ in the integrals with respect to $\mathbb L_{{\lambda}_m}(d{\omega})$ and $\mathbb L_{{\lambda}}(d{\omega})$ is not necessary. Finally we note that by assumption, and by the relations (\[show1\]) and (\[single\]), with ${\lambda}^a$ substituted for ${\lambda}$, for the first term in the identity (\[98\]) we obtain: $$\lim_{n\to\infty}\lim_{m\to\infty} \int F\ffi_n \mathbb L_{{\lambda}^m} = \int F\ffi \, \mathbb L_{{\lambda}^a}.$$ This proves (\[cor\_id\_modified\]). Let $[s,t]{\subset}[0,T]$ be an arbitrary subinterval, $X^{[s,t]}$ be the space of all functions on $[s,t]$. We introduce further notation: Let ${{\rm C}}^V[s,t]$ be the space of continuous functions of bounded variation on $[s,t]$, ${\mathcal}L_{\mathrm{step}}[s,t]$ be the space of right continuous step functions on $[s,t]$ taking a finite number of values. Below, on $X^{[s,t]}$ we define the distributions $\mathbb L^x_{{\lambda};s,t}$ and $\mathbb L_{{\lambda};s,t}$, ${\lambda}\in V[0,T]$, $x\in X$. Let $s < \tau_1 < \cdots < \tau_{k-1} < t$ be a partition of $[s,t]$, $f: X^{k+1}\to \Rnu$ be bounded and measurable. The finite dimensional distributions of $\mathbb L_{{\lambda};s,t}^x$ are defined by the formula: $$\mm{ {l} {\displaystyle}\int_{X^{[s,t]}} f({\omega}(s), {\omega}(\tau_1), \ldots, {\omega}(t))\,\mathbb L_{{\lambda};s,t}^x(d{\omega}) = \\ {\noalign{\vskip 0.2pc}}{\displaystyle}\intl_X \hspace{-1mm} dx_1 \, p_{\lambda}(s, x, \tau_1, x_1)) \cdots \intl_X \hspace{-1mm} dx_k \, p_{\lambda}(\tau_{k-1},x_{k-1}, t, x_k)\,f(x, x_1,\ldots, x_k). }$$ We extend $\mathbb L^x_{{\lambda};s,t}$ to ${\sigma}_c(X^{[s,t]})$ by Kolmogorov’s extension theorem, and notice that under assumptions that we made, this measure it concentrated on the right continuous paths without discontinuities of the second kind. Let $f$ be $\mathbb L_{{\lambda};s,t}^x$-integrable. We define \[Jarzynski\_theorem\_gen\_gen\] Let ${\lambda}\in V[0,T]$, and let $\{t_1 < \dots < t_k = T\}$ be the set of its discontinuity points. Further let the transition density function $p_{\lambda}$ of the Markov process ${\Gamma}^{\lambda}_t$ satisfy Assumptions \[asn1\], \[asn2\], and \[asn2’\], and the probability distribution $\mathbb L_{\lambda}$ be given by (\[prob\_dist\]). Let Assumptions \[asn4’\], \[asn4\], \[asn44\], and \[asn5\] of Theorem \[Jarzynski\_theorem\_gen\] be fulfilled. Additionally, we assume that 8. \[asn8\] If ${\lambda}^m \in {\mathcal}L_{\mathrm{step}}[0,T]$, and ${\lambda}^m \rightrightarrows {\lambda}$ on $[0,T]$, as $m\to\infty$, then for each $i$, $0 {\leqslant}i < n$, $\mathbb L_{{\lambda}^m; t_i,t_{i+1}} \to \mathbb L_{{\lambda}; t_i,t_{i+1}}$ weakly relative to the family of bounded continuous cylinder functions on $X^{[t_i,t_{i+1}]}$; 9. \[asn9\] For all bounded and continuous functions $f: X^{[t_i,t_{i+1}]} \to \Rnu$, the function $X \to \Rnu, \; x{\mapsto}\mathbb L^x_{{\lambda};t_i,t_{i+1}}[f]$ is continuous. Then, the function $e^{-\beta W_{\lambda}}$ is $\mathbb L_{\lambda}$-integrable, and The continuity of functions $X^{[t_i,t_{i+1}]} \to \Rnu$ is understood with respect to the topology of pointwise convergence. Let ${\lambda}= {\lambda_{\mathrm{c}}}+ {\lambda_{\mathrm{step}}}$, where ${\lambda_{\mathrm{c}}}\in {{\rm C}}^V[0,T]$, and ${\lambda_{\mathrm{step}}}= \sum_{i=0}^{n-1} {\lambda}_i {\mathbb I}_{[t_{i},t_{i+1})}+ {\lambda}_n \in {\mathcal}L_{\mathrm{step}}[0,T]$. On each interval $[t_{i-1},t_i]$ we define the function ${\omega}_i = {\omega}|_{[t_{i-1},t_i]}$, and identify each ${\omega}\in X^{[0,T]}$ with the sequence $({\omega}_1, \ldots, {\omega}_n)$. Since ${\Gamma}^{\lambda}_t$ is a Markov process, we obtain: $$\mm{ {l} {\displaystyle}\int_{X^{[0,T]}} e^{-\beta W_{\lambda}({\omega})} \mathbb L_{{\lambda}}(d{\omega}) = \int_X dx_0 \, q_{{\lambda}(t_0)}(x_0) \int_{X^{[t_0,t_1]}} \mathbb L_{{\lambda};t_0,t_1}^{x_0}(d{\omega}_1)\\ {\noalign{\vskip 0.2pc}}{\displaystyle}\int_{X^{[t_1,t_2]} }\mathbb L_{{\lambda};t_1,t_2}^{{\omega}_1(t_1)}(d{\omega}_2) \cdots \int_{X^{[t_{n-1},t_n]}} \mathbb L_{{\lambda};t_{n-1},t_n}^{{\omega}_{n-1}(t_{n-1})}(d{\omega}_n) e^{-\beta W_{\lambda}({\omega}_1, \ldots, {\omega}_n)} }$$ where $$\mm{ {ll} {\displaystyle}W_{\lambda}({\omega}_1, \ldots, {\omega}_n) & = {\displaystyle}\sum_{i=1}^{n} \int_{t_{i-1}}^{t_{i}} {\partial}_{\lambda}{\mathcal}H({\omega}_i(t),{\lambda}(t))d{\lambda_{\mathrm{c}}}(t)\\ {\noalign{\vskip 0.2pc}}& {\displaystyle}+ \sum_{i=1}^{n} \bigl({\mathcal}H({\omega}_i(t_i),{\lambda}(t_i)) - {\mathcal}H({\omega}_i(t_i),{\lambda}(t_i-0))\bigr). }$$ On each interval $[t_{i-1},t_i]$, we define continuous functions ${\lambda}_{(i)}(t) = {\lambda}(t)$, $t\in [t_{i-1},t_{i})$, and ${\lambda}_{(i)}(t_{i})= {\lambda}(t_{i}-0)$, $1{\leqslant}i {\leqslant}n$, and introduce a notation: $$W_{{\lambda}_{(i)}}({\omega}_i) = \int_{t_{i-1}}^{t_i} {\partial}_{\lambda}{\mathcal}H({\omega}_i(t),{\lambda}(t))d{\lambda}_{(i)}(t).$$ We obtain: $$\mm{ {l} {\displaystyle}\intl_{X^{[0,T]}} e^{-\beta W_{\lambda}({\omega})} \mathbb L_{{\lambda}}(d{\omega}) = {\displaystyle}\int_X dx_0 \, q_{{\lambda}(t_0)}(x_0)\\ {\noalign{\vskip 0.2pc}}{\displaystyle}\intl_{X^{[t_0,t_1]}} e^{-\beta W_{{\lambda}_{(1)}}({\omega}_1)} e^{-\beta \bigl( {\mathcal}H({\omega}_1(t_{1}),{\lambda}(t_{1})) - {\mathcal}H({\omega}_1(t_{1}), {\lambda}(t_{1}-0)) \bigr)} \mathbb L_{{\lambda};t_0,t_1}^{x_0}(d{\omega}_1)\\ {\noalign{\vskip 0.2pc}}{\displaystyle}\intl_{X^{[t_1,t_2]}} e^{-\beta W_{{\lambda}_{(2)}}({\omega}_2)} e^{-\beta \bigl( {\mathcal}H({\omega}_2(t_{2}),{\lambda}(t_{2})) - {\mathcal}H({\omega}_2(t_{2}), {\lambda}(t_{2}-0)) \bigr)} \mathbb L_{{\lambda};t_1,t_2}^{{\omega}_1(t_1)}(d{\omega}_2) \cdots \\ {\noalign{\vskip 0.2pc}}{\displaystyle}\intl_{X^{[t_{n-1},t_{n}]}} \hspace{-5mm} e^{-\beta W_{{\lambda}_{(n)}}({\omega}_{n})} e^{-\beta \bigl( {\mathcal}H({\omega}_n(t_{n}),{\lambda}(t_{n})) - {\mathcal}H({\omega}_n(t_{n}), {\lambda}(t_{n}-0)) \bigr)} \mathbb L_{{\lambda};t_{n-1},t_n}^{{\omega}_{n-1}(t_{n-1})}(d{\omega}_{n}). }$$ Let $\mathbb E_{\bar {\lambda}}$ denote the expectation relative to the measure $q_{\bar{\lambda}}(x)dx$. We can rewrite the above relation in terms of expectations: Let $F: X^{[t_i,t_{i+1}]} \to \Rnu$ be bounded and continuous. We obtain: Note that in (\[above\]) $F$ is always a function of the form where $f: X\to \Rnu$ is bounded and continuous by Assumptions \[asn5\] and \[asn9\]. Taking the expectation $\mathbb E_{\mathbb L_{{\lambda}; t_i,t_{i+1}}}$ of the function (\[measurable\]), and applying Corollary \[corollary22\] we obtain: Applying the identities (\[55\]) and (\[this\_id\]) to (\[above\]), we obtain: \[corollary2\] Let the assumptions of Theorem \[Jarzynski\_theorem\_gen\_gen\] be fulfilled, and let $f:~X\to \Rnu$ be bounded and continuous. Then, We repeat the argument that we used in the proof of Theorem \[Jarzynski\_theorem\_gen\_gen\] to obtain (\[above\]) in connection to the expression $\mathbb E_{\mathbb L_{\lambda}}[(f\circ\pi_T) \, e^{-\beta \, W_{\lambda}}]$. Instead of the very last expectation in (\[above\]), we obtain $$\mathbb E_{\mathbb L^{{\omega}_{n-1}(t_{n-1})}_{{\lambda}; t_{n-1},t_n}}[ e^{-\beta W_{{\lambda}_{(n)}}({\omega}_n)}\, f({\omega}_n(t_n)) \, e^{-\beta \bigl( {\mathcal}H({\omega}_n(t_n),{\lambda}(t_n)) - {\mathcal}H({\omega}_n(t_n), {\lambda}(t_n-0))\bigr)}].$$ Applying Corollary \[corollary22\], we obtain: Instead of the multiplier $\frac{Z_{{\lambda}(t_{n}-0)}}{Z_{{\lambda}(t_{n-1})}}\,\frac{Z_{{\lambda}(t_n)}}{Z_{{\lambda}(t_{n}-0)}}$ in the relation (\[last\]) we obtain $\frac{Z_{{\lambda}(t_n-0)}}{Z_{{\lambda}(t_{n-1})}}\, \frac{Z_{{\lambda}(t_n)}}{Z_{{\lambda}(t_n-0)}}\, \mathbb E_{{\lambda}(t_n)}[f]$ according to (\[29\]). This gives (\[corollary\_relation\_gen\]). Bochkov–Kuzovlev’s identity =========================== Here we give a rigorous mathematical proof of the identity announced in [@bochkov1]–[@bochkov4] where the authors used a different definition of work to that in the papers [@Jar97_1]–[@Jar97]. The difference between the two definitions of work was analyzed in the paper [@Jar07], and it was found that $$W({\omega})-W_0({\omega}) = {\mathcal}H({\omega}(T),{\lambda}(T)) - {\mathcal}H({\omega}(T),{\lambda}(0)),$$ where $W_0$ is the work in the Bochkov–Kuzovlev sense. We use this equality as the definition of the new work $W_0$, and will prove the following theorem: Let the assumptions of Theorem \[Jarzynski\_theorem\_gen\_gen\] be fulfilled. Then, Moreover, the identities (\[BK\_identity\]) and (\[J\_identity\_gen\]) are equivalent. Applying Corollary \[corollary2\] and the identity (\[J\_identity\_gen\]) we obtain: å[ E\_[L\_]{} \[ e\^[-W\_0]{}\] & = & E\_[L\_]{}\[ e\^[-W]{}\] E\_[(T)]{} \[e\^[-(H((T),(T)) - H((T),(0)))]{}\]\ & = & E\_[L\_]{}\[ e\^[-W]{}\] = = 1. ]{} This relation also shows that the identities (\[BK\_identity\]) and (\[J\_identity\_gen\]) are equivalent. The author thanks Professor C. Dellago for attracting her attention to Jarzynski’s identity, and Professor O. G. Smolyanov for useful discussions. This work was supported by the research grant of the Erwin Schrödinger Institute for mathematical physics, by the Austrian Science Fund (FWF) under START-prize-grant Y328, and by the research grant of the Portuguese Foundation for Science and Technology (FCT). [99]{} C. Jarzynski, Nonequilibrium equality for free energy differences, Phys Rev Lett [**78**]{}, 2690, pp. 2690–2693, 1997. C. Jarzynski, Equilibrium free-energy differences from nonequilibrium measurements: A master-equation approach, Phys. Rev. E [**56**]{}, 5018–5035, 1997. C. Jarzynski, Comparison of far-from-equilibrium work relations, Comptes Rendus Physique, [**8**]{}, 495–506, 2007. G. N. Bochkov, Yu. E. Kuzovlev, General theory of thermal fluctuations in nonlinear systems, Zh. Eksp. Teor. Fiz. [**72**]{}, 238–243, 1977; \[JETP [**45**]{}, 125–130, 1977\]. G. N. Bochkov, Yu. E. Kuzovlev, Fluctuation-dissipation relations for nonequilibrium processes in open systems, Zh. Eksp. Teor. Fiz. [**76**]{}, 1071–1079, 1979; \[JETP [**49**]{}, 543–551, 1979\]. G. N. Bochkov, Yu. E. Kuzovlev, Nonlinear fluctuation-dissipation relations and stochastic models in nonequilibrium thermodynamics, I. Generalized fluctuation-dissipation theorem, Physica [**106A**]{}, 443–479, 1981. G. N. Bochkov, Yu. E. Kuzovlev, Nonlinear fluctuation-dissipation relations and stochastic models in nonequilibrium thermodynamics, II. Kinetic potential and variational principles for nonlinear irreversible processes, Physica [**106A**]{}, 480, 1981. G. E. Crooks, Path-ensemble averages in systems driven far from equilibrium, Physical review E [**61**]{}, pp. 2361–2366, 2000. G. E. Crooks, Nonequilibrium Measurements of Free Energy. Differences for Microscopically Reversible Markovian Systems. Journal of Statistical Physics, [**90**]{}, Nos 5/6, pp. 1481–1487, 1998. G. E. Crooks, Entropy production fluctuation theorem and the nonequilibrium work relation for free energy differences, Physical Review E [**60**]{}, pp. 2721–2726, 1999. C. Dellago, E. Schöll-Paschinger, A proof of Jarzynski’s non-equilibrium work theorem for dynamical systems that conserve the canonical distribution, J. Chem. Phys. [**125**]{}, 2006. C. Dellago, P. L. Geissler, W. Lechner, H. Oberhofer, Equilibrium free energies from fast-switching simulations with larg time steps, Journal of Chemical Physics [**124**]{}, 044113, 2006. E. B. Dynkin, Markov processes, Academic press, 1965. G. Hummer, A. Szabo, Free energy reconstruction from nonequilibrium single-molecule pulling experiments, Proc. Nat. Acad. Sci. [**98**]{}, 3658, 2001. G. Hummer, A. Szabo, Free energy surfaces from single-molecule force spectroscopy. Acc Chem Res [**38**]{}, 504–513, 2005. J. Liphardt et al., Equilibrium information from nonequilibrium measurements in an experimental test of Jarzynski’s equality, Science [**296**]{}, 1832, 2002. N. I. Portenko, A. V. Skorokhod, B. M. Shurenkov, Markov processes, Itogi nauki i techniki. Sovremennye problemy matematiki. Fundamentalnye napravleniya. [**46**]{}, pp. 5–248, 1989 (in Russian).

ArXiv

--- abstract: 'The critical boundaries separating ordered from chaotic behavior in randomly wired $S$-state networks are calculated. These networks are a natural generalization of random Boolean nets and are proposed as on extended approach to genetic regulatory systems, sets of cells in different states or collectives of agents engaged into a set of $S$ possible tasks. A order parameter for the transition is computed and analysed. The relevance of these networks to biology, their relationships with standard cellular automata and possible extensions are outlined.' --- 16 cm 20 cm 0 cm -1cm Phase Transition in Random Networks with Multiple States [ Ricard V. Solé $^{1,2}$, Bartolo Luque $^{3}$ and Stuart Kauffman $^2$]{} \(1) Complex Systems Research Group, Department of Physics - FEN Universitat Politécnica de Catalunya Campus Nord, Mòdul B5 , 08034 Barcelona (Spain) \(2) Santa Fe Institute 1399 Hyde Park Road, Santa Fe, New Mexico (USA) \(3) Centro de Astrobiología (CAB), Ciencias del Espacio, INTA Carretera de Ajalvir km. 4, 28850 Torrejón de Ardoz, Madrid (Spain) =5 mm [**Submitted to Physical Review E**]{} Introduction ============ Cellular automata (CA) \[1\] and random Boolean networks (RBN) \[2\] are two key approaches in our understanding of complexity \[3,4\]. In both cases, discrete-time and discrete-space are used, together with a finite number of states. For CA, we have a dynamical evolution defined (for the one-dimensional case with a $(2r+1)$-neighborhood) by means of: $$\sigma_i(t+1) = \Phi \Bigl [ \; \sigma_{i-r} (t) , \sigma_{i-r+1}(t) , ... , \sigma_{i+r}(t) \; \Bigr ] \eqno (1)$$ where $i=1,...,N$ and $\sigma_i(t) \in \Sigma \equiv \{ 0, 1, ... , S-1 \}$. Here $\Phi$ is a given (previously defined) rule \[1\]. Identical rules are defined for each $(2r+1)$-neighborhood. In contrast with this homogenity in neighbors and rules, RBN’s are a particularly interesting class of dynamical systems which shares some properties with CA models, but where randomness is introduced at several levels. Now the dynamics is given by: $$\sigma_i(t+1) = \Lambda_i \; \Bigl [ \; \sigma_{i_1}(t) , \sigma_{i_2}(t) , ... , \sigma_{i_K}(t) \; \Bigr ] \eqno (2)$$ where $i=1,...,N$ and $\sigma_i(t) \in \Sigma \equiv \{0,1\}$. Each automaton is randomly connected with exactly $K$ others wich send inputs to him. Here $\Lambda_i$ is a Boolean function also randomly choosen from a set ${\cal F}_K $ of all possible $K$-inputs Boolean functions. In spite of the random selection both of neighbors and Boolean functions, it is well known that a critical connectivity $K_c$ exists where “spontaneous order crystallizes” \[2\]. In $K_c$ a small number of attractors $(\approx O(\sqrt N))$ is observed which show high homeostatic stability (i.e. high stability against minimal perturbations in single elements or Boolean functions), and low reachability among different attractors. These properties are consistent with some observations of the genome organization \[2,5\]. A considerable effort has been dedicated over the last years to the analysis (both theoretical and computational) of the scaling behavior of the numbers and size of cycles close to the critical boundary \[5-7\]. Alternative models to this problem are provided by neural network-like aproaches \[8\]. In this case we have a dynamical system given by the set: $$\sigma_i(t+1) = \Phi \lbrack J_{ii}\sigma_i(t)+ \sum_{l=1}^{K-1} J_{ij_{l}(i)}\sigma_{j_{l}(i)}(t)+\theta_i\rbrack \eqno (3)$$ where $\Phi(x)$ is a sigmoidal function which can be assimptotically approched by a sign function, $\Phi(x)= sgn (x)$ which is $sgn(x) =0$ if $x\leq 0$ and $sgn (x) = 1$ otherwise. Not surpringly, neural networks with random connectivities show some of the properties displayed by RBN. Two well-defined phases are also described, which also depend upon the input connectivity \[9\]. Though the picture of on-off genes is a simplified one, in some cases it can be justified from general arguments based on mutual inhibition among genes \[10\]. Let us consider two given genes which interact among them by means of two given proteins, whose concentrations are indicated by the pair $(x,y)$. This can be used as a toy model of some well known properties of the life cycle of viruses (like the $\lambda$-Phage \[3,10\]). A theoretical model of mutual inhibition is given by the following couple of cross-inhibitory equations: $${dx \over dt} = {\mu^n \over \mu^n + y^n} - x \eqno (4.a)$$ $${dy \over dt} = {\mu^n \over \mu^n + x^n} - y \eqno (4.b)$$ where $x,y \in \bf R^+$. Here $n \in \bf N$ is a measure of the strength of the interaction. For $\mu=1/2$, the stability of the fixed point $P^*=(1/2,1/2)$ is given by the eigenvalues of the Jacobi matrix: $$L_{\mu}(P^*) = \left ( \matrix{ -1 & {-{n \over 2}} \cr {-{n \over 2}} & -1 \cr} \right ) \eqno (5)$$ which leads to stability for $n<2$ and to instability otherwise. For $n>n_c=2$, $P^*$ becomes a saddle point and two new stable attractors $\{ P_+, P_-\}$ are formed (stable nodes). For strong interactions, one of the concentrations dominate over the other and, basically, the two new attractors are simply $P_+ \approx (1,0)$ and $P_- \approx (0,1)$. So the result of this interaction between continuous quantities leads to a basically binary outcome \[11\]. But gene activity can lead to more complex combinations (see figure (1) as example). And in fact experimental analysis of RNA levels in developing embryos shows a wide range of gene activity levels. Different genes interact in very complex ways, and activation or represion takes place in such a way that the same gene can be more or less active depending on the coupling of proteins to different regulatory sites \[12\]. To see this, let us consider a general network formed by $m$ elements showing mutual inhibition, i.e.: $${dx_i \over dt} = f_{\mu}^{(i)} (x_1, x_2, ... , x_n) = {\mu^n \over \mu^n + \sum_{j \ne i} x_j^n} - x_i \eqno (6)$$ With $i=1,...,m$. This network leads to the emergence of a (usually large) number of attractors when $n$ is large enough. An example of the phase space trajectories for $m=3$ and $\mu=1/2$ is shown in figure 2. We can see in figure (2) upper that for low $n$ values ($n=1$) only a coexistence point is reached. However, see figure (2) lower, once the nonlinearity is increased beyond a treshold, different attractors are obtained, indicated by the flow of the vector field (here $n=4$). Three corners are occupied by the simplest attractors (molecular switches with $P^* \in \{(1,0,0), (0,1,0), (0,0,1)\}$). But we also have $P^*=(x_1^*, x_2^*, x_3^*)$ with $x_1^* \neq x_2^* \neq x_3^* \in \{ \alpha, \beta, \gamma \}$ (here $\alpha=0.13, \beta=0.34, \gamma = 0.97$). So three different states are available for each variable. For a more complex network, many possible states will be possible displaying a range of intermediate states. In order to avoid this problem other approximations have been proposed. These approximations retain the discreteness of the units but consider a continuous range of values \[11, 14\] more close to the real picture, where gene activity can be measured in terms of RNA concentrations \[12\]. These results makes necessary a generalization of RBNs extending them into a more general framework. In particular, we can ask how are the critical boundaries defined for a more general situation when multiple states are available. Although the starting point of our study deals whith genomic regulatory circuits, a set of different, but related problems could be explored within this framework. If a given state is identified in terms of a cell type, for example, the interactions between differents cells (automata) with different cell type will eventually lead to transitions which are known to be rather complex and neighbour-dependent \[12\]. Another example could be a system where each unit can have a complex internal state roughly characterized by scalar quantity (or “state”) which can be, for example, the mean activity. A further scenario where our description applies is a system of machines (computers or agents) each one engaged in a given task. These machines might be connected in complex ways, and switch from a task to another by depending on their actual inputs. Discrete dynamical systems involving a given range of states per unit are well known in the physics literature. Cellular automata (as described by equation (1)) are just the best known example \[1\]. Another example from statistical mechanics is the $q-$ Potts model \[13\] which generalizes the classical Ising spin systems. It well known that this model shows phase transitions with a critical temperature $T_c(q)$ depending on $q$, the number of possible spin states. Our goal in this paper is to explore the properties of the critical points for $S$-state random networks. In this context, the analysis of transition phenomena in CA shows that, for large r-neighborhoods, a sharp (second order ) transition occurs for well-defined $\lambda$-values. However this result cannot be traslated to the genome because typically gene regulation is a non-local phenomenon \[8,12\]. We should expect, however, to find some agreement between the CA results and those derived from networks involving multiple states and random connectivity. The paper is organized the follows manner: in section 2, random networks with multiple states are define and the critical surface in the $(K,S,p)$ space is derived from Derrida’s annealed approximation. In the the Appendix I associated with this section a maximum entropy variational approach is used in order to compute the expected distribution of states at criticality. In section 3 the relationship with CA models for large number of states $S$ is analysed. In section 4 a variant of Flyvbjerg’s method of stable core computation is introduced in order to obtain an order parameter (the [*self-overlap*]{}) for the transition. In the Appendix II associated with this section we compute a Lyapunov exponent for the system. Finally, a discussion of the general results and possible extensions is given in section 5. The model and Derrida’s annealed approach ========================================= The previous definition of random Boolean network, where $\Sigma = \{ 0,1 \}$ can be generalized to a wider set of states, $\Sigma = \{ 0,1, ... , S-1 \}$ that is to say, to random networks with multiple states (RNS). As with RBN, each element $\sigma_i$ $(i=1,...,N)$ receives $K$ inputs; so we have: $$\sigma_i(t+1) = \Lambda_i^s \; \Bigl [ \; \sigma_{i_1} (t), \sigma_{i_2}(t) , ... , \sigma_{i_{K}}(t) \; \Bigr ] \eqno (7)$$ where the superindex $s$ of the functions $\Lambda_i^s $ indicate that its inputs (and outputs) can take $S$ values and the functions $\Lambda_i^s$ are randomly chosen from a set $ \cal S_K$. As with RBNs, this model exhibits three characteristic dynamical regimes, which we will explore. Three examples are shown in figures (3-5) for a $N=100$ network with $S=5$ states, $K=2$ and different $p$-values. Here $p$ is an additional which was introduced by B. Derrida et al. \[15\] in order to make the RBN transition smooth and continuous. The parameter $p$, known as the bias of $\Lambda_i$, is defined as the probability of having $0$ as output in the function $\Lambda_i$. This parameter is easy to extend in natural form for the RNS (see below). The plots correspond to the chaotic (p=0.60), critical (p=0.689) and frozen (p=0.8) regimes. Now in order to characterize the critical properties of our system, we will applie the well known Derrida’s method, based on the annealed approximation \[15\]. This approach is a way to avoid the dynamical correlations which appear as the system evolves in time. In the thermodynamic limit the original (quenched) model and the annealed counterpart share the same phase transition curves \[16\]. We start with two randomly chosen configurations $$C_1(t) \equiv (\sigma_1^{(1)}(t),...,\sigma_N^{(1)}(t))\eqno (8.a)$$ $$C_2(t) \equiv (\sigma_1^{(2)}(t),...,\sigma_N^{(2)}(t))\eqno (8.b)$$ which are also randomly taken from the set ${\cal C}_S(N)$ of all the possible $N$-strings (clearly $\sharp {\cal C}_S(N) = S^N$). Following Derrida’s method, it can be shown that the overlap $a_{12}(t)\in [0,1]$, defined as the normalized number $N a_{12}(t)$ of elements with common states in $C_1(t)$ and $C_2(t)$, will evolve in time following a nonlinear one-dimensional map. In short, let $N a_{12}(t+1)$ the net overlap after one iteration of $C_1(t)$ and $C_2(t)$ under (7). Then a new set of connections and $\Lambda_i^s $ functions is again taken from ${\cal S_K}$, and a new iteration is performed over $C_1(t+1)$ and $C_2(t+1)$. For our $S$-state random network and assuming that $1 \over S$ is the probability that two outputs of an arbitrary function $\Lambda_i^s $ are identical, following the arguments in \[15\], we have a nonlinear map for $a_{12}$ given by: $$a_{12}(t+1) = a_{12}^{K}(t) + {1 \over S} \Biggl [ 1 - a_{12}^{K}(t)\Biggr ]\eqno (9)$$ where: $a_{12}^K (t)$ is the fraction of elements with identical inputs and $ [ 1 - a_{12}^{K}(t)]/S$ is the probability for two automata, with at least one different input, being equal at $t+1$. Now the critical point where the phase transition takes place (separating the so called frozen and chaotic phases) is obtained from the stability condition: $$\left. {\partial a_{12}(t+1) \over \partial a_{12}(t) } \right\vert_{a^*=1}=K \Biggl (1 - {1 \over S} \Biggr ) \leq 1 \eqno (10)$$ or in other words the following relation between the average network connectivity and the number of available states is reached: $$K = {1 \over 1-{1\over S}} \eqno(11)$$ We see that for $S=2$ (Boolean net) we recover the well known critical point $K_c=2$. For $S\gg 2$ we can approximate the marginal stability relation to: $$K \approx 1+{1 \over S} \eqno (12)$$ We also see that when $S \rightarrow \infty$ the average connectivity moves to $K_c \rightarrow 1$ (i.e. the critical point is reached only at the lowest connectivity). The previous inequality (10) is related with the high-temperature value of damage spreading \[23\] of the Potts model \[13\]. In RNS it is possible to extend Derrida’s $p$-parameter definition: $p$ is defined as the probability of having a $0$ as output of the function $\Lambda_i^s$ ($0$ is the quiescent state) and the other states are equally likely to be present. Although this choice seems too limited, it can be justified from variational arguments (see Appendix I). Now the overlap will evolve following the 1D map: $$a_{12}(t+1) = a_{12}^K(t) + \Biggl [ p^2 + {(1-p)^2 \over S-1} \Biggr ] ( 1 - a_{12}^K (t) ) \eqno(13)$$ where we have: - $a_{12}^K (t)$ is the fraction of elements with identical inputs at time $t$ - $( 1 - a_{12}^K (t) )p^2$ is the probability of two automata with at least one different input, being equal to “0” at $t+1$. - $( 1 - a_{12}^K (t) )(1-p)^2/(S-1)$ gives the probability of finding two automata with at least one different input, being equal (but differents from $0$) at $t+1$. Now, the stability analysis for $a_{12}^*=1$ gives us a general condition for the critical points: $$p_c= {1 \over S} \Biggl [ 1 + \Bigl ( 1 - {S \over K} \Bigl [ (2-S) K + S - 1 \Bigr ] \Bigr )^{1/2} \Biggr ] \eqno(14)$$ Defining a critical surface $p_c=p_c(K,S)$ separating the frozen (ordered) phase from the chaotic one. Many possible characterizations of the two phases and the critical surface can be used. In figure (6a-c) we show three examples of the magnetization $m$ (for different RNS with $N=100, K=2$ and $S=5$ states). Here $m$ is defined as the number of times $n$ a given automaton $S_i$ is such that $S_i=0$ over $T$ steps i. e. $m=n/T$. Then $P(m)$ is a histogram giving the distribution of $m$ for the whole system i. e. it counts how many automata with the same average magnetization $m$. We can see, as expected, how $P(m)$ moves from a singular distribution with few peaks to a more continuous distribution at the chaotic regime. A numerical characterization fo these three regimes can be easily obtained by computing the frequency distribution $N(T)$ of cycles of different length $T$. Since these dynamical systems are finite, with a maximum number of $S^N$ states, the previous equations (7) lead to two types of attractors (cycles and steady states). However, the presence of a critical boundary separating the two regimes is made clear by the existence of well-defined scaling laws. In figure (6d-f) we show the frequency distribution of periods for the three regimes. At the chaotic phase, long-periodic orbits are rather common. This is due to the exponential increase of attractor lengths in this phase. The frozen regime shows the opposite tendency: the periods are very short, with an exponential decay with $T$. At the critical boundary, both short and long periods are present, in such a way that in spite of the frequent observation of short-period orbits, long-periodic orbits of length $T \approx O(N)$ are allways found. In random Boolean networks, the previous observations have been deeply explored through both numerical and theoretical approximations \[6,7\]. A specially important relation is the dependence of the average period $<T>$ with the system size $N$. When RBN are used with $p=0.5$, it is found that the critical phase ($K=2$) is characterized by a scaling $<T> \approx \sqrt{N}$ and an exponential growth $<T(K>2)> \approx 2^{BN}$ is obtained at the chaotic phase, with $B>1$ \[2\]. We have explored this scaling behavior in the RNS counterpart. Two examples of our calculations are shown in figure 8(a-b). The average period length has been calculated in two different cases, both varying $p$ (a) and the number of states $S$ (b). we can see how the scaling $<T(N)>\approx N^{\tau}$ is present at criticality and how an exponential increase in $<T>$ also appears as we enter into the chaotic phase. Numerical simulations show that for the chaotic phase an exponential law $<T(S,N)>\approx S^{BN}$ is obtained, with $B>1$. Numerical estimations of this value gave $B = 1.11 \pm 0.02$. CA and RNS ========== It can easily be checked from (14) that for $S=2$ (Boolean net) we have $$p_c= {1 \over 2}+ {1 \over 2}\sqrt{( 1 - {2 \over K})} \eqno(15)$$ which gives us the well known critical line in the $(K,p)$-phase space \[15,17\]. Equation (15) is reminiscent of those found in standard CA (with nearest-neighbor connectivity \[18\]). The existence of a transition domain for CA models has been characterized by means of several approaches. Langton’s $\lambda$-parameter gives a rough characterization of these critical points when complex CA rules (involving high $S$ and/or $r$ values) are used. If $S^{2r+1}$ is the total number of $(2r+1)$-configurations in the CA rule table, as described by (1), and $M$ out of the total map to a non-zero state, then $\lambda$ is defined as \[18\]: $$\lambda \equiv {M \over S^{2r+1}} \eqno(16)$$ Thus, $1-\lambda$ corresponds to $p$ under the RNS definition. This parameter has been shown to be related with -but not equivalent to- a temperature in statistical physics \[18\]. It was observed (particularly in the large-$r$ limit) that these CA exhibit some of the characteristics of second-order phase transitions (for a detailed study see \[18\], although the association between computation, complexity and critical phenomena has been shown to be flawed \[19\]). Li et al. \[18\] showed by means of mean field approximations that, for $S=2$, there is a simple relationship between the critical value $\lambda_c$ and $r$ (the neighborhood radius): $$1 - \lambda_c = {1 \over 2} + {1 \over 2} \sqrt{1 - {2 \over r+1}} \eqno(17)$$ This result is consistent with equation (15). But in (17), instead of $2r+1$ (the total connectivity of the CA model) we have $r+1$ (half of it plus one). In the RNS counterpart, we have $K$, the total connectivity. These differences are due to the spatial correlations intrinsic to CA models, which are destroyed in the RNS counterpart. This is consistent with the analysis of Li et at. \[18\] which is based in a mean-field estimation of the spreading rate of diferent CA patterns. Assuming symmetry in the average spreading rate $\gamma(r)$ in right and left directions, they obtain the onset to non-zero spreading at $\lambda_c$ given by (17). The corresponding critical curves for both the $1-\lambda$ parameter (squares) and $p$ (circles) as a function of the connectivity are shown in figure (8). Remember that a neighbor radius $r$ in CA is equivalent to $K=2r+1$ in RNS. Thus the first point represented for RNS whit $S=2$ is $K=3$ with $p_c=0.79$. Both curves share the same qualitative behavior, but larger values of the bias are required in RNS in relation with CA in order to reach the ordered regime, as expected. A similar relationship is found for the temperature in relation with $q$ in some $q$-Potts models \[13\]. Stable core and the self-overlap method ======================================= A well-defined critical phase transition should be characterized by an appropiate order parameter $\Omega$ such that $\Omega > 0$ at the disordered regime and zero otherwise. In fact, $d^*=1-a^*$ as like parameter of order for RBN and RNS \[15\]. In this context, Flybvjerg \[22\] explored a different approach than Derrida’s by defining the [*stable core*]{} at time $t$, $c(t)$, as the (normalized) fraction of automata $S_i$ (independent of the initial condition) reach stable values at a given step $t$, i.e. remain constant after $t$. The stable core asymptotic behavior, $c(t \rightarrow \infty)$, can be used as an order parameter for the frozen-chaos transition in RBN. Explicitly, if $\Omega \equiv 1 - c(t \rightarrow \infty)$, the previous requirements for the order parameter definition hold. The argument of Flybvjerg for finding an iterated equation for $c(t)$ is as follows. There are $K+1$ mutually exclusive reasons for a given automata $S_i$ to be part of the stable core from step $t$ to $t+1$. The probabilities of this $K+1$ reasons are: $(0)$ All the inputs, $\sigma_{i_1} (t), \sigma_{i_2}(t),...,\sigma_{i_K}(t)$, belong to the stable core at $t$. Given that the inputs are choosen at random, this situation occur with probability $c^K(t)$ (where $c(t)$ is interpreted as the probability of belonging to the stable core in $t$). $(1)$ $\Lambda_i$ is independent of one input and the rest of the $K-1$ variables belong to the stable core: $${K\choose 1} c^{K-1}(t) (1-c(t)) p_1 \eqno(18)$$ where $p_1$ is defined as the probability of $\Lambda_i$ to be independent of one input after fixing its $K-1$ remainder inputs. $\vdots$ $(j)$ $\Lambda_i$ is independent of $j$ inputs and the rest of the $K-j$ variables belong to the stable core: $${K\choose j} c^{K-j}(t) (1-c(t))^j p_j \eqno(19)$$ where $p_j$ is defined as the probability of $\Lambda_i$ to be independent of $j$ arguments after fixing its $K-j$ remainder arguments. $\vdots$ $(K)$ $\Lambda_i^s$ is a constant fuction (independent of its $K$ arguments), with all out the stable core: $$(1-c(t))^K p_K \eqno(20)$$ Adding the $K+1$ exclusive reasons: $$s(t+1) = \sum_{i=0}^{K} {K\choose i} c^{K-i}(t) (1-c(t))^K p_i \eqno(21)$$ where $p_i$ is the probability that the Boolean function is independent of a certain number $i$ of inputs after fixing its $K-i$ remainder inputs (a biologically relevant case are the canalizing functions that depend on a unique input and are independent of remainder \[2\]). For the Boolean functions with bias $p$, i.e., with value $0$ with probability $p$ and value 1 with probability $(1-p)$ it gives: $$p_i = p^{2^i} + (1-p)^{2^i} \eqno(22)$$ By substituying (22) into (21) and analyzing its stability the following transition critical curve is obtained: $K 2p(1-p) = 1$ according to \[15,23\], reached by different methods. How to extend this argument to RNS? Apparently, such an extension is straightforward. It seems that, again, there are $K+1$ exclusive reasons for a given automata to move from outside the stable core to be a member of it. Such a simple translation of the previous procedure is not possible, and we can show why by means of a simple example. Let us consider a RNS with $K=4$ and $S=3$, i.e.: $$\Lambda_i^3=\Lambda_i(\sigma_{i_1}, \sigma_{i_2},\sigma_{i_3}, \sigma_{i_4}) \; \; \; ; \; \; \; i=1,2,...,N \eqno(23)$$ and let us assume that at a given $t$ the first three inputs belong to the stable core , i.e: $\sigma_{i_1}, \sigma_{i_2},\sigma_{i_3} \in s(t)$ and $\sigma_{i_4} \notin s(t)$. The question is whether or not $\sigma_i (t+1) \in s(t+1)$ if $\sigma_i (t) \notin s(t)$. This will happen if $\Lambda_i^3$ is independent of the fourth input. But it can also occur (and this makes a big difference) if $\sigma_{i_4}$ can only reach a subset $\Sigma' \subset \Sigma$, say $\Sigma' \equiv \{ 0, 1 \}$. In such case, we have a reduction in the number of degrees of freedom for $\sigma_{i_4}$. In this case it can happen that $\Lambda_i^3$ is independent of $\sigma_{i_4}$ for $\sigma_{i_4} \in \Sigma'$ but not if $\sigma_{i_4}=2$ (for $S=2$ a reduction of one degree of freedom simply means that the input belongs to the stable core). These possible outcomes forces us to weight the probabilities that the input units have their degrees of freedom reduced. More generally, the exclusive reason (1) implies that $\Lambda_i^s$ is independent of a given input $\sigma_{i_j}$. This occurs with probability $p_1=p^2+(1-p)^2$ for the case of two states. For $S$-states, it can occur that $\sigma_{i_j} \in \Sigma' \subset \Sigma$ and $\sigma_{i_j} \notin s(t)$. In order to calculate the order parameter equation for networks with multiple states, we will compute the successive overlap (or [*self-overlap*]{} \[25\]) $a_{t+1,t}$, defined as the fraction of automata such that: $\sigma_i(t+1)=\sigma_i(t)$. This is in fact the overlap between the system configuration at $t$ with the next configuration at $t+1$. If the assymptotic value of the stable core is $1$ the self-overlap $a_{t+1,t}(t\to \infty)$ will be one too. This allows us to use the asymptotic value of the self-overlap as an appropiate order parameter and as an alternative to the standard overlap equation analysed in section 2. The iterated map for the self-overlap will be $$a_{t+1,t} = a_{t,t-1}^K + \Biggl ( p^2 + {(1-p)^2 \over S-1} \Biggr) (1 - a^K_{t,t-1}) \eqno (24)$$ where we desribe the self-overlap between the set at $t+1$ and $t$ as a function of the self-overlap for $(t-1,t)$. If we interpret $a_{t,t-1}$ as the probability for an arbitrary unit to remain in the same state at both $t-1$ and $t$, the term $a_{t,t-1}^K$ gives the probability that all the inputs of a given unit are the same from $t-1$ to $t$. Obviously $(1-a^K_{t,t-1})$ is the probability that at least one of the inputs will be different between $t$ and $t-1$. In that case, there is still a possibility of remaining in the same state at $t$ and at $t+1$ just by chance. This is given by $(p^2 + (1-p)^2)/(S-1)$. The stability analysis now gives: $$\left. {\partial a_{t+1,t} \over \partial a_{t,t-1} } \right\vert_{a^*}= K \Biggl [ 1- \Biggl (p^2 + {(1-p)^2 \over S-1}\Biggr) \Biggr ] \leq 1 \eqno(25)$$ which is easily interpreted from a perturbative point of view \[23\]: $p^2 + {(1-p)^2 \over S-1}$ is the probability of no-propagation for the minimal perturbation (one change of state $l$ to $m$ in one input). Then, $K\Bigl [ 1- \Bigl (p^2 + {(1-p)^2 \over S-1}\Bigr) \Bigr ]$ is the mean number of changes generated by the perturbation in one time step. This interpretation allows to define a Lyapunov exponent for this system (see Appendix II). In figures (9-10) we show (continous lines) the asymptotic behavior of both the overlap and the self-overlap for different $p$ and $K$ and $S=10$ from equation (24). As we can see, the dynamical equation for the self-overlap (24) is the same as the standard overlap equation (13), although it is conceptually different from it. The self-overlap, as it occurs with the stable core, can only be applied to the quenched system and fails with the annealed case. Besides, the self-overlap is computationally more efficient than the standard overlap, which requires the parallel running of the two replicas of the system. In figure (9) we have followed the time evolution (black symbols) of overlaps between two RNS quenched replicas of size $N=1000$ as in \[15\]. Each point is compute averaging $100$ experiments. The connectivity $K=2$ and the number of avalible states $S=10$ are fixed, and the bias $p$ change. The white symbols represent identical averages for self-overlap, except for the crosses where $N=10.000$ for critical $p=0.70$ in order to show the finite size effect of the critical transition. This effect is more acused for the self-overlap as for the overlap. In figure (12) we show the time evolution, this time fixint the states $S=10$ and the bias $p=0.81$, of the overlap and self-overlap. Again the continous lines are the theoretical evolution from equation (29). Black and white symbols correspond to overlap and self-overlap respectivily. In this case the variation of the connectivity despite the transition at $K=3$. Again the crosses ($N=10.000$ for self-overlap) show the finite size efecfct. All our simulations show a very good agreement whit the theoretical evolution of the overlap (eq.13) and the self-overlap (eq.24). Thus, any of them acts as good order parameter. Summary ======= In this paper we have proposed a simple, straighforward extension of random Boolean networks in terms of a discrete model with multiple states. The starting point of our analysis is the observation that several natural systems involving a wired set of entities of some kind often display a range of state values instead of simple on -off characteristics. In some cases, as the specific tasks (states) performed by ants in a colony, the discreteness of the states is rather well defined, although their transitions (as a consequence of interactions of ants engaged in different tasks) can be rather complex \[24\] suggesting that the underlying task-dependent rules are complicated ones. When cell types are considered, we also get a well-defined set of “states” although each one itself is the result of a complex dynamical system (the gene expression pattern) at a given stationary state. A classical set of experiments with different organisms at different levels of development reported the existence of complex interactions between cell types eventually leading to switches by depending of the cell types under interaction \[2,12\]. In other systems, like the genomic regulatory network, the activity of some genes is sometimes close to an on-off switch but it displays a range of activity levels. A first approximation to such continuous networks is the RNS described and analysed in this paper. Following previous studies, our interest was focused on to the critical features of these networks. It has been conjectured that a wide range of complex systems, from evolving ecosystems, biological regulation networks, traffic or ant colonies \[4\], to cite only a few, display patterns in space and time suggesting that they are close to phase transition points. In this context, we have found the explicit form of the transition domains as well as an order parameter equation for the transition. The critical boundaries for this model have been obtained from Derrida’s annealed approximation and the computation of $s(t)$ (the stable core equation) required the development of an alternative treatment to the Flyvbjerg derivations for standard RBN \[22\]. This model has been shown to share some common traits with other related dynamic models like cellular automata and the $q$-Potts model. Several extensions of this work are possible. Further quantities can be defined in order to characterize the damage spreading. This can be easily done by means of theoretical extensions of previous definitions of Lyapunov exponents for RBNs \[25\]. One of them is the analysis of the square-lattice counterpart of the RNS with nearest-neighbor connectivity. This type of network has been proved to be extremely useful in order to display the critical features of the RBN model in terms of percolation on a square lattice \[15\]. A different extension would be a re-definition of the network rules in order to make the transitions from different states closer to the continuous dynamics. If the output of the $\Lambda_i^s$ function is randomly chosen from $\Sigma$, we must expect to observe a time evolution of individual units in terms of jumps from a given state to any other state. But if we want to retain the relationship between the $S$-state model with the corresponding continuous counterpart (as described by sets of nonlinear differential equations) a consistent increase or decrease in individual activity should be observable. This problem can be easily solved through the introduction of a new set of rules describing the outputs of the $\Lambda_i^s$-function in terms of increases/decreases of gene activity and will be reported elsewhere. [Acknowledgments]{} The authors would like to thank Andy Wuensche, Brian Goodwin, Jordi Delgado and Mar Cabeza for useful discussions and help. This work has been partially supported by a grant DGYCIT PB-97-0693 and by the Santa Fe Institute (RVS and SK) and by the Centro de Astrobiología (BL). Appendix I: RNS and maximum entropy =================================== The relation (11) can be easily generalized: $K$ can be not unique, but an average connectivity $<K>$ is always able to be defined \[17\] and we can have a probability distribution of states, i.e. $P(\mu)=P[\Lambda_i^S=\mu]$ for $\mu \in \Sigma$. In this general case, the critical line transition is given by: $<K> = 1/(1 - {\sum {P^{2}(\mu)}})$ \[23\]. This result can be used in order to find the expected probability distribution of states at criticality \[17\]. Using the following maximum entropy constraints: $$\sum_{\mu} {P(\mu)} =1 \eqno (1)$$ $$\sum_{\mu} {P^2(\mu)}= 1 - {1 \over {<K>}} \eqno (2)$$ $P^*_c(\mu)$ can be derived from a variational procedure known as the maximum entropy formalism \[20\]. We first construct the Lagrangian: $${\cal {L}}({\bf P})= - {\sum_{\mu} {P(\mu) \log P(\mu)}} - \beta \biggl ({{\sum_{\mu} {P^2(\mu)}} - {1 \over <K>}}\biggr ) - \alpha \biggl ( {\sum_{\mu} {P(\mu)}} - 1 \biggr ) \eqno (3)$$ where ${\bf P} = (P(0), P(1), \dots , P(S-1))$ and the variation $${\partial {{\cal {L}}({\bf P})} \over \partial {{\bf P}} } = 0 \eqno (4)$$ is performed. This leads to a set of equations $$\log {P(\mu)} + 2 P(\mu) \beta = -(1+\alpha) \eqno (5)$$ which must be solved for $P(\mu)$. Let us introduce $P(0)=w$ and $P(\mu \ne 0)= \alpha_\mu w$ with $\mu = 1, 2, \dots , S-1$ with $w \in [0,1]$ and $\{ \alpha_\mu \}$ is a set of positive constants such that $$w (1+ \sum_{\mu = 1}^{S-1} {\alpha_\mu}) = 1 \eqno (6)$$ Then we get: $$\log w + 2 \beta w = - (1 + \alpha)$$ $$\log w + \log {\alpha_\mu} + 2 \beta {\alpha_\mu} w = -(1+\alpha) \; \; \; \; \; \mu = 1,2, \dots , S-1 \eqno (7)$$ So $\alpha_{\mu} = C$ (constant) $\forall \mu$ and as a consequence $$P(0) = w \eqno(8.a)$$ $$P(\mu)={1-w \over S-1} \; \; \; \; \mu=1,2,\dots ,S-1 \eqno (8.b)$$ The most likely probability distribution close to criticality will be a delta-shaped $P(\mu)$ with a given state present with probability $w$ and the rest equally distributed. We should note that this property (i.e. a sharply peaked distribution with one or a few predominant states) has been reported for the onset of chaos in low-dimensional chaotic dynamical systems \[21\]. Here the probability distributions maximize a previously defined complexity measure $\cal C$ and numerical simulations of RNS indeed show that such singular distribution with one or a few peaks is characteristic at criticality. Appendix II: Lyapunov exponents for RNS ======================================= If, in the distance method or Derrida’s aproximation, we interpret the replica of the one system as a perturbation on the original system \[25\], we can define a expansion rate of the perturbation in a time $t$ for RNS easilly: $$\eta(t) = {d(t+1) \over d(t)} \eqno (1)$$ In a natural way, we can define now the Lyapunov exponent as the temporal average of the logarithm of the expansion rate: $$\lambda (T) = {1\over T} \sum_{t=1}^T \log \eta (t) =\log{\sqrt[T] {\prod_{t=1}^T {\eta (t)}}}=\log{\bar{\eta}} \eqno (2)$$ If we use the equation for the distance between configurations: $$d_{12}(t+1)=\Biggl (p^2 + {(1-p)^2 \over S-1}\Biggr)[1-(1-d_{12}(t))^K] \eqno (3)$$ Then the expansion rate in the time $t$ of the perturbation, is: $$\eta(t) = { d_{12}(t+1) \over d_{12}(t)} = {\biggl (p^2 + {(1-p)^2 \over S-1}\biggr ) \{1-[1-d(t)]^K \} \over d(t)} \eqno (4)$$ and approximating for small $d(t)$: $${(1-d(t))^K} \approx {1-Kd(t)} \eqno (5)$$ we have: $${\eta(t)} \approx {\Biggl (p^2 + {(1-p)^2 \over S-1}\Biggr)K} \eqno (6)$$ an average constant expansion rate that give us that Lyapunov’s exponent: $$\lambda = \log {\Biggl [ \Biggl (p^2 + {(1-p)^2 \over S-1}\Biggr)K \Biggr]} \eqno(7)$$ that determines the two classics regimes: $\lambda <0$ (order) and $\lambda > 0$ (chaos), whit the marginal case $\lambda = 0$, in total concordance with the transition surface. There are diferent methods to compute the largest Lyapunov exponent. In order to show consistence we will demonstrate that it is possible compute Lyapunov exponents from the self-overlap in agreement with the previous result. The Wolf’s method \[26\] is used to estimate numerically Lyapunov exponents from time series. In short, the method is described as follows: get two points of the temporal serie, let us say ${\bf X}(t_1)$ and ${\bf X}(t_2)$ and calculate their distance: $\mid {\bf X}(t_2) - {\bf X}(t_1) \mid$. Assume that $\mid {\bf X}(t_2) - {\bf X}(t_1) \mid < \epsilon$, being $\epsilon > 0$ small. Next, compute the distance, after a time $T$, i.e: $\mid {\bf X}(t_2+T) - {\bf X}(t_1+T) \mid$. This time, usuallyly, is a fraction of the characteristic period or the time required for the autocorrelation function go to zero. Repeating for $n$ pairs of points and averaging, we obtain an estimation of the Lyapunov exponent: $$\lambda = {1 \over nT} \sum_{t_{2} \neq t_{1}}^{n}{ \log{ { \mid {\bf X}(t_2+T) - {\bf X}(t_1+T) \mid \over \mid {\bf X}(t_2) - {\bf X}(t_1) \mid } } } \eqno (8)$$ For RNS, we can write down an equation for the normalized Hamming distance between succesive time steps in our system, i.e. for the complementary probability of the auto-overlap: $d_{t,t-1}=1-a_{t,t-1}$. Thus the equation for the self-overlap becomes: $$d_{t+1,t} = \Biggl[ 1- \Biggl(p^2 + {(1-p)^2 \over S-1}\Biggr)\Biggr][1-(1-d_{t,t-1})^K] \eqno (9)$$ If we approximate linearly close to the fixed point $d^*=0$. The function becomes: $$d_{t+1,t} = K\Biggl[ 1- \Biggl(p^2 + {(1-p)^2 \over S-1}\Biggr)\Biggr]d_{t,t-1} \eqno (10)$$ The iterated equation now gives: $$d_{t+T,t+T-1} = \Biggl \{K\Biggl[ 1- \Biggl(p^2 + {(1-p)^2 \over S-1}\Biggr)\Biggr]\Biggr \}^Td_{t,t-1} \eqno (11)$$ To compute (8) in our terms, and taking $t_2=t_1+1$, i.e.: a natural step in our system, we have: $${d_{t+T,t+T-1} \over d_{t,t-1}}=\Biggl \{K\Biggl [ 1- \Biggl (p^2 + {(1-p)^2 \over S-1}\Biggr)\Biggr]\Biggl \}^T\eqno (12)$$ a constant value that, once introduced in the sum (8), determines the Lyapunov exponent (7) as sugested previously. References ========== \[1\] S. Wolfram, Physica D10 (1984) 1; A. Wuensche, [*The global dynamics of cellular automata*]{}, SFI studies in the sciences of complexity, Addison-Wesley, 1992; Wuensche has done a remarkable research on the basins of attraction of both cellular automata and random Boolean networks; the basic software, named DDLab, is available in the website: http://www.santafe.edu/wuensch/ddlab.html. \[2\] S. A. Kauffman, J. Theor. Biol. 22 (1969) 437 S. A. Kauffman, J. Theor. Biol. 44 (1974) 167; S. A. Kauffman, Physica D10 (1984) 145; S. A. Kauffman, Physica D42 (1990) 135; S. A. Kauffman , [*The Origins of Order*]{} Oxford U. Press (Oxford, 1993);R. Somogyi and C. A. Sniegoski, Complexity 4 (1996) 45 \[3\] D. Kaplan and L. Glass, [*Understanding Nonlinear Dynamics*]{}, Springer-Verlag (New York 1995) \[4\] R.V. Solé, S.C. Manrubia, B. Luque, J. Delgado and J. Bascompte, Complexity 1(4) (1995) 13 \[5\] R.J. Bagley and L. Glass, J. theor. Biol. 183 (1996) 269 \[6\] U. Bastolla and G. Parisi, Physica D 98 (1996) 1; U. Bastolla and G. Parisi, J. Phys. A: Math. Gen. 30 (1997) 5613 \[7\] A. Battacharjya and S. Liang, Physica D95 (1996) 29; A. Bhattacharijya and S. Liang, Phys. Rev. Lett. 77 (1996) 1644 \[8\] J.A. de Sales, M.L. Martins and D.A. Stariolo Phys. Rev. E 55 (1997) 3262. In this paper the authors study the overall behavior of gene networks by means of a model involving both short- and long-range couplings among genes. They found a rich variety of behaviors and their results consistently fitted known scaling laws for the number of differentiated cells and cell length cycles as a function of $N$. \[9\] K.E. Kurten, Phys. Lett. A 129, 157 (1992) \[10\] S.L. Adhya and D. F. Ward, Prog. Nucleic Acid Res. Mol. Biol. 26 (1981) 103 \[11\] L. Glass and S.A. Kauffman, J. theor. Biol. 39 (1973) 103; see also L. Glass and S. A. Kauffman, J. Theor. Biol. 34 (1972) 219 \[12\] S. F. Gilbert, [*Developmental Biology*]{}, Sinauer, Massachusetts (1996); B.C. Goodwin Temporal organization in cells London, Academis Press (1963) \[13\] F. Y. Wu, Rev. Mod. Phys. 54 (1982) 235; C. Tsallis and A. C. N. de Magalhaes, Phys. Rep. 268 (1996) 305 \[14\] J. E. Lewis and L. Glass, Int. J. Bif. Chaos 1 (1991) 477; T. Mestl, R. J. Bagley and L. Glass, Phys. Rev. Lett. 79 (1997) 653; A. Wagner, Proc. Natl. Acad. Sci. USA 91 (1994) 4387 . These authors consider a simple formulation of biological regulatory networks in terms of a neural-like dynamics, as defined by: $ {d x_i \over dt} = - x_i \Phi_i \Bigl ( \sum_j W_{ij} x_j - \Theta_i \Bigr )$, with $i,j=1,2,...,N$. As usual $\Phi(z)$ is a sigmoidal function and $\theta$ a threshold. The matrix $W_{ij} \in \Re$ describes the specific features of the interactions among different elements (genes). We can see that equations (4.a-b) are just a specific example with $N=2, \Theta_i=0, W_{ij}=1$ and $\Theta(z)=\mu^n/(\mu^n + z^n)$. \[15\] B. Derrida and Y. Pomeau, Europhys. Lett. 1 (1986) 45 B. Derrida and Y. Pomeau, Europhys. Lett. 2 (1986) 739 G. Weisbuch and D. Stauffer, J. Physique 48 (1987) 11-18 \[16\] H.J. Hilhorst and M. Nijmajer, J. Physique 48 (1987) 185 \[17\] R. V. Solé‚ and B. Luque, Phys. Lett. A196 (1995) 331 \[18\] C. G. Langton, Physica D52 (1990) 12 W. Li, N. H. Packard and C. G. Langton, Physica D45 (1990) 77 W. K. Wootters and C. G. Langton, Physica D45 (1990) 95 \[19\] A. Dhar, P. Lakdawala, G. Mandal and S. R. Wadia. Phys. Rev. E51 (1995) 3032 \[20\] H. Haken, Information and self-organization, Springer, Berlin, 1988; J. N. Kapur, Maximum Entropy models in Science and Engineering, Wiley, New Delhi, 1993 \[21\] C. Anteneodo and A. R. Plastino, Phys. Lett. A223 (1996) 348 \[22\] H. Flybjerg, J. Phys. A: Math. Gen. 21 L955 (1988) \[23\] B. Luque and R. V. Solé, Phys. Rev. 55, 1 (1997) \[24\] R. V. Solé and O. Miramontes, Physica D80 (1995) 171 and references cited \[25\] B. Luque and R. V. Solé, adap-org/9907001 \[26\] A. Wolf, J. B. Swift, H. L. Swinney and J. A. Vastano, Physica D16 (1985) 285

ArXiv

--- abstract: 'We have used the ACS camera on HST to obtain $(V,I)$ photometry for 57,000 red-giant stars in the halo of the Leo elliptical NGC 3377, an intermediate-luminosity elliptical. We use this sample of stars to derive the metallicity distribution function (MDF) for its halo field stars, and comment on its chemical evolution history compared with both larger and smaller E galaxies. Our ACS/WFC field spans a radial range extending from 4 to 18 kpc projected distance from the center of NGC 3377 and thus covers a significant portion of this galaxy’s halo. We find that the MDF is broad, reaching a peak at log $(Z/Z_{\odot}) \simeq -0.6$, but containing virtually no stars more metal-poor than log $(Z/Z_{\odot}) = -1.5$. It may, in addition, have relatively few stars more metal-rich than log $(Z/Z_{\odot}) = -0.3$, although interpretation of the high-metallicity end of the MDF is limited by photometric completeness that affects the detection of the reddest, most metal-rich stars. NGC 3377 appears to have an enrichment history intermediate between those of normal dwarf ellipticals and the much larger giants. As yet, we find no clear evidence that the halo of NGC 3377 contains a significant population of “young” ($< 3$ Gy) stars.' author: - 'William E. Harris' - 'Gretchen L. H. Harris' - 'Andrew C. Layden' - 'Peter B. Stetson' title: 'HST Photometry for the Halo Stars in the Leo Elliptical NGC 3377[^1] ' --- Introduction {#intro} ============ The imaging cameras on the HST have provided powerful tools to study large photometric samples of individual stars in nearby galaxies that were quite out of reach from ground-based instruments. Although even the best photometry cannot provide the level of detail that is achievable by high-resolution spectroscopy, these deep stellar samples open a valuable route to understanding the early chemical evolution of their host galaxies, compared with the much cruder information from integrated-light photometry and spectroscopy. Among the possible target galaxies within reach of the HST, the ellipticals are of special interest because they may result from the widest possible range of formation histories, from hierarchical merging at very early times, to recent major mergers, to later growth by satellite accretion. Representative dwarf ellipticals are readily accessible within the Local Group (particularly NGC 147, 185, 205), but the Local Group gives us no examples of giant or intermediate-sized E galaxies to work with. In a series of previous papers [@h99; @h00; @h02; @rej05], we presented analyses of samples of stars in the closest giant elliptical, NGC 5128. These studies cover four locations in its halo at projected distances ranging from 8 to 40 kpc. In all four of these studies the red-giant stars are clearly resolved, and in the deepest one [@rej05], even the old horizontal-branch population is reached. We found that in all four locations in the halo of this giant elliptical, the metallicity distribution of the giant stars is clearly metal-rich (with $\langle$ Fe/H $\rangle \simeq -0.4$) and broad, but with extremely small proportions of classically metal-poor stars (i.e. those with \[Fe/H\] $< -1$). Generalizing the results for NGC 5128 to all large ellipticals may, however, be compromised by the possibility that NGC 5128 could be a major merger remnant in which the majority of its halo stars were actually formed within the metal-rich disks of its progenitor galaxies. In fact, both a major-merger or hierarchical-merging approach are capable of creating an MDF in the halo with the same basic characteristics as we observe [@bekki03; @bea03]. To extend the range of information we have to work with, clearly it is of great interest to obtain the metallicity distribution function (MDF), [*based directly on samples of individual stars*]{}, in other galaxies of the widest possible range of properties. NGC 5128, at a distance of only 4 Mpc, is by far the closest easily accessible big elliptical. To reach other ellipticals, we must step to the Leo group objects at $D \sim 10$ Mpc, and then outward to Virgo ($D \sim 16$ Mpc) and Fornax ($D \sim 19$ Mpc), which have the nearest accessible large samples of target galaxies. The technological gains afforded by the step from the HST WFPC2 camera to the newer ACS, with its much higher sensitivity, spatial resolution, and field of view, now bring these more distant targets within reach (see, for example, the recent photometric work of [@will06] to resolve the brightest $\sim 1.5$ mag of the red-giant branches in four globular clusters at the Virgo distance). Roughly speaking, observing the halo of NGC 5128 with WFPC2 is comparable (for the same number of HST orbits) with observing the Leo group ellipticals with ACS. In this paper, we present new color-magnitude photometry for the halo stars in the intermediate-sized E5 elliptical NGC 3377 in the Leo group, and briefly compare its metallicity distribution function (MDF) with those of both a dwarf and a giant. NGC 3377 has a luminosity of $M_V^T \simeq -19.9$ falling between those of the Local Group dwarfs and the giant NGC 5128 that have been observed in previous studies; it therefore provides an attractive bridge in the parameter space of galaxy properties that we can add to the discussion of chemical evolution models. The key parameters for the galaxy are summarized in Table 1. Observations and Data Reduction {#observations} =============================== The imaging data we use for this study were obtained in our HST program 9811. NGC 3377 is at relatively high galactic latitude and low foreground reddening ($b = 58\fdg3$, $E_{B-V}$ = 0.03) and is thus well placed for deep photometry of its halo stars, as free from field contamination and foreground absorption as we can hope to have. We used the Advanced Camera for Surveys in its Wide Field Channel, with image scale $0\farcs05$ per pixel, to reach the maximum photometric depth and field coverage. A single field was targeted at $\alpha = 10^h 47^m 49\fm00, \delta = +13\fdg 55\farcm 40\farcs0$ (J2000). This field is $3\farcm82$ southeast of the galaxy center, equivalent to $\simeq$12 kpc projected linear radius. Since the effective radius of the galaxy light profile is $R_e = 1\farcm1$, our field is well beyond the central bulge and can plausibly be considered as giving us a first look at the genuine halo. Fields further in would obviously be of interest to track any metallicity gradient that the system might have, but would also be much more difficult to measure because of considerably increased crowding. The filters used were the ACS/WFC “wide V” ($F606W$) and “wide I” ($F814W$), the same ones as in our previous NGC 5128 studies. The $(V-I)$ color index is a very useful metallicity indicator particularly for old red-giant stars in the range \[Fe/H\] $\sim -1.5$ to $\sim 0$, which as will be seen below, is our main range of interest. Total exposure times were 38500 sec over 15 full-orbit exposures in $F606W$, and 22260 sec over 9 full-orbit exposures in $F814W$, adding up to a total of 24 orbits on this single field. The individual exposures were dithered over steps from 0 to 20 pixels, allowing elimination of most cosmic rays, bad pixels, and other artifacts on the detector. To prepare the images for photometry, we extracted the drizzled individual images from the HST Data Archive, registered them to within 0.05 pixels, and median-combined them. This procedure gave us a single very deep exposure in each filter. In Figure 1, we show the combined $I-$band image, demonstrating that the halo is well resolved into stars throughout. A small inset portion of the field is shown in Figure 2. Many faint background galaxies can be seen through the halo, but the overwhelming majority of detected objects in this field are the red giant stars of NGC 3377 itself. To carry out the photometry on the two combined frames we used the standalone version of [*DAOPHOT*]{} in its latest ([*daophot 4*]{}) version as written by one of us (PBS). The normal sequence of [*find/phot/allstar*]{} was used, with point spread functions defined from the average of 50 to 60 stars spread across the field. The pixel coordinate systems on the two images were registered, and then the detected stars on each one were matched up to find those measured in both colors. The number density of stars on the frame was reasonably high (the average separation between detected stars is $\simeq 7$ px, while the FWHM of the point spread function is 2.3 px or $0\farcs115$); also, differences in magnitude are not useful for deciding on matchup of stars between the $V$ and $I$ frames since the target stars have a very large true range in color. Matchup was therefore done by coordinates alone in a careful iterative procedure: five iterations were done with successively increasing matching radius, starting with $\Delta r = 0.2$ px and increasing to 1.5 px. In practice, more than 95% of the 75000 matches between the $V$ and $I$ images fell well within 1 pixel of each other. The next step was to define the cleanest possible subset of these matches. We rejected any detected objects that fell in the regions of 5 bigger-than-average background galaxies and one bright star, simply by masking out those regions. We also rejected objects with very poor goodness-of-fit to the PSF ($\chi_V > 2.5, \chi_I > 2.0$); and any with much larger than normal [*ALLSTAR*]{} magnitude uncertainties ($e_V, e_I > 0.25$). Although there is a very noticeable gradient in the density of stars from the lower right portion of the field to the upper left (see Fig.1), none of the areas are excessively crowded to the point where photometry via normal PSF fitting is unusually difficult. Our data do not enter the high-crowding regime that can be found in other examples in the literature, such as (among others) the inner halo of NGC 5128 [@h02], the bulge of the Local Group dwarf M32 [@gri96], or the halo of the Virgo dE VCC1104 [@har98]. Nevertheless, to extract a conservatively “uncrowded” sample of stars to work with, we calculated the nearest-neighbor distance (NND) for every measured star on both frames. We define the NND for a star simply as the distance to the closest neighboring star on the list *regardless of relative brightness*. The distribution of the NND values is shown in Figure \[crowding\], showing that the typical separation between objects is $\sim 5$ px or about twice the FWHM. To cull the photometry files of stars that are even mildly crowded, we decided to reject *any* object for which $NND < 3$ px, on either the $V$ or $I$ frames. This step removed about 17% of the $I$ detections and 12% of the $V$ detections. We regard the NND cutoff of 3 px as an extremely stringent choice; in fact, *allstar* has no difficulty handling cases down to $NND \sim FWHM$ or even less. The final result of the photometry, after removal of (a) the masked regions, (b) objects with high $\chi$ and high magnitude uncertainties, and (c) stars for which any concern existed about crowding, left a total of 57039 well measured, uncrowded stars. For calibration of the photometry, we chose to transform the filter-based magnitudes $F606W, F814W$ into $V, I$ in order to facilitate comparisons with previous work [@h99; @h00; @h02; @rej05]. First, we used the [*daophot/substar*]{} routine to subtract all but the brighter stars from the final averaged pair of images. We then obtained aperture photometry of these now-isolated bright stars to correct the [*allstar*]{} instrumental magnitudes to aperture magnitudes, and extrapolated these to large radius following the standardized prescriptions of @sir05. These gave us the magnitudes $F606W$ and $F814W$ on the natural ACS VEGAMAG filter system. Finally, we independently derived transformations of these to the standard $VI$ system by measuring an extensive set of images of the NGC 2419 standard field from the HST Archive, and comparing these with ground-based standard data in the same field. The ground-based data are part of PBS’s ongoing programme [@ste00] to maintain and upgrade an all-sky system of photometric standards on the [*BVRI*]{} system of @lan92. The photometry of the NGC 2419 field in particular is discussed by @ste05. This remote globular cluster has the advantages (for calibration purposes) of a blue horizontal branch at V$\,\sim\,$21, a red giant branch tip at ($V$, [*V-I*]{}) = (17.3, 1.50), and a wealth of foreground stars. The 1,257 stars adopted as standards in this field have a median of 46 groundbased observations per star in each of the $V$ and $I$ filters, a median standard error of 0.0054$\,$mag in $V$ and 0.0058$\,$mag in $I$, and a magnitude and color range $ 12.66 < V < 22.64$, $-0.05 < (V-I) < 3.04$. Keeping quadratic terms in color index, we found for these $$\begin{aligned} F435W \, = \, B + 0.135 (B-V) - 0.44 (B-V)^2 \\ F606W \, = \, V - 0.265 (V-I) + 0.025 (V-I)^2 \\ F814W \, = \, I + 0.028 (V-I) - 0.008 (V-I)^2\end{aligned}$$ These equations reproduce the NGC 2419 standard stars in all bands to within a scatter $\sigma = \pm 0.03$ mag. The equation for $B$ is not used in this paper but is listed for information. These transformations are very close to those published in @sir05, and also to those derived for our NGC 5128 outer-halo field [@rej05], which were taken during the same Cycle. The final color-magnitude diagram for the complete sample of 57039 stars is shown in Figure \[cmd1\] and will be discussed below. We estimated the internal photometric uncertainties and the detection completeness through a series of artificial-star tests with the [*addstar*]{} component of [*DAOPHOT*]{}. Stars were added to the combined $V$ and $I$ images in groups of 1000 over a wide range of magnitudes; these experiments were done independently on the $F606W$ and $F814W$ images. The images were remeasured in the same way as the original frames. The fraction $f$ of stars recovered, as a function of instrumental magnitude, is shown in Figure \[completeness\]. (The magnitudes here are the filter-based ones $F606W$ and $F814W$, discussed below.) The limits of our data, defined as the magnitudes at which $f$ drops to 0.5, are $F606W(lim) = 28.95$ and $F814W(lim) = 27.70$. The trend of $f$ with magnitude is well described in each case by a Pritchet interpolation function [@fl95], an analytic function with two parameters: the limiting magnitude, and a parameter $\alpha$ giving the steepness of the dropoff. For these images we find $\alpha=2.5$ for $V$ and $\alpha=2.7$ for $I$. Because of the color terms in the transformations (Eqs. 2,3), these limits in the native filter-defined magnitudes do not correspond to single $V$ or $I$ values, but for a giant star with a typical color of $(V-I) \simeq 2$, the limits are $V(lim) \simeq 29.4$ and $I(lim) \simeq 27.7$. In the color-magnitude diagram of Fig. \[cmd1\], the 50% completeness lines are shown. An important feature of these lines is that the limiting curve for $V$ (the upward-slanting line on the right side of the CMD) cuts off our ability to see any extremely red stars that might actually be present; these would fall at the most metal-rich end of our metallicity distribution function. Considerably deeper exposures in $V$ will be needed to explore the true “red limit” of the giant stars in this galaxy. Within the limits imposed by the photometry, we explicitly take into account the completeness fraction $f$ in our derivation (below) of the metallicity distribution. As will be seen later, the completeness cutoff may affect how much of the metal-rich end of the MDF is ultimately detectable. We find that the 50% completeness level does not change significantly with radius $R$ from galaxy center, except perhaps marginally in the region $R < 2'$ of highest crowding. A visual confirmation is shown in Figure \[4panel\_cmd\], where we subdivide the CMD data into four radial regions. The placements of the completeness cutoffs do not change relative to the distribution of stars (we will further justify this point in the later discussion on the radial gradient of the metallicity distribution). A further indication of the relative importance of crowding is shown in Figure \[skylevel\], where we show the local sky intensity around each measured star (in digital units, directly as returned by [*allstar*]{}) as a function of radius from the center of NGC 3377. The inward increase becomes steeper within $R = 2'$, but the net change over the entire frame is quite modest. However, in the discussion that follows (for the distance measurement and the metallicity distribution function), to remove any residual concerns about crowding issues we further restrict our analysis to the “safest” region $R > 2'$. The random uncertainties of the photometry, as derived from the artificial-star tests, are shown in Figure \[random\_errors\]. The rms uncertainty rises smoothly with magnitude in a roughly exponential manner, shown by the curves in the Figure. These have equations $\sigma(V) = 0.05$ exp$((F606W-26.00)/1.8)$ and $\sigma(I) = 0.05$ exp$((F814W-25.5)/1.3)$. In the analysis, we make no use of the data fainter than the 50% completeness level (marked by the vertical line in the figures). Finally, in Figure \[photom\_errors\] we show the mean trends for systematic bias in the photometry: a sample of the artificial-star tests is plotted, showing the median difference $\Delta m = (input - measured)$ for both filters. The median lines, plotted in 0.1-mag bins, show a consistent trend for stars to be measured $0.01 - 0.02$ mag too bright, a common feature of photometry of faint stars within moderately crowded fields. Fainter than the 50% completeness level, the mean curves diverge strongly to positive $\Delta m$ and the photometry becomes systematically unreliable. It is important to note, however, that for the stars brighter than the completeness limit, the mean biasses in the two filters run closely parallel, and the net bias in mean [*color*]{} $(V-I)$ – an important consideration since the color index is our main metallicity indicator – is negligibly small and has no effect on the derivation of the MDF. In addition, the random uncertainties have no significant effect on the inferred properties of the MDF, because we use only the brightest $\simeq 1$ mag of the RGB where the intrinsic color range of the stars is far larger than the photometric scatter. At the very top of the RGB, the intrinsic spread is $\Delta(V-I) \simeq 1.6$ mag while the photometric scatter is $\pm$0.1 mag or less. Distance Calibration ==================== NGC 3377 is part of the Leo I group along with several other large galaxies, most of which are spiral-type. Numerous measurements of distance to these individual Leo members are in the literature from a variety of well established distance indicators including Cepheids, planetary nebula luminosity function (PNLF), surface brightness fluctuation (SBF), and the tip of the old-red-giant branch (TRGB). [*NGC 3377:*]{} For NGC 3377 itself, the SBF method [@tonry01] gives $(m-M)_0 \equiv \mu = 30.25$, while the PNLF method [@ciar89] gives $\mu = 30.07 \pm 0.18$. [*NGC 3379:*]{} For the single giant elliptical in Leo I, the TRGB method has been applied through HST photometry in both the optical $I$ band and the near-infrared. The former [@sakai97] gives $\mu = 30.30 \pm 0.27$, and the latter [@gregg04] gives $\mu = 30.17 \pm 0.12$. The PNLF method [@ciar89] gives $\mu = 29.96 \pm 0.16$, and the SBF method [@tonry01] $\mu=30.12$. [*Large Spirals:*]{} Cepheid-based distances have been published for some of the major spirals in the group. These include $\mu = 30.01 \pm 0.19$ for NGC 3351 [@graham97], $30.25 \pm 0.18$ for NGC 3368 [@tanvir99], and $30.10 \pm 0.14$ [unweighted average of 27 Cepheids; @saha99] or $29.71 \pm 0.08$ [@freedman01] for NGC 3627. PNLF distances for these galaxies include $\mu=30.05 \pm 0.16$ for NGC 3351 and $29.99 \pm 0.08$ for NGC 3627 [@ciar02]. Treating all these Leo members as if they are at the same true distance from us and taking an indiscriminate average of all these measurements suggests a rough consensus near $\mu \simeq 30.1 \pm 0.05$, or $D \simeq 10.4$ Mpc for the group as a whole. The galaxy-to-galaxy dispersion of these measurements, which is $\sigma_{\mu} = 0.17$ mag, is quite similar to the typical internal uncertainties of each one and gives no strong evidence that the distance depth of the group is an important factor. Our new photometry penetrates well into the the old-halo red giant branch of NGC 3377, with a large sample of stars, and thus provides a new opportunity to use the TRGB distance indicator rather precisely. The key parameter is the “tip magnitude” which is, physically, the luminosity of the helium flash in the core of the red giant as it reaches the top of its first ascent along the giant branch. The bolometric luminosity of the RGB tip is, fortunately, only mildly dependent on metallicity for old stars, allowing it to be turned into an accurate standard candle. Our analysis follows the methods used in @sakai96 [@sakai97] and @h99, among others: we plot up the luminosity function of the RGB stars and use the detailed shape of the LF in the $I$ band to define the onset of the RGB. For stars more metal-poor than \[Fe/H\] $\simeq -0.7$ (which include the majority of the ones we measure here; see next section), the $I$ band has the strong advantage that the differential bolometric correction across the top of the RGB is almost cancelled by the dependence of $M_{bol}(tip)$ on metallicity, leaving $M_I(tip)$ with only a gradual slope with increasing color. The luminosity function is shown in Figure \[lf\], based on 49380 stars beyond $R>2'$ from the galaxy center. It has been smoothed with a Gaussian kernel of $\sigma_I = 0.02$ mag. Completeness corrections are quite unimportant here, since the $f=0.5$ completeness level is $I \simeq 27.6$, much fainter than the well resolved top of the RGB. The fact that the LF rises gradually upward at the TRGB (rather than the ideal case of an abrupt jump) is a normal consequence of an intrinsically steep LF at that point, convolved with the observational photometric scatter, and adding in the smoothing effect of the slight downward slope of the tip toward redder colors, and a small amount of field contamination, all of which act to blur out the TRGB point to some extent [see @har98; @h99 for detailed discussion and methodology]. The actual tip is clearly somewhere near $I = 26$, but to define the true TRGB precisely, we look for a strong change in the slope of the LF as we go to fainter magnitudes across the tip. The numerically calculated first derivative of the LF, $dn/dI$, is shown in the second panel. A sharp upturn is present past $I = 26$ with well defined peaks that appear at $I = 26.2$ and 26.3. These are independent of the precise value of the smoothing kernel within broad limits. To decide which of these to pick we use the “edge response filter” or ERF (the numerical second derivative of the LF), plotted in the lower panel of Fig. \[lf\]. Here, we adopt the first and most prominent peak at $I(tip)= 26.2 \pm 0.1$. The distance modulus follows immediately once we apply a fiducial value for $M_I(tip)$. The most well established recent calibration from within the Milky Way is from a very large sample of stars in $\omega$ Cen [@bell04], giving $M_I(tip) = -4.05 \pm 0.12$. Calibrations based on theoretical RGB models [e.g. @sal02] yield values in the range $-3.95$ to $-4.22$ depending on both the details of the stellar physics and on observational constraints, but are entirely consistent with the $\omega$ Cen value, which we adopt here. We therefore obtain $(m-M)_I = 30.25 \pm 0.15$ for NGC 3377. This must be corrected for a foreground absorption of $A_I = 0.07\pm0.02$, giving a final TRGB distance measurement $\mu = 30.18 \pm 0.16$. Averaging the TRGB distance in with the SBF and PNLF measurements listed above, and giving the three methods equal weights, we arrive at an average $(m-M)_0 = 30.17 \pm 0.10$, or $D = 10.8 \pm 0.5$ Mpc for NGC 3377 itself. These three methods – all based, necessarily, on the properties of the old stellar population – are in excellent mutual agreement, and we consider the distance to this system to be as well established as any in the local galactic neighborhood. The Nature of the Brighter Stars ================================ In Figs. \[cmd1\] and \[4panel\_cmd\] there are numerous stars scattered above the TRGB. There are $N_{bright} \simeq 1360$ such stars within $24 < I < 26$. Could these be a signature of a younger, intermediate-age population ($\tau \lesssim 3$ Gy), which would indicate some more recent star formation in the galaxy’s history? in Figure \[xy2\], we show the spatial distribution of these stars across our ACS field, compared with the distribution of the stars in the brightest half-magnitude of the RGB itself. Both types of stars show an obvious gradient decreasing from galaxy center, and so a significant fraction of $N_{bright}$ must be genuine members of NGC 3377. Before placing limits on the number of such stars that might genuinely be present, we need to rule out other possible contributors, including (a) field contamination, (b) accidental blends of two normal RGB stars, or (c) giants or AGB stars in highly evolved or temporary states including long-period variables (LPVs), which can contribute noticeable numbers of supra-TRGB stars even in an old population [see @rej03; @gregg04 for more extensive discussion and modelling]. We dealt with a similar issue for the halo stars in NGC 5128 [@h99] and found that only $\sim$1% of the population there could be ascribed to the intermediate-age category. The first option (field-contaminating objects) will include both foreground stars and faint, very small background galaxies that are near-starlike in appearance. The number of foreground stars in the direction of NGC 3377 and over the area of ACS/WFC should be $N \lesssim 20$ from Galactic models [e.g. @bah81], but the number of faint, misidentified background galaxies is almost certainly larger and would best be measured from an adjacent ‘control’ field, which we do not have. However, we can specify a reasonable upper limit to $N_{field}$ by using the outermost regions of our measured field. For $R_{GC} > 4\farcm5$ (an area of $2.78$ arcmin$^2$) there are 94 objects within $24 < I < 26$, suggesting that over the entire 11.33 arcmin$^2$ field $N_{field}$ should be at most 376. The number density of stars is still declining at the outer edges of the field, and so this number must be a generous upper limit. Since we see 1360 such objects over the whole field, this suggests $\sim 1000 - 1200$ of them are intrinsic to NGC 3377. Some objects could appear in this bright range because of accidental blends of two stars that are both near the upper end of the RGB. Blends like this are very likely to be responsible for much of the scatter of stars appearing just above the nominal RGB tip[^2], but not for the ones $\sim 0.7$ mag or more above it. Statistically the number of blended pairs will increase as the square of the number of stars per unit area, $$N_{blend} \, \simeq \, {N_{\star}^2 \over 2} {{\pi q^2} \over d^2}$$ where $N_{\star}$ is the number of stars on the frame capable of generating a blended pair brighter than the TRGB, $q$ is the radius of one resolution element, and $d^2$ is the area of the field. Adopting $q \simeq 2$ px, and using the fact that there are $\simeq 30000$ stars within 1 magnitude of the RGB tip, we expect $N_{blend} \sim 350$, or about one-quarter of all the objects appearing above the TRGB. The presence of LPV-type stars is favored in an intrinsically metal-rich old population [see @rej03], and their numbers should be proportional to the total luminosity of the whole population contained in our field. Extrapolating from the stars brighter than our completeness limit with a standard Population II luminosity function, we estimate very roughly that $V(int) = 14.7$ or $L \sim 8 \times 10^7 L_{\odot}$. From [@ren98] we then find $N_{LPV} \simeq 400$, similar to the expected number of accidental blends. An observational confirmation that LPV-type variables should indeed be there to be found is discussed by @gregg04, who present $HST/NICMOS$ near-infrared photometry of the RGB stars in inner-halo fields for NGC 3379, the other Leo elliptical. They find (see their Table 3) that between one-quarter and one-half of the detected stars above the TRGB are variable, consistent with our rougher estimate. @rej03 found a total of more than 1100 LPVs in two fields around NGC 5128 covering an area of 10.46 arcmin$^2$, and state that these make up roughly half of the total number of stars above the TRGB, again consistent with the estimates made here. In summary, we suggest that more than half the population of stars that lie clearly above the RGB tip are due to a combination of blends, temporary high-luminosity states of the giants, and a small amount of field contamination. The remainder not accounted for by these effects is thus $N_{bright} \sim 400 \pm 100$, or $\simeq$1% of our observed total on the field. We conclude tentatively that there is no evidence for any significant “young” ($\tau < 3$ Gy) population of stars in our field. These rough estimates are not intended to replace a more comprehensive population synthesis analysis, based on simulating the observed RGB with stars drawn from model isochrones over a complete range of ages and metallicities [see, e.g. @will06 for a recent example]. In a later stage of this work we will develop such simulations and constrain the age range of the halo stars more completely. The Metallicity Distribution ============================ Deriving the MDF ---------------- We are now in a position to use the colors of the RGB stars to gain a first direct measurement of the metallicity distribution function (MDF) in this galaxy. To enable direct comparisons with other systems, we follow the same method as used previously for NGC 5128 and selected dwarf ellipticals [@h99; @h00; @h02]. A finely spaced grid of RGB evolutionary tracks is superimposed on the measured CMD, and we interpolate between these to tag every star with a heavy-element abundance $Z$ based on its location within the tracks. As before, we use the $\alpha-$enhanced tracks of @van00, and calibrate them in terms of absolute $(V-I)$ color by force-fitting them to the observed RGB sequences for Milky Way globular clusters. This approach is described in detail in @h02 and we do not repeat it here. However, we emphasize that the metallicity scale we derive this way is an [*observationally calibrated one*]{} based on real globular clusters. [*The theoretical RGB tracks are used only to aid interpolation between the observed sequences for real clusters,*]{} and thus our results do not depend critically on the particular choice of stellar models. The one unavoidable and model-dependent assumption that underlies this method is that the mean age of the NGC 3377 stars is taken to be the same as for our local globular clusters (that is, about 12 Gy). If in fact they are systematically younger, then this method would slightly underestimate their $Z-$abundances since the RGB locus shifts blueward at lower age. However, as long as the age differences are not severe, the age correction to the MDF is small; for example, in @rej05 we note that a shift in mean age from 12 Gy to 8 Gy would produce only a $0.1-$dex underestimate in log $Z$. See also the RGB track comparisons for different ages in @h99, which demonstrate similar conclusions. In Figure \[cmd\_fiducial\] we show again the composite color-magnitude diagram, but now with the RGB tracks added. The solid lines are those of the @van00 grid, whereas the two dashed lines at right are ones for Solar ($Z=Z_{\odot}$) and $\simeq 3 Z_{\odot}$ metallicities. It is clear, as we noted above, that the photometric detection limit for red stars set by the $F606W$ exposures would eliminate any giant stars at Solar metallicity or above from our sample, even if they were present. What is not obvious from this Figure is whether or not this detection cutoff imposes any more basic limit on our interpretation of the MDF; that is, are many high$-Z$ stars actually likely to be present in this galaxy? To gain a better reply to this question, we need to look at the details of the MDF itself as far as we can gauge them. Our derived MDF is shown in Figure \[feh\_3panel\], where we divide the sample into half-magnitude bins by approximate luminosity $M_{bol}$. By doing this, we test for any systematic errors in the interpolation procedure that might result from incorrect placement of the RGB model grid on the data; we also use it to reveal any broadening of the deduced MDF from photometric scatter toward the faint end of the data. Encouragingly, we see no systematic shift in the shape of the MDF: in all three bins, the peak occurs at log $(Z/Z_{\odot}) \simeq -0.6$, with a broad tail extending to lower metallicity and a much steeper ramp-down to higher metallicity. As we found for the giant elliptical NGC 5128 [@h02], remarkably few stars are more metal-poor than log $(Z/Z_{\odot}) = -1.5$. As a visual confirmation of this point, note directly from the CMD how few stars there are near the bluest RGB tracks, particularly at the high-luminosity end where the track color is most sensitive to metallicity. In other words, NGC 3377 – [*like the giant ellipticals, but unlike the smallest dwarfs and globular clusters*]{} – contains few stars indeed that are in the same regime as the “classic metal-poor” halo of the Milky Way. This comparison (see also the discussion below) indicates that NGC 3377 did not form simply by the amalgamation of fully-formed dwarfs. Fig. \[feh\_3panel\] explicitly shows the MDF with, and without, photometric completeness corrections, since these are important to our assessment of the MDF. Any extremely red stars falling to the right of the $f=$50% line in the CMD have been rejected from the sample, since in this region the completeness correction itself becomes dangerously large and the random and systematic errors of the photometry increase rapidly. For all stars brighter than this cutoff level, the completeness-corrected samples (the open histograms in Fig. \[feh\_3panel\]) have been constructed by weighting each star individually as $(1/f)$ where $f = f_I \cdot f_V$ is the combined completeness fraction at its particular location in the CMD. For comparison, the unweighted MDF (based only on counting up all stars with $f > 0.5$) is shown in the hatched regions. Because the 50% completeness line in $V$ runs nearly parallel to the RGB tracks in its region, it essentially has the effect of cutting off the MDF rather abruptly at log $(Z/Z_{\odot}) = -0.3$. Redward of this point, we have no reliable knowledge of the MDF shape. But by the time we reach this photometrically-driven cutoff, the MDF is already well past its peak frequency and declining steeply. The evidence is thus strongly suggestive that we have seen the majority of the total MDF, and that only a small fraction of its stars are at Solar-type abundance or higher. By contrast, the giant NGC 5128 has an MDF extending well up past Solar abundance, even in the outer reaches of its halo [@h02; @rej05]. The fact that the $V$ completeness cutoff is so close to the MDF peak, however, leaves lingering doubts that we have indeed seen the full shape of the metallicity distribution. Could there, in fact, be significant numbers of more metal-rich stars in the galaxy that perhaps belonged to a later star-forming episode and are still waiting to be identified? Such stars would be almost as bright in $I$ as the other RGB stars, but would have been too red to have been detected in $V$. To look into this a bit further, we have used the $I-$band measurements alone to isolate stars that were (a) detected and measured on the $I-$band image; (b) [*not*]{} detected on the $V-$band image; and (c) comparably bright in $I$ to the other RGB stars. If there are large numbers of these, it would be suggestive that we have missed much of the actual RGB. The luminosity function for the stars measured [*only*]{} in $I$, and passing the same photometric culling described above, is shown as the dotted line in Figure \[2lf\]. This LF rises to a peak at the completeness limit $I \simeq 27.7$ and then falls steeply. By contrast, the LF for the stars measured in [*both*]{} $V$ and $I$ (solid line in Fig. \[2lf\], replotted from Fig. \[lf\]) has a broad peak determined by the $V-$band completeness limit. Very metal-rich RGB stars (those with log $Z/Z_{\odot} > -0.3$) would belong to the “$I$ only” LF and would lie in the approximate range from $I \simeq 26.6$ (for the very top of the RGB at $(V-I) \sim 3$) down to the $I$ completeness limit. We see that the LF for the stars in that range (dotted line in the Figure) stays well below the LF for the stars with both $V,I$ (solid line). In the range $26.6 < I < 27.6$, 20008 objects were measured [*only*]{} in $I$ but not $V$. But this total should be considered as only an [*upper*]{} limit to the actual number of metal-rich RGB stars, because close visual inspection of the “$I$ only” candidates shows that a high fraction of them are small clumps of bad pixels, or faint, very small background galaxies. These passed the $\chi$ and $e_I$ selection steps, but did not match objects appearing at the same place on the $V$ image and thus were rejected in the construction of the CMD. Taking these into account, we make a [*very rough*]{} estimate that there are fewer than $\sim 10000$ genuine RGB stars redder than the completeness cutoff in our CMD. By comparison, if we employ a simple chemical evolution model to predict the shape of the entire MDF and use it to extrapolate the observed MDF up past the photometric completeness cutoff (see next section), we find that 10% of the MDF should lie at $Z > 0.5 Z_{\odot}$, corresponding to $\sim 5700$ ‘missed’ stars in addition to our observed sample of 57000 over all radii. These very metal-rich stars would not change the MDF in a major way if (for example) they formed only the high-metallicity tail of the MDF that we already see. Unfortunately, any more definitive statement will have to await deeper photometry than is currently available. A clear answer would require doubling the $V$ exposure time; that is, adding another 15 orbits to what we already have in hand. The incremental cost of HST time would thus be significant. A Chemical Evolution Model -------------------------- To step beyond the raw MDF into a simple physical interpretation, we use the basic chemical evolution model outlined in @h02 (hereafter HH02) that we applied to the NGC 5128 halo and bulge stars. Similar models have also been used for the halo of the Milky Way [@prantzos03], and the globular cluster systems of large galaxies [@van04], among other situations. This is the so-called “accreting box” model in which we assume that a region of primordial ($Z \simeq 0$) gas turns itself into stars through a long succession of star-forming episodes, all the while that more gas is flowing into the region. The rate of gas infall is assumed to die away gradually and smoothly with time, so that in the late stages of the region’s history, its chemical evolution asymptotically approaches the classic “closed box” or “simple” model. In a general sense, this model is a first-order description of what would be expected to happen during hierarchical merging of a large set of initial, pristine gas clouds, wherein star formation within the clouds is happening simultaneously with the merging into a bigger final galaxy. In HH02 we postulate that the gas infall rate starts at a rather high level and then dies away as an exponential decay with time. We also assume that at each star formation step, the same (small) fraction of ambient gas gets turned into stars. The abundance $Z$ of the stars forming at any given moment is then determined by the $Z(gas)$ left behind by the previous steps, in addition to the amount and composition of the new gas entering the region just before the next star formation step occurs. The sequence of timesteps can be numerically integrated to give the final distribution $n(Z)$ of the stars (see HH02 for details). The model in its simplest form has three essential and unavoidable parameters: the effective yield $y_{eff}$ of nucleosynthesis; the initial gas infall rate relative to the amount of gas initially present in the region; and the exponential decay time for the infall. An extremely effective way to exhibit the match between model and data is through the linear form of the MDF, as number of stars per unit heavy-element abundance $Z/Z_{\odot}$. The key results are shown in Figure \[zmodel\], where we include stars brighter than $M_{bol} \simeq -2.5$, i.e. ones in the upper two panels of Fig. \[feh\_3panel\]. As before, we include only stars at radii $R_{GC} > 2'$. The raw (shaded) and completeness-corrected (unshaded) MDF are both shown. One particular accreting-box model solution is shown as the curved line. For $Z \gtrsim 0.5 Z_{\odot}$, the observed MDF is cut off by photometric incompleteness, but as noted above, this cutoff is well past the MDF peak at $Z \simeq 0.15 Z_{\odot}$, and the MDF is seen to decline strongly up till the photometric cutoff. The model fit shown in Fig. \[zmodel\] is for a triad of parameters that match the [*completeness-corrected*]{} data. The effective yield is $y_{eff} = 0.0039 \simeq 0.26 Z_{\odot}$; the exponential decay time for infall is 30 timesteps; and the total amount of gas infall is equal to 4 times the initial gas present in the region. The numerical integration is carried through several hundred timesteps until almost no gas remains. In each timestep we assume following HH02 that 5% of the ambient gas is converted to stars, and that the infalling gas has $Z=0$. As we commented in HH02, the absolute value of the timestep $\delta t$ is arbitrary in this model and must be supplied by other methods; however, through various arguments about the actual rate of hierarchical merging (see again HH02), it is plausibly near $\delta t \sim $ 30 Myr give or take factors of two. If so, the main period where infall is strongly influencing the MDF shape is thus the first $1 -3$ Gyr. Each of the three critical parameters in the model controls three distinctly different features of the final MDF. First, the rapid rise in $n(Z)$ from $Z = 0$ to $Z = 0.1$ is set by the initial infall rate: the less the infall of pristine gas, the smaller the rise, until in the limit of zero infall, the $n(Z)$ curve approaches the simple exponential decline of the closed-box model, $n \sim e^{-Z/y}$. Second, the location of the peak, where the MDF turns over and begins to decline, is driven by the infall decay time: a longer decay time produces a higher$-Z$ peak point since there is more gas to allow the enrichment process to continue driving upward. Finally, the long exponential tail to higher $Z$, where infall has essentially stopped and the only process left is the conversion of the remaining gas into stars (that is, a closed-box model), is shaped by the effective yield $y_{eff}$. We have restricted the present comparisons to the simplest set of parameters possible within the context of the model, and the three that are the least co-dependent. But additional model parameters are possible: for example, ones introduced in our previous analysis of NGC 5128 included an initial, very short, time period $\tau_1$ over which the gas infall is constant or rapidly ramps up; and also the assumed abundance $Z_i$ of the infalling gas. However, these prove to be much less critical to the solution than the three primary parameters defined above. The MDF data do not have enough resolution at extremely low $Z$ to constrain $\tau_1$; and any plausibly small abundance for the infalling gas in the range $Z_i \lesssim 0.1 Z_{\odot}$ can be used to generate adequate matches to the data (see HH02 and Vandalfsen & Harris 2004 for more extensive discussion). The value of the effective nucleosynthetic yield at about one-quarter Solar is a factor of two or three smaller than standard $y-$values that would hold in a closed-box model where all the gas remains held [*in situ*]{} and where the stellar IMF and nucleosynthesis rate are normal [e.g. @pagel75]. But if some fraction of the gas escapes in stellar and SN-driven winds (leaky-box) then the effect is to lower $y_{eff}$ in proportion [@binney98]. Formally, our results indicate that about half the gas present in the NGC 3377 region was ejected during the successive rounds of star formation, preventing the enrichment from proceeding upward past about one-half Solar levels. For comparison, in HH02 we found $y_{eff} \simeq 0.3 Z_{\odot}$ for the outer halo of NGC 5128, similar to the solution for NGC 3377. For the inner region of NGC 5128, much deeper in its potential well where virtually all the gas could be held, we found $y_{eff} \simeq 0.85 Z_{\odot}$. For systems such as these, with broad MDFs and obviously extended star formation histories, much more complex models can be clearly brought to bear. Numerous examples of such models can be found in the literature [e.g. @naab06; @rom05; @wor05; @scan05; @valle05; @laf04; @oey00 among many recent discussions]. These have been developed all the way to full hierarchical-merging and chemodynamical-evolution codes [e.g. @kawata03; @cole00; @bea03; @som99]. In some cases very different parametric approaches are used; for example, @wor05 advocate the use of a closed-box model in which $y_{eff}$ changes with time. In such a model, quite a strong decrease in $y$ with increasing $Z$ is necessary to reproduce the rising part of the MDF at low $Z$. In our view, the use of gas infall combined with $y_{eff} \simeq const$ is much more physically plausible, and has the strong advantage of employing a process (inflow of low-metallicity gas) which is almost certainly happening anyway at early times. It is notable, however, that the minimally simple set of parameters we use here already provides an accurate first-order fit to the MDF. The interpretation we offer for the validity of our first-order model is simply that in galaxies this large, we are looking at the combined results of so many star-forming events that much of the individual detail has been averaged out and is now no longer distinguishable. Gradients and Fine Structure? ----------------------------- We can also use the MDF to search for any indication of a radial metallicity gradient. Our ACS/WFC field has stars with projected radial distances from the center of NGC 3377 from $1\farcm3$ to $5\farcm7$ (4.1 kpc to 17.7 kpc), a large enough range to be worth exploring for trends. The MDF, again in its linear form, is shown in Figure \[z\_3panel\] for an inner zone ($1\farcm3 - 2\farcm0$), a middle zone ($2\farcm0 - 3\farcm5$), and an outer zone ($3\farcm5 - 5\farcm7$). In each case, we show the [*same chemical evolution model*]{} applied to all three radial bins that we deduced for the total population in Fig. \[zmodel\] above, simply renormalized for the different total numbers of stars. All three panels are similar. We see no evidence for a radial gradient in metallicity, either in the peak position of the MDF or in its detailed shape. Although we deleted the innermost region $R < 2'$ in the earlier analysis to minimize random scatter in the CMD, LF, and MDF, it has not led to any systematic bias in our understanding of the metallicity structure. This is, perhaps, not a surprising result since the effective radius of the galaxy is $1\farcm1$ and thus all the stars we sample are well outside it. In NGC 5128, we found that the MDF is virtually unchanged over a radial range from $\sim 10$ kpc out to 40 kpc [@h99; @h00; @rej05]; it is only within the bulge region at $r < 8$ kpc (for NGC 5128, $r_{eff} = 5.5$ kpc) that significant changes began showing up and the MDF became systematically more metal-rich (HH02). Finally, we note that in either logarithmic or linear form, the MDF is very broad but contains a tantalizing hint of fine structure. The “notches” at $(Z/Z_{\odot}) \simeq 0.2$ and at $\simeq$0.3 are statistically significant at $\gtrsim 5 \sigma$ in a nominal sense (the bin at the first gap has 1000 stars whereas the smooth model would predict 1500; and the second gap has 820 stars versus 1120 expected from the model curve). Furthermore, the same gaps appear in all three of the radial bins plotted in Fig. \[z\_3panel\]. These small gaps are suggestive, perhaps, of distinct episodes of star formation that were incompletely averaged out over the whole galaxy. The extreme low-luminosity systems $\omega$ Cen and Carina show MDFs that are more sharply divided into a small number of distinct episodes [e.g. @koch06; @sollima05 among many others], whereas no such gaps appear in the much larger NGC 5128 [@h02; @rej05]. In this respect as in many others, NGC 3377 appears to occupy an intermediate position in the whole range of old, composite stellar systems: the smaller the system, the more likely it is that its MDF will be dominated by a smaller number of star-forming events and will take on a discontinuous morphology. Comparisons with Smaller and Larger Systems ------------------------------------------- Adding this study to some of the previously cited papers, we can now make a first, admittedly rudimentary, attempt to put together a comparison of MDFs for the full range of normal E galaxies, from dwarfs to giants. An illustrative sample is shown in Figure \[z\_4gal\]. To represent the low-luminosity end of the normal E sequence, we use the “typical” Local Group dwarf NGC 147 ($M_V^T = -15.6$), with photometric $(V,I)$ RGB data from @han97. We also include two spatial regions from the giant NGC 5128, at $M_V^T = -22.1$: an outer-halo field at 40 kpc projected distance [@rej05], and the 8-kiloparsec “inner halo” or outer-bulge field analyzed by HH02. In all cases, exactly the same interpolation code and RGB model grid has been used to derive the MDF, so the results are internally homogeneous. The sequence in Fig. \[z\_4gal\] runs from the small potential well of the dwarf, through the intermediate-mass NGC 3377, to the outer regions of a giant, and finally to the deep inner regions of the giant. The change in the MDF clearly shows up as a progressively increasing spread in heavy-element abundance. We repeat that the upper tail for NGC 3377 is cut off by photometric incompleteness for $Z > 0.6 Z_{\odot}$ and it is quite possible that a thin extension to higher $Z$ is present. By contrast, the NGC 147 and NGC 5128 color-magnitude diagrams are more deeply sampled and less affected by incompleteness. In terms of physical trends, however, it is interesting to note that the [*peak*]{} of the MDF changes only modestly, from $\simeq 0.1 Z_{\odot}$ in the dwarf up to $\simeq 0.3 Z_{\odot}$ in the bulge of the giant. Instead, the most prominent effect shows up in the relative number of stars at high metallicity: in the dwarf, nothing is present above $0.4 Z_{\odot}$, while the MDF for the giant extends far above Solar. This is the most dramatic effect of the depth of the potential well within which their stars were forming: in the deeper wells, almost all of the gas will be retained, permitting the enrichment to continue on through many generations and go up to high $Z$. A secondary feature worth noting is the steepness of the rise at low $Z$ in the two smaller galaxies, showing the importance of infalling gas. This pattern is suggestive of an interpretation that the smaller galaxies took somewhat longer to assemble than the giants, during which time the inflow of pristine gas could continue longer and have a more important effect on the final MDF. Summary ======= We have used deep HST/ACS images of a field in the intermediate-luminosity elliptical galaxy NGC 3377, a member of the Leo group, to derive the metallicity distribution function of its stars. More than 57000 stars were measured over a radial region extending from $1\farcm6$ to $5\farcm7$ from galaxy center. The $(V,I)$ data reach deep enough to reveal almost 2 magnitudes of the red-giant branch and show the expected large spread in color that accompanies a high internal dispersion in metallicity. No significant presence of any young stellar population (stars younger than a few Gyr) can be seen in the color-magnitude diagram and the interpretation of the system is relatively simple. Using a finely spaced grid of RGB models calibrated against the Milky Way globular clusters, we derive the heavy-element abundance distribution $n(Z)$ for the NGC 3377 stars. No apparent radial gradient in metallicity shows up; in all parts of the halo, the distribution peaks at log $(Z/Z_{\odot}) \simeq -0.6$ and is quite broad. Very few stars are, however, more metal-poor than log $(Z/Z_{\odot}) = -1.5$, and the data suggest that few stars are more metal-rich than $-0.3$. It is possible that this galaxy did not reach Solar abundance during its enrichment history, at least in its halo. The overall shape of its MDF in all respects lies between the more metal-poor dwarf ellipticals on the one hand, and the very much more metal-rich giants on the other. WEH and GLHH thank the Natural Sciences and Engineering Research Council of Canada for financial support. We also thank an anonymous referee for constructive comments on the first version of the paper. Bahcall, J., & Soneira, R. 1981, , 47, 357 Beasley, M.A., Harris, W.E., Harris, G.L.H., & Forbes, D.A. 2003, , 340, 341 Bekki, K., Harris, W.E., & Harris, G.L.H. 2003, , 338, 587 Bellazzini, M., Ferraro, F.R., Sollima, A., Pancino, E., & Origlia, L. 2004, , 424, 199 Binney, J., & Merrifield, M. 1998, Galactic Astronomy (Princeton: Princeton Univ. Press) Chiappini, C., Matteuci, F., & Gratton, R. 1997, , 477, 765 Ciardullo, R., Jacoby, G.H., & Ford, H.C. 1989, , 344, 715 Ciardullo, R., Feldmeier, J.J., Jacoby, G.H., Kuzio de Naray, R., Laychak, M.B., & Durrell, P.R. 2002, , 577, 31 Cole, S., Lacey, C.G., Baugh, C.M., & Frenk, C.S. 2000, , 319, 168 Fleming, D.E.B., Harris, W.E., Pritchet, C.J., & Hanes, D.A. 1995, , 109, 1044 Freedman, W.L. et al. 2001, , 553, 47 Graham, J.A. et al. 1997, , 477, 535 Gregg, M.D., Ferguson, H.C., Minniti, D., Tanvir, N., & Catchpole, R. 2004, , 127, 1441 Grillm,air, C.J. et al. 1996, , 112, 1975 Han, M. et al. 1997, AJ, 113, 1001 Harris, G.L.H., Harris, W.E., & Poole, G.B. 1999, , 117, 855 Harris, G.L.H., & Harris, W.E. 2000, , 120, 2423 Harris, W.E. 1996, , 112, 1487 Harris, W.E. 2001, in Star Clusters, Saas-Fee Advanced Course 28 (New York: Springer), ed. L.Labhardt & B.Binggeli Harris, W.E., Durrell, P.R., Pierce, M.J., & Secker, J. 1998, Nature, 395, 45 Harris, W.E., & Harris, G.L.H. 2002, , 123, 3108 (HH02) Kawata, D., & Gibson, B.K. 2003, , 340, 908 Koch, A., Grebel, E.K., Wyse, R.F.G., Kleyna, J.T., Wilkinson, M.I., Harbeck, D.R., Gilmore, G.F., & Evans, N.W. 2006, , 131, 895 Lafranchi, G.A., & Matteucci, F. 2004, , 351, 1338 Landolt, A.U. 1992, , 104, 340 Naab, T., & Ostriker, J.P. 2006, , 366, 899 Oey, S. 2000, , 542, L25 Pagel, B.E.J., & Patchett, B.E. 1975, , 172, 13 Prantzos, N. 2003, ,404, 211 Rejkuba, M., Minnite, D., Silva, D.R., & Bedding, T.R. 2003, , 411, 351 Rejkuba, M., Greggio, L., Harris, W.E., Harris, G.L.H., & Peng, E.W. 2005, , 631, 262 Renzini, A. 1998, , 115, 2459 Romano, D., Tosi, M., & Matteucci, F. 2005, , 365, 759 Saha, A., Sandage, A., Tamman, G.A., Labhardt, L., Maccheto, F.D., & Panagia, N. 1999, , 522, 802 Sakai, S., Madore, B.F., & Freedman, W.L. 1996, , 461, 713 Sakai, S., Madore, B.F., Freedman, W.L., Lauer, T.R., Ajhar, E.A., & Baum, W.A. 1997, , 478, 49 Salaris, M., Cassisi, S., & Weiss, A. 2002, , 114, 375 Scannapieco, C., Tissera, P.B., White, S.D.M., & Springel, V. 2005, , 364, 552 Sirianni, M. et al. 2006, , 117, 1049 Sollima, A., Pancino, E., Ferraro, F.R., Bellazzini, M., Straniero, O., & Pasquini, L. 2005, , 634, 332 Somerville, R.S., & Primack, J.R. 1999, , 310, 1087 Stetson, P.B. 2000, , 112, 925 Stetson, P.B. 2005, , 117, 563 Tanvir, N.R., Ferguson, H.C., & Shanks, T. 1999, , 310, 175 Tonry, J.L., Dressler, A., Blakeslee, J.P., Ajhar, E.A., Fletcher, A.B., Luppino, G.A., Metzger, M.R., & Moore, C.B. 2001, , 546, 681 Valle, G., Shore, S.N., & Galli, D. 2005, , 435, 551 VanDalfsen, M.L., & Harris, W.E. 2004, , 127, 368 VandenBerg, D.A., Swenson, F.J., Rogers, F.J., Iglesias, C.A., & Alexander, D.R. 2000, , 532, 430 Williams, B.F. et al. 2007, , 654, 835 Worthey, G., España, A., MacArthur, L., & Courteau, S. 2005, , 631, 820 [ccccccccc]{} $Type$ & E5\ $\alpha$ (J2000) & $10^h 47^m 42\fs4$\ $\delta$ (J2000) & $13\fdg 59' 08''$\ $v_r$ (helio) & 665 km s$^{-1}$\ $A_V$ & 0.10\ $(m-M)_0$ & $30.17 \pm 0.10$\ $V_T^0$ & 10.23\ $M_V^T$ & $-19.9$\ [^1]: Based on observations made with the NASA/ESA Hubble Space Telescope, obtained at the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5-26555. These observations are associated with program \#9811. Support for this work was provided in part by NASA through grant number HST-GO-09811.01-A from the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5-26555. [^2]: There are $\simeq 1300$ stars in our CMD in the small interval $I = 24.0 - 24.2$ just above the RGB tip that are responsible for blurring out the definition of the TRGB and creating the smooth rolloff in the luminosity function just above the tip; see Figure 10. Most of these objects can be understood as due to blends of the brighter RGB stars with the huge number of RGB stars $\gtrsim 1.5$ mag [*fainter*]{} than the tip.

ArXiv

--- abstract: 'We produce two families of rank zero quadratic twists of the elliptic curve $E:~y^2=x^3+2$. Moreover, we prove that there are infinitely many members with rank zero in each of these families.' address: 'Azizul Hoque@Harish-Chandra Research Institute, HBNI, Chhatnag Road, Jhunsi, Allahabad-211019, India.' author: - Azizul Hoque title: Rank zero quadratic twists of $y^2 = x^3 + 2$ --- Introduction ============ It is well known that the (abelian) group $E(\mathbb{Q})$ of the rational points on an elliptic curve $E$ defined over $\mathbb{Q}$ is finitely generated. The rank of $E$ is the minimal number of generators of the free abelian group $E(\mathbb{Q})/E(\mathbb{Q})_{\text{tors}}$. Thus it is positive if and only if there are infinity many rational points on $E$. The determination of the rank of $E$ is an important problem. Specially producing elliptic curves with arbitrarily high rank is still an unsolved problem. A geometric analogue of this problem has been shown to be true by Shafarevich and Tate [@ST]. In most cases where the rank has been obtained, it is actually very small. In fact, it is widely believed that half of the curves have rank zero, while the other half are of rank one. Thus there is not much room left for the curves of higher rank. Let $E$ be an elliptic curve defined over $\mathbb{Q}$ by the Weierstrass equation $$E:~y^2=x^3+ax+b,$$ where $a$ and $b$ are integers. For a square-free integer $D$, the quadratic twist of $E$ by $D$ is given by $$E_D:~ y^2 = x^3 + aD^2x + bD^3.$$ Then one can ask the following question: \[q1\] What can be said about the variation of the rank of $E_D$ as $D$ varies over the square-free integers? Let $\mathcal{N}$ be the conductor of $E$ and $\chi_D$ the quadratic character associated to $D$. The parity conjecture states that if $\gcd(D, 2\mathcal{N}) = 1$, then the ranks of both $E$ and $E_D$ have the same parity if and only if $\chi_D(-\mathcal{N}) = 1$. Under truth of this conjecture, Gouvea and Mazur [@GM] proved that the number of square-free integer $D$ with $|D|\leq X$ such that the rank of $E_D$ is positive and even is $\gg X^{\frac{1}{2}-\varepsilon}$ for each $\varepsilon>0$. It is also conjectured that there are infinitely many primes $p$ for which $E_p$ has rank zero and there are infinitely many primes $q$ for which $E_q$ has positive rank. Along this direction, Mai and Murty [@MM] proved that there are infinitely many square-free integers $D\equiv 1\pmod {4N}$ for which $E_D$ has rank zero, where $N$ denotes the conductor of $E$ over $\mathbb{Q}$. Ono [@ON] improved this result by showing that there is a set $S$ of primes $p$ with density $\frac{1}{3}$ for which if $D = \prod\limits_i p_i$ is a square-free, where $p_i\in S$, then $E_D$ has rank zero for some special elliptic curves $E$. We consider the elliptic curve $E(k):~x^3+k$ for some non-zero square-free integer $k$. Kihara [@KI] proved that there are infinitely many integers $k$ for which $E(k)$ has rank at least $7$. Chang [@CH] proved that $E(k)$ has rank zero for some values of $k$. On the other hand, the quadratic twist of $E(k)$ by a square-free integer $D$ is given by $E_D(k): y^2+kD^3$. Recently, Wu and Qin [@WQ] proved that $E_D(1)$ has rank zero when $D\equiv 3\pmod 4$ is a negative square-free integer satisfying some conditions on the selection of $D$, the class number of $\mathbb{Q}(\sqrt{2D})$ and solutions of associated Pell equation. In this paper, we prove that the ranks of $E_{-D}(2)$ and $E_{3D}(2)$ are zero for any square-free odd positive integer $D(\ne 3)$ by supplying some sufficient conditions on the class number of $\mathbb{Q}(\sqrt{-2D})$ and the coefficients of the fundamental unit of $\mathbb{Q}(\sqrt{6D})$. We also produce infinitely many positive integers $D$ for which $E_{-D}(2)$ and $E_{3D}(2)$ have rank zero. More precisely, we prove the following: \[thm\] Let $D(\ne 3)$ be a square-free odd positive integer satisfying the following conditions: - $3$ does not divide the class number of $\mathbb{Q}(\sqrt{-2D})$. - $3$ does not divide the coefficients of the fundamental unit of $\mathbb{Q}(\sqrt{6D})$. Then the ranks of $E_{-D}(2)$ and $E_{3D}(2)$ are zero. \[thmi\] Let $D\equiv 2\pmod 3$ be a square-free odd positive integer and $3$ does not divide the class number of $\mathbb{Q}(\sqrt{-2D})$, then the ranks of $E_{-D}(2)$ and $E_{3D}(2)$ are zero. Furthermore, there are infinitely many such $D$ for which the ranks of $E_{-D}(2)$ and $E_{3D}(2)$ are zero. Proof of Theorem \[thm\] ======================== We begin the following reflection theorem of Scholz [@SC] which will be needed to prove the next two propositions. \[thmsc\] Let $D>1$ be a square-free integer. Let $r$ and $s$ be the $3$-ranks of the class groups of the imaginary quadratic field $\mathbb{Q}(\sqrt{-D})$ and the real quadratic field $\mathbb{Q}(\sqrt{3D})$. Then $s\leq r\leq s+1$. We first give the proof of the following crucial result. \[prop1\] Let $D$ be as in Theorem \[thm\]. Then $$\#\{(x,y)\in E_{-D}(2)(\mathbb{Q}):ord_p(y)\leq 0~~\forall p\mid 6D,~ p\text{ prime}\}=0.$$ To prove this proposition, it is sufficient to show that the equation $$\label{eq2.1} y^2=x^3-2D^3z^6$$ has no integer solutions in $x,y,z$ with $\gcd(x, y, z)=1$, $\gcd(y,D)=1$ and $z\ne 0$. Without loss of generality, we assume that $(x,y,z)$ is an integer solution of such that $y$ and $z$ are positive as well as $z$ is minimal. We can exclude the cases where one or both of $x$ and $y$ are even since these cases would imply that $z$ is even too, which is a contradiction. Thus the only remaining possibility is that both $x$ and $y$ are odd. Since $D$ is square-free and $\gcd(y,D)=1$, so that $\gcd(x,2D)=1$ and $\gcd(y,2D)=1$. We now rewrite as follows: $$\label{eq2.2} (y+Dz^3\sqrt{-2D})(y-Dz^3\sqrt{-2D})=x^3.$$ Utilizing $\gcd(x,y,z)=1$ and $\gcd(x,2D)=1$, we observe that $\gcd(y+Dz^3\sqrt{-2D}, y-Dz^3\sqrt{-2D})=1$. Therefore gives $$\label{eq2.3} y+Dz^3\sqrt{-2D}=(a+b\sqrt{-2D})^3$$ for some integers $a$ and $b$ satisfying $\gcd(a,b)=1$ and $$\label{eq2.4} a^2+2b^2D=x.$$ This shows that $a$ is odd as $x$ is odd. Equating the real and imaginary part from , we get $$\label{eq2.5} y=a^3-2ab^2D,$$ $$\label{eq2.6} Dz^3=3a^2b-2b^3D.$$ Since $D\ne 3$ and it is square-free, so that reading modulo $D$, we get $ab\equiv 0\pmod D$. This implies either $a\equiv 0\pmod D$ or $b\equiv 0\pmod D$ as $\gcd(a,b)=1$. If $a\equiv 0\pmod D$, then utilizing we get $x\equiv 0\pmod D$. This contradicts to the fact $\gcd(x,D)=1$. Thus $b\equiv 0\pmod D$ and we write $b=Db_1$ for some integer $b_1$. Hence gives $$\label{eq2.7} z^3=b_1(3a^2-2b_1^2D^3).$$ Reading modulo $3$, we get $z\equiv bD\pmod 3$. Utilizing this in and then reading modulo $9$, we get $$b_1^3D^3\equiv 3a^2b_1-2b_1^3D^3\pmod 9.$$ If $3\mid D$ and $3\nmid b_1$, then $3\mid a$ and thus gives $3\mid y$ which contradicts to $\gcd(y,D)=1$. Therefore $3\mid b_1$ and we write $b_1=3b_2$ for some integer $b_2$. We utilize this in to get $3\mid z$, and put $z=3z_1$. Therefore takes the form: $$\label{eq2.8} 3z_1^3=b_2(a^2-6b_2^2D^3).$$ Since $\gcd(a,b)=1$ and $3\mid a$, so that gives $3\mid b_2$. We put $b_2=3b_3$ for some integer $b_3$. Thus becomes $$z_1^3=b_3(a^2-54b_3^2D^3).$$ It is clear that $\gcd(b_3, a^2-54b_3^2D^3)=1$ since $\gcd(a,b_3)=1$ due to $\gcd(a,b)=1$. Thus there exist two integers $A$ and $B$ such that $b_3=B^3$ and $a^2-54b_3^2D^3=A^3$. These together give rise to $$\label{eq2.9} a^2-54B^6D^3=A^3.$$ It is clear that $3\nmid A$; otherwise $3\mid a$ which is a contradiction as $3\mid b$ too but $\gcd(a,b)=1$. Since $a$ is odd, so that $A$ is odd too. Also if a prime $p\mid \gcd(a,A)$ then by , $p\mid D$ and thus by we get $p\mid x$ which contradicts to $\gcd(x, D)=1$. Therefore $\gcd(a, A)=1$. We now rewrite as follows: $$(a+3DB^3\sqrt{6D})(a-3DB^3\sqrt{6D})=A^3.$$ It is clear that $\gcd(a+3DB^3\sqrt{6D},a-3DB^3\sqrt{6D})=1$ since $\gcd(a,2A)=1$. Since the class number of $\mathbb{Q}(\sqrt{-2D})$ is not divisible by $3$, so that by Theorem \[thmsc\] the class number of $\mathbb{Q}(\sqrt{6D})$ is not divisible by $3$. Therefore, $$\label{eq2.10} a+3DB^3\sqrt{6D}=u(\alpha+\beta\sqrt{6D})^3,$$ where $u$ is unit in $\mathbb{Q}(\sqrt{6D})$ and $\alpha, \beta$ are integers such that $\gcd(\alpha, \beta)=1$ as $\gcd(a, 3BD)=1$. Since $6D\equiv 2\pmod 4$, so that the fundamental unit, $\varepsilon$ in $\mathbb{Q}(\sqrt{6D})$ is of the form $ \varepsilon=T+U\sqrt{6D}$. Therefore the only possibilities for $u$ are $1, \varepsilon$ and $\varepsilon^2$ since the higher powers of $\varepsilon$ can be absorbed in $(\alpha+\beta\sqrt{6D})^3$. We first assume that $u=1$. Then gives: $$\label{eq2.11} a=\alpha^3+18\alpha\beta^2D,$$ $$\label{eq2.12} DB^3=\alpha^2\beta+2\beta^3 D.$$ Since $D$ is square-free, so that shows that $D\mid \alpha\beta$. Clearly, $\gcd(D,\alpha)=1$; otherwise by $\gcd(D,\alpha)\mid a$ which contradicts to $\gcd(a,D)=1$. Thus $D\mid \beta$, and we put $\beta=D\beta_1$ for some rational integer $\beta_1$. Therefore becomes $$B^3=\beta_1(\alpha^2+2\beta_1^2D^3).$$ It is clear that $\gcd(\beta_1, \alpha^2+2\beta_1^2D^3)=1$ as $\gcd(\alpha, \beta_1)=1$. Therefore there exist two integers $\beta_2$ and $\alpha_1$ such that $\beta_1=\beta_2^3$ and $\alpha^2+2\beta_1^2D^3=\alpha_1^3$. These further imply $$\alpha^2=\alpha_1^3-2D^3\beta_2^6.$$ This shows that $(\alpha_1, \alpha, \beta_2)$ is another solution of $\eqref{eq2.1}$ with $\gcd(\alpha,D)=1$ and $\beta_2\ne 0$. Furthermore, $|\beta_2|<|\beta_1|^{\frac{1}{3}}<|B|<|b_3|^{\frac{1}{3}}<|\frac{b_2}{3}|^{\frac{1}{3}}<|\frac{b_1}{9}|^{\frac{1}{3}}$ and thus by $|\beta_2|<z$. This contradicts the minimality of $z$. We now consider the case where $u=\varepsilon$. Then gives the following: $$\label{eq2.13} a=T\alpha(\alpha^2+18\beta^2D)+18DU\beta(\alpha^2+2D\beta^2),$$ $$\label{eq2.14} 3DB^3=3T\beta(\alpha^2+6\beta^2D)+U\alpha(\alpha^2+18D\beta^2).$$ We read modulo $3$ to get $\alpha^3T\equiv a\pmod 3$. This imples $3\nmid\alpha$ and $3\nmid T$ as $3\nmid a$. Reading modulo $3$, we get $U\alpha^3\equiv 0\pmod 3$. This implies $3\mid U$ since $3\nmid \alpha$. This contradicts the assumption (II). Finally if $u=\varepsilon^2$, then gives $$\label{eq2.15} a=(T^2+6DU^2)(\alpha^3+18\alpha\beta^2D)+12DTU(3\alpha^2\beta+6D\beta^3),$$ $$\label{eq2.16} 3DB^3=(T^2+6DU^2)(3\alpha^2\beta+6D\beta^3)+2TU(\alpha^3+18\alpha\beta^2D).$$ Reading modulo $3$, we get $3\nmid T$ and $3\nmid \alpha$ since $3\nmid a$. We finally read to get $TU\alpha\equiv 0\pmod 3$ which implies $U\equiv 0\pmod 3$. This contradicts the assumption (II). \[prop2\] Let $D$ be as in Theorem \[thm\]. Then $$\#\{(x,y)\in E_{3D}(2)(\mathbb{Q}):ord_p(y)\leq 0~~\forall p\mid 6D,~ p\text{ prime}\}=0.$$ Analogous to the proof of Proposition \[prop1\], it is sufficient to prove that the equation $$\label{eq2.17} y^2=x^3+2(3D)^3z^6$$ has no integer solutions in $x,y,z$ with $\gcd(x,y,z)=1$, $\gcd(y,D)=1$ and $z\ne 0$. Without loss of generality, we assume that $(x,y,z)$ is an integer solution of such that $y$ and $z$ are positive as well as $z$ is minimal. We can exclude the cases where $\gcd(x,6)=1$ or $\gcd(y,6)=1$ as these cases would imply that $\gcd(x,y,z)\ne 1$. Thus the only remaining possibility is that both $x$ and $y$ are odd as well as $3\nmid xy$. Since $D$ is square-free, $\gcd(y,D)=1$ and $\gcd(x,y,z)=1$, so that $\gcd(x,6D)=1$ and $\gcd(y,6D)=1$. We now rewrite as $$\label{eq2.18} (y+3Dz^3\sqrt{6D})(y-3Dz^3\sqrt{6D})=x^3.$$ It is clear that $\gcd(y+3Dz^3\sqrt{6D}, y-3Dz^3\sqrt{6D})=1$ as $\gcd(x,y,z)=\gcd(x,2D)=1$. Since $3$ does not divide the class number of $\mathbb{Q}(\sqrt{-2D})$ so that by Theorem \[thmsc\], $3$ does not divide the class number of $\mathbb{Q}(\sqrt{6D})$. Therefore from we can write $$\label{eq2.19} y+3Dz^3\sqrt{6D}=u(a+b\sqrt{6D})^3,$$ where $u$ is unit in $\mathbb{Q}(\sqrt{6D})$ and $a, b$ are integers such that $\gcd(a, b)=1$ as $\gcd(y, 3Dz)=1$. Since $6D\equiv 2\pmod 4$, so that the fundamental unit in $\mathbb{Q}(\sqrt{6D})$ is of the form $ T+U\sqrt{6D}$. Therefore $u$ is given by $(T+U\sqrt{6D})^\delta$ with $\delta=0,1,2$ as the higher powers can be absorbed in $(a+b\sqrt{6D})^3$. First consider the case when $\delta=0$. Then implies $$\label{eq2.20} y=a^3+18ab^2D,$$ $$\label{eq2.21} Dz^3=a^2b+2b^3 D.$$ shows that $D\mid ab$ as $D$ is square-free. It is clear that $\gcd(D,a)=1$; otherwise by $\gcd(D,a)\mid y$ which contradicts to $\gcd(y,D)=1$. Hence $D\mid b$, and we write $b=Db_1$ for some integer $b_1$. Thus implies $$z^3=b_1(a^2+2b_1^2D^3).$$ Since $\gcd(a,b)=1$, so that $\gcd(a, b_1)=1$ and hence $\gcd(b_1, a^2+2b_1^2D^3)=1$. Therefore we can find two integers $B$ and $A$ satisfying $b_1=B^3$ and $a^2+2b_1^2D^3=A^3$. These further give rise to $$a^2=A^3-2D^3B^6.$$ This shows that $(A, a, B)$ is another solution of $\eqref{eq2.17}$ satisfying $\gcd(a,D)=1$ and $B\ne 0$. Moreover, $|B|<|b_1|^{\frac{1}{3}}<z$. This contradicts the minimality of $z$. We now consider the case where $\delta=1$. In this case, gives: $$\label{eq2.22} y=aT(a^2+18Db^2)+18DUb(a^2+2b^2D),$$ $$\label{eq2.23} 3Dz^3=3Tb(a^2+2Db^2)+Ua(a^2+18b^2D).$$ We read modulo $3$ to get $a^3T\equiv y\pmod 3$. This imples $3\nmid a$ and $3\nmid T$ as $3\nmid y$. Reading modulo $3$, we get $Ua^3\equiv 0\pmod 3$. This implies $3\mid U$ since $3\nmid a$ which contradicts the assumption (II). Finally if $\delta=2$, then provides $$\label{eq2.24} y=(T^2+6DU^2)(a^3+18ab^2D)+36DTUb(a^2+2Db^2),$$ $$\label{eq2.25} 3Dz^3=(T^2+6DU^2)(3a^2b+6Db^3)+2TUa(a^2+18b^2D).$$ Reading modulo $3$, we get $3\nmid aT$ since $3\nmid y$. We finally read to get $TUa\equiv 0\pmod 3$ which implies $U\equiv 0\pmod 3$. This contradicts the assumption (II). We also need the following two results in order to complete the proof of Theorem \[thm\]. The first result recalls from [@SI Ex. 10.19, p. 323]. \[lsi1\] For a sixth-power-free integer $m$, let $E(m):~ y^{2}=x^{3}+m$. Then $E(m)(\mathbb{Q})_{\text{tors}}|6$. More precisely, $$E(m)(\mathbb{Q})_{\text{tors}}\cong \begin{cases} \mathbb{Z}/6\mathbb{Z} & \text{ if } m=1, \\ \mathbb{Z}/ 3\mathbb{Z} & \text{ if } m\neq 1 \text{ is a cube}, \text{ or } m=-432, \\ \mathbb{Z}/ 2\mathbb{Z} & \text{ if } m\neq 1 \text{ is a square}, \\ 1& \text{ otherwise}. \end{cases}$$ The following result comes from [@SI p. 203]. \[lsi2\] Let $E(m)$ be as in Lemma \[lsi1\]. Let $P(x,y) \in E(m)$. Then $$\left(x([2]P),~ y([2]P)\right)=\left( \frac{9x^{4}-8y^{2}x}{4y^{2}},~ \frac{-27x^{6}+36y^{2}x^{3}-8y^{4}}{8y^{3}}\right).$$ We assume that $m=-2D^3,~ 6D^3$. Then $E_{-D}(2)$ and $E_{3D}(2)$ can be represented by $E(m)$. Thus utilizing Proposition \[prop1\] and Proposition \[prop2\], we can conclude that $$E(m)(\mathbb{Q})=\left\lbrace (x,y) \in \mathbb{Q}^2:~y^2=x^3+m,~\text{ ord}_p(y)\geq 1~\forall p\mid 3m, ~p\text{ prime} \right\rbrace.$$ Now in order to complete the proof, it suffices to show that $E(m)(\mathbb{Q})$ is finite. For a prime divisor $p$ of $3m$, $y^2=x^3+m$ provides the following: - $\text{ord}_p(y)\geq 1$ if and only if $\text{ord}_p(x)=1$ when $p\ne 3$. - $\text{ord}_3(y) \geq 1$ if and only if $ \text{ord}_3(x)=\begin{cases} 1 & \text{if } 3\mid m, \\ 0 & \text{if } 3\nmid m. \end{cases} $ Applying Lemma \[lsi1\], we obtain $E(m)(\mathbb{Q})_{tors} = O$.\ We assume on the contrary that $E(m)(\mathbb{Q}) \neq E(m)(\mathbb{Q})_{tors}$. Then we can find $P(x,y) \in E(m)(\mathbb{Q})\setminus E(m)(\mathbb{Q})_{tors}$ and a prime divisor $p$ of $3m$ such that $ord_p(y)\geq 1$. Applying Lemma \[lsi2\] and utilizing induction on $n$, one gets $$\text{ord}_p(y([2^{n}]P))\leq 0 \begin{cases} \forall n\geq 1 & \text{ if }p\ne 3,\\ \forall n\geq 2 & \text{ if }p= 3.\end{cases}$$ Assume that $m$ has $t$ distinct primes factors and we put $n=2^{t+1}$. Then for any prime factor $p$ of $3m$, one gets $\text{ord}_p(y([2^{n}]P)) \leq 0$. This is a contradiction. Thus we complete the proof. Proof of Theorem \[thmi\] ========================= We denote by $h(d)$ the class number of $\mathbb{Q}(\sqrt{d})$ for any square-free integer $d$. Ankeny, Artin and Chowla [@AAC Theorem II] (also see, [@HS]) gave the following congruence relationship relating the class numbers of real and imaginary quadratic fields. \[thmB\] Assume that $d = 3q$, where $q$ is square-free positive integer and $q\equiv 1\pmod 3$. Then $Th(-q)+Uh(d)\equiv 0\pmod 3$, where $T$ and $U$ are the coefficients of the fundamental unit of $\mathbb{Q}(\sqrt{d})$. To prove this theorem, it suffices to show that the assumption (II) in Theorem \[thm\] is not necessary in this case. Suppose that $\varepsilon=T+U\sqrt{6D}$ is the fundamental unit in $\mathbb{Q}(\sqrt{6D})$. We have already seen in the proof of Theorem \[thm\] that $T\not\equiv 0\pmod 3$. We utilize the assumption $h(-2D)\not\equiv 0\pmod 3$ and Theorem \[thmsc\] to obtain $h(6D)\not \equiv 0\pmod 3$. Since $D\equiv 2\pmod 3$, so that $2D\equiv 1 \pmod 3$. Also $2D$ is square-free as $D$ is odd and square-free. Therefore by Theorem \[thmB\], we get $Th(-2D)+ Uh(6D)\equiv 0\pmod 3$. This concludes that $U\not\equiv0\pmod 3$. We can say by the Cohen-Lenstra heuristics that there exist infinitely many square-free positive integers $D$ such that $D\equiv 2\pmod 3$ and $h(-2D)\not\equiv 0\pmod 3$. For each of these $D$, ranks of $E_{-D}(2)$ and $E_{3D}(2)$ are zero. This complete the proof. acknowledgement {#acknowledgement .unnumbered} =============== The author is grateful to Prof. Kalyan Chakraborty for his careful reading, helpful comments and suggestions. The author would like to thank Dr. Kalyan Banerjee for fruitful discussion on this manuscript. The author is supported by the SERB-NPDF (PDF/2017/001958), Govt. of India. [25]{} N. C. Ankeny, E. Artin and S. Chowla, [*The class number of real quadratic fields*]{}, Ann. Math. [**56**]{} (1952), 479–493. H. Cohen and H. W. Lenstra Jr., [*Heuristics on class groups of number fields*]{}, in: Number Theory, Noordwijkerhout, 1983, in: Lecture Notes in Math., [**1068**]{}, Springer, Berlin, 1984, 33–62. K. -L. Chang, [*One some Diophantine equations $y^{2}=x^{3}+k$ with no rational solutions*]{}, Quart. J. Math., Oxford Ser. [**19**]{} (1948), 181–188. F. Gouvea and B. Mazur, [*The square-free sieve and the rank of elliptic curves*]{}, J. Amer. Math. Soc. [**4**]{} (1991), no. 1, 1–23. A. Hoque and H. K. Saikia, [*A note on quadratic fields whose class numbers are divisible by $3$*]{}, SeMA J. [**73**]{} (2016), no. 1, 1–5. S. Kihara, [*On the rank of the elliptic curve $y^2=x^3+k$. II*]{}, Proc. Japan Acad. Ser. A Math. Sci. [**72**]{} (1996), no. 10, 228–229. L. Mai and M. R. Murty, [*A notion on quadratic twists of an elliptic curve*]{}, in: H. Kisilevsky, M. R. Murty (Eds.), Elliptic Curves and Related Topics, CRM [**4**]{} (1994), 121–124. K. Ono, [*Rank zero quadratic twists of modular elliptic curves*]{}, Compositio Math. [**104**]{} (1996), n0. 3, 293–304. A. Scholz, [*Über die Beziehung der Klassenzahl enquadratischer Körper zueinander*]{}, J. Reine Angew. Math. [**166**]{} (1932), 201–203. J. H. Silverman, [*The Arithmetic of Elliptic Curves*]{}, 2nd ed., Graduate Texts in Mathematics, [**106**]{}, Springer-Verlag, New York, 2009. I. R. Shafarevich and J. T. Tate, [*The rank of elliptic curves*]{}, Dokl. Akad. Nauk. USSR, [**175**]{} (1967), 770–773. X. Wu and Y. Qin, [*Rational points of elliptic curve $y^2=x^3+k^3$*]{}, Algebra Colloq. [**25**]{} (2018), no. 1, 133–138.

ArXiv

--- abstract: 'We define oscillating sequences which include the Möbius function in the number theory. We also define minimally mean attractable flows and minimally mean-L-stable flows. It is proved that all oscillating sequences are linearly disjoint from minimally mean attractable and minimally mean-L-stable flows. In particular, that is the case for the Möbius function. Several minimally mean attractable and minimally mean-L-stable flows are examined. These flows include the ones defined by all $p$-adic polynomials, all $p$-adic rational maps with good reduction, all automorphisms of $2$-torus with zero topological entropy, all diagonalized affine maps of $2$-torus with zero topological entropy, all Feigenbaum zero topological entropy flows, and all orientation-preserving circle homeomorphisms.' address: - | Fan Ai-Hua: LAMFA UMR 7352, CNRS\ Faculté des Sciences\ Université de Picardie Jules Verne\ 33, rue Saint Leu\ 80039 Amiens CEDEX 1, France - | Yunping Jiang: Department of Mathematics\ Queens College of the City University of New York\ Flushing, NY 11367-1597\ and\ Department of Mathematics\ Graduate School of the City University of New York\ 365 Fifth Avenue, New York, NY 10016 author: - Aihua Fan and Yunping Jiang title: 'Oscillating Sequences, Minimal Mean Attractability and Minimal Mean-Lyapunov-Stability' --- [^1] **Introduction** ================ Consider a pair $\mathcal{X} = (X,T)$, where $X$ is a compact metric space with metric $d (\cdot, \cdot)$ and $T: X\to X$ is a continuous map. We call $\mathcal{X}$ a [*flow*]{} or a [*dynamical system*]{} because we will consider iterations $\{ T^{n}\}_{n=0}^{\infty}$. A sequence of complex numbers $\xi(n)$, $n=1, 2,\cdots$, is said to be [*observed*]{} in $\mathcal{X}$ if there is a continuous function $f: X\to \mathbb{C}$ and a point $x \in X$ such that $\xi(n) = f(T^n x)$. Given a sequence of complex numbers ${\bf c}: =(c_{n})$, $n=1, 2, \cdots$. We say that it is [*linearly disjoint*]{} from the flow $\mathcal{X}$ if we have $$\label{disjointness} \lim_{N\to \infty} \frac{1}{N}\sum_{n=1}^N c_n \xi(n) =0$$ for any observable $\xi(n)$ in $\mathcal{X}$. An interesting sequence is the Möbius function $c_{n}:=\mu (n)$. Recall that by definition $\mu(1) =1$, $\mu(n)=(-1)^r$ if $n$ is a product of $r$ distinct primes, and $\mu(n)=0$ if $n$ is not square-free. Sarnak has conjectured that the Möbius function $\mu(n)$ is linearly disjoint from all flows with zero topological entropy (see [@Sa1; @Sa2]). This is now called Sarnak’s conjecture, which remains open in its generality although it is proved in several special cases [@Bou; @BSZ; @GT; @Kar; @LS; @MR]. As pointed out in [@Sa1; @Sa2], this conjecture has a connection with many important problems in number theory, for example, the Riemann hypothesis. More recently, in [@Ge], Ge studied Sarnak’s conjecture in association with a $C^*$ algebra and its maximal ideal space and in [@EKLR], which contains a good survey of the subject and a rather complete list of refereces, El. Abdalaoui et al. considered the Sarnak conjecture for sequences of numbers in $\{-1,0,1\}$ by comparing several natural generalizations. An important property satisfied by the Möbius function is $$~\label{pnt} \sum_{n=1}^{N} \mu (n) =o(N)$$ which is equivalent to the prime number theorem that the number of prime numbers less than or equal to $N$ is approximately $N/\ln N$. A stronger statement (conjecture) is that for $\epsilon >0$ $$~\label{pnt} \sum_{n=1}^{N} \mu (n) =O_{\epsilon}(N^{\frac{1}{2}+\epsilon})$$ which is equivalent to the Riemann hypothesis. Another property satisfied by the Möbius function is Davenport’s theorem [@Da] that $$\label{daven} \sum_{n=1}^{N} \mu (n) e^{2\pi i \alpha n} = o(N)$$ for any real number $0\leq \alpha <1$ and Hua’s theorem [@Hua] $$\label{hua} \sum_{n=1}^{N} \mu (n) e^{2\pi i \alpha n^{k}} = o(N)$$ for any real number $0\leq \alpha <1$ and any integer $k\geq 2$. The reader can refer to [@LZ] for more delicate estimates for these sums for $k\geq 1$. In this paper, we will not only consider the Möbius function but also a sequence of complex numbers ${\bf c}=(c_n)$ satisfying the oscillating condition $$\label{oscillating} \sum_{n=1}^{N} c_n e^{-2\pi i n t} =o_{t} (N), \quad \forall t \in [0, 1),$$ and the growth condition $$\label{growth} \sum_{n=1}^{N} |c_n|^{\lambda} =O(N)$$ for some $\lambda>1$ (such a sequence ${\bf c}$ will be called an oscillating sequence). There are many arithmetic functions, including the Möbius function and the Liouville function, which are oscillating sequences (see [@DD; @De]). Recall that the Liouville function $l(n)$ is defined as $l(n) =(-1)^{\Omega (n)}$ where $\Omega(n)$ is the number of prime factors of $n$, counted with multiplicity. On the other hand, we will define a minimally mean attractable flow and a minimally mean-L-stable flow. We will prove that all oscillating sequences are linearly disjoint from all minimally mean attractable and minimally mean-L-stable flows. We will prove that $p$-adic polynomial flows, $p$-adic rational flows with good reduction, automorphism flows with zero topological entropy on $2$-torus, all diagonalized affine flows of $2$-torus with zero topological entropy, all Feigenbaum zero topological entropy flows and all orientation-preserving circle homeomorphism flows are all minimally mean attractable and minimally mean-L-stable flows. We organize our paper as follows. In Section 2, we define oscillating sequences (Definition \[os\]). In Section 3, we define minimally mean-L-stable flows (Definition \[mmls\]) and minimal mean attractable flows (Definition \[mma\]). We will also review mean-L-stable flows and its equivalent notion, mean-equicontinuous flows. In Section 4, we prove that all oscillating sequences are linearly disjoint from all minimal mean attractable and minimally mean-L-stable flows (Theorem \[main1\]). In particular, as an example, we have that the Möbius function is linearly disjoint from all minimal mean attractable and minimally mean-L-stable flows (Corollary \[scmmammls\]). As another example, we have that all oscillating sequences are linearly disjoint from all equicontinuous flows (Corollary \[M2\]). In Section 5, we discuss $p$-adic polynomial flows and $p$-adic rational flows. All polynomials with $p$-adic integral coefficients define equicontinuous flows on the ring of $p$-adic integers. Thus they are minimal mean attractable and minimally mean-L-stable flows. All $p$-adic rational maps with good reduction define minimal mean attractable and minimally mean-L-stable flows on the projective line $\mathbb{P}^1(\mathbb{Q}_p)$. Therefore, all oscillating sequences are linearly disjoint from all these $p$-adic polynomial flows and $p$-adic rational flows (Corollary \[pp\] and Corollary \[rp\]). In Section 6, we explain how to use our method to study affine maps and automorphisms of $2$-torus with zero topological entropy. We first prove that in the diagonalized case, all oscillating sequences are linearly disjoint from all flows defined by affine maps of $2$-torus with zero topological entropy (Corollary \[diagcoro\]) since these flows are equicontinuous (Proposition \[diag\]). We then prove that all flows defined by automorphisms of $2$-torus with zero topological entropy are minimal mean attractable and minimally mean-L-stable flows \[nondiagthm\] (they are not equicontinuous). Thus all oscillating sequences are linearly disjoint from all flows defined by automorphisms of $2$-torus with zero topological entropy (Corollary \[nondiagcoro\]). Furthermore, we give an affine map of $2$-torus with zero topological entropy and an oscillating sequence such that this oscillating sequence is not linearly disjoint from the flow defined by this affine map. Notice that Liu and Sarnak has proved that the Möbius function is linearly disjoint from this example in [@LS]. We further note that in order to be linearly disjoint from all flows with zero topological entropy, an oscillating sequence should not only be oscillating in the first order but also be oscillating for any higher order (Remark \[higher\]) and the Möbius function has this property as shown by Hua in [@Hua] (see Equation (\[hua\])). In Section 7, we prove that all Feigenbaum zero topological entropy flows are minimal mean attractable and minimally mean-L-stable flows. Thus all oscillating sequences are linearly disjoint from all Feigenbaum zero topological entropy flows (Theorem \[main2\]). In Section 8, we prove that Denjoy counter-examples are minimal mean attractable but not equicontinuous even when restricted to their non-wandering sets, which are minimal subsets. However, we prove that Denjoy counter-examples are minimally mean-L-stable (Theorem \[ceq\]). In the proof of this theorem, we first prove that the flow defined by a Denjoy counter-example is minimally mean-L-stable when its non-wandering set has zero Lebesgue measure, and then we show that every Denjoy counter-example is conjugate to a Denjoy counter-example whose non-wandering set has zero Lebesgue measure (an obseravtion pointed out to us by Davit Karagulyan). Thus all oscillating sequences are linearly disjoint from all flows defined by Denjoy counter-examples (Corollary \[scdenjoy\]). The case of Möbius sequence was already considered by Karagulyan [@Kar] and now is one of the consequences of Theorems \[main1\] and \[ceq\] (Corollary \[Kar\]). [*Acknowledgement.*]{} We would like to thank Professors Hedi Daboussi, Liming Ge, Davit Karagulyan, Jianya Liu, Jörg Schmeling, Jie Wu, Xiangdong Ye and Enrique Pujals for sharing information and having many interesting discussions with us. We would also like to thank Professor Peter Sarnak for his encouragement comments on our initial version of this paper. The work was started when the first author was visiting Lund University and he would like to thank the Knut and Alice Wallenberg Foundation for its support. Both authors would like to thank the Academy of Mathematics and Systems Science at the Chinese Academy of Sciences for its hospitality when they visited there and worked on this project. **Oscillating Sequences** ========================= In classical analysis (see [@Kahane]), a sequence of complex numbers $(u_n)_{n\in \mathbb{Z}}$ is called a generalized almost periodic sequence if for any $t\in [0, 1)$ the following limit exists $$c(t) =\lim_{N\to \infty} \frac{1}{2N+1}\sum_{n=-N}^N u_n e^{- 2\pi i n t}.$$ Almost periodic and generalized almost periodic sequence were studied by Bohr, Hartman, et al. The spectrum of a generalized almost periodic sequence is the set of $t\in[0, 1)$ such that $c(t)\not=0$. In general, the spectrum is countable. We are interested in generalized almost periodic sequence with empty spectrum defined on positive integers. Thus we introduce the following definition.  \[os\] Let ${\bf c}:=(c_n)$, $n=1, 2, \cdots$, be a sequence of complex numbers. We say that ${\bf c}$ is an [*oscillating sequence*]{} if the Cesaro means $$\sigma_N(t):= \frac{1}{N}\sum_{n=1}^{N} c_n e^{-2\pi i n t}$$ converge to zero as $N$ tends to the infinity for every $0\leq t <1$. All oscillating sequences in this paper are also assumed satisfying the growth condition (\[growth\]). If we only assume that $\lim \sigma_N(\alpha)$ exists but is not necessarily zero for every $0\le t <1$, we denote $$\mathcal{Z}({\bf c}) = \left\{t \in [0,1)\;|\; \lim_{N\to \infty} \sigma_N(t)=0\right\}.$$ So, the statement that ${\bf c}$ is an oscillating sequence means $ \mathcal{Z}({\bf c})=[0, 1). $ The complementary of $\mathcal{Z}({\bf c})$ is defined to be the spectrum of [**c**]{}. Let us give several examples of oscillating sequences.  \[ex1\] The sequence of complex numbers $ {\bf c} =\big( e^{ 2\pi i n \alpha}\big) $ for a fixed $0\leq \alpha <1$ is not an oscillating sequence, because $ \mathcal{Z}({\bf c})= [0, 1)\setminus \{ \alpha\}. $ Note that the spectrum is one-point set $\{\alpha\}$.  \[ex2\] The sequence of complex numbers $ {\bf c}= \big( e^{2\pi i c n \log n}\big), \;\; c >0, $ is an oscillating sequence because the Cesaro means $$\sigma_N(t)= \frac{1}{N}\sum_{n=1}^N e^{2\pi i c n \log n} e^{-2\pi i n t} = O\Big(\frac{1}{\sqrt{N}}\Big)$$ uniformly on $0\leq t <1$ ([@Z Vol. 1, p. 197]).  \[ex3\] The sequence of complex numbers $ {\bf c} = \big( e^{2\pi i n^2 \alpha}\big) $ for any fixed irrational $\alpha$ is an oscillating sequence because the sequence $$\{n^2\alpha - n t \pmod{1} \}_{n=0}^{\infty}$$ is uniformly distributed on the unit interval $[0,1]$ for any fixed $0\leq t <1$. So the Cesaro means $$\sigma_N(t)= \frac{1}{N}\sum_{n=1}^N e^{2\pi i (n^2\alpha - n t)}=o_{t} (1).$$ \[n2\] Consider the sequence of complex numbers $ {\bf c} : = \big( e^{2\pi i n^2 \alpha}\big) $ where $0\leq \alpha =p/q<1$ is a rational number with $(p, q)=1$. The spectrum of ${\bf c}$ is the set of rational numbers $\frac{r}{s}$ such that $s|q$ and $$\sum_{k=0}^{q-1} e^{2\pi i (k^2p/q + k r/s)}\not=0.$$ We have $$\lim_{N\to \infty} \frac{1}{N} \sum_{n=1}^N e^{2\pi i n^2 \alpha} e^{-2\pi i n t} =\sum_{k=0}^{q-1} e^{2\pi i (k^2p/q + k t)} \lim_{N\to \infty} \frac{1}{N} \sum_{m: mq + k \le N} e^{2\pi i mq t}.$$ If $t$ is irrational or if $t=r/s$ with $s\not|q$, the last sum is bounded and the above limit is zero. Then such a $t$ is not in the spectrum. If $s|q$, we have $$\lim_{N\to \infty} \frac{1}{N} \sum_{n=1}^N e^{2\pi i n^2 \alpha} e^{-2\pi i n r/s} =\frac{1}{q}\sum_{k=0}^{q-1} e^{2\pi i (k^2p/q + k r/s)}.$$ It is easy to check that $$\mathcal{Z}(\{ e^{2\pi i n^2 \cdot \frac{1}{2}} \}) =[0, 1)\setminus \big\{ \frac{1}{2}\big\},$$ $$\mathcal{Z}(\{ e^{2\pi i n^2 \cdot \frac{1}{3}} \})= \mathcal{Z}(\{ e^{2\pi i n^2 \cdot \frac{2}{3}} \}) =[0, 1)\setminus \big\{ 0, \frac{1}{3}, \frac{2}{3}\big\},$$ $$\mathcal{Z}(\{ e^{2\pi i n^2 \cdot \frac{1}{4}} \}) = \mathcal{Z}(\{ e^{2\pi i n^2 \cdot \frac{3}{4}} \}) =[0, 1)\setminus \big\{0, \frac{1}{2}\big\}.$$ The following example is due to Davenport [@Da] (see also [@GT2]).  \[ex4\] The Möbius function $\big(\mu (n)\big)$ is an oscillating sequence. Actually a stronger result holds for the Cesaro means $$\sigma_N (t) =\frac{1}{N} \sum_{n=1}^{N} \mu(n) e^{- 2\pi i n t} = O\Big(\frac{1}{\log^h N}\Big)$$ for any $h>0$ where the estimate is uniform on $0\leq t <1$. Another example can be obtained from the following proposition. Let $(\xi_n)$ be a sequence of independent and identically distributed real random variables such that $$%\mathbb{E} e^{\lambda \xi_{n}}= \mathbb{E} e^{\lambda \xi_1}\le e^{\frac{\lambda^2}{2}}, \quad \forall \lambda \in \mathbb{R}. %\;\;\forall n\geq 1$$ Such a sequence is called a subnormal sequence.  \[ex5\] Suppose $(\xi_{n}$) is a subnormal sequence. Let $(u_n)$ be a sequence of real numbers such that $u_n = O(n^\tau)$ for some $0<\tau<1/2$. Then almost surely the random sequence ${\bf c}: = (u_n \xi_n)$ is an oscillating sequence. Actually, as a consequence of the Salem-Zygmund inequality on the uniform estimate of random trigonometric polynomials and as we refer to [@Kahane p.73], where the Rademacher random sequence was considered but the proof for subnormal sequences is the same, we have almost surely the Cesaro means $$\sigma_N (t) =\frac{1}{N} \sum_{n=1}^{N} u_n \xi_n e^{-2\pi i n t} = O\Big(\frac{\sqrt{\log N}}{N^{\frac{1}{2}-\tau}}\Big)$$ where the estimate is uniform on $0\leq t <1$. This proves the proposition. From [@FS] we can get other oscillating sequences similar to those in Proposition \[ex5\]. The next example is due to Daboussi-Delange [@DD p. 254].  \[ex6\] Let ${\bf c}=(c_n)$ be a multiplicative arithmetical function such that $|c_n|\le 1$ for all $n$. Then ${\bf c}$ is an oscillating sequence if and only if for any Dirichlet character $\chi$ and and real number $u$, we have $$\sum_{p } \frac{1}{p} \Big( 1 - \mbox{\rm Re} \ \big( \chi(p) f(p) p^{-i u} \big) \Big) =\infty$$ where the sum is taken over prime numbers $p$. This classical result of Daboussi and Delange is generalized to asymptotic orthogonality of multiplicative functions to “irrational ” nilsequences by N. Frantzikinakis and B. Host [@FH2014]. The oscillating sequences shares the following property that arithmetic subsequences of an oscillating sequence are oscillating. Suppose that ${\bf c}= (c_{n})$ is an oscillating sequence. Then for any $q\ge 2$ and any $r=1, 2, \cdots, q$ we have $$\label{c-mod-q} \lim_{N\to \infty} \frac{1}{N} \sum_{\stackrel{1\le n\le N}{n\equiv r\, (\!\!\!\!\!\!\mod q)}} c_{n}e^{-2\pi i t n} =0, \quad \forall t \in [0, 1).$$ The idea comes from [@DD]. The proof is based on the fact that $$\frac{1}{q}\sum_{j=1}^q e^{\frac{2\pi i j a}{q}} = 1 \ \mbox{\rm or}\ 0$$ according to $a \equiv 0$ or $\not\equiv 0$ ($\!\!\!\!\mod q$). It follows that $$\begin{aligned} \frac{1}{N} \sum_{\stackrel{1\le n\le N}{n\equiv r\, (\!\!\!\!\!\!\mod q)}} c_{n} e^{-2\pi i t n} &=& \frac{1}{N} \sum_{1\le n\le N} c_{n} e^{-2\pi i t n} \frac{1}{q}\sum_{j=1}^q e^{\frac{-2\pi i j (n-r)}{q}}\\ &=& \frac{1}{q}\sum_{j=1}^q e^{\frac{2\pi i jr}{q}} \frac{1}{N} \sum_{1\le n\le N} c_{n} e^{-2\pi i \big(t+\frac{j}{q}\big) n} \to 0 \end{aligned}$$ as $N\to \infty$, since ${\bf c}$ is an oscillating sequence. The proof shows that for (\[c-mod-q\]) to hold for a fixed $t$ it suffices that $t+ j/q\in \mathcal{Z}({\bf c})$ for all $j=1, 2, \cdots, q$. In [@KS], Kahane and Saias studied completely multiplicative functions with zero sum. These functions share another kind of oscillating property. **Minimal Mean Attractability and Minimal Mean-L-Stabity** ========================================================== In this section we define a class of flows from which all oscillating sequences will be proved to be linearly disjoint. We start with recalling the classical definition of equicontinuous flows.  \[eqcn\] A flow $\mathcal{X}=(X,T)$ is said to be [*equicontinuous*]{} if the family $\{T^n\}_{n=0}^{\infty}$ is equicontinuous. That is to say, for every $\epsilon > 0$ there is a $\delta > 0$ such that whenever $x, y \in X$ with $d(x, y) <\delta$, we have $d(T^n x,T^n y) <\epsilon$ for $n = 0,1,2, \cdots$. It is well known that a flow $\mathcal{X}=(X,T)$ with $T$ being surjective is equicontinuous if and only if there exists a compatible metric $\rho$ on $X$ such that $T$ acts on $X$ as an isometry, i.e., $\rho(T x,Ty)=\rho (x, y)$ for all $ x, y \in X$. Thus, if $T$ is surjective and equicontinuous, then $T$ must be a homeomorphism. Moreover, it is also known that for an equicontinuous flow $\mathcal{X}$, if $T$ is a homeomorphism, then $T$ acts as a minimal system on the closure $\overline{O(x)}$ of the forward orbit $O(x)=\{T^{n}\}_{n=0}^{\infty}$ of any $x\in X$ (see  [@Pe]). Thus, for an equicontinuous flow $\mathcal{X}$, when $T$ is a transitive homeomorphism, it is always conjugate to a minimal rotation on a compact abelian metric group. In this case, let $m$ be the unique Haar probability measure on $X$, then the measurable dynamical system $(X,T,m)$ has discrete spectrum. An equicontinuous flow $\mathcal{X}$ such that $T$ is a homeomorphism can be decomposed into minimal subsystems. Sarnak’s conjecture holds for such a flow. Actually, we will prove that Sarnak’s conjecture holds for a much larger class of flows which we will call minimally mean attractable and minimally mean-L-stable flows. We give a more detailed description below. We first weaken the equicontinuity condition following Fomin in [@Fomin]. Suppose $\mathbb{N}=\{ 1,2, \cdots, \}$ is the set of natural numbers. Let $E$ be a subset of $\mathbb{N}$. The upper density of $E$ is, by definition, $$\overline{D}(E) = \limsup_{n\to\infty} \frac{\sharp (E\cap[1, n])}{n}.$$  \[mma\] A flow $\mathcal{X}=(X,T)$ is said to be [*mean-L-stable*]{} ([**MLS**]{} for short) (here $L$ recalls the sense of Lyapunov) if for every $\epsilon > 0$, there is a $\delta > 0$ such that $d(x, y) <\delta$ implies $d(T^nx,T^ny) <\epsilon$ for all $n=0, 1, 2, \cdots$ except a subset of natural numbers with the upper density less than $\epsilon$. Fomin proved that a minimal and [**MLS**]{} flow is uniquely ergodic in [@Fomin]. Oxtoby proved that a transitive and [**MLS**]{} flow is uniquely ergodic in [@Oxtoby1952]. Li, Tu and Ye further proved that every ergodic invariant measure on a [**MLS**]{} flow has discrete spectrum. Any flow $\mathcal{X}$ admits a minimal sub-flow, for example, the restriction on the $\omega$-limit set of any point.  \[mmls\] We say that a flow $\mathcal{X}=(X, T)$ is [*minimally [**MLS**]{}*]{} ([**MMLS**]{} for short) if for every minimal subset $K\subseteq X$, the sub-flow $\mathcal{K}=(K, T)$ is [**MLS**]{}. Following [@LTY], we say the flow ${\mathcal X}$ is [*mean-equicontinuous*]{} at a point $x \in X$ if for every $\epsilon > 0$, there is $\delta > 0$ such that for every $y \in X$ with $d(y, x)<\delta$ we have $$\label{MEC} \limsup_{n\to\infty} \frac{1}{n} \sum_{k=0}^{n-1} d(T^kx, T^k y) <\epsilon.$$ We say that the flow ${\mathcal X}$ is mean-equicontinuous if it is mean-equicontinuous at every point $x\in X$. That is, by the compactness of $X$, the above inequality (\[MEC\]) holds for all $x, y\in X$ such that $d(y, x)<\delta$. The following proposition is proved by Li, Tu, and Ye in [@LTY]  \[eq\] A flow ${\mathcal X}$ is mean-equicontinuous if and only if it is [**MLS**]{}. Let $C(X)$ be the space of all continuous functions $f: X\to \mathbb{C}$ with maximum norm $$\|f\|_{\infty} =\max_{x\in X} |f(x)|.$$  \[meaneq\] Suppose the flow ${\mathcal X}$ is [**MLS**]{}. Suppose ${\bf c}=(c_n)$ is a sequence satisfying the growth condition (\[growth\]). Then for any continuous function $f\in C(X)$, $$S_Nf(x) =\frac{1}{N} \sum_{n=1}^{N} c_n f(T^n x), \quad (N=1, 2, \cdots)$$ is an equicontinuous sequence in $C(X)$. Suppose $\|f\|_{\infty}\not= 0$, otherwise it is trivial. Take a $\lambda > 1$ satisfying (\[growth\]). Suppose $\lambda'> 1$ be the number such that $ \frac{1}{\lambda} + \frac{1}{\lambda'}=1. $ There is a positive constant $C$ (we can take $C$ to be the supremum of $(\frac{1}{N} \sum_{n=1}^{N} |c_{n}|^{\lambda} )^{1/\lambda}$) such that for any $x, y\in X$, we have $$~\label{Holder1} |S_Nf(x) - S_N f(y)| \leq C \Big(\frac{1}{N} \sum_{n=0}^{N-1} |f(T^n x) - f(T^n y)|^{\lambda'}\Big)^{1/\lambda'}.$$ This is a direct consequence of the Hölder inequality and (\[growth\]). By the uniform continuity of $f$, for any $\epsilon >0$ there exists $\eta >0$ such that $$\label{uniform_con} d(x, y)< \eta \Rightarrow |f(x)- f(y)|<\frac{\epsilon}{2 C}.$$ We can take $\eta \le \frac{\epsilon}{4C\|f\|_\infty}$. Since $\mathcal{X}$ is [**MLS**]{}, there is a $\delta' > 0$ such that $x, y\in X$ with $d(x, y) <\delta'$ implies $d(T^nx,T^ny) <\eta$ for all positive integer $n$ except a set, denoted $E$, of upper density less than $\eta$. There exists an integer $N^*$ such that $$\frac{\sharp (E \cap [1, N])}{N} < \eta, \qquad \forall N \ge N^*$$ Thus by (\[Holder1\]), if $N\ge N^*$ and if $ d(x, y)<\delta'$ we have $$\label{difference*} |S_Nf(x) - S_N f(y)| < C \big(2 \|f\|_\infty \eta + \frac{\epsilon}{2C}\big)= \epsilon.$$ In the above inequality, we first split the sum in (\[Holder1\]) into two sums according to $d(T^nx, T^n y)\ge \eta$ or $<\eta$. The finite many family $\{S_1f , \cdots, S_{N^*}f\}$ being equicontinuous, there exists $\delta''$ such that $$\label{difference**} d(x, y)<\delta'' \Rightarrow \max_{1\le N\le N^*} |S_Nf(x) -S_Nf(y)|<\epsilon.$$ Finally we conclude the proposition from (\[difference\*\]) and (\[difference\*\*\]). In [@Auslander p. 575], Auslander had decomposed a [**MLS**]{} flow into a star closed decomposition. In this decomposition each component contains precisely one minimal set and all invariant measures concentrated on minimal sets. This implies that from measure-theoretical point of view, the union of these minimal sets is an attractor: all points outside the union of these minimal sets are attracted into minimal sets eventually. We will develop this idea into a new concept called the minimal mean attractablity.  \[mma\] Suppose $\mathcal{X}=(X,T)$ is a flow. Suppose $K$ is a closed subset of $X$ and $T: K\to K$ is minimal. We say $x\in X$ is [*mean attracted*]{} to $K$ if for any $\epsilon >0$ there is a point $z=z_{\epsilon, x}\in K$ (depending on $x$ and $\epsilon$) such that $$\label{MA} \limsup_{N\to\infty} \frac{1}{N} \sum_{n=1}^{N} d(T^n x, T^n z) <\epsilon.$$ The [*basin of attraction*]{} of $K$, denoted $\hbox{\rm Basin}(K)$, is defined to be the set of all points $x$ which are mean attracted to $K$. It is trivial that $K \subset \hbox{\rm Basin}(K)$. We call $\mathcal{X}$ [*minimally mean attractable*]{} ([**MMA**]{} for short) if $$\label{Decomp} X= \bigcup_{K} \mbox{\rm Basin}(K)$$ where $K$ varies among all minimal subsets of $X$. Recall that a point $x\in X$ is attracted to $K$ if $$\lim_{n\to \infty} d(T^n x, K)=0.$$ In general, this can not imply to that $x$ is mean attracted to $K$. However, if $K$ is a periodic cycle, it does. We have even the following proposition. Let $x\in X$. Suppose that $\omega(x) =\{x_0,x_1, \cdots, x_{k-1}\}$ is a periodic cycle of period $k\geq 1$. Then for any $\epsilon >0$ sufficiently small, there exists an integer $\tau \ge 0$ and $z\in \omega(x)$ such that $$d(T^{\tau+m}x, T^{\tau +m} z) <\epsilon, \quad \forall m\ge 0.$$ Let $K=\omega(x)$. Take any $\epsilon >0$ sufficient small so that the balls $B(x_j, \epsilon)$, $j=0,1,\cdots, k-1$, are disjoint. For any $j$ and any $\tau \ge 0$, consider the subset of the natural numbers $$R_\tau(x_j) =\{n\ge \tau: T^n x \in B(x_j, \epsilon)\}.$$ For any fixed $0\leq j<k$, $$\lim_{R_\tau(x_j) \ni n \to \infty} T^n x = x_j.$$ By the continuity of $T$, we have $$\lim_{R_\tau(x_j)+1 \ni n \to \infty} T^n x = x_{j+1 \!\!\!\!\pmod{k}}.$$ This implies that we can find a $\tau_{j}\geq 1$ such that $$\label{recurrence} R_\tau (x_j)+1 \subset R_\tau(x_{j+1}), \quad \forall \tau\geq \tau_{j}.$$ Let $\tau_{\max}=\max_{0\leq j<k} \tau_{j}$. Take $\tau \geq \tau_{\max} \in R_{\tau_{\max}} (x_{0})$. Then $$d(T^\tau x , x_0) <\epsilon$$ and from (\[recurrence\]), $$d(T^{m+\tau}x, T^m x_0) = d(T^{m+\tau}x, x_{m \!\!\!\! \pmod{k}})<\epsilon, \quad \forall m\geq 0.$$ We take $z\in K$ such that $T^\tau z =x_0$, then we have that $$d(T^{\tau+m}x, T^{\tau+m} z) = d(T^{\tau+m}x, x_{m \pmod{k}})<\epsilon, \quad \forall m\geq 0.$$  \[1ex\] The flow $$([-1, 1], T(x)=-x^{2})$$ is [**MMA**]{} and [**MMLS**]{}. The point $-1$ and $0$ are only two fixed points. Since $T'(-1)=-2$ and $T'(0)=0$, $-1$ is the repelling fixed point and $0$ is the attracting fixed point. We have that $$\mbox{\rm Basin}(\{-1\}) = \{-1, 1\}, \quad \hbox{and}\quad \mbox{\rm Basin}(\{0\}) = (-1, 1).$$ Thus we have $$[-1,1]= {\rm Basin}(\{-1\})\cup {\rm Basin}(\{0\}).$$ This implies that $\mathcal{X}$ is [**MMA**]{} and [**MMLS**]{}. This flow is actually equicontinuous on any closed interval $[-a,a]\subset (-1,1)$. The next example is different (see Remark \[neq\]). Let $\mathbb{T}^{2}=\mathbb{R}^{2}/\mathbb{Z}^{2}$ be the standard 2-torus. Consider the automorphism $$T_{t}(x,y) = (x+ty,y): \mathbb{T}^{2}\to \mathbb{T}^{2}$$ for a fixed integer $t$.  \[2ex\] The flow $$\mathcal{X}= (\mathbb{T}^{2}, T_{t})$$ is [**MMA**]{} and [**MMLS**]{}. Actually we have that all automorphisms on $\mathbb{T}^2$ with zero topological entropy are [**MMA**]{} and [**MMLS**]{}. Let $\mathbb{T}=\mathbb{R}/\mathbb{Z}$ be the circle. Every circle $\mathbb{T}_{y}=\mathbb{T}\times \{y\}$ ($y$ being fixed) is $T_{t}$-invariant. The restriction of $T_{t}$ on each $\mathbb{T}_{y}$ is a one-dimensional rotation with the rotation number $ty$. If $ty$ is irrational, then $T_{t}: \mathbb{T}_{y}\to \mathbb{T}_{y}$ is minimal. If $ty$ is rational, then every point $(x,y)\in \mathbb{T}_{y}$ is a periodic point. Then the restriction of $T_{t}$ on each of these periodic orbits is minimal. In this case, $\mathbb{T}_{y} = \cup_{(x,y)\in \mathbb{T}_{y}} Orb ((x,y))$. Furthermore, we have that the minimal decompsition $$\mathbb{T}^{2}=\Big( \cup_{ty: irrational} \mathbb{T}_{y}\Big) \cup \Big( \cup_{ty: rational}\cup_{(x,y)\in \mathbb{T}_{y}} Orb\big( (x,y)\big)\Big) .$$ So this flow is [**MMA**]{}. Since $T_{t}$ on each $\mathbb{T}_{y}$ is isometric, so it is equicontinuous on each of its minimal subset. So it is [**MMLS**]{}. As we shall see in Section 6, any automorphism $T$ on $\mathbb{T}^2$ with zero topological entropy is (linearly) topologically conjugate to a $T_{t}$ for some integer $t$. So the flow $\mathcal{X}=(\mathbb{T}^{2}, T)$ is also [**MMA**]{} and [**MMLS**]{}.  \[neq\] The flow $(\mathbb{T}^{2}, T_{t})$ is not mean-equicontinuous (equivalently, not [**MLS**]{}) because the rotation number of $T_{t}|_{\mathbb{T}_{y}}$ varies with $y$. **Disjointness** ================ Recall that $\mathcal{X}=(X,T)$ is a flow if $X$ is a compact metric space with metric $d$ and $T:X\to X$ is a continuous map. A sequence $(\xi(n))$ is said to be [*observable*]{} in $\mathcal{X}$ if there exists a continuous function $f\in C(X)$ and a point $x\in X$ such that $\xi(n) = f(T^n x)$ for all $n\ge 1$. Following the idea of Sarnak [@Sa1], we say that a given sequence of complex numbers ${\bf c}=(c_n)$ is [*linearly disjoint*]{} from $\mathcal{X}$ if for any observable $(\xi(n))$ in $\mathcal{X}$, we have $$~\label{ld} \lim_{N \to \infty} \frac{1}{N}\sum_{n=1}^{N} c_n \xi(n)=0.$$ The first main result in this paper is the following one.  \[main1\] Any oscillating sequence ${\bf c}=(c_n)$ satisfying the growth condition (\[growth\]) is linearly disjoint from all [**MMA**]{} and [**MMLS**]{} flows $\mathcal{X}=(X, T)$. More precisely, for any continuous function $f\in C(X)$, we have that $$\lim_{N \to \infty} \frac{1}{N}\sum_{n=1}^{N} c_n f(T^n x)=0$$ for every $x\in X$ and the limit is uniform on each minimal subset. By the hypothesis, $X$ has the decomposition in Equality (\[Decomp\]). So, we need only prove that for any minimal subset $K\subset X$, we have $$\label{limsup-K} \forall x\in \mbox{\rm Basin}(K), \quad \lim_{N \to \infty} \frac{1}{N}\sum_{n=1}^{N} c_n f(T^n x)= 0.$$ For this purpose, we define $$S_N f(x) =\frac{1}{N}\sum_{n=1}^{N} c_n f(T^{n}x), \quad N=1, 2, \cdots.$$ We first prove that the sequence $\mathcal{S}= \{ S_{N} f(x)\}_{N=1}^{\infty}$ converges to $0$ uniformly on $K$. The sequence $\mathcal{S}$ is uniformly bounded in $C(K)$ by $C \|f\|_{\infty}$ for some constant $C>0$ and is equicontinuous by Proposition \[meaneq\]. So, it is precompact in $C(K)$ by the Arzela-Ascoli theorem. Therefore, every subsequence of $\mathcal{S}$ has a convergent subsequence. What we have to prove is the claim that every convergent subsequence converges uniformly on $K$ to zero. Let us prove this claim by contradiction. Suppose that some subsequence $S_{N_j} f$ converges uniformly to a continuous function $g \in C(K)$ which is not identically zero. Assume $|g(x_0)|>0$ for some $x_0\in K$ and we can actually assume that $$\label{contradiction2} |g(x)|>0, \ \ \ \forall x\in B(x_0, r)=\{x\in K\;|\; d(x, x_{0}) <r\}$$ for some $r>0$ by the continuity of $g$. On the other hand, we consider the probability measures on $K$: $$\nu_{N} =\frac{1}{N}\sum_{n=1}^{N} \delta_{T^n x_0}$$ where $\delta_y$ denote the Dirac measure concentrated at $y$. Let $\nu$ be a weak limit of a subsequence of the sequence $(\nu_{N})$. Then it is $T$-invariant. Since $K$ is minimal, by the Gottschalk theorem (see [@Gottschalk1944 Lemma 4]), $x_0$ is almost periodic in the sense that for every $\epsilon >0$, the set of $n$ such that $d(x_0, T^n x_0)<\epsilon$ is relatively dense. This implies that $\nu(B(x_0, \epsilon))>0$ for any $\epsilon >0$ (the support of $\nu$ is actually whole minimal set $K$). Now consider the measure-preserving dynamical system $$(K, \mathcal{B}(K), T, \nu)$$ where $\mathcal{B}(K)$ is the Borel $\sigma$-field of $K$. Let $\sigma_f$ be the spectral measure of $f$ with respect to the $T$-invariant probability measure $\nu$ on $K$, $f$ being considered as in $L^2(\nu)$ and $\sigma_f$ being defined on the unite circle $$S^{1}=\{ z\in {\mathbb C}\;|\; |z|=1\}.$$ By the spectral lemma (see [@Krengel p. 94-95]), we have $$\|S_{N}f\|^2_{L^2(\nu)} =\int_{\mathbb T} \Big| \frac{1}{N}\sum_{n=1}^{N} c_n e^{2\pi i n t}\Big|^2 d\sigma_f(t).$$ Since ${\bf c}=(c_{n})$ is an oscillating sequence, by the Lebesgue dominated convergence theorem, the right hand of the above equation converges to $0$. This implies that $S_Nf$ converges to zero in $L^2(\nu)$-norm. In particular, $S_{N_j}f$ converges to zero in $L^2(\nu)$-norm. Consequently, there is a subsequence $N_j'$ of $N_j$ such that $S_{N_j'}f$ converges to zero $\nu$-almost everywhere, which contradicts (\[contradiction2\]) and the fact that $\nu(B(x_0, r))>0$. The contradiction implies that the sequence $\mathcal{S}$ converges uniformly on $K$ to $0$. We have just proved the limit (\[limsup-K\]) for $x\in K$. Now we are going to prove (\[limsup-K\]) for a general point $x \in \mbox{\rm Basin}(K)$. Let $\eta >0$ be an arbitrarily small number. By the uniform continuity of $f$, there is a $\epsilon >0$ such that $d(u, v)<\epsilon$ implies $|f(u) - f(v)| <\eta$. We assume that $2\|f\|_\infty \epsilon \le \eta$. For $x\in \mbox{\rm Basin}(K)$, by the definition of $\mbox{\rm Basin}(K)$, there exists $z=z_{\epsilon,x}\in K$ such that $$\label{MA2} \limsup_{N\to\infty} \frac{1}{N} \sum_{n=1}^{N} d(T^n x, T^n z) <\epsilon^2.$$ which implies that $\overline{D}(E)\le \epsilon$ where $E=\{n\ge 0: d(T^n x, T^n z)\ge \epsilon\}$, because $$\epsilon \sharp (E\cap [1, N]) \le \sum_{n=1}^{N} d(T^n x, T^n z).$$ Write $$S_Nf(x) = S_Nf(x) - S_Nf(z) + S_Nf(z).$$ As $z\in K$, we have proved that $S_Nf(z)$ tends to zero. On the other hand, following the exact same proof of Proposition \[meaneq\], we show that $$\label{difference2} |S_Nf(x) - S_N f(z)| \leq C (2 \|f\|_\infty \epsilon + \eta)\le 2 C \eta.$$ Since $\eta>0$ is arbitrary, we have thus proved (\[limsup-K\]) for every $x\in \mbox{\rm Basin}(K)$. A special case of Theorem \[main1\] is that  \[scmmammls\] The Möbius function is linearly disjoint from all [**MMA**]{} and [**MMLS**]{} flows. This is because the Möbius function gives an oscillating sequence due to Davenport’s theorem (see Example \[ex4\]). Another consequence of Theorem \[main1\] is that \[M2\] Any oscillating sequence satisfying the growth condition (\[growth\]) is linearly disjoint from all equicontinuous flows. Let $\mathcal{X}=(X, T)$ be an equicontinuous flow. The sequence of compact sets $\{ T^n X\}_{n=1}^{\infty}$ decrease to a non-empty compact set, which we denote by $X_\infty$. The sub-flow $T: X_\infty \to X_\infty$ is surjective and equicontinuous and thus a homeomorphism. The set $X_\infty$ is decomposed into minimal subsets. We claim that each point $x\in X$ is mean attracted to some minimal set. In fact, the sequence $T^nx$ admits a subsequence $T^{n_j} x$ converging to a limit point $x_\infty$, which must belong to $X_\infty$. Recall the definition of equicontinuity for $T$: for any $\epsilon >0$ there exists $\eta >0$ such that $$d(u, v)<\eta \Rightarrow d(T^n u, T^n v) <\frac{\epsilon}{2} \ \ \ (n=1, 2, \cdots).$$ Take $j_0$ sufficiently large such that $d(T^{n_{j_0}} x, x_\infty) < \eta$. As $T: X_\infty \to X_\infty$ is a homeomorphism, there exist $z\in X_\infty$ such that $x_\infty = T^{n_{j_0}} z$ so $$d(T^{n_{j_0}} x, T^{n_{j_0}} z) < \eta.$$ By the equicontinuity, we have $$d(T^{n_{j_0} +n} x, T^{n_{j_0} +n} z) <\frac{\epsilon}{2} \ \ \ (n=1, 2, \cdots).$$ It follows that $$\limsup_{N\to \infty} \frac{1}{N}\sum_{n=1}^{N} d(T^n x, T^n z)\le \frac{\epsilon}{2} < \epsilon.$$ Thus we have proved that $x$ is mean attracted to $$K:=\overline{O(z)}=\overline{O(x_\infty)},$$ where $O(\cdot)$ means the forward orbit. So we have proved that this flow is [**MMA**]{}. But it is equicontinuous on each minimal subset, so it is also [**MMLS**]{}.  \[R1b\] As one can see in the proof of Theorem \[main1\], we don’t really need that the spectrum of the sequence ${\bf c}=(c_{n})$ is empty. What we really need is that the spectrum of ${\bf c}$ as a subset in the circle $S^1$ is disjoint from the support of $\sigma_f$ in the circle $S^1$. In other words, the oscillating condition on ${\bf c}$ in Theorem \[main1\] can be relaxed to that $$\mathcal{Z}({\bf c}) \subseteq \hbox{supp}(\sigma_f) .$$ As a direct consequence of the proof of Theorem \[main1\] and Remark \[R1b\], we have that  \[M2\] Suppose $S^{1} =\{ z\in \mathbb{C}\;|\; |z|=1\}$ is the unit circle. Let $R_\alpha (z) =e^{2\pi i \alpha} z$ be the rigid rotation where $\alpha \in (0, 1)$ is irrational. Any sequence ${\bf c}=(c_{n})$ satisfying the growth condition (\[growth\]) such that $\alpha \mathbb{Z}\subset \mathcal{Z}({\bf c})$ is linearly disjoint from the flow $\mathcal{X}=(S^{1}, R_\alpha)$. We use the notation in the proof of Theorem \[main1\]. The rotation $R_\alpha$ admits the Lebesgue measure as the unique invariant measure and the support of $\sigma_f$ is contained in $\alpha \mathbb{Z}$. Actually it is easy to check that $$\sigma_f = \sum_{n\in \mathbb{Z}} |\widehat{f}(n)|^2 \delta_{e^{2\pi i\alpha n}}.$$ \[R1\] The idea of using the spectral lemma in the proof of Theorem \[main1\] comes from [@FS1] where the ergodic Hilbert transforms associated to irrational rotations are studied. We will apply Theorem \[main1\] to several different flows. In Section 5, we apply to all $p$-adic polynomial flows and $p$-adic rational flows. In Section 6, we apply to all automorphisms and some affine maps of $2$-torus with zero topological entropy. In Section 7, we apply to all Feigenbaum zero topological entropy flows. In Section 8, we apply to all orientation-preserving irrational circle homeomorphisms. **$p$-Adic Polynomial Flows and $p$-Adic Rational Flows** ========================================================= We give here some examples of equicontinuous flows in the fields of $p$-adic numbers. These flows share a very nice minimal decomposition ([@FL11], [@FFLW]). We can compare this decomposition with Auslander’s star closed decomposition [@Auslander]. Let $p\ge 2$ be a prime number and let $\mathbb{Z}_p$ be the ring of $p$-adic integers. Let $\mathbb{Z}_p[x]$ be the ring of all polynomials with coefficients in $\mathbb{Z}_p$. Then every polynomial $P\in \mathbb{Z}_p[x]$ defines a flow $\mathcal{X}=(\mathbb{Z}_p, P)$. It is trivial that $P$ is $1$-Lipschitz function, i.e. $$|P(x) - P(y)|_p\le |x - y|_{p} \quad (\forall x, y\in \mathbb{Z}_p)$$ where $|\cdot|_p$ is the $p$-adic norm on $\mathbb{Z}_p$. So, the flow $\mathcal{X}=(\mathbb{Z}_p, P)$ is equicontinuous. In particular, the adding machine $add (x)=x +1 \in \mathbb{Z}_p[x]$ is a special case and it is actually an equicontinuous homeomorphism. Corollary \[M2\] gives us the following result.  \[pp\] Any oscillating sequence ${\bf c} =(c_{n})$ satisfying the growth condition (\[growth\]) is linearly disjoint from all $p$-adic polynomial flows $\mathcal{X}=(\mathbb{Z}_p, P)$ for $P\in \mathbb{Z}_p[x]$. It was proved in [@FL11] that for a polynomial flow $\mathcal{X}=(\mathbb{Z}_p, P)$ as in Corollary \[pp\], we have the minimal decomposition $$\mathbb{Z}_p = \mathcal{P} \sqcup\mathcal{M}\sqcup \mathcal{B}$$ where $ \mathcal{P} $ is the finite set consisting of all periodic points of $P$, $ \mathcal{M} = \bigsqcup_i \mathcal{M} _i$ is the union of all (at most countably many) clopen invariant sets such that each $\mathcal{M}_i$ is a finite union of balls and each subsystem $P: \mathcal{M}_i \to \mathcal{M}_i$ is minimal, and points in $\mathcal{B}$ lie in the attracting basin of a periodic orbit or of a minimal sub-flow. As we proved in Corollary \[M2\], points in $\mathcal{B}$ are mean attracted to a periodic cycle or a minimal set $M_i$. More generally, we can consider $p$-adic rational flow in $p$-adic field. Let $\mathbb{Q}_p$ be the field of $p$-adic numbers. Any point in the projective line $\P$ may be given in homogeneous coordinates by a pair $[x_1 : x_2]$ of points in $\Qp$ which are not both zero. Two such pairs $[x_1 : x_2]$ and $[\lambda x_1 : \lambda x_2]$ with nonzero factor $\lambda \in \Q_p^*$ are identified. The field $\Qp$ may be identified with the subset of $\P$ given by $$\left\{[x : 1] \in \mathbb P^1(\Qp) \mid x \in \Qp\right\}.$$ This subset covers all points in $\P$ except one: the point of infinity, which may be given as $\infty = [1 : 0].$ The spherical metric defined on $\P$ is analogous to the standard spherical metric on the Riemann sphere. If $u=[x_1,y_1]$ and $v=[x_2,y_2]$ are two points in $\P$, we define $$\rho(u,v)=\frac{|x_1y_2-x_2y_1|_p}{\max\{|x_1|_{p},|y_1|_{p}\}\max\{|x_2|_{p},|y_{2}|_{p}\}}$$ or, viewing $\mathbb{P}^{1}(\Qp)$ as $\Qp\cup\{\infty\}$, for $z_1,z_2 \in \Qp\cup \{\infty\}$ we define $$\rho(z_1,z_2)=\frac{|z_1-z_2|_{p}}{\max\{|z_1|_{p},1\}\max\{|z_2|_{p},1\}} \qquad\mbox{if~}z_{1},z_{2}\in \Qp,$$ and $$\rho(z,\infty)=\left\{ \begin{array}{ll} 1, & \mbox{if $|z|_{p}\leq 1$;} \\ 1/|z|_{p}, & \mbox{if $|z|_{p}> 1$.} \end{array} \right.$$ Remark that the restriction of the spherical metric on the ring $\Zp:=\{x\in \Q_p, |x|\leq 1\}$ of $p$-adic integers is same to the metric induced by the absolute value $|\cdot|_p$. A rational map $R\in \Qp(z)$ induces a map on $\P$, which we still denote as $R$. Rational maps are always Lipschitz continuous on $\P$ with respect to the spherical metric ([@SilvermanGTM241 Theorem 2.14]). Rational maps with good reduction are $1$-Lipschitz continuous ([@SilvermanGTM241 p. 59]) in the sense that $$\rho(R(u),R(v))\leq \rho(u,v), \quad \forall u,v \in \P.$$  \[rp\] Any oscillating sequence ${\bf c}=(c_{n})$ satisfying the growth condition (\[growth\]) is linearly disjoint from all $p$-adic rational flows $\mathcal{X}=(\P, R)$, where $R \in \mathbb{Q}_p(x)$ is a rational map with good reduction. It was proved in [@FFLW] that for a rational function with good reduction of order at least $2$, we have a minimal decomposition, $$\P = \mathcal{P} \sqcup\mathcal{M}\sqcup \mathcal{B}.$$ which is similar to the case of $p$-adic polynomial flow. One special case of Corollary \[rp\] is that  \[psarnak\] The Möbius function is linearly disjoint from all $p$-adic polynomial flows $(\mathbb{Z}_p, P)$ and all $p$-adic rational flow with good reduction $(\mathbb{P}^1(\mathbb{Q}_p), R)$. **Automorphisms on $2$-Torus with Zero Topological Entropy** ============================================================ Let $$\mathbb{T}^2=\mathbb{R}^2/\mathbb{Z}^2$$ be the $2$-torus. Let $SL(\mathbb{Z},2)$ be the space of all matrices $A$ of integral entries such that $\det A=\pm 1$. Then all $A{\bf x}$ of $\mathbb{T}^2$ for $A\in SL(\mathbb{Z},2)$ represent all automorphisms of $\mathbb{T}^{2}$. Affine maps on $\mathbb{T}^{2}$ are of the form $$T({\bf x}) = A{\bf x} + {\bf b}$$ where $A$ is an automorphism of $\mathbb{T}^2$ and ${\bf b}$ is an element in $\mathbb{T}^2$. Considered $T$ as a flow on $\mathbb{T}^2$, the entropy is equal to the logarithm of the sum of the modulus of all eigenvalues of $A$ of modulus strictly larger than $1$. Thus $T$ is of zero topological entropy if and only if all eigenvalues of $A$ are of modulus $1$. Let $\|\cdot\|_{\mathbb{R}^2}$ be the Euclidean norm on $\mathbb{R}^2$. Recall that the induced metric on $\mathbb{T}^2$ is defined by $$\|{\bf x}-{\bf y}\|_{\mathbb{T}^2} = \inf_{{\bf n}\in \mathbb{Z}^2} \|{\bf x}-{\bf y}-{\bf n}\|_{\mathbb{R}^2}$$ Also recall that the Hermitian norm on $\|\cdot\|_{\mathbb{C}^2}$ on $\mathbb{C}^2$ is defined by $$\|{\bf z}\|_{\mathbb{C}^{2}} = z_1 \overline{z}_1 + z_2 \overline{z}_2, \quad {\bf z} = (z_1, z_2)\in \mathbb{C}^2.$$ The space $\mathbb{R}^2$ is considered as a subspace of $\mathbb{R}^2$. It is clear that for ${\bf x}\in \mathbb{R}^2$, we have $$\|{\bf x}\|_{\mathbb{T}^2}\le \|{\bf x}\|_{\mathbb{R}^2}= \|{\bf x}\|_{\mathbb{C}^2}.$$ The following proposition actually hods in higher dimension, but we just state in the dimension $2$. \[diag\] Let $A\in SL(\mathbb{Z}, 2)$. Suppose that $A$ is diagonalizable in the field of complex numbers and have all its eigenvalues of modulus $1$. Then there exists a constant $C$ such that for all ${\bf n}\in \mathbb{Z}$ and all ${\bf x}\in \mathbb{T}^2$ we have $$\|A^n {\bf x}\|_{\mathbb{T}^2}\le C \|{\bf x}\|_{\mathbb{T}^2}.$$ Thus the flow $\mathcal{X}= (\mathbb{T}^{2}, x \mapsto A{\bf x} +{\bf b})$ for every ${\bf b}\in \mathbb{T}^{2}$ is equicontinuous. Let $A=PDP^{-1}$ where $D$ is the diagonal matrix with the eigenvalues $\lambda_1, \lambda_{2}$ of $A$ on the diagonal and $P$ is an invertible complex matrix. Then $$\|A^n {\bf x}\|_{\mathbb{R}^2}= \|P D^n P^{-1}{\bf x}\|_{\mathbb{R}^2} = \|P D^n P^{-1}{\bf x}\|_{\mathbb{C}^2} \le \|P\|_{\mathbb{C}^2} \|D^n {\bf z}\|_{\mathbb{C}^2}$$ where ${\bf z}= P^{-1}{\bf x}=(z_{1},z_{2})$. However $$\|D^n {\bf z}\|_{\mathbb{C}^2}^2= \sum_{j=1}^{2} |\lambda_j^n z_j|^2 = \sum_{j=1}^{2} |z_j|^2 = \|P^{-1} {\bf x}\|_{\mathbb{C}^2}^2.$$ Thus $$\|A^n {\bf x}\|_{\mathbb{R}^2} \le \|P\|_{\mathbb{C}^2} \|P^{-1}\|_{\mathbb{C}^2} \|{\bf x}\|_{\mathbb{R}^2}.$$ Let ${\bf n}\in \mathbb{Z}^{2}$ such that $\|{\bf x}\|_{\mathbb{T}^2} = \|{\bf x}-{\bf n}\|_{\mathbb{R}^2}$. Replace ${\bf x}$ by ${\bf x} -{\bf n}$ in the above inequality, we get $$\|A^n ({\bf x} -{\bf n})\|_{\mathbb{R}^2} \le \|P\|_{\mathbb{C}^2} \|P^{-1}\|_{\mathbb{C}^2} \|{\bf x}\|_{\mathbb{T}^2}.$$ Since $A^n {\bf n} \in \mathbb{Z}^2$ for all $n\geq 1$, finally we get for $C= \|P\|_{\mathbb{C}^2} \|P^{-1}\|_{\mathbb{C}^2}$, $$\|A^n {\bf x}\|_{\mathbb{T}^2} \le C \|{\bf x}\|_{\mathbb{T}^2}.$$ Combining this proposition and Corollary \[M2\], we have that  \[diagcoro\] Suppose $A\in SL(\mathbb{Z},2)$ is diagonalizable in complex field. And all its eigenvalues have modulus $1$. Then any oscillating sequence ${\bf c}=(c_{n})$ satisfying the growth condition (\[growth\]) is linearly disjoint from the flow $$\mathcal{X} = (\mathbb{T}^{2}, T{\bf x}: =A{\bf x}+{\bf b})$$ for every ${\bf b}\in \mathbb{T}^{2}$. Now suppose $A$ is not diagonalizable. In this case, $A$ has a double eigenvalue $1$ or $-1$ and $\det A=1$. A $2\times 2$-square matrix is called a [*module matrix*]{} if all entries are integers and its determinant is $1$. \[nondiag\] Let $M$ be a modular matrix with $\pm 1$ as a double eigenvalue. There exists a modular matrix $P$ such that $$P^{-1} M P=\pm T_{t}$$ where $T_{t}$ is a modular matrix of the form $$T_{t} = \left( \begin{array}{cc} 1 & t \\ 0 & 1 \\ \end{array} \right), \quad t \in \mathbb{Z}.$$ If the double eigenvalue of $M$ is $-1$. Then $-M$ has a double eigenvalue $1$. So we only need to prove the case that the double eigenvalue of $M$ is $1$. In this case the trace of $M$ is $2$. So we can write $$M = \left( \begin{array}{cc} a & b \\ c & 2-a \\ \end{array} \right), \quad M -I = \left( \begin{array}{cc} a-1 & b \\ c & 1-a \\ \end{array} \right).$$ The fact $\det M =1$ means $$\label{det} -bc = (a-1)^2.$$ If $b=0$, we must have $a=1$ by (\[det\]). So $$M = \left( \begin{array}{cc} 1 & 0 \\ c & 1 \\ \end{array} \right) = P \left( \begin{array}{cc} 1 & - c \\ 0 & 1 \\ \end{array} \right) P^{-1}, \quad \mbox{\rm with} \ \ P= \left( \begin{array}{cc} 0 & 1 \\ -1 & 0 \\ \end{array} \right)$$ Thus we are done. If $c=0$, we must have $a=1$ and there is nothing to do. Assume $b\not=0$ and $c\not=0$. By (\[det\]), we have $a\not=1$ and $c = -\frac{(a-1)^2}{b}$. We are looking for $${\bf x} = \left( \begin{array}{c} x_1 \\ x_2 \\ \end{array} \right)\in \mathbb{Z}^2$$ which is an eigenvector associated to $1$, i.e. $(M-I){\bf x}=0$. This equation is equivalent to $$(a-1) x_1 + b x_2 = 0\quad \hbox{and} \quad c x_{1}+ (1-a) x_{2}=0.$$ Let $d$ be the gcd of $(a-1)$ and $b$. The above two equations become one equation $$\label{EE1} \frac{a-1}{d} x_1 + \frac{b}{d} x_2=0$$ We choose an integral solution to the first equation of (\[EE1\]): $${\bf x} = \left( \begin{array}{c} x_1 \\ x_2 \\ \end{array} \right) = \left( \begin{array}{c} \frac{b}{d} \\ -\frac{a-1}{d} \\ \end{array} \right).$$ Then we are looking for a generalized eigenvector $${\bf y} = \left( \begin{array}{c} y_1 \\ y_2 \\ \end{array}\right)\in \mathbb{Z}^2$$ such that $(M-I) {\bf y} = t {\bf x}$ with $t$ an integer to be determined. The first line of the last equation is equivalent to $$\label{EE2} (a-1) y_1 + b y_2 = t \frac{b}{d}.$$ We claim that $b|d^2$. In fact, $b|(a-1)^2$. Let us factorize $a-1$ into primes: $$a-1 = \pm p_1^{\alpha_1} p_2^{\alpha_2} \cdots p_r^{\alpha_r}.$$ Then we can write $$b = \pm p_1^{\beta_1} p_2^{\beta_2} \cdots p_r^{\beta_r}, \quad (0\le \beta_j \le 2\alpha_j, 1\le j \le r).$$ So $$d = p_1^{\alpha_1 \wedge \beta } p_2^{\alpha_2\wedge \beta_2} \cdots p_r^{\alpha_r\wedge \beta_r},$$ where $\alpha\wedge \beta=\min \{ \alpha, \beta\}$. Then $b|d^2$ is equivalent to say that $\beta_j \le 2 (\alpha_j \wedge \beta_j)$ for all $1\le j \le r$, which is true because $0\le \beta_j \le 2\alpha_j$. Take $t = b^{-1} d^2 \in \mathbb{Z}$. Then the equation (\[EE2\]) becomes $$\label{EE3} \frac{a-1}{d} y_1 + \frac{b}{d} y_2 = 1.$$ By the Bézout theorem, the equation (\[EE3\]) admits a solution $(y_1, y_2)$ in $\mathbb{Z}^2$ because $\big( (a-1)/d, b/d\big)=1$. Let $P=({\bf x}, {\bf y})$ be the $2\times 2$ matrix with columns ${\bf x}$ and ${\bf y}$. Let ${\bf u}=(x_1, y_1)$ and ${\bf v} =(x_2, y_2)$ be the two rows of $P$. The first equation of (\[EE1\]) and (\[EE3\]) can be rewritten as $$\label{EE4} {\bf e}_{2}:= \frac{a-1}{d} {\bf u} + \frac{b}{d}{\bf v} = (0,1)$$ Then $$\det P = \det ({\bf u}, {\bf v}) = \frac{d}{b} \det ({\bf u}, d^{-1} b {\bf v}) = \frac{d}{b}\det ({\bf u}, {\bf e}_2)= \frac{d}{b} \cdot \frac{b}{d}=1$$ where the relation (\[EE4\]) is used for the third equality. This implies that $P$ (as well as $P^{-1}$) is a positive modular matrix and $$P^{-1} MP = T_{t}$$ for some $t=b^{-1}d^{2} \in \mathbb{Z}$. Following Proposition \[diag\], Proposition \[nondiag\], Example \[2ex\] and Theorem \[main1\], we have that  \[nondiagthm\] Suppose $A\in SL(\mathbb{Z},2)$ and suppose all its eigenvalues have modulus $1$. Then the flow $\mathcal{X} = (\mathbb{T}^{2}, T{\bf x}=A{\bf x})$ is [**MMA**]{} and [**MMLS**]{}. A consequence of Theorem \[main1\] and Theorem \[nondiagthm\] is the corollary.  \[nondiagcoro\] Suppose $A\in SL(\mathbb{Z},2)$ and suppose all its eigenvalues have modulus $1$. Any oscillating sequence ${\bf c}=(c_{n})$ satisfying the growth condition (\[growth\]) is linearly disjoint from the flow $\mathcal{X}=(\mathbb{T}^{2}, T{\bf x}=A{\bf x})$. The disjointness of the Möbius function from the flow in Corollary \[nondiagcoro\] was proved by Liu and Sarnak in [@LS]. Let us make a remark about the number $t\in \mathbb{Z}$ in Proposition \[nondiag\]. It may not be unique but can not also arbitrary. Suppose we have another $t'\in \mathbb{Z}$ such that Proposition \[nondiag\] holds. Then we have a modular matrix $P$ such that $$T_{t}P=PT_{t'}.$$ Let $(x,y)$ be the first row of $P$ and $(u, v)$ be the second row of $P$. Then we have that $$tu=0, \;\; tv=t'x,\;\; t'u=0.$$ This implies that $t=0$ if and only if $t'=0$. If $t\not=0$ and $t'\not=0$, then $v=(t'/t)x$ and $u=0$. Since $det P=1$, this implies that $x^{2} =t/t'$. That is $t/t'$ must be the square of an integer. This implies that $T_{t}$ and $T_{t'}$ are (linearly) conjugate dynamical systems on $\mathbb{T}^{2}$ if and only if $t/t'$ is the square of an integer. For example, $$\left( \begin{array}{cc} 1 & 2 \\ 0 & 1 \\ \end{array} \right) \quad \hbox{and}\quad \left( \begin{array}{cc} 1 & 1 \\ 0 & 1 \\ \end{array} \right).$$ are not conjugate flows on $\mathbb{T}^{2}$. Furthermore, let us give an example to which Proposition \[nondiag\] applies: Consider the modular matrix $$M= \left( \begin{array}{cc} -5 & 6 \\ -6 & 7 \\ \end{array} \right), \quad P= \left( \begin{array}{cc} 1 & 0 \\ 1 & 1 \\ \end{array} \right), \quad T_{6}= \left( \begin{array}{cc} 1 & 6 \\ 0 & 1 \\ \end{array} \right),$$ Then we have $$P^{-1}MP =T_{6}.$$ When $A$ is not diagonalizable, we show in the following example that there is an oscillating sequence and an affine map ${\bf x}:\to A{\bf x} +{\bf b}$ on $2$-tours such that this oscillating sequence is not linearly disjoint from the flow defined by this affine map. Let $$A = \left( \begin{array}{cc} 1 & 0 \\ 1 & 1 \\ \end{array} \right) , \qquad {\bf b}=\left( \begin{array}{c} \alpha \\ 0 \\ \end{array} \right)$$ where $0<\alpha<1$ is an irrational number. Define $$T_{A,b}({\bf x}) = A{\bf x} +{\bf b}: {\mathbb T}^{2}\to {\mathbb T}^{2}.$$ The sequence ${\bf c}=(e^{- \pi i n^2 \alpha})$ is an oscillating sequence (refer to Example 3) satisfying the growth condition (\[growth\]). But it is not linearly disjoint from the flow ${\mathcal X}=({\mathbb T}^{2}, T_{A,b})$ (whose topological entropy is zero). Let $ {\bf x}= (x, y)^{t} \in {\mathbb T}^{2}$. Observe that $$T^n {\bf x} = \left( \begin{array}{c} x + n \alpha \\ \frac{1}{2}n(n-1) \alpha + n x + y \\ \end{array} \right).$$ If we take the continuous function $f(x, y)=e^{2\pi i y}$, then we have $$\frac{1}{N}\sum_{n=1}^N e^{- \pi i n^2 \alpha} f(T^n(\alpha/2, 0)) = \frac{1}{N}\sum_{n=1}^N e^{- \pi i n^2 \alpha} e^{ \pi i n^2 \alpha} =1$$ which does not tend to zero. As a skew product of an irrational rotation, the map $T_{A,b}$ is uniquely ergodic with the Lebesque measure as the invariant measure having full support. This implies that $T_{A,b}$ is minimal, so [**MMA**]{}. But it is not [**MMLS**]{} (equivalently [**MLS**]{}). Otherwise, Theorem \[main1\] would imply that the sequence ${\bf c}$ defined by $c_{n}=e^{- \pi i n^2 \alpha}$ is linearly disjoint from the flow ${\mathcal X} =({\mathbb T}^{2}, T_{A,b})$.  \[higher\] In [@LS], Liu and Sarnak has proved that the Möbius function $\mu (n)$ is linearly disjoint from all affine flows $\mathcal{X} =( \mathbb{T}^{2}, T{\bf x} =A{\bf x} +{\bf b})$ whose topological entropy is zero, where $A\in SL(\mathbb{Z},2)$ and ${\bf b}\in \mathbb{T}^{2}$ (the result holds in high dimensions). This result and the above example indicate that in order for an oscillating sequence ${\bf c}=(c_{n})$ to be linearly disjoint from all zero entropy flows, it may not be oscillating only in the first order but also oscillating in any higher order, that is, $$\lim_{N \to \infty} \frac{1}{N} \sum_{n=1}^N c_n e^{2\pi i P(n)}=0$$ for all real polynomials $P$. The Möbius function and the von Mangoldt function, which is defined as $\Lambda (n)= \log p$ if $n=p^{k}$ for some prime number and integer $k\geq 1$ and $0$ otherwise, share the above oscillating property for any high order as shown in [@Hua]. **Feigenbaum Zero Topological Entropy Flows** ============================================= A famous zero topological entropy flow has been studied extensively for last fourty years is the so-called Feigenbaum quadratic-like map. It is a non-linear dynamical system and does not have the mean-equicontinuous property on the whole space. In this section, we will prove that a Feigenbaum zero topological entropy flow is minimally mean attractable ([**MMA**]{}) and minimally mean-L-stable ([**MMLS**]{}). Thus Sarnak’s conjecture is true for such a dynamical system. We use $ent (T)$ to denote the topological entropy of a dynamical system $T: X\to X$. For the definition of the topological entropy, the reader can refer to [@Bo; @Si; @Pa]. Let $I=[-1,1]$. Suppose $$T_{t} (x) =t-(1+t) x^{2}: I\to I.$$ It is a family of quadratic polynomials with parameter $-1/2 \leq t \leq 1$. Then $-1$ is a fixed point of $T_{t}$ and $T_{t}$ maps $1$ to $-1$. The map $T_{t}$ has a unique critical point $0$ and it is non-degenerate, that is, $T_{t}'(0)=0$ and $T_{t}''(0)\not=0$. When $t_{0}=-1/2$, $T_{t_{0}}$ has $-1$ as a parabolic fixed point, that is, $T_{t_{0}}'(-1)=1$ (see Figure 1). And $T_{t_{0}}$ has no other fixed point. One can see easily that $$T_{t_{0}}^{n} (x) \to -1 \quad \hbox{ as $n\to \infty$}, \quad \forall \; x\in I .$$ So $ent (T_{t_{0}}) =0$ and in this case, $K_{0}=\{-1\}$ is the only minimal set and $\hbox{\rm Basin}(K_{0})=[-1,1]$. Thus the flow $\mathcal{X}=([-1,1], T_{t_{0}})$ is [**MMA**]{} and [**MMLS**]{}.  \[fig1\] When $t>t_{0}$ but close to $t_{0}$, $-1$ becomes a repelling fixed point of $T_{t}$, that is, $T_{t}' (-1)>1$, but $T_{t}$ has another fixed point $p_{t}$ which is an attracting fixed point, that is, $T_{t} (p_{t}) =p_{t}$ and $|T_{t}' (p_{t})| <1$ (see Figure 2). Furthermore, we have that $$T_{t}^{n} (x) \to p_{t} \quad \hbox{as $n\to \infty$}, \;\;\forall \; x\in (-1,1) .$$  \[fig2\] This phenomenon keeps until $p_{t}$ becomes a parabolic fixed point at $t_{1}$, that is, $T_{t_{1}} '(p_{t_{1}}) =-1$. So, for $t_{0}< t\leq t_{1}$, we have that $ent (T_{t}) =0$ and $T_{t}$ has two minimal subsets $K_{0}=\{-1\}$ and $K_{1} =\{ p_{t}\}$ and $\hbox{\rm Basin}(K_{0})=\{-1,1\}$ and $\hbox{\rm Basin}(K_{1})=(-1,1)$. This proves that the flow $\mathcal{X}=([-1,1], T_{t})$ is [**MMA**]{} and [**MMLS**]{}. When $t>t_{1}$ but close to $t_{1}$, $p_{t}$ becomes a repelling fixed point of $T_{t}$, that is, $T_{t} (p_{t}) =p_{t}$ and $|T_{t} '(p_{t})|>1$, but there is a periodic cycle of period $2$ appearing near $p_{t}$, which we denote as $\{ p_{t,2,0}, p_{t,2,1}\}$. This cycle is attractive, that is, $$T_{t} (p_{t,2.0})= p_{t, 2,1}, \quad T_{t} (p_{t,2.1})= p_{t, 2,0}$$ and $$|(T_{t}^{2})' (p_{t,2.0})| = |(T_{t}^{2})' (p_{t,2.1}) |<1$$ (see Figure 2). Moreover, $$T_{t}^{n} (x) \to \{ p_{t,2,0}, p_{t,2,1}\}\;\; \hbox{ as $n\to \infty$}, \;\;\; \forall x\in I\setminus \big( \cup_{m=1}^{\infty} T_{t}^{-m} (p_{t}) \cup \{-1,1\}\big).$$ This phenomenon keeps until the cycle $\{ p_{t,2,0}, p_{t,2,1}\}$ becomes a parabolic periodic cycle of period $2$ at $t_{2}$, that is, $$|(T_{t_{1}}^{2})' (p_{t_{2},2,0})| =|(T_{t_{1}}^{2})' (p_{t_{2},2,1})|=1.$$ So, for $t_{1}< t\leq t_{2}$, we have that $ent (T_{t}) =0$ and $T_{t}$ has three minimal subsets $K_{0}=\{-1\}$ and $K_{1}=\{p_{t}\}$ and $K_{2} =\{ p_{t,2,0}, p_{t,2,1}\}$ with $$\hbox{\rm Basin}(K_{0})=\{-1,1\}, \;\;\; \hbox{\rm Basin}(K_{1})=\cup_{n=0}^{\infty} T^{-n} (p_{t}),$$ and $$\hbox{\rm Basin}(K_{2})=(-1,1)\setminus \hbox{\rm Basin}(K_1).$$ We have proved that the flow $\mathcal{X}=([-1,1], T_{t})$ is [**MMA**]{} and [**MMLS**]{} for $t_{1}<t\leq t_{2}$.  \[fig3\] To see how the periodic cycle of period $2$ was born, we can use the period doubling operator ${\mathcal R}$ as follows. When $t_{1}\leq t\leq t_{2}$, let $\beta = p_{t}>0$. Consider the map $${\mathcal R} (T_{t}) (x) = -\beta^{-1} T_{t}^{2} (-\beta x): I\to I.$$ It is not a quadratic polynomial but has a graph similar to $T_{t}$ for $t_{0} \leq t\leq t_{1}$. So when $t=t_{1}$, ${\mathcal R} (T_{t})$ has a parabolic fixed point $-1$ and when $t_{1}<t<t_{2}$, ${\mathcal R} (T_{t})$ has $-1$ as a repelling fixed point and there is an attractive fixed point $q_{t}$. When $t=t_{2}$, ${\mathcal R} (T_{t})$ has $-1$ as a repelling fixed point and $q_{t}$ becomes a parabolic fixed point. Note that $$p_{t,2,0} =-\beta q_{t} \quad \hbox{and}\quad p_{t,2,1}=T_{t} (p_{t,2,0})$$ become a cycle of periodic points of period $2$ for $T_{t}$. In general, repeat the above process, we have the following dichotomy. There is an increasing sequence $$t_{0} < t_{1}< \cdots < t_{n} < t_{n+1} <\cdots <1,$$ for any $t_{n-1}<t \leq t_{n}$, $T_{t}$ has a repelling periodic cycle $$Orb (i) =\{ p_{t, 2^{i}, k}\}_{k=0}^{2^{i-1}}$$ of period $2^{i}$ for any $0\leq i<n$ and an attractive (or parabolic but semi-attractive) periodic cycle $$Orb (n) =\{ p_{t, 2^{n}, k}\}_{k=0}^{2^{n-1}}$$ such that $$T_{t}^{n} (x) \to Orb (n)\;\; \hbox{ as $n\to \infty$}, \;\;\; \forall x\in I\setminus P_{t}$$ where $P_{t}= \cup_{m=0}^{\infty} T^{-m}_{t} (\cup_{i=0}^{n-1} Orb (i))$ is a countable subset of $I$. So we have that $ent (T_{t}) =0$ for $t_{0}\leq t<t_{\infty}$, where $$t_{\infty}=\lim_{n\to \infty} t_{n}.$$ In the case $t_{n-1}<t\leq t_{n}$, $T_{t}$ has $n+1$ minimal subset sets, $K_{0} =\{-1\}$, $K_{1} =\{p_{t}\}$, and $$K_{i+1} =\{ p_{t,2^i, k}\}_{k=0}^{2^{i}-1}, \quad i=1, 2, \cdots, n-1.$$ with $\hbox{\rm Basin}(K_{0})=\{-1,1\}$ and $$\hbox{\rm Basin}(K_{i})=\cup_{n=0}^{\infty} T^{-n} (K_{i}), \quad i=1, 2, \cdots, n-1$$ and $$\hbox{\rm Basin}(K_{n})=[-1,1]\setminus \cup_{i=0}^{\infty} \hbox{\rm Basin}(K_{i}).$$ Thus we have proved that the flow $\mathcal{X}=([-1,1], T_{t})$ is [**MMA**]{} and [**MMLS**]{}. Concluding from the above, we have that  \[easyone\] For any $-1/2\leq t<t_{\infty}$, the flow $$\mathcal{X} =(I, T_{t} (x) =t-(1+t)x^{2})$$ is [**MMA**]{} and [**MMLS**]{}.  \[prefeig\] Any oscillating sequence ${\bf c}=(c_{n})$ satisfying the growth condition (\[growth\]) is linearly disjoint from $$\mathcal{X}= (I, T_{t} (x) =t-(1+t)x^{2})$$ for any $-1/2\leq t<t_{\infty}$ A special case in Corollary \[prefeig\] is the Möbius function. The quadratic polynomial $T_{t_{\infty}}$ is called the [*Feigenbaum quadratic polynomial*]{}. This polynomial has been studied for the past 40 years since it was discovered by Feigenbaum [@Fe1; @Fe2] and, independently, by Coullet and Tresser [@CT], in 1970s. The following theorem can be found in the literature [@CEBook; @JBook; @MBook; @MVBook].  \[fei\] The quadratic polynomial $T_{t_{\infty}}$ on $I$ as a dynamical system has two repelling fixed points $-1$ and $p_{t_{\infty}}$ and a repelling periodic cycle $$Orb (n) =\{ p_{t, n, k}\}_{k=0}^{2^{n-1}}$$ of period $2^{n}$ for every $n\geq 1$. There is a Cantor set $\Lambda$ such that $$T_{t_{\infty}}^{n} (x) \to \Lambda \;\; \hbox{ as $n\to \infty$},\;\;\; \forall \; x\in I\setminus P_{t_{\infty}},$$ where $$P_{t_{\infty}}=\cup_{m=0}^{\infty} T^{-m} \Big( \{ -1, p_{t_{\infty}}\} \cup \big(\cup_{n=0}^{\infty} Orb (n)\big) \Big)$$ is a countable subset of $I$. Moreover, the topological entropy $$ent (T_{t_{\infty}})=0.$$ In the rest of this section, we will give a rigorous proof of that $T_{t_{\infty}}$ on $I\setminus P_{t_{\infty}}$ is [**MLS**]{} (that is, mean-equicontinuous). Actually, we will prove all Feigenbaum zero topological entropy dynamical systems are [**MLS**]{} (that is, mean-equicontinuous) on their corresponding set. Our strategy in the proof is as follows. First we will give a rigorous proof of that the restriction of a Feigenbaum zero topological entropy flow to its Cantor set attractor is topologically conjugate to the adding machine on the $2$-adic symbolic flow (refer to Section 5). This fact was first observed by Feigenbaum in [@Fe1; @Fe2] when he proposed his famous universality theory in chaos (and, independently, by Coullet and Tresser [@CT]). Then it was emphasized by Sullivan in his many lectures attended by the second author in 1986-1989 and in our paper and book [@JMS; @JBook]. Our rigorous proof is based on two facts. The first one is the existence of the fixed point of the periodic doubling operator which was hot topic during the past 40 years in the research of complex dynamics and chaos. The existence of the fixed point was first observed by Feigenbaum in [@Fe1; @Fe2] and used to explain his university theory in chaos. Then was proved by Lanford in [@La] with computer-assistant (but still a legitimate mathematical proof). Furthermore, a conceptual understanding and more rigorous mathematical proof by using Teichmüller theory and following idea of Moscow rigidity theorem was given by very recent work of many mathematicians (refer to [@Su; @MBook2; @JBook; @Ly]. The other is the connection between Cantor set attractor of the Feigenbaum fixed point and the non-escaping set of an induced expanding flow as studied in our previous paper [@JMS]. To state our main result, we first define the space of all Feigenbaum quadratic-like maps. Suppose $T: I\to I$ is a $C^{3}$-map. We call it a [*quadratic-like map*]{} if 1. $T(-1) =T(1)=-1$; 2. $T'(0)=0$ and $T''(0)\not=0$; 3. $T'(x) >0$ for $x\in [-1, 0)$ and $T' (x) <0$ for $x\in (0, 1]$; 4. the Schwarzian derivative $$S(T) (x):=\frac{T'''(x)}{T'(x)} -\frac{3}{2} \Big(\frac{T''(x)}{T'(x)}\Big)^{2}<0, \quad \forall x\in I.$$ Let ${\mathcal Q}$ be the space of all quadratic-like maps. Then the topological entropy $$ent (T): {\mathcal Q}\to {\mathbb R}^{+}=\{ x\geq 0\}$$ defines a functional. Using the theory of kneading sequences of Milnor-Thurston [@MT; @CEBook], the space ${\mathcal Q}$ is divided into two parts, $${\mathcal Q}_{0} =\{T\in {\mathcal Q} \;|\; ent (T) =0\}$$ and $${\mathcal Q}_{+} =\{ T\in {\mathcal Q} \;|\; ent (T) >0\}$$ The boundary of both spaces $${\mathcal F}_{\infty} =\partial {\mathcal Q}_{0}=\partial {\mathcal Q}_{+}$$ is called the [*space of Feigenbaum quadratic-like maps*]{}. The space of Feigenbaum quadratic-like maps can be defined much more general. For examples, we can assume that the type of critical point is $|x|^{\alpha}$ for some $\alpha >1$ and that the smoothness is $C^{1+Z}$ where $Z$ means the Zygmund modulus of continuity. See [@JBook] for a more general definition of Feigenbaum folding maps. To avoid to lost the reader from our main purpose, we use an easy definition of a Feigenbaum map. What we will prove in Theorem \[main2\] works for any Feigenbaum folding map but the argument will be more complicate. We would like to note that the space ${\mathcal F}_{\infty}$ is all quadratic-like maps with the same Feigenbaum kneading sequence which is the kneading sequence of $T_{t_{\infty}}$. Moreover, we have the following result from the literature [@JBook; @MBook; @MVBook].  \[fei1\] Any $T\in {\mathcal F}_{\infty}$ has two fixed points $-1$ and $\beta$ and a repelling periodic cycle $$Orb (n) =\{ p_{T, 2^n, k}\}_{k=0}^{2^{n-1}}$$ of period $2^{n}$ for every $n\geq 1$. There is a Cantor set $\Lambda_{T}$ such that $$T^{n} (x) \to \Lambda_{T}\;\; \hbox{ as $n\to \infty$}, \;\;\; \forall x\in I\setminus P_{T}$$ where $$P_{T}=\cup_{m=0}^{\infty} T^{-m} \Big(\{-1,\beta\}\cup \big(\cup_{n=1}^{\infty} Orb (n)\big)\Big)$$ is a countable subset of $I$. Furthermore, there is a homeomorphism $h$ of $I$ such that $$h\circ T=T_{t_{\infty}} \circ h \quad \hbox{on $I$}.$$ Thus the topological entropy $$ent(T)=0.$$ We call ${\mathcal X} =(I, T)$ for $T\in {\mathcal F}_{\infty}$ a Feigenbaum zero topological entropy flow.  \[main2\] Any Feigenbaum zero topological entropy flow $\mathcal{X}$ is [**MMA**]{} and [**MMLS**]{}. We first define the period doubling operator on ${\mathcal F}_{\infty}$ as follows. Any $T\in {\mathcal F}_{\infty}$ has two fixed points $\alpha =-1$ and $\beta \in (0,1)$. Define $${\mathcal R} (T) = -\beta^{-1} T^{2} (-\beta x), \quad x\in I.$$ Then ${\mathcal R} (T) \in {\mathcal F}_{\infty}$. Thus $${\mathcal R} : {\mathcal F}_{\infty}\to {\mathcal F}_{\infty}$$ defines an operator which is called the period doubling operator. As we have remarked in the paragraph after Theorem \[fei\], the period doubling operator has a unique fixed point $G\in {\mathcal F}_{\infty}$, that is, $${\mathcal R} (G) =G.$$ With the existence of this fixed point, we continue our proof as follows. Consider the critical value $c_{1}= G(0)$ and $c_{2} =G^{2} (0)$, $c_{3} =G^{3}(0)$, $c_{4}=G^{4} (0)$. Then we have that $$c_{2} < 0 <c_{4} <c_{3} <c_{1}.$$ Let $J=[c_{2}, c_{1}]$ and let $J_{0} = [c_{2}, c_{4}]$ and $J_{1}=[c_{3}, c_{1}]$.  \[fig4\] We now define a $C^{3}$ expanding map $E(x)$. Let $\beta \in (0,1)$ be the fixed point of $G$. Define $$E(x) =\left\{ \begin{array}{ll} -\beta^{-1} x, &x\in J_{0};\cr -\beta^{-1} g(x), & x\in J_{1} \end{array}\right.$$ It has been shown in [@JMS] that $E(x)$ is expanding, that is, $|E'(x)| >1$. Therefore, the non-escaping set $$NE = \cap_{n=0}^{\infty} E^{-n} (J)$$ is a Cantor set (see [@JBook Chapter 1] for the detailed proof of this fact). Let $$PC= \{ c_{n} =G^{n} (0)\}_{n=1}^{\infty}$$ be the set of all post-critical points for $G$. Then the attracting Cantor set for $G$ satisfies that $$\Lambda_{G} =\overline{PC}.$$ Since ${\mathcal R}(G)=G$, we can see that endpoints of all intervals in $\cap_{n=0}^{N} E^{-n} (J)$ are contained in $PC$. This implies that $$NE = \Lambda_{G}.$$ Since $E(x)$ is an expanding map of degree $2$, we can conjugate $E|NE$ to the shift map $\sigma$ of the symbolic space (the $2$-adic space in Section 5) $${\mathbb Z}_{2} = \prod_{n=0}^{\infty} \{0, 1\}=\{ w=i_{0}i_{1}\cdots i_{n-2}i_{n-1}\cdots\;|\; i_{n-1} \in \{0, 1\}, \; n=1, 2, \cdots\}$$ as follows. Consider the initial partition $$\eta_{0} = \{ J_{0}, J_{1}\}.$$ We can pull back it to get the $n^{th}$-partition $$~\label{part} \eta_{n} = E^{-n} (\eta_{0}) =\{ J_{w_{n}}\}$$ where $w_{n}=i_{0}i_{1}\cdots i_{n-2}i_{n-1}$. We label intervals in $\eta_{n}$ inductively as that $$E(J_{w_{n}}) = J_{\sigma (w_{n})}\in \eta_{n-1},$$ where $\sigma (w_{n}) = i_{1}\cdots i_{n-2}i_{n-1}$. From this labelling, we have that $$\cdots \subset J_{i_{0}i_{1}\cdots i_{n-2}i_{n-1}}\subset J_{i_{0} i_{1}\cdots i_{n-2}}\subset \cdots \subset J_{i_{0}}.$$ Thus for any $w=i_{0}i_{1}\cdots i_{n-2}i_{n-1} \cdots \in {\mathbb Z}_{2}$, we have that the set $$\cap_{n=1}^{\infty} J_{i_{0}i_{1}\cdots i_{n-1}} =\{ x_{w}\}$$ contains one point. Set $\pi (w) =x_{w}$. Suppose the Cantor set $NE$ is a metric space with the metric induced from the Lebesgue metric on $I$ and suppose ${\mathbb Z}_{2}$ is a metric space with the metric defined as $$d (w, w') = \frac{1}{2^{n}}$$ if $w=i_{0}i_{1}\cdots i_{n-1} i_{n}\cdots$ and $w'=i_{0}i_{1}\cdots i_{n-1} i_{n}'\cdots$ and if $i_{n}\not= i_{n}'$. Then we have that $$\pi : {\mathbb Z}_{2} \to NE$$ is a homeomorphism such that $$\pi \circ E = \sigma \circ \pi.$$ Now we define the add machining $add(w)=w+1$ on ${\mathbb Z}_{2}$ as follows. For any $w=i_{0}i_{1}\cdots i_{n-1}\cdots$, if $i_{0}=0$, then $add (w) =1i_{1}\cdots i_{n-1}\cdots$ and if $i_{0}=1$, then the first digit of $add (w)$ is $0$ and consider $i_{1} +1$, etc. From the construction of $NE=\Lambda_{G}$, we have that $$G(x_{w}) =x_{add (w)}.$$ This implies that $$G (x) = \pi \circ add\circ \pi^{-1} (x)$$ and, furthermore, $ G^{n} (x) =\pi \circ add^{n} \circ \pi^{-1} (x) $ for $x\in NE$. Since the adding machine $add$ is isometric then equicontinuous and $\pi$ is a conjugacy, we have that $\mathcal{G}=(NE, G)$ is equicontinuous. Now for any $T\in {\mathcal F}_{\infty}$, we have a homeomorphism $h: I\to I$ such that $$T^{n} =h\circ G^{n}\circ h^{-1} \quad \hbox{on $I$}.$$ Both $h|\Lambda_{G}$ and $h^{-1}|\Lambda_{T}$ are uniformly continuous. So we have that $\mathcal{T}=(\Lambda_{T}, T)$ is equicontinuous. Now what we have proved is that $T$ has countably many minimal subsets $K_{0}=\{-1\}$, $K_{1}=\{p\}$, $K_{n+1}=Orb(n)$ for all $n\geq 1$, and $K_{\infty}=\Lambda_{T}$. Since $K_{0}$ or $K_{1}$ contains only one fixed point, it is easy to see that $$\hbox{\rm Basin} (K_{0}) =\{-1,1\}, \quad \hbox{\rm Basin} (K_{1}) =\cup_{m=0}^{\infty} T^{-n} (K_{1}).$$ Similarly, since $K_{n+1}$ contains only one periodic cycle of period $2^{n}$ for $n\geq 1$, $$\hbox{\rm Basin} (K_{n+1}) =\cup_{m=0}^{\infty} T^{-n} (K_{n+1}), \quad n=1, 2, \cdots.$$ Now we claim that $$\hbox{\rm Basin} (K_{\infty}) =[-1,1]\setminus \cup_{n=0}^{\infty} \hbox{\rm Basin} (K_{n}).$$ To prove this claim, we consider $G$ and the induced expanding map $E$. Let $\eta_{n}$ be the $n^{th}$-partition in (\[part\]). Since $E$ is expanding, we have a constant $0<\tau<1$ such that $$\tau_{n} = \max_{J\in \eta_{n}} |J| \leq \tau^{n}, \quad \forall n\geq 1.$$ The map $G$ permutes intervals in $\eta_{n}$ (see Fig. 4). For any $$x\in [-1,1]\setminus \cup_{n=0}^{\infty} \hbox{\rm Basin}_{G} (K_{n}),$$ we know that $G^{n}(x)\to NE=\Lambda_{G}$ as $n\to \infty$. Note that $$\Lambda_{G}=NE=\cap_{n=1}^{\infty} \cup_{J\in \eta_{n}} J.$$ Therefore, for any $\epsilon >0$, we have an integer $N'>0$ such that $\tau_{N'}\leq \tau^{N'}<\epsilon$. We also have an integer $N''\geq N'$ such that $G^{N''} (x)$ is in some interval $J\in \eta_{N'}$. Take a point $z_{0}\in J\cap \Lambda_{G}$. Then we have that $$|G^{N''}(x) - z_{0}| <\epsilon.$$ Since $G:\Lambda_{G}\to \Lambda_{G}$ is a homeomorphism, we have a point $z\in J'\cap \Lambda_{G}$ for $J'\in \eta_{N'}$ such that $z_{0}=G^{N''} (z)$ (refer to Fig. 4). This implies that $$|G^{N''}(x)-G^{N''} (z)|<\epsilon.$$ Since $G$ permutes intervals in $\eta_{N}'$, for any $n\geq N''$, we have that $G^{n}(x)$ and $G^{n}(z)$ are in the same interval in $\eta_{N'}$, this further implies that $$|G^{n}(x)-G^{n}(z)| <\epsilon, \quad \forall n\geq N''.$$ Thus we have that $$\limsup_{N\to\infty} \frac{1}{N} \sum_{n=1}^{N} |G^n (x)-G^n (z)| <\epsilon.$$ This implies that $$\hbox{\rm Basin}_{G} (K_{\infty}) =[-1,1]\setminus \cup_{n=0}^{\infty} \hbox{\rm Basin}_{G} (K_{n}).$$ We proved the claim for $G$. Since $T^{n} =h\circ G^{n}\circ h^{-1}$ and since both $h$ and $h^{-1}$ are uniformly continuous on $I$, we proved also the claim for $T$. Now we knew that $T: K_{\infty}\to K_{\infty}$ is minimal and equicontinuous. For each $n\geq 0$, the map $T: K_{n}\to K_{n}$ is minimal and equicontinuous because it is just a periodic cycle. Thus we finally proved that ${\mathcal X}=(I,T)$ is [**MMA**]{} and [**MMLS**]{}. One consequence of Theorem \[main1\] and Theorem \[main2\] is the following corollary with the Möbius function as a special example.  \[scfeig\] Any oscillating sequence is linearly disjoint from all Feigenbaum zero topological entropy flows. After this paper was completed, we learnt that Karagulyan [@Kar] has proved that the Möbius function is linearly disjoint from all continuous interval maps with zero topological entropy. It can be proved that any oscillating sequence is linear disjoint from all continuous interval maps with zero topological entropy and the special property of Möbius function is not needed in Karagulyan’s proof, but only the oscillating property. We expect that all continuous interval maps with zero topological entropy are [**MMA**]{} and [**MMLS**]{}. **Circle Homeomorphisms** ========================= In this section, we discuss flows of zero topological entropy on the unit circle which are [**MMA**]{} and [**MMLS**]{} but not equicontinuous on its minimal set. Recall that we use $S^{1}=\{ z\in {\mathbb C} \;|\; |z|=1\}$ to denote the unit circle. Suppose $T: S^1\to S^1$ is a continuous map. Then from [@CMY; @Va], we know that if ${\mathcal X} =(S^{1}, T)$ is equicontinuous, then either - $\deg (T)=0$ and for every $x\in S^1$, $T^{n}x\to E$ as $n\to \infty$, where $E$ is the fixed point set of $T^{2}$, - $\deg(T)=1$ and $T$ is conjugate to a rigid rotation $R_{\rho}(z) =e^{2\pi i \rho} z$ for some $0\leq \rho <1$, or - $\deg(T)=-1$ and $T^{2}$ is the identity. The above three cases are simple and Theorem \[main1\] applies. Moreover, the flow defined by an orientation-preserving circle homeomorphism is equicontinuous if and only if it is conjugate to a rigid rotation. Therefore, we know that all oscillating sequences are linearly disjoint from all orientation-preserving rational circle homeomorphisms and all orientation-preserving irrational circle homeomorphisms which are not Denjoy counter-example. We are now concerning that if Theorem \[main1\] applies to all Denjoy counter-examples of orientation-preserving irrational circle homeomorphisms. A Denjoy counter-example is an orientation-preserving circle homeomorphism which is only semi-conjugate to a rigid rotation $R_{\rho}(z)=e^{2\pi i \rho}z$ for an irrational rotation number $0<\rho <1$. That is, we have a continuous onto (but not 1-1) map $h: S^1\to S^1$ such that $$~\label{conj} h\circ T =R_{\rho}\circ h$$ on $S^1$. Thus ${\mathcal X} =(S^{1}, T)$ is not equicontinuous according to [@CMY; @Va]. Let us first state the Poincaré theorem on orientation-preserving circle homeomorphisms with irrational rotation numbers.  \[pt\] Suppose $T: S^1\to S^1$ is an orientation-preserving circle homeomorphism with irrational rotation number $\rho$. Then there is a continuous onto map $h: S^1\to S^1$ satisfying (\[conj\]). Moreover, the $\omega$-limit set $\omega (z)$ for a point $z$ in the circle is either a Cantor set $\Lambda$, which is independent of $z$, or the whole circle $S^{1}$. In the case $\omega (z)=S^{1}$, $h$ is a homeomorphism. In the case $\omega (z)=\Lambda$, a Cantor set, $S^{1}\setminus \Lambda =\cup J$ is the union of all wandering intervals. Here $\Lambda$ is also called the non-wandering set for $T$. Denjoy constucted an example $T$ of an orientation-preserving circle homeomorphism with irrational rotation number such that for any $z \in S^{1}$, $\omega (z)=\Lambda$ is a Cantor set. In this case, $T: \Lambda \to \Lambda$ is minimal. We call such a map a Denjoy counter-example. Associated to a Denjoy counter-example, there are two flows ${\mathcal X}=(S^{1}, T)$ and ${\mathcal A}=(\Lambda, T)$. We have that the following theorem.  \[ceq\] Suppose $T: S^{1}\to S^{1}$ is a Denjoy counter-example. The flow $\mathcal{X}= (S^{1}, T)$ is [**MMA**]{} but the flow $\mathcal{A} =(\Lambda, T)$ is not equicontinuous. However, the flow $\mathcal{A} =(\Lambda, T)$ is [**MLS**]{}, thus, the flow $\mathcal{X}= (S^{1}, T)$ is [**MMA**]{} and [**MMLS**]{}. For any $x, y\in \Lambda$ close to each other, let $[x,y]$ be the shortest arc with endpoints $[x, y]$ in the circle. The arc length of $[x,y]$ is $|x-y|$. Assume the total length of $S^{1}$ is $1$. Since $T$ is a Denjoy counter-example, $\Lambda$ is a Cantor set and $$S^{1}\setminus \Lambda=\cup I$$ is the union of all wondering intervals. We have that $$\sum_{I} |I| \leq 1.$$ For any $x\in S^{1}\setminus \Lambda$, it is in a wandering interval $I=(a,b)$, where $a, b\in \Lambda$. Let $I_{n} =T^{n}(I)$, $n=0, 1, \cdots$. Then we have that $$\sum_{n=0}^{\infty} |I_{n}| \leq 1.$$ Thus $|I_{n}|\to 0$ as $n\to \infty$. Since $T^{n}(x) \in I_n$ and $T^{n}(a)\in \overline{I}_{n}$, we have that $$\lim_{n\to +\infty} |T^{n} (a)- T^{n}(x)| = \lim_{n\to +\infty} |I_n| = 0,$$ which implies that $\mathcal{X}=(S^{1}, T)$ is [**MMA**]{}. Consider the set ${\mathcal W}$ of all wandering intervals, which is countable. We partition it into full orbits $${\mathcal W} =\{ I_{m,n}\; |\; I_{m, n} =T^{n} (I_{m, 0}),\;\; 0\leq m< L,\; n\in {\mathbb Z}\},$$ where $L$ is the number of full orbits of wandering intervals. Here $L$ is either a positive integer or $\infty$. Since $\Lambda$ is a Cantor set, we have $$S^{1} =\overline{\bigcup_{0\le m < L} \bigcup_{n\in {\mathbb Z}} I_{m,n}}$$ Let $h$ be the semi-conjugacy between $T$ and the rigid rotation $R_{\rho}$. See (\[conj\]). Then the image $ h(I_{m,n})$ is a singleton $\{x_{m,n}\}$ and we write $x_{m, n}=h(I_{m, n})$. By (\[conj\]), $$\label{hh1} \{ x_{m,n} =h(I_{m, n}) = R^{n}_{\rho} (x_{m,0}) \;|\; n\in {\mathbb Z}\}$$ is the full orbit of $x_{m, 0}$ under iterations of $R_{\rho}$ for any fixed $0\leq m<L$. Also notice that $h$ is monotone increasing. So, if $x$ and $y$ are two points in the Cantor set $\Lambda$， we have $$\label{hh2} I_{m,n} \subset [x, y] \quad\mbox{\rm iff}\quad x_{m,n} \in [h(x), h(y)].$$ The rigid rotation $R_{\rho}$ is uniquely ergodic with the Lebesgue measure as the unique invariant measure. This fact, together with (\[hh1\]), implies that for any $z, w\in S^{1}$, $$\label{hh3} \lim_{n\to \infty} \frac{\#(\{ k\;|\; x_{m, k}\in [z, w], \; -n \leq k\leq 0\})}{n} = |z-w|.$$ This allows us to prove that ${\mathcal A}=(\Lambda, T)$ is not equicontinuous. In fact, given any $x\not= y\in \Lambda$, by (\[hh2\]) and (\[hh3\]) we have that $$\begin{aligned} & & \lim_{n\to \infty} \frac{\#(\{ k\;|\; I_{0, k}\subset [x, y], \; -n \leq k\leq 0\})}{n} \\ &=& \lim_{n\to \infty} \frac{\#(\{ k\;|\; x_{0,k}\in [h(x), h(y)], \; -n \leq k\leq 0\})}{n} = |h(x)-h(y)| >0.\end{aligned}$$ Thus there are infinitely many negative integers $k$ such that $I_{0,k}\subset [x,y]$. For such $k$’s we have $$I_{0,0} =T^{|k|} (I_{0, k}) \subset [T^{|k|}(x), T^{|k|}(y)]$$ so that $$|T^{|k|}(x)-T^{|k|}(y)|\geq |I_{0}|.$$ for infinitely many positive integers $|k|$. Therefore, ${\mathcal A} =(\Lambda, T)$ is not equicontinuous. Next we prove that ${\mathcal A} =(\Lambda, T)$ is [**MLS**]{}. We first assume that the non-wandering set $\Lambda$ has zero Lebesgue measure. Under this assumption, we have that $$\sum_{0\le m< L} \sum_{n\in {\mathbb Z}} |I_{m,n}| =1.$$ Moreover, for any two points $x\not= y\in \Lambda$, $$~\label{zeq} |x-y| =\sum_{I\in {\mathcal W}, I\subset [x,y]} |I|.$$ For any $N\ge 1$, consider the rectangle $$R_N=\{ (m, n)\in \mathbb{N}\times \mathbb{Z}\; |\; 0\le m \le N \wedge (L-1), |n|\le N\},$$ where $N\wedge (L-1)= min\{ N, L-1\}$. For any $\epsilon >0$, we have a large integer $N_{0}\ge 1$ such that $$\sum_{(m, n )\in \mathbb{N}\times \mathbb{Z} \setminus R_{N_0}} |I_{m,n}| < \epsilon.$$ Since $h$ is uniformly continuous, we can find a $0< \delta\leq \epsilon$ such that for any $x\not= y\in \Lambda$ with $|x-y|<\delta$, $$0< |h(x)-h(y)| < \frac{\epsilon}{(N_{0}\wedge (L-1)+1)(2N_{0}+1)}.$$ Let $$E=\bigcup_{(m, l)\in R_{N_0}} E_{m,l}$$ where $$E_{m,l}=\{0\leq k<\infty \;|\; I_{m,-k+l} \subset [x,y]\}.$$ For any fixed $(m, l)\in R_{N_0}$, we have $$\begin{aligned} \lim_{n\to \infty} \frac{\#(E_{m,l}\cap [1,n])}{n} & = & \lim_{n\to \infty} \frac{\#(\{ k\;|\; I_{m, k+l}\subset [x, y], \; -n \leq k\leq 0\})}{n} \\ &= & \lim_{n\to \infty} \frac{\#(\{ k\;|\; x_{m,k+l}\in [h(x), h(y)], \; -n \leq k\leq 0\})}{n} \\ & = & |h(x)-h(y)| >0.\end{aligned}$$ Therefore the density of $E$ exists and is equal to $$\overline{D} (E)= \limsup_{n\to \infty} \frac{\#(E\cap [1,n])}{n} = (N_{0}\wedge(L-1)+1) (2N_{0}+1) |h(x)-h(y)| < \epsilon$$ For any $n\not\in E$, the interval $[T^{n}(x), T^{n}(y)]$ contains only wandering intervals from $${\mathcal W}_{N_{0}} =\{ I_{m,n} =T^{n} (I_{m, 0})\; |\; (m,n)\in \mathbb{N}\times \mathbb{Z} \setminus R_{N_0}\}.$$ From (\[zeq\]), we have that $$|T^{n}(x)-T^{n}(y)| \leq \sum_{(m,n) \in \mathbb{N}\times \mathbb{Z} \setminus R_{N_0}} |I_{m,n}| <\epsilon.$$ This implies that ${\mathcal A}=(\Lambda, T)$ is [**MLS**]{}. Thus it is mean-equicontinuous. Since the flow ${\mathcal X} =(S^{1}, T)$ has only one mean minimal attractor $\Lambda$, so it is [**MMLS**]{}. When the non-wandering set $\Lambda$ has a positive Lebesgue measure, there is another Denjoy counter-example $\widetilde{T}$ whose non-wandering set $\widetilde{\Lambda}$ has zero Lebesgue measure such that both of them are semi-conjugate to the same rigidity rotation $R_{\rho}$. That is, we have two continuous onto maps $h, \widetilde{h}: {\mathbb T}\to {\mathbb T}$ such that $$h\circ T= R_{\rho}\circ h \quad \hbox{and}\quad \widetilde{h}\circ \widetilde{T}= R_{\rho} \circ \widetilde{h}.$$ Now we can define $$\widehat{h} : {\mathbb T}\to {\mathbb T}$$ which formally can be written as $\widetilde{h}^{-1} \circ h$. It maps every wandering interval to the corresponding wandering interval bijectively and preserves orders of wandering intervals. And it maps $\Lambda$ to $\widetilde{\Lambda}$ bijectively. Thus it is a homeomorphism of the circle such that $$\widehat{h} \circ T = \widetilde{T} \circ \widehat{h}.$$ Since $\widetilde{T}$ is [**MMLS**]{} and since $T$ is topologically conjugate to $\widetilde{T}$, we get that $T$ is [**MMLS**]{}. One consequence of Theorem \[main1\] and Theorem \[main2\] and Theorem \[ceq\] is the following corollary with the Möbius function as a special example.  \[scdenjoy\] Any oscillating sequence satisfying the growth condition (\[growth\]) is linearly disjoint from any Denjoy counter-example. In particular, we have  \[Kar\] The Möbius function is linearly disjoint from all orientation-preserving circle homeomorphisms. [99]{} J. Auslander, [*Mean-L-stable systems*]{}, Illinois J. Math., 3 (1959), 566-579. J. Bourgain, [*On the correlation of the Möbius function with random rank one systems*]{}. arXiv:1112.1031, 2011. J. Bourgain, P. Sarnak, and Ziegler, [*Disjointness of Möbius from horocycle flows*]{}. From Fourier analysis and number theory to Radon transforms and geometry, 67-83, Dev. Math. 28, Springer, New York, 2013. R. Bowen, [*Equilibrium States and the Ergodic Theory of Anosov Diffeomorphisms*]{}. Lecture Notes in Mathematics, no. 470: (A. Dold and B. Eckmann, editors). Springer-Verlag (Heidelberg, 1975), 108 pp. A. M. Bruckner and T. Hu, [*Equicontinuity of iterates of an interval map*]{}, Tamkang J. Math. 21 (1990), no. 3, 287-294. P. Coullet and C. Tresser, [*Itération dèndomorphismes et groupe de renormalisation*]{}. C. R. Acad. Sci. Paris Ser., A-B [**287**]{}, 1978, pp. A577-A580. (J. Phys. Coll. 39:C5 (1978), 25-28; supplément au 39:8). P. Collet and J.-P. Eckmann, [*Iterated Maps on the Interval as Dynamical Systems*]{}. Progress in Physics, Vol. [**1**]{}, Birkhäuser, Boston, 1980. S. Cho, K. Min, and S. Yang, [*Equicontinuity of iterates of a map on the circle*]{}. Bull. Korean Math. Soc. 30 (1993), No. 2, pp. 239-244. H. Daboussi and H. Delange, [*On multiplicative arithmetical functions whose modulus does not exceed one*]{}, J. London Math. Soc. (2) 26 (1982), no. 2, 245-264. H. Davenport, [*On some infinite series involving arithmetical functions (II)*]{}, Quart. J. Math. Oxford, 8 (1937), 313-320. H. Delange, [*Sur les fonctions multiplicatives complexes de module*]{}, Ann. Inst. Fourier, 44 (1994), no. 5, 1323-1349. El H. El Abdalaoui, J. Kulaga-Przymus, M. Lemanczyk and Th. De La Rue, [ *The Chowla and the Sarnak conjectures from ergodic theory point of view*]{}, arXiv:1410.1673. A. H. Fan, S. L. Fan, L. M. Liao and Y. F. Wang, [*On minimal decomposition of $p$-adic rational functions with good reduction*]{}, preprint 2015. A. H. Fan and L. M. Liao, [*On minimal decomposition of p-adic polynomial dynamical systems*]{}, [ Adv. Math.]{}, 228:2116–2144, 2011. A. H. Fan and J. Schmeling, [ *Everywhere divergence of one-sided ergodic Hilbert transform*]{}, preprint. A. H. Fan and D. Schneider, [*On an inequality of Littlewood-Salem*]{}, Ann. I. H. Poincar,39, 2 (2003) 193-216 M. Feigenbaum, [*Quantitative universality for a class of non-linear transformations*]{}. J. Stat. Phys. [**19**]{} (1978), 25-52. M. Feigenbaum, [*The universal metric properties of non-linear transformations*]{}. J. Stat. Phys. [**21**]{} (1979), 669-706. S. Fomin, [*On dynamical systems with a purely point spectrum*]{}, Dokl. Akad. Nauk SSSR, vol. 77 (1951), 29-32 (In Russian). N. Frantzikinakis and B. Host, [ *Higher order Fourier analysis of multiplicative functions and applications*]{}. arXiv:1403.0945. L. M. Ge, [ *Topology of natural numbers and entropy of arithmetic functions*]{}, preprint, 2015. W. H. Gottschalk, [*Orbit-closure decompositions and almost periodic properties*]{}, Bull. Amer. Math. Soc. 50, (1944), 915-919. B. Green and T. Tao, [*The Möbius function is strongly orthorgonal to nilsequences*]{}. Ann. Math. (2), 175 (2012), 541-566 B. Green and T. Tao, [*Quadratic uniformity of the Möbius function*]{}. arXiv:math/0606087v2 \[math.NT\] 22 Sep 2007. L. G. Hua, [*Additive Theory of Prime Numbers*]{} (Translations of Mathematical Monographs : Vol 13). Amer Mathematical Society. 1966. Y. Jiang, T. Morita, and D. Sullivan, [*Expanding direction of period doubling operator*]{}. Comm. in Math. Phys., [**144**]{} (1992), no. 3, 509-520. Y. Jiang, [*Renormalization and Geometry in One-Dimensional and Complex Dynamics*]{}. Advanced Series in Nonlinear Dynamics, Vol. 10 (1996) World Scientific Publishing Co. Pte. Ltd., River Edge, NJ. J. P. Kahane, [*Some random series of functions*]{}, Cambride University Press, 1985. J.P.Kahane and E. Saias, [*Fonction complètement multiplicatives de somme nulle*]{}, preprint 2015. https://hal.archives-ouvertes.fr/hal-01177065. D. Karagulyan, [*On Möbius orthogonality for interval maps of zero entropy and orientation-preserving circle homeomorphisms.*]{} Ark. Mat., to appear (DOI: 10.1007/s11512-014-0208-5). U. Krengel, [*Ergodic theorems*]{}, Walter de Gruyter, Berlin, New York, 1982. O.E. Lanford III, [*A Computer-assisted proof of Feigenbaum conjectures*]{}. Bull. Am. Math. Sco. 6, 427-434 (1982) J. Li, S. Tu and X. D. Ye, [*mean-equicontinuity and mean sensitivity*]{}, arXiv: 1312.7663v3. J. Y. Liu and P. Sarnak, [*The Möbius function and distal flows*]{}, Duke Math. J. 164 (2015), no. 7, 1353-1399. J. Y. Liu and T. Zhan, [*Exponential sums involving the Möbius function*]{}. Indagationes Mathematicae Volume [**7**]{} (1996), Issue 2, 271-278. M. Lyubich, [*Feigenbaum-Coullet-Tresser universality and Milnor’s hairiness conjecture*]{}. Annals of Mathematics, v. [**149**]{} (1999), 319 - 420. C. Mauduit and J. Rivat, [*Sur un problème de Gelfond: la somme des chiffres des nombres premiers*]{}. Ann. Math. (2) [**171**]{} (2010), 1591-1646. C. McMullen, [*Complex Dynamics and Renormalization*]{}. Ann. of Math. Stud., vol [**135**]{}, Princeton Univ. Press, Princeton, NJ, 1994. C. McMullen, [*Renormalization and $3$-Manifolds which Fiber over the Circle*]{}. Ann. of Math. Stud., vol [**135**]{}, Princeton Univ. Press, Princeton, NJ, 1996. W. de Melo and S. van Strien, [*One-Dimensional Dynamics*]{}. SpringerVerlag, Berlin, Heidelberg, 1993. W. Parry, [*Topics in Ergodic Theory*]{}. Cambridge University Press, 1981. J. Milnor and W. Thurston, [*On iterated maps of the interval: I and II*]{}. In Lecture Notes in Mathematics, Vol. 1342, pp. 465-563, Springer, New York, Berlin, 1988, pp. 465-563. J. C. Oxtoby, [*Ergodic sets*]{}, Bull. Amer. Math. Soc., 58 (1952), 116-136. K. Petersen, [*Ergodic Theory*]{}. Cambridge Studies in Advanced Mathematics. P. Sarnak, [*Three lectures on the Möbius function, randomness and dynamics*]{}, IAS Lecture Notes, 2009;\ http://publications.ias.edu/sites/default/files/MobiusFunctionsLectures(2).pdf. P. Sarnak, [*Möbius randomness and dynamics*]{}, Not. S. Afr. Math. Soc. 43 (2012), 89-97. J. Silverman, [*The arithmetic of dynamical systems*]{}, volume 241 of [ Graduate Texts in Mathematics]{}. Springer, New York, 2007. Ya. Sinai, [*Introduction to Ergodic Theory*]{}. Princeton Press, 1966. D. Sullivan, [*Bounds, quadratic differentials, and renormalization conjectures*]{}. American Mathematical Society Centennial Publications, Volume 2: Mathematics into the Twenty-First Century, AMS, Providence, RI, 1992, pp. 417-466. A. Valaristos, [*Equicontinuity of iterates of circle maps*]{}. Internat. J. Math. Math. Sci. 21 (1998), no. 3, 453-458. Zygmund, [*Trigonometric Series*]{}. Cambridge University Press, Cambridge, London, New York, Melbourne, 1959. [^1]: The first author is partially supported by NSFC \[grant number 11471132\]. The second author is partially supported by the collaboration grant from the Simons Foundation \[grant number 199837\] and the CUNY collaborative incentive research grants \[grant number 2013\] and awards from PSC-CUNY and a grant from NSFC \[grant number 11171121\].

ArXiv

--- abstract: 'Tauvel’s height formula, which provides a link between the height of a prime ideal and the Gelfand-Kirillov dimension of the corresponding factor algebra, is verified for quantum nilpotent algebras.' author: - 'K R Goodearl[^1],  S Launois[^2]  and T H Lenagan[^3]' title: 'Tauvel’s height formula for quantum nilpotent algebras' --- Introduction ============ Quantum nilpotent algebras, originally introduced under the name Cauchon-Goodearl-Letzter extensions [@llr-ufd], are iterated Ore extensions with special properties which cover a wide variety of algebras, including many of the algebras that appear as quantised coordinate rings. Examples include quantum Schubert cell algebras, quantum matrix algebras, generic quantised coordinate rings of affine, symplectic and euclidean spaces, and generic quantised Weyl algebras. The precise definition of a quantum nilpotent algebra is recalled in Section \[section-cgl\]. It is designed to allow application of both Cauchon’s deleting derivations algorithm and the Goodearl-Letzter stratification theory for prime ideals. In studying the prime spectrum of an algebra $R$, key invariants for a prime ideal $P$ are its height, ${{\rm ht}}(P)$, and the Gelfand-Kirillov dimension of the factor algebra, ${{\rm GK}}(R/P)$. Tauvel’s height formula $${{\rm GK}}(R/P)+{{\rm ht}}(P) = {{\rm GK}}(R)$$ provides a useful connection between these two invariants and it is of interest to know when this formula holds. The purpose of this paper is to show that Tauvel’s height formula does hold for all quantum nilpotent algebras. Several verifications of the height formula for particular classes of algebras have proceeded by first proving that the algebras are catenary, see, for example, [@c2; @gl-cat; @h1; @llr2; @oh; @y1; @y2]. In these papers, catenarity is demonstrated by first establishing certain homological conditions and showing that normal separation holds for the prime spectrum. While the homological conditions can easily be established for quantum nilpotent algebras, normal separation remains elusive at the moment, although we do conjecture that this condition holds for all quantum nilpotent algebras. The approach taken in the present paper is to exploit Cauchon’s deleting derivations algorithm to establish Tauvel’s height formula for torus-invariant prime ideals of a quantum nilpotent algebra, then extend this to primitive ideals by using the Goodearl-Letzter stratification theory by virtue of the fact that the primitive ideals are identifiable as the maximal members in individual strata. Finally, the formula is established for arbitrary prime ideals via the link between the prime spectrum of a given stratum and the prime spectrum of an associated commutative Laurent polynomial algebra, where the formula is well-known. As far as we are aware, the approach we use (from torus-invariant primes to primitive ideals, then arbitrary primes) has not been used before; so this result advertises the approach. Quantum nilpotent algebras and stratification theory {#section-cgl} ==================================================== Fix a base field ${{\mathbb K}}$ throughout the paper. Let $N$ denote a positive integer and $R$ an iterated Ore extension of the form $$R\ = \ {{\mathbb K}}[x_1][x_2;\sigma_2,\delta_2]\cdots[x_N;\sigma_N,\delta_N],$$ where $\sigma_j$ is an automorphism of the ${{\mathbb K}}$-algebra $R_{j-1}:={{\mathbb K}}[x_1][x_2;\sigma_2,\delta_2]\dots[x_{j-1};\sigma_{j-1},\delta_{j-1}]$ and $\delta_j$ is a ${{\mathbb K}}$-linear $\sigma_j$-derivation of $R_{j-1}$ for all $j\in { [ \hspace{-0.65mm} [}2 ,N {] \hspace{-0.65mm} ]}$. (When needed, we denote $R_0 := {{\mathbb K}}$ and set $\sigma_1 := {\text{id}}_{{\mathbb K}}$, $\delta_1 := 0$.) In other words, $R$ is a skew polynomial ring whose multiplication is determined by: $$x_j a = \sigma_j(a) x_j + \delta_j(a)$$ for all $j\in { [ \hspace{-0.65mm} [}2 ,N {] \hspace{-0.65mm} ]}$ and $a \in R_{j-1}$. Thus $R$ is a noetherian domain. Henceforth, we assume that $R$ is a quantum nilpotent algebra, as in the following definition. \[CGLdef\] The iterated Ore extension $R$ is said to be a *quantum nilpotent algebra* or a *CGL extension* [@llr-ufd Definition 3.1] if it is equipped with a rational action of a ${{\mathbb K}}$-torus $\ch = ({{\mathbb K}}^*)^d$ by ${{\mathbb K}}$-algebra automorphisms satisfying the following conditions: 1. The elements $x_1, \ldots, x_N$ are $\ch$-eigenvectors. 2. For every $j \in { [ \hspace{-0.65mm} [}2,N {] \hspace{-0.65mm} ]}$, $\delta_j$ is a locally nilpotent $\sigma_j$-derivation of $R_{j-1}$. 3. For every $j \in { [ \hspace{-0.65mm} [}1,N {] \hspace{-0.65mm} ]}$, there exists $h_j \in \ch$ such that $(h_j\cdot)|_{R_{j-1}} = \sigma_j$ and $h_j \cdot x_j = q_j x_j$ for some $q_j \in {{\mathbb K}}^*$ which is not a root of unity. (We have omitted the condition $\sigma_j \circ \delta_j = q_j \delta_j \sigma_j$ from the original definition, as it follows from the other conditions; see, e.g., [@gy0 Eq. (3.1); comments, p.694].) From (i) and (iii), there exist scalars ${\lambda}_{j,i} \in {{\mathbb K}}^*$ such that $\sigma_j(x_i) = {\lambda}_{j,i} x_i$ for all $i < j$ in ${ [ \hspace{-0.65mm} [}1,N {] \hspace{-0.65mm} ]}$. A two-sided ideal $I$ of $R$ is said to be [*$\ch$-invariant*]{} if $h\cdot I=I$ for all $h \in \ch$. An [*$\ch$-prime ideal*]{} of $R$ is a proper $\ch$-invariant ideal $J$ of $R$ such that if $J$ contains the product of two $\ch$-invariant ideals of $R$ then $J$ contains at least one of them. We denote by $\ch$-$\spec(R)$ the set of all $\ch$-prime ideals of $R$. Observe that if $P$ is a prime ideal of $R$ then $$(P:\ch)\ := \ \bigcap_{h\in \ch} h\cdot P$$ (namely, the largest $\ch$-invariant ideal contained in $P$) is an $\ch$-prime ideal of $R$. For any $\ch$-prime ideal $J$ of $R$, we denote by $\spec_J (R)$ the [*$\ch$-stratum*]{} associated to $J$; that is, $$\spec_J (R) :=\{ P \in \spec(R) \mbox{ $\mid$ } (P:\ch)=J \}.$$ Then the $\ch$-strata of $\spec(R)$ form a partition of $\spec(R)$; that is, $$\label{eq:Hstratification} \spec(R) := \bigsqcup_{J \in \ch\mbox{-}\spec(R)}\spec_J(R).$$ This partition is the so-called [*$\ch$-stratification*]{} of $\spec(R)$. It follows from work of Goodearl and Letzter [@gl2 Proposition 4.2] that every $\ch$-prime ideal of $R$ is completely prime, so $\ch$-$\spec(R)$ coincides with the set of $\ch$-invariant completely prime ideals of $R$. Moreover there are at most $2^N$ $\ch$-prime ideals in $R$. As a consequence, the prime spectrum of $R$ is partitioned into a finite number of parts, the $\ch$-strata. In case $R$ is *torsionfree*, meaning that the subgroup $\langle \lambda_{j,i} \mid 1\le i<j \le N \rangle$ of ${{\mathbb K}}^*$ is torsionfree, all prime ideals of $R$ are completely prime [@gl3 Theorem 2.3]. For each $\ch$-prime ideal $J$ of $R$, the space $\spec_J(R)$ (equipped with the relative Zariski topology inherited from $\spec(R)$) is homeomorphic to ${\rm Spec}({{\mathbb K}}[z_1^{\pm 1},\ldots ,z_d^{\pm 1}])$ for some $d$ which depends on $J$ [@bg Theorems II.2.13 and II.6.4], and the primitive ideals of $R$ are precisely the prime ideals that are maximal in their $\ch$-strata [@bg Theorem II.8.4]. Cauchon’s deleting derivations algorithm {#canonicalembedding} ======================================== As we have seen in the previous section, the $\ch$-prime ideals of a quantum nilpotent algebra $R$ are key in studying the whole prime spectrum. Cauchon’s deleting derivations algorithm [@c1], which we summarise below, provides a powerful way of studying the $\ch$-prime ideals of $R$. Deleting derivations algorithm ------------------------------ In order to describe the prime spectrum of $R$, Cauchon [@c1 Section 3.2] has constructed an algorithm called the *deleting derivations algorithm*. This algorithm constructs, for each $j = N+1, N, \dots, 2 $, an $N$-tuple $(x_1^{(j)},\dots,x_N^{(j)})$ of elements of the division ring of fractions $\operatorname{Fract}(R)$ defined as follows: 1. When $j=N+1$, we set $(x_1^{(N+1)},\dots,x_N^{(N+1)}) :=(x_1,\dots,x_N)$. 2. Assume that $j<N+1$ and that the $x_i^{(j+1)}$ ($i \in { [ \hspace{-0.65mm} [}1,N {] \hspace{-0.65mm} ]}$) are already constructed. Then it follows from [@c1 Théorème 3.2.1] that $x_j^{(j+1)} \neq 0$ and, for each $i\in { [ \hspace{-0.65mm} [}1,N {] \hspace{-0.65mm} ]}$, we set $$x_i^{(j)} := \begin{cases} x_i^{(j+1)} & \quad \mbox{ if }i \geq j \\ \displaystyle{\sum_{k=0}^{+\infty } \frac{(1-q_j)^{-k}}{[k]!_{q_j}} } \delta_j^k \circ \sigma_j^{-k} (x_i^{(j+1)}) (x_j^{(j+1)})^{-k} & \quad \mbox{ if }i < j, \end{cases}$$ where $[k]!_{q_j}=[0]_{q_j} \times \dots \times [k]_{q_j}$ with $[0]_{q_j}=1$ and $[i]_{q_j}=1+q_j+\dots+q_j^{i-1}$ when $i \geq 1$. For all $j \in { [ \hspace{-0.65mm} [}2,N+1 {] \hspace{-0.65mm} ]}$, we denote by $R^{(j)}$ the subalgebra of $\operatorname{Fract}(R)$ generated by the $x_i^{(j)}$; that is, $$R^{(j)}:= {{\mathbb K}}\langle x_1^{(j)},\dots,x_N^{(j)} \rangle .$$ The following results were proved by Cauchon [@c1 Théorème 3.2.1 and Lemme 4.2.1]. For $j \in { [ \hspace{-0.65mm} [}2,N+1 {] \hspace{-0.65mm} ]}$, we have 1. $R^{(j)}$ is isomorphic to an iterated Ore extension of the form $${{\mathbb K}}[Y_1]\cdots[Y_{j-1};\sigma_{j-1},\delta_{j-1}][Y_j;\tau_j]\cdots[Y_N;\tau_N]$$ by an isomorphism that sends $x_i^{(j)}$ to $Y_i$ ($1 \leq i \leq N$), where $\tau_j,\dots,\tau_N$ denote the ${{\mathbb K}}$-linear automorphisms such that $\tau_{\ell}(Y_i)=\lambda_{\ell,i} Y_i$ ($1 \leq i \leq \ell$). 2. Assume that $j \neq N+1$ and set $S_j:=\{(x_j^{(j+1)})^n ~|~ n\in \mathbb{N} \}= \{(x_j^{(j)})^n ~|~ n\in \mathbb{N} \}$. This is a multiplicative system of regular elements of $R^{(j)}$ and $R^{(j+1)}$, that satisfies the Ore condition in $R^{(j)}$ and $R^{(j+1)}$. Moreover we have $$\label{matchloc} R^{(j)}S_j^{-1}=R^{(j+1)}S_j^{-1}.$$ It follows from these results that $R^{(j)}$ is a noetherian domain, for all $j\in { [ \hspace{-0.65mm} [}2,N+1 {] \hspace{-0.65mm} ]}$. As in [@c1], we use the following notation. We set ${\overline{R}}:=R^{(2)}$ and $T_i:=x_i^{(2)}$ for all $i\in { [ \hspace{-0.65mm} [}1,N {] \hspace{-0.65mm} ]}$. Note that $x_i^{(i+1)} = x_i^{(i)} = \cdots = x_i^{(2)} = T_i$ for $i \in { [ \hspace{-0.65mm} [}1,N {] \hspace{-0.65mm} ]}$. Hence, the structure of $R^{(j)}$ as an iterated Ore extension can be expressed as $$\label{R(j)} R^{(j)} = {{\mathbb K}}[x_1^{(j)}] \cdots [x_{j-1}^{(j)}; \sigma_{j-1}, \delta_{j-1}] [T_j; \tau_j] \cdots [T_N; \tau_N].$$ It follows from [@c1 Proposition 3.2.1] that ${\overline{R}}$ is a quantum affine space in the indeterminates $T_1,\dots,T_N$, that is, ${\overline{R}}$ is an iterated Ore extension twisted only by automorphisms. It is for this reason that Cauchon used the expression “effacement des dérivations”. More precisely, let $\Lambda=\left( \lambda_{i,j} \right) \in M_N({{\mathbb K}}^*)$ be the multiplicatively antisymmetric matrix where the $\lambda_{j,i}$ with $i<j$ come from the quantum nilpotent algebra structure of $R$ (Definition \[CGLdef\]). Thus, $$\lambda_{j,i}= \begin{cases} 1 & \mbox{ if } i=j \\ \lambda_{i,j}^{-1} & \mbox{ if }i> j. \end{cases}$$ Then we have $$\label{Rbar.qtorus} {\overline{R}}= {{\mathbb K}}\langle T_1,\dots,T_N \mid T_iT_j = \lambda_{i,j} T_j T_i \;\; \forall\, i,j \in { [ \hspace{-0.65mm} [}1,N {] \hspace{-0.65mm} ]}\rangle =\co_{\Lambda}({{\mathbb K}}^N).$$ Canonical embedding ------------------- The deleting derivations algorithm was used by Cauchon in order to relate the prime spectrum of a quantum nilpotent algebra $R$ to the prime spectrum of the associated quantum affine space ${\overline{R}}$. More precisely, he has used this algorithm to construct embeddings $$\varphi_j:{\rm Spec}(R^{(j+1)}) \longrightarrow {\rm Spec}(R^{(j)}) \qquad {\rm for ~} j \in { [ \hspace{-0.65mm} [}2,N {] \hspace{-0.65mm} ]}.$$ Recall from [@c1 Section 4.3] that these embeddings are defined as follows. Let $P \in {\rm Spec}(R^{(j+1)})$. Then $$\varphi_j (P) := \left\{\begin{array}{ll} PS_j^{-1} \cap R^{(j)} & \mbox{ if } x_j^{(j+1)} \notin P \\ g_j^{-1} \left( P/\langle x_j^{(j+1)}\rangle \right) & \mbox{ if } x_j^{(j+1)} \in P, \\ \end{array}\right.$$ where $g_j$ denotes the surjective homomorphism $$\label{gjdef} g_j:R^{(j)}\rightarrow R^{(j+1)}/\langle x_j^{(j+1)}\rangle \quad \text{defined by} \quad g_j(x_i^{(j)}):=x_i^{(j+1)} + \langle x_j^{(j+1)}\rangle \ \ \forall\, i \in { [ \hspace{-0.65mm} [}1,N {] \hspace{-0.65mm} ]}.$$ (For more details see [@c1 Lemme 4.3.2].) It was proved in [@c1 Proposition 4.3.1] that $\varphi_j$ induces an inclusion-preserving and -reflecting homeomorphism from the topological space $$\{P \in {\rm Spec}(R^{(j+1)}) \mid x_j^{(j+1)} \notin P \}$$ onto $$\{Q \in {\rm Spec}(R^{(j)}) \mid x_j^{(j)} \notin Q \};$$ also, $\varphi_j$ induces an inclusion-preserving and -reflecting homeomorphism from $$\{P \in {\rm Spec}(R^{(j+1)}) \mid x_j^{(j+1)} \in P \}$$ onto its image under $\varphi_j$. Note however that, in general, $\varphi_j$ is not a homeomorphism from ${\rm Spec}(R^{(j+1)})$ onto its image. Composing these embeddings, we get an embedding $$\varphi:=\varphi_2 \circ \dots \circ \varphi_N : {\rm Spec}(R) \longrightarrow {\rm Spec}({\overline{R}}),$$ which is called the *canonical embedding* from ${\rm Spec}(R)$ into ${\rm Spec}({\overline{R}})$. Cauchon diagrams ---------------- For any subset $w$ of $\{1,\dots,N\}$, let $K_w$ denote the $\ch$-prime ideal of $\widebar{R}$ generated by the $T_i$ with $i\in w$. A subset $w\subseteq\{1,\dots,N\}$ is said to be a [*Cauchon diagram*]{} for $R$ if $$K_w= \langle T_i\mid i\in w\rangle \in \varphi(\hspec(R)),$$ in which case we denote by $J_w$ the unique $\ch$-prime ideal of $R$ such that $$\varphi(J_w) = K_w \,.$$ A useful way to represent a Cauchon diagram $w$ is as follows. Draw $N$ boxes in a row, and colour the $i$-th box black if and only $i\in w$; the remaining boxes are coloured white. For example, if $N=5$ and $w=\{1,2,5\}$ we draw the diagram (0,2) rectangle (1,3); (1,2) rectangle (2,3); (2,2) rectangle (3,3); (3,2) rectangle (4,3); (4,2) rectangle (5,3); (0,2) rectangle (1,3); (1,2) rectangle (2,3); (4,2) rectangle (5,3); (In fact, the term “Cauchon diagram” originates from Cauchon’s use of a related representation in the case of quantum matrices.) We write $\#{\rm black}(w)$ and $\#{\rm white}(w)$ for the number of black and white boxes, respectively, in a Cauchon diagram $w$. That is, $\#{\rm black}(w) = |w|$ and $\#{\rm white}(w) = N - |w|$. In Section \[building\] we will investigate a way in which we can recolour boxes in a given Cauchon diagram so that the recoloured diagram is still a Cauchon diagram. This will provide us with a way of constructing descending chains of $\ch$-prime ideals of $R$. We shall need the fact that the deleting derivations process is $\ch$-equivariant, as we now indicate. The action of $\ch$ on $R$ of course extends to an action of $\ch$ on $\operatorname{Fract}(R)$ by ${{\mathbb K}}$-algebra automorphisms (although this action is not rational). Given any $\ch$-eigenvector $v \in \operatorname{Fract}(R)$, denote by $\chi(v)$ the $\ch$-eigenvalue of $v$, so that $h\cdot v = [\chi(v)(h)] v$ for all $h \in \ch$. \[chi.xji\] For all $j \in { [ \hspace{-0.65mm} [}2,N+1 {] \hspace{-0.65mm} ]}$ and $i \in { [ \hspace{-0.65mm} [}1,N {] \hspace{-0.65mm} ]}$, the element $x_i^{(j)}$ is an $\ch$-eigenvector with $\chi(x_i^{(j)}) = \chi(x_i)$. We proceed by induction on $j = N+1,\dots,2$, the case $J=N+1$ holding trivially. Assume that $j<N+1$ and that the statement holds for all $x_i^{(j+1)}$. The statement for $x_i^{(j)}$ then holds trivially in case $i \ge j$, so assume that $i<j$. By [@c1 Proposition 2.2], the map $\theta : {{\mathbb K}}\langle x_1^{(j+1)}, \dots, x_{j-1}^{(j+1)} \rangle \rightarrow \operatorname{Fract}(R)$ given by $$\theta(a) = \displaystyle{\sum_{k=0}^{+\infty } \frac{(1-q_j)^{-k}}{[k]!_{q_j}} } \delta_j^k \circ \sigma_j^{-k} (a) (x_j^{(j+1)})^{-k}$$ is a ${{\mathbb K}}$-algebra homomorphism, and $\theta$ is $\ch$-equivariant by [@llr-ufd Lemma 2.6]. Therefore $x_i^{(j)} = \theta(x_i^{(j+1)})$ is an $\ch$-eigenvector with $\chi(x_i^{(j)}) = \chi(x_i^{(j+1)}) = \chi(x_i)$, as required. For any $j \in { [ \hspace{-0.65mm} [}2,N {] \hspace{-0.65mm} ]}$, the algebra $R^{(j+1)}$ is generated by $\ch$-eigenvectors and its ideal $\langle x_j^{(j+1)} \rangle$ is $\ch$-invariant, so $R^{(j+1)}/\langle x_j^{(j+1)} \rangle$ inherits an induced $\ch$-action. In view of Lemma \[chi.xji\], we obtain the following \[gjHequi\] For each $j \in { [ \hspace{-0.65mm} [}1,N {] \hspace{-0.65mm} ]}$, the homomorphism $g_j : R^{(j)} \rightarrow R^{(j+1)}/\langle x_j^{(j+1)} \rangle$ of is $\ch$-equivariant. Gelfand-Kirillov dimension and transcendence degree =================================================== We denote the Gelfand-Kirillov dimension of a ${{\mathbb K}}$-algebra $A$ by ${{\rm GK}}(A)$. A standard reference for results concerning Gelfand-Kirillov dimension is [@kl]. Three key results that we need are the following. \[theorem-gkdim\] Let $R\ = \ {{\mathbb K}}[x_1][x_2;\sigma_2,\delta_2]\cdots[x_N;\sigma_N,\delta_N]$ be a quantum nilpotent algebra. Then ${{\rm GK}}(R) = N$. This follows easily from [@lr Lemma 2.3]. \[theorem-gkdrop\] Let $A$ be a noetherian ${{\mathbb K}}$-algebra and let $P$ be a prime ideal of $A$. Then $${{\rm GK}}(A/P) + {{\rm ht}}(P) \le {{\rm GK}}(A).$$ Noetherianness is more than is needed here – the result holds if all prime factor rings of $A$ are right Goldie [@kl Corollary 3.16]. \[modfinext\] If $B \subseteq A$ are ${{\mathbb K}}$-algebras such that $A$ is finitely generated as a right $B$-module, then $${{\rm GK}}(A) = {{\rm GK}}(B).$$ [@kl Proposition 5.5]. We shall also make use of the Gelfand-Kirillov transcendence degree of ${{\mathbb K}}$-algebras $A$, denoted $\operatorname{{\rm Tdeg}}(A)$. See [@kl] or [@z1], for instance, for the precise definition. \[def.tdegstable\] A ${{\mathbb K}}$-algebra $A$ is said to be $\operatorname{{\rm Tdeg}}$-*stable* [@z1] if the following hold: 1. ${{\rm GK}}(A)=\operatorname{{\rm Tdeg}}(A)$. 2. For every multiplicative system $S$ of regular elements of $A$ that satisfies the Ore condition, we have: $\operatorname{{\rm Tdeg}}(S^{-1}A)=\operatorname{{\rm Tdeg}}(A)$. A key instance of this property is \[qtorTdegstable\] Every quantum torus is Tdeg-stable. [@Lor Corollary 2.2] or [@z1 Proposition 7.2]. We excerpt the following key result from [@z1]. \[transfer.Tdeg\] [[@z1 Proposition 3.5(4)]]{} Let $A$ be a semiprime Goldie ${{\mathbb K}}$-algebra and let $B \subseteq A$ be a semiprime Goldie subalgebra such that $\operatorname{Fract}B = \operatorname{Fract}A$. If $A$ is $\operatorname{{\rm Tdeg}}$-stable, then ${{\rm GK}}(B) = {{\rm GK}}(A)$ and $B$ is $\operatorname{{\rm Tdeg}}$-stable. An immediate application is that any quantum nilpotent algebra $R$ is $\operatorname{{\rm Tdeg}}$-stable, since, by [@gy0 Theorem 4.6], $R$ is trapped between a quantum affine space $\co_{\mathbf{q}}({{\mathbb K}}^N)$ and the corresponding quantum torus $\co_{\mathbf{q}}(({{\mathbb K}}^*)^N)$. In fact, $\operatorname{{\rm Tdeg}}$-stability holds for all $\ch$-prime factors of $R$, by the same argument used in [@stephane Proposition 1.3.2.2] for the case of quantum matrices. This relies on a result of Cauchon [@c1 Théorème 5.4.1] which shows that for any Cauchon diagram $w$ of $R$, the algebras $R/J_w$ and ${\overline{R}}/K_w$ have isomorphic localizations. For later use, we require the following $\ch$-equivariant version of the result. Recall that the algebra ${\overline{R}}/K_w$ is a quantum affine space with canonical generators given by the cosets of those $T_i$ with $i \in { [ \hspace{-0.65mm} [}1,N {] \hspace{-0.65mm} ]}\setminus w$, where $N$ is the length of $R$. Let ${\mathcal{E}}_w$ denote the multiplicative set in ${\overline{R}}/K_w$ generated by $\{ T_i \mid i \in { [ \hspace{-0.65mm} [}1,N {] \hspace{-0.65mm} ]}\setminus w \}$. Observe that ${\mathcal{E}}_w$ is an Ore set in ${\overline{R}}/K_w$ and that $({\overline{R}}/K_w){\mathcal{E}}_w^{-1}$ is a quantum torus of rank $N-|w|$. \[matchloc2\] Let $R$ be a quantum nilpotent algebra of length $N$ and let $w$ be a Cauchon diagram for $R$. There exists an Ore set ${\mathcal{F}}_w$ of regular $\ch$-eigenvectors in $R/J_w$ such that [(a)]{} There is an $\ch$-equivariant ${{\mathbb K}}$-algebra isomorphism $(R/J_w){\mathcal{F}}_w^{-1} \rightarrow ({\overline{R}}/K_w){\mathcal{E}}_w^{-1}$. [(b)]{} $\spec_{J_w}(R) = \{ P \in \spec(R) \mid P\supseteq J_w \ \text{and} \ (P/J_w) \cap {\mathcal{F}}_w = \varnothing \}$. The Ore set we label ${\mathcal{F}}_w$ is denoted $\Sigma_{N+1}$ in [@c1 Subsection 5], where we take $P=J_w$. It is obtained as the end result of a sequence of Ore sets $\Sigma_2, \dots, \Sigma_{N+1}$ in subalgebras $A_2\cong {\overline{R}}/K_w,\dots,A_{N+1} = R/J_w$ of $\operatorname{Fract}(R/J_w)$, where $\Sigma_2$ is the image of ${\mathcal{E}}_w$ in $A_2$. Specifically, $\Sigma_{j+1} = \Sigma_j \cap A_{j+1}$ for $j \in { [ \hspace{-0.65mm} [}2,N{] \hspace{-0.65mm} ]}$. The action of $\ch$ on $R/J_w$ by automorphisms extends to an action on $\operatorname{Fract}(R/J_w)$ by automorphisms (although no longer rational). As one notes, the elements of $\Sigma_2$ are regular $\ch$-eigenvectors, so the same holds for $\Sigma_{N+1} = {\mathcal{F}}_w$. That $\Sigma_{N+1}$ satisfies the Ore condition in $A_{N+1}$ is proved in [@c1 Proposition 5.4.4(2)]. \(a) This isomorphism corresponds to an equality in [@c1], due to identifications made in that paper. In [@c1], ${\overline{R}}/K_w$ is identified with $A_2$ via a ${{\mathbb K}}$-algebra epimorphism $f_2 : R_2 = {\overline{R}}\rightarrow A_2$ with kernel $K_w$ [@c1 Proposition 5.4.1(2)]. It follows from Lemma \[chi.xji\] that $f_2$ is $\ch$-equivariant. Therefore the induced isomorphism $({\overline{R}}/K_w){\mathcal{E}}_w^{-1} \rightarrow A_2 \Sigma_2^{-1}$ is $\ch$-equivariant. Since $A_2 \Sigma_2^{-1} = A_{N+1} \Sigma_{N+1}^{-1} = (R/J_w) {\mathcal{F}}_w^{-1}$ [@c1 Proposition 5.4.4(3)], we obtain (the inverse of) the desired $\ch$-equivariant isomorphism. \(b) If $P \in \spec_{J_w}(R)$, then $P/J_w$ contains no nonzero $\ch$-invariant ideals, so it cannot contain any $\ch$-eigenvectors. Hence, $(P/J_w) \cap {\mathcal{F}}_w = \varnothing$. Conversely, suppose $P \in \spec(R)$ with $P \supseteq J_w$ and $(P/J_w) \cap {\mathcal{F}}_w = \varnothing$. Then $I = (P:\ch)$ is a proper $\ch$-invariant ideal of $R$ such that $I \supseteq J_w$ and $(I/J_w) \cap {\mathcal{F}}_w = \varnothing$, and so $(I/J_w){\mathcal{F}}_w^{-1}$ is a proper $\ch$-invariant ideal of $(R/J_w){\mathcal{F}}_w^{-1}$. Under the isomorphism in (a), $(I/J_w){\mathcal{F}}_w^{-1}$ corresponds to a proper $\ch$-invariant ideal $I'$ of $({\overline{R}}/K_w){\mathcal{E}}_w^{-1}$. However, by [@c1 Lemme 5.5.3], $({\overline{R}}/K_w){\mathcal{E}}_w^{-1}$ is an $\ch$-simple ring, i.e., it has no nonzero proper $\ch$-invariant ideals. Consequently, $I' = 0$, whence $(I/J_w){\mathcal{F}}_w^{-1} = 0$. This forces $I/J_w = 0$, that is, $(P:\ch) = I = J_w$. Therefore $P \in \spec_{J_w}(R)$. \[GKRmodJw\] If $R$ is a quantum nilpotent algebra of length $N$ and $w$ is a Cauchon diagram for $R$, then $R/J_w$ is $\operatorname{{\rm Tdeg}}$-stable and ${{\rm GK}}(R/J_w) = N-|w| = \#{\rm white}(w)$. By Lemma \[qtorTdegstable\], the quantum torus $({\overline{R}}/K_w){\mathcal{E}}_w^{-1}$ is $\operatorname{{\rm Tdeg}}$-stable, and we observe that this algebra has GK-dimension $N-|w|$. In view of Theorem \[matchloc2\], the algebra $(R/J_w){\mathcal{F}}_w^{-1}$ has the same properties. Since $\operatorname{Fract}R/J_w = \operatorname{Fract}(R/J_w){\mathcal{F}}_w^{-1}$, Proposition \[transfer.Tdeg\] yields the desired conclusions. Building height of $\ch$-primes by using black boxes {#building} ==================================================== Let $R\ = \ {{\mathbb K}}[x_1][x_2;\sigma_2,\delta_2]\cdots[x_N;\sigma_N,\delta_N]$ be a quantum nilpotent algebra of length $N$. In Theorem \[theorem-gkdim\] we have seen that ${{\rm GK}}(R)=N$, and Theorem \[theorem-gkdrop\] says that $${{\rm GK}}(R/P)+{{\rm ht}}(P) \leq {{\rm GK}}(R)$$ for each prime ideal $P$ of $R$. We aim to prove that these inequalities are actually equalities. In this section, we establish the height formula for $\ch$-prime ideals of $R$. As an abbreviation, we write $P{\vartriangleleft'_\ch}A$ to denote that $P$ is an $\ch$-prime ideal in $A$ for any ring $A$ equipped with a group action by $\ch$. We will show that if we take the Cauchon diagram for an $\ch$-prime ideal $P$ of $R$ and change the final black box to a white box then we again get the Cauchon diagram for an $\ch$-prime ideal and that this new $\ch$-prime ideal is necessarily contained in the original $\ch$-prime ideal. In this way, we can build a descending chain of $\ch$-prime ideals with length equal to the number of black boxes. This provides a lower bound for the height of $P$. Combining this with Corollary \[GKRmodJw\] and Theorem \[theorem-gkdim\], we see that Tauvel’s height formula for the $\ch$-primes of $R$ then follows easily, and that the height of an $\ch$-prime $P$ of $R$ is equal to the number of black boxes in the Cauchon diagram for $P$. Black box removal ----------------- Fix an $\ch$-prime ideal $J_w$ of $R$, with $w$ nonempty. Let $k$ be the maximal member of $w$. We want to show that $w':= w\backslash\{k\}$ is a Cauchon diagram, so that there is an $\ch$-prime ideal $J_{w'}$ in $R$ with $\varphi(J_{w'})= K_{w'}$. In order to do this, we need to reverse the procedure described in Section \[canonicalembedding\]. We keep the notation of that section; in particular, $\widebar{R}$ denotes the quantum affine space that is reached at the end of the deleting derivations process, and $\varphi:\spec(R)\hookrightarrow \spec(\widebar{R})$ is the canonical embedding. Recall the iterated Ore extension presentation of $R^{(j)}$ from . When $j\leq k$, because of the nature of the automorphisms $\tau_i$ with $i\geq j$, we may write (with a slight abuse of notation) $$R^{(j)} = A^{(j)}[T_k;\tau_k],$$ where $$A^{(j)}:= {{\mathbb K}}[x_1^{(j)}]\cdots[x_{j-1}^{(j)};\sigma_{j-1},\delta_{j-1}][T_j;\tau_j]\cdots [T_{k-1};\tau_{k-1}][T_{k+1};\tau_{k+1}]\cdots [T_N;\tau_N].$$ Here for $i>k$ we have written $\tau_i$ for the restriction of the original $\tau_i$ to the algebra $${{\mathbb K}}\langle x_1^{(j)}, \dots, x_{j-1}^{(j)}, T_j, \dots, T_{k-1}, T_{k+1}, \dots, T_{i-1} \rangle.$$ The following technical lemma, needed in the next result, gives a sufficient criterion for recognising when an $\ch$-ideal is induced from the base ring in an Ore extension endowed with a suitable $\ch$-action. \[lemma-induced\] Let $B=A[X;\sigma]$ be an Ore extension of ${{\mathbb K}}$-algebras, and assume that $\sigma$ extends to an automorphism ${\widehat{\sigma}}$ of $B$ such that ${\widehat{\sigma}}(X) = q X$ for some $q\in {{\mathbb K}}^*$ which is not a root of unity. Let $I$ be a ${\widehat{\sigma}}$-invariant ideal of $B$, and suppose that $aX^n \in I$ implies $a\in I$, for any $a\in A$ and $n \in \mn$. Then 1. $I= (I\cap A)B$. 2. The natural map $(A/(I\cap A))[\widebar{X};\widebar{\sigma}] \rightarrow B/I$ is an isomorphism. \(i) We adapt the proof of [@llr-ufd Lemma 2.2]. If (i) fails, there exists an element $b = \sum_i b_iX^i \in I$ with all $b_i \in A$ but some $b_j \notin I$. Set $m := \min \{ i \mid b_i \notin I \}$ and $n := \max \{ i \mid b_i \notin I \}$. Since $b_i \in I$ for $i \notin { [ \hspace{-0.65mm} [}m,n {] \hspace{-0.65mm} ]}$, we may remove terms with these indices from $b$, that is, there is no loss of generality in assuming $b = \sum_{i=m}^n b_i X^i$. If $m=n$, then $b_n X^n \in I$ with $b_n \notin I$, contradicting our hypotheses. Thus, $m<n$. Without loss of generality, $m-n$ is minimal among instances of elements with the properties of $b$. Now $I$ contains the elements $$Xb = \sum_{i=m}^n \sigma(b_i) X^{i+1} \qquad \text{and} \qquad {\widehat{\sigma}}(b) = \sum_{i=m}^n q^i \sigma(b_i) X^i,$$ and so it also contains the element $${\widehat{\sigma}}(b) X - q^n X b = \sum_{i=m}^{n-1} (q^i - q^n) \sigma(b_i) X^{i+1}.$$ The minimality of $m-n$, together with the assumption that $q$ is not a root of unity, implies that $\sigma(b_i) \in I$ for all $i \in { [ \hspace{-0.65mm} [}m,n-1 {] \hspace{-0.65mm} ]}$. But then $b_m = {\widehat{\sigma}}^{-1}(\sigma(b_m)) \in I$, a contradiction. \(ii) This follows easily from (i), e.g. by using the comment in [@gl1 2.1(vi)]. \[largerCauchon\] Let $w$ be a nonempty Cauchon diagram for $R$ and let $k := \max w$. The set $w':=w\backslash\{k\}$ is a Cauchon diagram with $J_{w'} \subsetneqq J_w$. If $w = \{k\}$, then $w' = \varnothing$. Clearly $\varphi(0) = 0 = K_\varnothing \in {{\rm Spec}}({\overline{R}})$, so $w'$ is a Cauchon diagram and $J_{w'} = 0 \subsetneqq J_w$. Hence, we may assume that $w \supsetneqq \{k\}$; in particular, $k\geq 2$. Note that if $s\neq k$ then $s\in w'$ if and only if $s\in w$. As $w$ is a Cauchon diagram, there exist $\ch$-prime ideals $J_w^{(i)}{\vartriangleleft'_\ch}R^{(i)}$ for $i \in { [ \hspace{-0.65mm} [}2,N+1 {] \hspace{-0.65mm} ]}$ such that $\varphi_i(J_w^{(i+1)})=J_w^{(i)}$ for $i \in { [ \hspace{-0.65mm} [}2,N {] \hspace{-0.65mm} ]}$, where $J_w^{(N+1)} = J_w$ and $J_w^{(2)} = K_w$. We aim to construct a corresponding sequence of $\ch$-prime ideals $J_{w'}^{(i)} {\vartriangleleft'_\ch}R^{(i)}$. First, we show that there are $\ch$-prime ideals $J_{w'}^{(2)}{\vartriangleleft'_\ch}R^{(2)},\dots,J_{w'}^{(k)}{\vartriangleleft'_\ch}R^{(k)}$ with $A^{(i)}\cap J_{w}^{(i)}\subseteq J_{w'}^{(i)}$ for each $i \in { [ \hspace{-0.65mm} [}2,k {] \hspace{-0.65mm} ]}$ and $\varphi_i(J_{w'}^{(i+1)})=J_{w'}^{(i)}$ for each $i \in { [ \hspace{-0.65mm} [}2,k-1 {] \hspace{-0.65mm} ]}$. To begin, set $J_{w'}^{(2)} :=K_{w'}$ and observe that $A^{(2)}\cap J_w^{(2)} = \langle T_j \mid j\in w' \rangle_{A^{(2)}}\subseteq J_{w'}^{(2)}$. Now assume that $2\leq s<k$ and that $\ch$-prime ideals $J_{w'}^{(2)}{\vartriangleleft'_\ch}R^{(2)},\dots,J_{w'}^{(s)}{\vartriangleleft'_\ch}R^{(s)}$ have been defined with $A^{(i)}\cap J_{w}^{(i)}\subseteq J_{w'}^{(i)}$ for each $i \in { [ \hspace{-0.65mm} [}2,s {] \hspace{-0.65mm} ]}$ and $\varphi_i(J_{w'}^{(i+1)})=J_{w'}^{(i)}$ for each $i \in { [ \hspace{-0.65mm} [}2,s-1 {] \hspace{-0.65mm} ]}$. We need to find an $\ch$-prime ideal $J_{w'}^{(s+1)}$ in $R^{(s+1)}$ such that $A^{(s+1)}\cap J_{w}^{(s+1)}\subseteq J_{w'}^{(s+1)}$ and $\varphi_s(J_{w'}^{(s+1)})=J_{w'}^{(s)}$. We distinguish between two possible cases, depending on whether or not $s \in w'$. [**Case (i)**]{} Assume that $s\not\in w'$. As $s<k$, we know that $s\not\in w$. In this case, set $J_{w'}^{(s+1)}:= J_{w'}^{(s)}[T_s^{-1}]\cap R^{(s+1)}$. It is easy to see that $J_{w'}^{(s+1)}{\vartriangleleft'_\ch}R^{(s+1)}$ (observe that $T_s \notin J^{(2)}_{w'}$ implies $T_s \notin J^{(s)}_{w'}$) and we can check (using ) that $\varphi_s(J_{w'}^{(s+1)})=J_{w'}^{(s)}$. Now, $$\begin{aligned} A^{(s+1)}\cap J_{w}^{(s+1)} &= A^{(s+1)}\cap \bigl( J_{w}^{(s)}[T_s^{-1}]\cap R^{(s+1)} \bigr) \subseteq A^{(s)}[T_{s}^{-1}]\cap J_{w}^{(s)}[T_s^{-1}]\cap R^{(s+1)} \\ &\subseteq (A^{(s)}\cap J_{w}^{(s)})[T_s^{-1}]\cap R^{(s+1)} \subseteq J_{w'}^{(s)}[T_s^{-1}]\cap R^{(s+1)} = J_{w'}^{(s+1)},\end{aligned}$$ as required to finish Case (i). (Here, the first containment and the last equality follow from , and the last containment is given by the inductive hypothesis.) [**Case (ii)**]{} Assume that $s\in w'$. In this case, $s\in w$. We have $T_s \in J^{(2)}_{w'}$ and $T_s \in J^{(2)}_w$. It follows that $T_s \in J^{(s)}_{w'}$ as well as $T_s \in J^{(s)}_w$ and $T_s \in J^{(s+1)}_w$. In order to set up an application of Lemma \[lemma-induced\], let $$\begin{aligned} A^{(s+1)}_{k-1} &:= {{\mathbb K}}\langle x_1^{(s+1)}, \dots, x_{s}^{(s+1)}, T_{s+1}, \dots, T_{k-1} \rangle \subset A^{(s+1)} \\ R^{(s+1)}_k &:= {{\mathbb K}}\langle x_1^{(s+1)}, \dots, x_s^{(s+1)}, T_{s+1}, \dots, T_k \rangle \subset R^{(s+1)},\end{aligned}$$ so that $R^{(s+1)}_k = A^{(s+1)}_{k-1}[T_k; \tau_k]$. In view of Lemma \[chi.xji\], the restriction of $(h_k\cdot)$ to $R^{(s+1)}_k$ (where $h_k$ is the element of $\ch$ occurring in part (iii) of Definition \[CGLdef\]) yields an automorphism ${\widehat{\tau}}_k$ which extends $\tau_k$ and satisfies ${\widehat{\tau}}_k(T_k) = q_k T_k$. [**Claim**]{} Suppose that $a\in A^{(s+1)}_{k-1}$ and $aT_k^n \in \langle T_s\rangle_{R^{(s+1)}_k}$ for some $n \ge 0$. Then $a\in \langle T_s\rangle_{A^{(s+1)}_{k-1}}$. [*Proof of Claim*]{} Recall from that $T_kT_s=\lambda T_sT_k$ where $\lambda = \lambda_{k,s} \in {{\mathbb K}}^*$. Suppose that $aT_k^n =\sum_i\, c_iT_sd_i$, with $c_i,d_i\in R^{(s+1)}_k$. Write $c_i=\sum_{\alpha}\, c_{i\alpha}T_k^\alpha$ and $d_i =\sum_\beta\,d_{i\beta}T_k^\beta$ with $c_{i\alpha}, d_{i\beta}\in A^{(s+1)}_{k-1}$. Thus, $$\begin{aligned} aT_k^n &= \sum_{i,\alpha,\beta}\, (c_{i\alpha}T_k^\alpha )T_s(d_{i\beta}T_k^\beta) = \sum_{i,\alpha,\beta}\, c_{i\alpha}\lambda^\alpha T_sT_k^\alpha d_{i\beta}T_k^\beta = \sum_{i,\alpha,\beta}\, c_{i\alpha}\lambda^\alpha T_s\tau_k^\alpha(d_{i\beta})T_k^\alpha T_k^\beta \\ &= \sum_{i,\alpha,\beta}\, c_{i\alpha}\lambda^\alpha T_s\tau_k^\alpha(d_{i\beta})T_k^{\alpha+\beta} = \sum_{m=0}^\infty\, \biggl(\, \sum_i\, \sum_{\alpha+\beta = m}\, c_{i\alpha}\lambda^\alpha T_s\tau_k^\alpha(d_{i\beta}) \biggr) T_k^m.\end{aligned}$$ Since $R^{(s+1)}_k =A^{(s+1)}_{k-1}[T_k;\tau_k]$ is an Ore extension, it follows that $$a= \sum_i\, \sum_{\alpha=0}^n c_{i\alpha}\lambda^\alpha T_s\tau_k^\alpha(d_{i, n-\alpha}) \in \langle T_s\rangle_{A^{(s+1)}_{k-1}} \,,$$ as required to establish the truth of the claim. The case $n=0$ of the claim implies that $\langle T_s\rangle_{R^{(s+1)}_k} \cap A^{(s+1)}_{k-1} = \langle T_s\rangle_{A^{(s+1)}_{k-1}}$. Applying Lemma \[lemma-induced\], we obtain $$\langle T_s\rangle_{R^{(s+1)}_k} = \langle T_s\rangle_{A^{(s+1)}_{k-1}} R^{(s+1)}_k \,.$$ Since the ideals $\langle T_s\rangle_{R^{(s+1)}_k}$ and $\langle T_s\rangle_{A^{(s+1)}_{k-1}}$ are invariant under $\tau_{k+1}, \dots, \tau_N$, it follows that $$\langle T_s\rangle_{R^{(s+1)}} = \langle T_s\rangle_{R^{(s+1)}_k} R^{(s+1)} \qquad \text{and} \qquad \langle T_s\rangle_{A^{(s+1)}} = \langle T_s\rangle_{A^{(s+1)}_{k-1}} A^{(s+1)},$$ whence $\langle T_s\rangle_{R^{(s+1)}} = \langle T_s\rangle_{A^{(s+1)}} R^{(s+1)}$. Consequently, $$\label{R/T.isom} \text{the natural map} \;\; \left( A^{(s+1)}/\langle T_s\rangle_{A^{(s+1)}}\right) [\widebar{T}_k;\widebar{\sigma}_k] \longrightarrow R^{(s+1)}/\langle T_s\rangle \;\; \text{is an isomorphism}.$$ Consider the map $g_s:R^{(s)}\longrightarrow R^{(s+1)}/\langle T_s\rangle$ that arises in the deleting derivations process . This map induces an isomomorphism from $R^{(s)}/ \ker(g_s)$ to $R^{(s+1)}/ \langle T_s\rangle$, and we know that $\ker(g_s)\subseteq J_{w}^{(s)} = g_s^{-1}(J_{w}^{(s+1)}/\langle T_s\rangle)$. In fact, there is the isomorphism $$R^{(s)}/J_{w}^{(s)}\cong R^{(s+1)}/J_{w}^{(s+1)}$$ that is induced by $g_s$. We shall prove that $\ker(g_s)\subseteq J_{w'}^{(s)}$. Let $x\in\ker(g_s)$, and write $x=\sum_i a_i^{(s)}T_k^i$ with each $a_i^{(s)}\in A^{(s)}$. Now by , $g_s(x)=0$ implies $g_s(a_i^{(s)}) = 0$, for each $i$; so $a_i^{(s)}\in\ker(g_s)\cap A^{(s)} \subseteq J_w^{(s)}\cap A^{(s)} \subseteq J_{w'}^{(s)}$, where the final containment is given by the inductive hypothesis. Hence, $\ker(g_s)\subseteq J_{w'}^{(s)}$, as required. Since $g_s$ is $\ch$-equivariant (Corollary \[gjHequi\]), it follows that there is an $\ch$-prime ideal $J_{w'}^{(s+1)}{\vartriangleleft'_\ch}R^{(s+1)}$ such that $J^{(s+1)}_{w'} \supseteq \langle T_s\rangle$ and $g_s$ induces an isomorphism $$R^{(s)}/J_{w'}^{(s)}\cong R^{(s+1)}/J_{w'}^{(s+1)}.\\[2ex]$$ In particular, $J^{(s)}_{w'} = g_s^{-1}(J^{(s+1)}_{w'}/\langle T_s\rangle) = \varphi_s(J^{(s+1)}_{w'})$. Let $z\in A^{(s+1)}\cap J_w^{(s+1)}$. There exists $z'\in A^{(s)}\cap J_w^{(s)}$ such that $g_s(z')=\widebar{z}\in R^{(s+1)}/\langle T_s\rangle$. Now, $z'\in J_{w'}^{(s)}$, by the inductive hypothesis; so $\widebar{z}=g_s(z')\in g_s(J_{w'}^{(s)}) = J_{w'}^{(s+1)}/\langle T_s\rangle$. Since $T_s\in J_{w'}^{(s+1)}$, this establishes the required inclusion $A^{(s+1)}\cap J_{w}^{(s+1)}\subseteq J_{w'}^{(s+1)}$ in this case. This finishes Case (ii). At this stage, we have constructed $\ch$-prime ideals $J_{w'}^{(i)}{\vartriangleleft'_\ch}R^{(i)}$ for $i \in { [ \hspace{-0.65mm} [}2,k {] \hspace{-0.65mm} ]}$ with $\varphi_i(J_{w'}^{(i+1)})=J_{w'}^{(i)}$ for $i= { [ \hspace{-0.65mm} [}2,k-1 {] \hspace{-0.65mm} ]}$. We still need to construct $\ch$-prime ideals $J_{w'}^{(i)}{\vartriangleleft'_\ch}R^{(i)}$ for $i \in { [ \hspace{-0.65mm} [}k+1,N+1 {] \hspace{-0.65mm} ]}$ with $\varphi_i(J_{w'}^{(i+1)})=J_{w'}^{(i)}$ for $i \in { [ \hspace{-0.65mm} [}k,N {] \hspace{-0.65mm} ]}$. However, for $s\geq k$, we know that $s\not\in w'$; so the same reasoning as in the first part of Case (i) above does what is required. Now $J_{w'}^{(N+1)}$ is an $\ch$-prime ideal of $R$ with $\varphi(J_{w'}^{(N+1)}) = J^{(2)}_{w'} = K_{w'}$, whence $w'$ is the Cauchon diagram for $J_{w'}^{(N+1)}$ and $J_{w'}^{(N+1)} = J_{w'}$. The fact that $J_{w'}\subsetneqq J_w$ follows from [@bl Theorem 1.4] \[htHprime\] Let $J$ be an $\ch$-prime ideal in the quantum nilpotent algebra $R$. If $w$ is the Cauchon diagram of $J$, then ${{\rm ht}}(J)\geq \#{\rm black}(w)$. Height formula for $\ch$-primes ------------------------------- \[Tauvel-Hprimes\] Let $J$ be an $\ch$-prime ideal of the quantum nilpotent algebra $R$ with Cauchon diagram $w$. Then $${{\rm GK}}(R/J) + {{\rm ht}}(J)={{\rm GK}}(R);$$ that is, Tauvel’s height formula holds for the $\ch$-prime ideals of any quantum nilpotent algebra. Furthermore, ${{\rm GK}}(R/J)=\#{\rm white}(w)$ and ${{\rm ht}}(J)=\#{\rm black}(w)$. We already have ${{\rm GK}}(R/J)=\#{\rm white}(w)$, by Corollary \[GKRmodJw\]. By using Corollary \[htHprime\] and Theorem \[theorem-gkdrop\], we see that $$N={{\rm GK}}(R)\geq {{\rm GK}}(R/J)+{{\rm ht}}(J)\geq \#{\rm white}(w)+\#{\rm black}(w)= N.$$ Tauvel’s height formula for $J$ follows, as does the claim about ${{\rm ht}}(J)$. In order to extend this result to arbitrary prime ideals, we need to further employ Tdeg-stability. The necessary details are given in the next section. The proof of Corollary \[htHprime\] shows that for any $\ch$-prime ideal $J$ of $R$ with Cauchon diagram $w$, there is a strictly descending chain $J_0 = J \supsetneq J_1 \supsetneq \cdots \supsetneq J_m = 0$ of $\ch$-primes of $R$ with $m = \#{\rm black}(w)$. In other words, the height of $J$ within the poset $\hspec R$ is at least $\#{\rm black}(w)$. Since this value, which we denote ${{\rm ht}}_{\hspec R}(J)$, is dominated by ${{\rm ht}}_{\spec R}(J) := {{\rm ht}}(J)$, we obtain the following from Theorem \[Tauvel-Hprimes\]. \[Hht=ht\] If $J$ is any $\ch$-prime ideal of the quantum nilpotent algebra $R$, then $$\label{HhtJ=htJ} {{\rm ht}}_{\hspec R}(J) = {{\rm ht}}_{\spec R}(J).$$ Equation had previously been established only under the hypothesis that $\hspec R$ has *$\ch$-normal separation*, meaning that for any $\ch$-prime ideals $J \subsetneq K$ of $R$, there is an $\ch$-eigenvector $u \in K \setminus J$ such that $u+J$ is normal in $R/J$ and the corresponding automorphism of $R/J$ is given by some element of $\ch$ (see [@y1 Proposition 5.9]). As a consequence, was known for quantum nilpotent algebras of the form $U^w_-({{\frak g}})$ (i.e., *quantum Schubert cell algebras*) [@y1 Proof of Theorem 5.8] and for cocycle twists of the $U^w_-({{\frak g}})$ [@y2 Section 5]. T-degree stability for primitive quotients of quantum nilpotent algebras ======================================================================== We now require some more precise information about $\operatorname{{\rm Tdeg}}$ and $\operatorname{{\rm Tdeg}}$-stability. This will be obtained using the *lower transcendence degree* (over ${{\mathbb K}}$) of a ${{\mathbb K}}$-algebra $A$, as defined in [@z2]. This degree, denoted $\operatorname{Ld}(A)$, is a value in $\mr_{\ge0} \cup \{\infty\}$; we refer to [@z2] for the definition. (We do not require lower transcendence degrees over division subalgebras of $A$.) \[finiteextLd\] Let $B \subseteq A$ be prime Goldie ${{\mathbb K}}$-algebras such that all regular elements of $B$ are also regular in $A$. If $A$ is finitely generated as a right $B$-module, then $\operatorname{Ld}(A) = \operatorname{Ld}(B)$. By [@z2 Theorem 0.3(2)], $\operatorname{Ld}(A) = \operatorname{Ld}(\operatorname{Fract}(A))$ and $\operatorname{Ld}(B) = \operatorname{Ld}(\operatorname{Fract}(B))$, so it remains to show that $\operatorname{Ld}(\operatorname{Fract}(A)) = \operatorname{Ld}(\operatorname{Fract}(B))$. This will follow from [@z2 Theorem 0.3(1)] once we show that $\operatorname{Fract}(A)$ is finitely generated as a right $\operatorname{Fract}(B)$-module, since $\operatorname{Fract}(B)$ is artinian. Due to the assumption on regular elements, we can identify $\operatorname{Fract}(B)$ with a subalgebra of $\operatorname{Fract}(A)$. We have $A = \sum_{i=1}^n a_iB$ for some $a_i \in A$. Set $D := A\cdot\operatorname{Fract}(B) = \sum_{i=1}^n a_i \operatorname{Fract}(B)$, a right $\operatorname{Fract}(B)$-submodule of $\operatorname{Fract}(A)$ which is finitely generated and thus artinian. Any regular element $a\in A$ is invertible in $\operatorname{Fract}(A)$, whence $aD \cong D$. Since $aD \subseteq D$ and $D$ is artinian on the right, $aD = D$, whence $a^{-1}D = D$. Now $D$ is a left ideal of $\operatorname{Fract}(A)$, and it contains $A$, so $D = \operatorname{Fract}(A)$. Therefore $\operatorname{Fract}(A)$ is a finitely generated right $\operatorname{Fract}(B)$-module, as desired. \[qtorusmodmax\] If $T$ is a quantum torus over ${{\mathbb K}}$ and $M$ is a maximal ideal of $T$, then $T/M$ is $\operatorname{{\rm Tdeg}}$-stable. As noted in [@z2 pp. 159–60], it suffices to show that $T/M$ is *$\operatorname{Ld}$-stable* in the sense that $\operatorname{Ld}(T/M) = {{\rm GK}}(T/M)$. Extension and contraction give inverse bijections between the sets of ideals in $Z(T)$ and $T$ (e.g., [@bg Proposition II.3.8]), so ${{\frak m}}:= M \cap Z(T)$ is a maximal ideal of $Z(T)$ and $M = {{\frak m}}T$. Since $Z(T)$ is a Laurent polynomial algebra over ${{\mathbb K}}$ (e.g., [@bg Lemma II.3.7(e)]), the field $Z(T)/{{\frak m}}$ is finite dimensional over ${{\mathbb K}}$. Write $T = {{\mathcal O}_{\mathbf{q}}}(({{\mathbb K}}^*)^n)$ for some $n \in \mn$ and some multiplicatively skewsymmetric matrix ${\mathbf{q}}= (q_{ij}) \in M_n({{\mathbb K}}^*)$. Let $y_1^{\pm1}, \dots, y_n^{\pm1}$ be a standard set of generators for $T$, so that $$T = {{\mathbb K}}\langle\, y_1^{\pm1}, \dots, y_n^{\pm1} \mid y_i y_j = q_{ij} y_j y_i \ \forall\; i,j \in { [ \hspace{-0.65mm} [}1,n {] \hspace{-0.65mm} ]}\, \rangle.$$ Then let $\{ y^a \mid a \in \mz^n \}$ be the corresponding ${{\mathbb K}}$-basis for $T$, where $$y^a := y_1^{a_1} y_2^{a_2} \cdots y_n^{a_n} \qquad \forall\; a = (a_1,\dots,a_n) \in \mz^n.$$ Set $Z := \{ a \in \mz^n \mid y^a \in Z(T) \}$, so that $Z(T) = \bigoplus_{a\in Z} {{\mathbb K}}y^a$ (e.g., [@bg Lemma II.3.7(a)]). Choose a subgroup $W \subseteq \mz^n$ maximal with respect to the property $W \cap Z = 0$. Then $\mz^n/(Z \oplus W)$ is finite, and $Z \oplus W$ is free abelian of rank $n$. Choose bases $(b_1,\dots,b_l)$ and $(b_{l+1},\dots,b_n)$ for $Z$ and $W$, respectively, and set $z_i := y^{b_i}$ for all $i$. Then $$Z(T) = \bigoplus_{a\in Z} {{\mathbb K}}y^a = {{\mathbb K}}[z_1^{\pm1}, \dots, z_l^{\pm1}].$$ Next, set $$C := \bigoplus_{a\in W} {{\mathbb K}}y^a = {{\mathbb K}}\langle z_{l+1}^{\pm1}, \dots, z_n^{\pm1} \rangle,$$ a quantum torus over ${{\mathbb K}}$ of rank $r := n-l$. Finally, set $$B := \bigoplus_{a\in Z\oplus W} {{\mathbb K}}y^a = {{\mathbb K}}\langle z_1^{\pm1}, \dots, z_n^{\pm1} \rangle,$$ a ${{\mathbb K}}$-subalgebra of $T$. Observe that the multiplication map $Z(T) \otimes_{{\mathbb K}}C \rightarrow T$ gives a ${{\mathbb K}}$-algebra isomorphism of $Z(T) \otimes_{{\mathbb K}}C$ onto $B$. We identify $B$ with $Z(T) \otimes_{{\mathbb K}}C$ via this isomorphism. We next show that $C$ is a central simple ${{\mathbb K}}$-algebra (meaning only that $C$ is a simple ring with center ${{\mathbb K}}$). Simplicity will follow from [@McPe Proposition 1.3] once we show that $Z(C) = {{\mathbb K}}$. We know that $Z(C)$ is spanned by those $y^a$ with $a\in W$ and $y^a \in Z(C)$. Let $s := |\mz^n/(Z \oplus W)|$, so that $sb \in Z \oplus W$ for all $b\in \mz^n$. If $a \in W$ and $y^a \in Z(C)$, then $y^a$ commutes with $y^{sb}$ for any $b \in \mz^n$. There is a scalar $\lambda_{a,b} \in {{\mathbb K}}^*$ such that $y^a y^b = \lambda_{a,b} y^b y^a$, whence $y^a y^{sb} = \lambda_{a,b}^s y^{sb} y^a = \lambda_{a,b}^s y^a y^{sb}$. But then $\lambda_{a,b}^s = 1$ and so $y^{sa} y^b = y^b y^{sa}$. Consequently, $y^{sa} \in Z(T)$ and $sa \in Z$. Since also $a\in W$, we must have $a=0$. Therefore $Z(C) = {{\mathbb K}}$, as required. Since $C$ is a central simple ${{\mathbb K}}$-algebra, ${{\frak m}}B = {{\frak m}}\otimes_{{\mathbb K}}C$ is a maximal ideal of $B$. Now $B/{{\frak m}}B = (Z(T)/{{\frak m}}) \otimes_{{\mathbb K}}C$, and we identify $C$ with a subalgebra of $B/{{\frak m}}B$. Note that any ${{\mathbb K}}$-basis for $Z(T)/{{\frak m}}$ provides a finite basis for $B/{{\frak m}}B$ as a free right and left $C$-module. In particular, it follows that all regular elements of $C$ are also regular in $B/{{\frak m}}B$. Thus, Lemma \[finiteextLd\] implies $\operatorname{Ld}(B/{{\frak m}}B) = \operatorname{Ld}(C)$. Now ${{\rm GK}}(C) = r$, and $C$ is $\operatorname{Ld}$-stable by [@z2 Corollary 6.3(1)], so $\operatorname{Ld}(C) = r$. Taking account of Proposition \[modfinext\], we therefore have $$\label{LdB/mB} \operatorname{Ld}(B/{{\frak m}}B) = {{\rm GK}}(B/{{\frak m}}B) = r.$$ Since ${{\frak m}}B$ is a maximal ideal of $B$, it must equal $B\cap M$. We then identify $B/ {{\frak m}}B$ with its image in $T/M$. Observe that $T$ is a free right and left $B$-module with a basis $\{ y^{u_1}, \dots, y^{u_s} \}$, where $\{ u_1,\dots,u_s \}$ is a complete set of coset representatives for $\mz^n/(Z \oplus W)$. Consequently, $T/M = T/{{\frak m}}T$ is a free right and left $(B/{{\frak m}}B)$-module with a basis $\{ y^{u_1}+M, \dots, y^{u_s}+M \}$, so $T/M$ is finitely generated as a right $(B/{{\frak m}}B)$-module and all regular elements of $B/{{\frak m}}B$ are also regular in $T/M$. Therefore Lemma \[finiteextLd\] and Proposition \[modfinext\], in combination with , yield $$\operatorname{Ld}(T/M) = {{\rm GK}}(T/M) = r. \qedhere$$ In [@bln Theorem 1.6], Tdeg-stability is proved for primitive quotients of uniparameter quantum nilpotent algebras. Here, we extend the result to all quantum nilpotent algebras. \[theorem-torsionfree\] Let $R$ be a quantum nilpotent algebra and let $P$ be a primitive ideal of R. Suppose that $J = J_w$ is the $\ch$-prime ideal of $R$ such that $P\in\spec_J(R)$. Then [(a)]{} There is an Ore set ${\mathcal{F}}_w \subseteq R/J$ of regular $\ch$-eigenvectors such that $(R/J){\mathcal{F}}_w ^{-1}$ is a quantum torus over ${{\mathbb K}}$ and $(P/J){\mathcal{F}}_w^{-1}$ is a maximal ideal of $(R/J){\mathcal{F}}_w^{-1}$. [(b)]{} $R/J$ and $R/P$ are $\operatorname{{\rm Tdeg}}$-stable. [(c)]{} ${{\rm GK}}\bigl( (R/J){\mathcal{F}}_w ^{-1} \bigr) = {{\rm GK}}(R/J)$ and ${{\rm GK}}\bigl( (R/P){\mathcal{F}}_w ^{-1} \bigr) = {{\rm GK}}(R/P)$. \(a) By Theorem \[matchloc2\], there is an Ore set ${\mathcal{F}}_w \subset R/J$ consisting of regular $\ch$-eigenvectors such that $$\begin{gathered} \spec_{J}(R) = \{ P \in \spec(R) \mid P\supseteq J \ \text{and} \ (P/J) \cap {\mathcal{F}}_w = \varnothing \} \\ (R/J){\mathcal{F}}_w^{-1} \cong ({\overline}{R}/K_w){\mathcal{E}}_w^{-1} \cong \co_{\Lambda_w}(({{\mathbb K}}^*)^{N-|w|}),\end{gathered}$$ where $\Lambda_w$ is a submatrix of $\Lambda$ (recall ). In view of [@gl2 Theorem 4.4], the primitive ideal $P$ is maximal in $\spec_J(R)$, and consequently $(P/J){\mathcal{F}}_w^{-1}$ is a maximal ideal of $(R/J){\mathcal{F}}_w^{-1}$. (b)(c) We have already shown in Corollary \[GKRmodJw\] that $R/J$ is $\operatorname{{\rm Tdeg}}$-stable. Note that the image of ${\mathcal{F}}_w$ in $R/P$ consists of regular elements (e.g., [@gw Lemma 10.19]), so that $(R/P){\mathcal{F}}_w^{-1}$ is naturally isomorphic to a subalgebra of $\operatorname{Fract}R/P$. In view of part (a) and Proposition \[qtorusmodmax\], the second part of (b) and part (c) follow from Proposition \[transfer.Tdeg\]. Tauvel’s height formula ======================= Let $P$ be a prime ideal of the quantum nilpotent algebra $R$. Then $${{\rm GK}}(R/P) + {{\rm ht}}(P)={{\rm GK}}(R);$$ that is, Tauvel’s height formula holds for all quantum nilpotent algebras. The height formula has been established for the $\ch$-prime ideals of $R$ in Theorem \[Tauvel-Hprimes\]. Next, we deal with the case where $P$ is a primitive ideal. Suppose that $J=J_w$ is the $\ch$-prime ideal such that $P\in\spec_J(R)$. Let ${\mathcal{F}}_w$ be the Ore set contained in $R/J$ that is mentioned in Theorem \[theorem-torsionfree\](a). Since $(R/J){\mathcal{F}}_w^{-1}$ is a quantum torus, its prime spectrum is catenary and Tauvel’s height formula holds in this algebra (e.g., [@bg Theorem II.9.14]). Now ${{\rm GK}}\bigl( (R/J){\mathcal{F}}_w^{-1} \bigr) = {{\rm GK}}(R/J)$ and ${{\rm GK}}\bigl((R/P){\mathcal{F}}_w^{-1} \bigr)= {{\rm GK}}(R/P)$ by Theorem \[theorem-torsionfree\](c). Hence, $${{\rm GK}}(R/P)={{\rm GK}}\bigl((R/P){\mathcal{F}}_w^{-1} \bigr) =^1 {{\rm GK}}\bigl( (R/J){\mathcal{F}}_w^{-1} \bigr) -{{\rm ht}}\bigl( (P/J){\mathcal{F}}_w^{-1} \bigr) ={{\rm GK}}(R/J) - {{\rm ht}}(P/J),$$ where $(=^1)$ follows from Tauvel’s height formula in $(R/J){\mathcal{F}}_w^{-1}$, and so we obtain ${{\rm GK}}(R/J)={{\rm GK}}(R/P)+{{\rm ht}}(P/J)$. Consequently, $$\begin{aligned} N &\geq{{\rm GK}}(R/P)+{{\rm ht}}(P)\geq \bigl( {{\rm GK}}(R/J)-{{\rm ht}}(P/J)\bigl)+ \bigl({{\rm ht}}(P/J)+{{\rm ht}}(J)\bigr) \\ &={{\rm GK}}(R/J)+{{\rm ht}}(J)=N;\end{aligned}$$ and so ${{\rm GK}}(R/P)+{{\rm ht}}(P)=N$, as required. Finally, let $P$ be an arbitrary prime ideal belonging to the $J$-stratum of $\spec(R)$, and let $Q$ be a maximal element of that stratum with $J\subseteq P\subseteq Q$. Then $Q$ is primitive by [@gl2 Theorem 4.4]. Within the stratum, we have catenarity due to the fact that $$\bigl( \spec_J(R), \subseteq \bigr) \cong \bigl( \spec \bigl( (R/J){\mathcal{F}}_w^{-1} \bigr), \subseteq \bigr).$$ Consequently, $$\begin{aligned} N &\geq& {{\rm GK}}(R/P)+{{\rm ht}}(P)\\ &\geq& \bigl({{\rm GK}}(R/Q)+{{\rm ht}}(Q/P)\bigr)+ \bigl({{\rm ht}}(P/J)+{{\rm ht}}(J)\bigr)\\ &=^1& {{\rm GK}}(R/Q)+\bigl({{\rm ht}}(Q/J)-{{\rm ht}}(P/J)\bigr)+ {{\rm ht}}(P/J)+{{\rm ht}}(J)\\ &=& {{\rm GK}}(R/Q)+{{\rm ht}}(Q/J)+ {{\rm ht}}(J)\\ &=^2& {{\rm GK}}(R/Q)+\bigl({{\rm GK}}(R/J)-{{\rm GK}}(R/Q)\bigr)+{{\rm ht}}(J)\\ &=& {{\rm GK}}(R/J)+{{\rm ht}}(J) =N,\end{aligned}$$ where in $(=^1)$ we are using catenarity within the stratum, and $(=^2)$ holds by the equality established in the primitive case above. Therefore $${{\rm GK}}(R/P)+{{\rm ht}}(P)=N,$$ as required. Examples ======== We have shown that quantum nilpotent algebras satisfy Tauvel’s height formula, but the question as to whether or not they are catenary remains open. We have also seen that in the presence of suitable homological conditions, normal separation implies catenarity and Tauvel’s height formula. However, for algebras that are not quantum nilpotent algebras, the notion of catenarity and Tauvel’s height formula are independent, as we see in the following examples. [ Several examples are known of algebras which are catenary but do not satisfy Tauvel’s height formula, such as the group algebra of the Heisenberg group (over any field) [@bell-sig Example 3.8] and the enveloping algebra of $\frak{sl}_2({{\mathbb K}})$ (for ${{\mathbb K}}$ algebraically closed of characteristic zero) [@bell-sig p.411]. We also point to [@bell-sig Example 2.9]: Let $A:={{\mathbb K}}[x,y]$ where $\operatorname{char}{{\mathbb K}}=0$. Let $\delta$ be the ${{\mathbb K}}$-linear derivation given by $\delta = (2y)\partial/\partial x +(x+y^2)\partial/\partial y$. Set $R:=A[z;\delta]$. The ring $R$ has Gelfand-Kirillov dimension three. However, the ideal $xR+yR$ is a prime ideal of height one, but $R/(xR+yR)\cong{{\mathbb K}}[z]$ has Gelfand-Kirillov dimension one; so Tauvel’s height formula fails for this ideal. This example was originally constructed by Jordan in [@jordan]. ]{} A modification of the previous example produces an example that is not catenary and does not satisfy Tauvel’s height formula. [ Let $A:={{\mathbb K}}[x,y,z]$ where $\operatorname{char}{{\mathbb K}}=0$ and $\delta$ is the ${{\mathbb K}}$-linear derivation given by $\delta = (2yz)\partial/\partial x +(x+y^2)\partial/\partial y$. Set $R:=A[w;\delta]$. See [@bell-sig Example 2.10] for details. ]{} It is easily seen that if Tauvel’s height formula holds in all prime factors of an algebra $R$, then $R$ is catenary. In fact, it suffices to know that for any prime ideals $P \supsetneq Q$ of $R$ with ${{\rm ht}}(P/Q)=1$, the equality ${{\rm GK}}(R/Q) = {{\rm GK}}(R/P)+1$ holds. Namely, this assumption implies that for any prime ideals $P \supsetneq Q$ of $R$, all saturated chains of prime ideals between $P$ and $Q$ have length ${{\rm GK}}(R/Q) - {{\rm GK}}(R/P)$. It appears unlikely that Tauvel’s height formula alone (holding just in an algebra rather than in all prime factors) implies catenarity, but no examples of non-catenary algebras with finite GK-dimension which satisfy Tauvel’s height formula are known. [ One might imagine that there would be a complementary result to Proposition \[largerCauchon\] about replacing a white box by a black box. Specifically, if $w$ is a Cauchon diagram for $R$ and $w \ne { [ \hspace{-0.65mm} [}1,N {] \hspace{-0.65mm} ]}$, one could conjecture that there exists a Cauchon diagram $w' \supsetneqq w$ such that $|w'\setminus w| = 1$ and $J_{w'} \supsetneqq J_w$. If such a result did hold then applying it iteratively starting with $w$ empty would yield a chain of $\ch$-primes of length equal to ${{\rm GK}}(R)$. That fails, e.g. for the first quantised Weyl algebra, $R = A_1^q({{\mathbb K}}) := {{\mathbb K}}\langle x_1,x_2 \mid x_2x_1 - qx_1x_2 = 1 \rangle$, with $q \in {{\mathbb K}}^*$ not a root of unity. Here ${{\rm GK}}(R) = 2$ but there are only two $\ch$-primes altogether, namely $0$ and $J := \langle x_1x_2 - x_2x_1 \rangle$. The Cauchon diagrams of $0$ and $J$ are $\varnothing$ and $\{1\}$, respectively. What goes wrong with the conjectured result is that $\{1,2\}$ is not a Cauchon diagram for $R$. ]{} Acknowledgement {#acknowledgement .unnumbered} =============== We thank James Zhang for useful discussions about transcendence degrees. [99]{} A D Bell and G Sigursson, [*Catenarity and Gelfand-Kirillov dimension in Ore extensions*]{}, J. Algebra **127** (1989), 409–425. J A Bell and S Launois, [*On the dimension of $H$-strata in quantum algebras*]{}, Algebra and Number Theory **4** (2010), 175–200. J A Bell, S Launois, and N Nguyen, [*Dimension and enumeration of primitive ideals in quantum algebras*]{}, J. Alg. Combin. **29** (2009), 269–294. K A Brown and K R Goodearl, [*Lectures on Algebraic Quantum Groups*]{}, Advanced Courses in Mathematics CRM Barcelona, Birkhäuser, Basel, 2002. G Cauchon, [ *Effacement des dérivations et spectres premiers des algèbres quantiques*]{}, J. Algebra **260** (2003), 476–518. G Cauchon, *Spectre premier de $O_q(M_n(k))$. Image canonique et séparation normale*, J. Algebra **260** (2003), 519–569. K R Goodearl, S Launois and T H Lenagan, [*Totally nonnegative cells and matrix Poisson varieties*]{}, Advances in Math. **226** (2011), 779–826. K R Goodearl, S Launois and T H Lenagan, [*Torus-invariant prime ideals in quantum matrices, totally nonnegative cells and symplectic leaves*]{}, Math. Zeitschr. **269** (2011), 29–45. K R Goodearl and T H Lenagan, [*Catenarity in quantum algebras*]{}, J. Pure Applied Algebra **111** (1996), 123–142. K R Goodearl and E S Letzter, [*Prime ideals in skew and q-skew polynomial rings*]{}, Memoirs Amer. Math. Soc. **109** (1994), no. 521. K R Goodearl and E S Letzter, *Prime factor algebras of the coordinate ring of quantum matrices*, Proc. Amer. Math. Soc. **121** (1994), 1017–1025. K R Goodearl and E S Letzter, [*The Dixmier-Moeglin equivalence in quantum matrices and quantum Weyl algebras*]{}, Trans. Amer. Math. Soc. **352** (2000), 1381–1403. K R Goodearl and R B Warfield, Jr., *An Introduction to Noncommutative Noetherian Rings*, Second Ed., London Math. Soc. Student Texts 61, Cambridge (2004) Cambridge University Press. K R Goodearl and M T Yakimov, *From quantum Ore extensions to quantum tori via noncommutative UFDs*, Advances in Math. **300** (2016), 672–716. K R Goodearl and M Yakimov, [*Quantum cluster algebra structures on quantum nilpotent algebras*]{}, Memoirs Amer. Math. Soc. **247** (2017), no. 1169, vii + 119pp. K L Horton, *The prime and primitive spectra of multiparameter quantum symplectic and Euclidean spaces*, Communic. in Algebra **31** (2003), 2713–2743. D A Jordan, [*Noetherian Ore extensions and Jacobson rings*]{}, J. London Math. Soc. (2) **10** (1975), 281–291. G R Krause and T H Lenagan, [*Growth of Algebras and Gelfand-Kirillov dimension*]{}, Graduate Studies in Mathematics, 22. American Mathematical Society, Providence, RI, 2000. S Launois, [*Idéaux premiers H-invariants de l’algèbre des matrices quantiques*]{}, Thèse de Doctorat, Université de Reims, 2003. S Launois, T H Lenagan and L Rigal, [*Quantum unique factorisation domains*]{}, J. London Math. Soc. (2) **74** (2006), 321–340. S Launois, T H Lenagan and L Rigal, *Prime ideals in the quantum grassmannian*, Selecta Math. (N.S.) **13** (2008), 697–725. T H Lenagan and L Rigal, [*The maximal order property for quantum determinantal ideals*]{}, Proc. Edinburgh Math Soc **46** (2003), 513–529. M Lorenz, *On the transcendence degree of group algebras of nilpotent groups*, Glasgow Math. J. **25** (1984), 167–174. J C McConnell and J C Pettit, *Crossed products and multiplicative analogues of Weyl algebras*, J. London Math. Soc. (2) **38** (1988), 47–55. S-Q Oh, *Catenarity in a class of iterated skew polynomial rings*, Communic. in Algebra **25** (1997), 37–49. P Tauvel, [*Sur les quotients premiers de l’algèbre enveloppante d’un algèbre de Lie résoluble*]{}, Bull. Soc. Math. France **106** (1978), 177–205. M Yakimov, *A proof of the Goodearl-Lenagan polynormality conjecture*, Internat. Math. Res. Notices (2013), no. 9, 2097–2132. M Yakimov, *Spectra and catenarity of multi-parameter quantum Schubert cells*, Glasgow Math. J. **55A** (2013), 169–194. J J Zhang, [*On Gelfand-Kirillov Transcendence degree*]{}, Trans. Amer. Math. Soc. **348** (1996), 2867–2899 . J J Zhang, *On lower transcendence degree*, Advances in Math. **139** (1998), 157–193. K R Goodearl: Department of Mathematics, University of California, Santa Barbara, CA 93106, USA Email: [goodearl@math.ucsb.edu]{} S Launois: School of Mathematics, Statistics and Actuarial Science, University of Kent Canterbury, Kent, CT2 7FS, UK Email: [S.Launois@kent.ac.uk]{} T H Lenagan: Maxwell Institute for Mathematical Sciences, School of Mathematics, University of Edinburgh, James Clerk Maxwell Building, King’s Buildings, Peter Guthrie Tait Road, Edinburgh EH9 3FD, Scotland, UK E-mail: [tom@maths.ed.ac.uk]{} [^1]: The research of the first named author was supported by US National Science Foundation grant DMS-1601184. [^2]: The research of the second named author was supported by EPSRC grant EP/N034449/1. [^3]: The research of the third named author was partially supported by a Leverhulme Trust Emeritus Fellowship.

ArXiv

--- abstract: 'The idea of a multiverse – an ensemble of universes – has received increasing attention in cosmology, both as the outcome of the originating process that generated our own universe, and as an explanation for why our universe appears to be fine-tuned for life and consciousness. Here we carefully consider how multiverses should be defined, stressing the distinction between the collection of all possible universes, and ensembles of really existing universes that are essential for an anthropic argument. We show that such realised multiverses are by no means unique. A proper measure on the space of all really existing universes or universe domains is needed, so that probabilities can be calculated, and major problems arise in terms of realised infinities. As an illustration we examine these issues in the case of the set of Friedmann-Lemaître-Robertson-Walker (FLRW) universes. Then we briefly summarise scenarios like chaotic inflation, which suggest how ensembles of universe domains may be generated, and point out that the regularities which must underlie any systematic description of truly disjoint multiverses must imply some kind of common generating mechanism. Finally, we discuss the issue of testability, which underlies the question of whether multiverse proposals are really scientific propositions.' author: - | G. F. R. Ellis,$^1$ U. Kirchner,$^1$ and W. R. Stoeger$^{1,2}$\ \ $^1$ Department of Mathematics and Applied Mathematics, University of Cape Town\ 7700 Rondebosch, South Africa\ $^2$ Permanent Address: Vatican Observatory Research Group, Steward Observatory\ The University of Arizona, Tucson, Arizona 85721, USA date: '20. June 2003' title: '**Multiverses and physical cosmology**' --- Introduction ============ The idea of a multiverse has been proposed as the only scientifically based way of avoiding the fine-tuning required to set up the conditions for our seemingly very unlikely universe to exist. Stephen Weinberg (2000), for example, uses it to explain the value of the cosmological constant, which he relates to anthropic issues. Martin Rees (2001) employs it to explain the whole set of anthropic coincidences (Barrow and Tipler 1986), that is, to explain why our universe is a congenial home for life. These and similar proposals have been triggered by the dawning awareness among many researchers that there may be many other existing universes besides ours. This possibility has received strong stimulation from proposals like Andrei Linde’s (1983, 1990) chaotic inflationary scenario, in which the origin of our own observable universe region naturally involves the origin of many other similar expanding universe regions. There is however a vagueness about the proposed nature of multiverses. They might occur in various ways, discussed by Weinberg (2000) and Tegmark (2003). They might originate naturally in different times and places through meta-cosmic processes like chaotic inflation, or in accord with Lee Smolin’s (1999) cosmic Darwinian vision. In the latter case, an ensemble of expanding universe regions grow from each other following gravitational collapse and re-expansion, where natural selection of universes through optimisation of black hole production leads to bio-friendly universe regions. This is an intriguing idea, but with many uncertain steps – in particular no proof has been given of the last step, that the physics that maximises black hole production also favours life. They might be associated with the multi-universe Everett-Wheeler-type interpretation of quantum mechanics. Or perhaps multiverses can be truly disjoint collections of universes (see Sciama 1993, Rees 2001, Tegmark 1998, 2003). Some refer to the separate expanding universe regions in chaotic inflation as ‘universes’, even though they have a common causal origin and are all part of the same single spacetime. In our view (as ‘uni’ means ‘one’) *the Universe* is by definition the one unique connected [^1] existing spacetime of which our observed expanding cosmological domain is a part. We will refer to situations such as in chaotic inflation as a *Multi-Domain Universe*, as opposed to a completely causally disconnected *Multiverse*. Throughout this paper, when our discussion pertains equally well to disjoint collections of universes (multiverses in the strict sense) and to the different domains of a Multi-Domain Universe, we shall for simplicity simply use the word *ensemble*. When an ensemble of universes are all sub-regions of a larger connected spacetime - the universe as a whole- we have the multi-domain situation, which should be described as such. Then we could reserve multiverse for the collection of genuinely disconnected universes – those which are not locally causally related. So far, none of these ideas have been developed to the point of actually describing such ensembles of universes in detail, nor has it been demonstrated that a generic well-defined ensemble will admit life. Some writers tend to imply that there is only one possible multiverse (characterised by “all that can exist does exist”). This vague prescription actually allows a vast variety of different realisations with differing properties, leading to major problems in the definition of the ensembles and in averaging, due to the lack of a well-defined measure and the infinite character of the ensemble itself. Furthermore it is not at all clear that we shall ever be able to accurately delineate the class of all possible universes. The aim of this paper is to help clarify what is involved in a full description of an ensemble of universes. Our first contribution is clarifying what is required in order to describe the space of possible universes, where much hinges on what we regard as ‘possible’. However that is only part of what is needed. It is crucial to recognise that anthropic arguments for *existence* based on ensembles of universes with specific properties require an *actually existing ensemble* with all the required properties. For purposes of providing an explanation of existence, simply having a conceptually possible ensemble is not adequate - one needs a link to objects or things that actually exist, or to mechanisms that make them exist. The second contribution of this paper is to show how an actually existing ensemble may be described in terms of a space of possible universes, by defining a distribution function (discrete or continuous) on the space of possible universes. This characterizes which of the theoretically possible universes have been actualised in the ensemble - it identifies those that have actually come into existence. This leads us to our third point: the problems arising when it is claimed that there is an actually existing ensemble containing an infinite number of universes or of expanding universe regimes. Actually existing infinities are very problematic. There are fundamental issues that arise in considering ensembles of actually existing universes: what would explain the existence of an ensemble, and its specific properties? Why should there be this particular ensemble, rather than some other one? Why should there be any regularity at all in its properties? The fourth point we make is that if all the universes in an ensemble show regularities of structure, then that implies some common generating mechanism. Some such structuring is necessary if we are to be able to describe a multiverse with specified properties - a coherent description is only possible through the existence of such regularities. Hence a multiverse consisting of completely causally disconnected universes is a problematic concept. The issue of testability is a further important consideration: Is there any conceivable direct or indirect way of testing for existence of an ensemble to which our universe belongs? Our fifth point is that there is no way we can test any mechanism proposed to impose such regularities: they will of necessity always remain speculative. The sixth point is to argue that existence of multiverses or ensembles is in principle untestable by any direct observations, and the same applies to any hypothesized properties we may suppose for them. However certain observations would be able to disprove existence of some multi-domain ensembles. It is only in that sense that the idea is a testable proposition. It is clear that in dealing with multiverses one inevitably runs up against philosophical and metaphysical issues, for example concerning the ability to make scientific conclusions in the absence of observational evidence, and in pursuing the issue of realised infinities. A companion more philosophically oriented paper will pursue those issues. Describing Ensembles: Possibility ================================= To characterise an ensemble of existing universes, we first need to develop adequate methods for describing the class of all possible universes. This requires us to specify, at least in principle, all the ways in which universes can be different from one another, in terms of their physics, chemistry, biology, etc. The Set of Possible Universes ----------------------------- The basis for describing ensembles or multiverses is contained in the structure and the dynamics of a space $\mathcal{M}$ of all possible universes $m$, each of which can be described in terms of a set of states $s$ in a state space $\mathcal{S}$. Each universe in $\mathcal{M}$ will be characterised by a set $\mathcal{P}$ of distinguishing parameters $p$, which are coordinates on $\mathcal{S}$. Some will be logical parameters, some will be numerical constants, and some will be functions or tensor fields defined in local coordinate neighbourhoods for $s$. Each universe $m$ will evolve from its initial state to some final state according to the dynamics operative, with some or all of its parameters varying as it does so. The course of this evolution of states will be represented by a path in the state space $\mathcal{S}$, depending on the parametrisation of $\mathcal{S}$ . Thus, each such path (in degenerate cases a point) is a representation of one of the universes $m$ in $\mathcal{M}$. The coordinates in $\mathcal{S}$ will be directly related to the parameters specifying members of $\mathcal{M} $. The parameter space $\mathcal{P}$ has dimension $N$ which is the dimension of the space of models $\mathcal{M}$; the space of states $ \mathcal{S}$ has $N+1$ dimensions, the extra dimension indicating the change of each model’s states with time, characterised by an extra parameter, e.g., the Hubble parameter $H$ which does not distinguish between models but rather determines what is the state of dynamical evolution of each model. Note that $N$ may be infinite, and indeed will be so unless we consider only geometrically highly restricted sets of universes. It is possible that with some parameter choices the same physical universe $ m $ will be multiply represented by this description; thus a significant issue is the equivalence problem – identifying which different representations might in fact represent the same universe model. In self-similar cases we get a single point in $\mathcal{S}$ described in terms of the chosen parameters $\mathcal{P}$: the state remains unchanged in terms of the chosen variables. But we can always get such variables for any evolution, as they are just comoving variables, not necessarily indicating anything interesting is happening dynamically. The interesting issue is if this invariance is true in physically defined variables, e.g., expansion normalised variables, then physical self-similarity is occurring. The very description of this space $\mathcal{M}$ of possibilities is based on an assumed set of laws of behaviour, either laws of physics or meta-laws that determine the laws of physics, which all universes $m$ have in common; without this, we have no basis for setting up its description. The detailed characterisation of this space, and its relationship to $\mathcal{S}$, will depend on the matter description used and its behaviour. The overall characterisation of $\mathcal{M}$ therefore must incorporate a description both of the geometry of the allowed universes and of the physics of matter. Thus the set of parameters $\mathcal{P}$ will include both geometric and physical parameters. The space $\mathcal{M}$ has a number of important subsets, for example: 1. $\mathcal{M}_{\mathrm{FLRW}}$ – the subset of all possible exactly Friedmann-Lemaître-Robertson-Walker (FLRW) universes, described by the state space $\mathcal{S}_{\mathrm{FLRW}}$ (in the case of dust plus non-interacting radiation a careful description of this phase space has been given by Ehlers and Rindler1989). 2. $\mathcal{M}_{\mathrm{almost-FLRW}}$ – the subset of all perturbed FLRW model universes. These need to be characterised in a gauge-invariant way (see e.g. Ellis and Bruni 1989) so that we can clearly identify those universes that are almost-FLRW and those that are not. 3. $\mathcal{M}_{\mathrm{anthropic}}$ – the subset of all possible universes in which life emerges at some stage in their evolution. This subset intersects $\mathcal{M}_{\mathrm{almost-FLRW}}$, and may even be a subset of $\mathcal{M}_{\mathrm{almost-FLRW}},$ but does not intersect $ \mathcal{M}_{\mathrm{FLRW}}$ (realistic models of a life-bearing universe like ours cannot be exactly FLRW, for then there is no structure). 4. $\mathcal{M}_{\mathrm{Observational}}$ – the subset of models compatible with current astronomical observations. Precisely because we need observers to make observations, this is a subset of $\mathcal{M}_{\mathrm{\ anthropic}}$. If $\mathcal{M}$ truly represents all possibilities, one must have a description that is wide enough to encompass *all* possibilities. It is here that major issues arise: how do we decide what all the possibilities are? What are the limits of possibility? What classifications of possibility are to be included? All that can happen happens  must imply all possibilities, as characterised by our description in terms of families of parameters: all allowed values must occur, and they must occur in all possible combinations. The full space $ \mathcal{M}$ must be large enough to represent all of these possibilities, along with many others we cannot even conceive of, but which can nevertheless in principle also be described by such parameters. An interesting related point has been pointed to us by Jean-Phillipe Uzan: it may be that the larger the possibility space considered, the more fine-tuned the actual universe appears to be - for with each extra possibility that is included in the possibility space, unless it can be shown to relate to already existing parameters, the actual universe and its close neighbours will live in a smaller fraction of the possibility space. For example if we assume General Relativity then there is only the parameter $G$ to measure; but if we consider scalar-tensor theories, then we have to explain why we are so close to General Relativity now. Hence there is a tension between including all possibilities in what we consider, and giving an explanation for fine tuning. From these considerations we have the first key issue: **Issue 1:** What determines $\mathcal{M}$? Where does this structure come from? What is the meta-cause that delimits this set of possibilities? Why is there a uniform structure across all universes $m$ in $ \mathcal{M}$? The meta-question is whether any of these questions can be answered scientifically. We return to that at the end. Adequately Specifying Possible Anthropic Universes -------------------------------------------------- When defining any ensemble of universes, possible or realised, we must specify all the parameters which differentiate members of the ensemble from one another at any time in their evolution. The values of these parameters may not be known or determinable initially in many cases – some of them may only be set by transitions that occur via processes like symmetry breaking within given members of the ensemble. In particular, some of the parameters whose values are important for the origination and support of life may only be fixed later in the evolution of universes in the multiverse. We can separate our set of parameters $\mathcal{P}$ for the space of all possible universes $\mathcal{M}$ into different categories, beginning with the most basic or fundamental, and progressing to more contingent and more complex categories. Ideally they should all be independent of one another, but we will not be able to establish that independence for each parameter, except for the most fundamental cosmological ones. In order to categorise our parameters, we can doubly index each parameter $p$ in $\mathcal{P}$ as $ p_j(i)$ such that those for $j=1-2$ describe basic physics, for $j=3-5$ describe the cosmology (given that basic physics), and $j=6-7$ pertain specifically to emergence and life (we must include the latter if we seriously intend to address anthropic issues). Our characterisation is as follows: 1. $p_1(i)$ are the basic physics parameters within each universe, excluding gravity - parameters characterising the basic non-gravitational laws of physics in action, related constants such as the fine-structure constant $\alpha $, and including parameters describing basic particle properties (masses, charges, spins, etc.) These should be logical parameters or dimensionless parameters, otherwise one may be describing the same physics in other units. 2. $p_2(i)$ are basic parameters describing the nature of the cosmological dynamics, e. g., $p_2(1)=1$ indicates Einstein gravity dominates, $p_2(1)=2$ indicates Brans-Dicke theory dominates, $p_2(1)=3$ indicates Electro-magnetism dominates, etc. Associated with each choice are the relevant parameter values, e.g., $p_2(2)=G$, $p_2(3)=\Lambda $, and in the Brans-Dicke case $p_2(4)=\omega $. If gravity can be derived from more fundamental physics in some unified fundamental theory, these will be related to $p_1(i);$ for example the cosmological constant may be determined from quantum field theory and basic matter parameters. 3. $p_3(i)$ are cosmological parameters characterising the nature of the matter content of a universe. These parameters encode whether radiation, baryons, dark matter, neutrinos, scalar fields, etc. occur, in each case specifying the relevant equations of state and auxiliary functions needed to determine the physical behaviour of matter (e.g. barotropic equations of state and the potential function for scalar fields). These are characterisations of physical possibilities for the macro-states of matter arising out of fundamental physics, so the possibilities here will be related to the parameters in $p_1(i)$. Realistic representations will include all the above, but simplified ensembles considered for exploratory purposes may exclude some or many of them. 4. $p_4(i)$ are physical parameters determining the relative amounts of each kind of matter present in the specific cosmological solutions envisaged, for example the density parameters $\Omega _{i}$of various components at some specific stage of its evolution (which then for example determine the matter to anti-matter ratio and the entropy to baryon ratio). The matter components present will be those characterised by $p_3(i). $ 5. $p_5(i)$ are geometrical parameters characterising the spacetime geometry of the cosmological solutions envisaged- for example the scale factor $a(t),$ Hubble parameter $H(t)$, and spatial curvature parameter $k$ in FLRW models. These will be related to $p_4(i)\,$by the gravitational equations set in $p_2(i),$ for example the Einstein Field Equations. 6. $p_6(i)$ are parameters related to the functional emergence of complexity in the hierarchy of structure, for example allowing the existence of chemically complex molecules. Thus $p_6(1)$ might be the number of different types of atoms allowed (as characterised in the periodic table), $ p_6(2)$ the number of different states of matter possible (solid, liquid, gas, plasma for example), and $p_6(3)$ the number of different types of molecular bonding. These are emergent properties arising out of the fundamental physics in operation, and so are related to the parameters set in $p_1(i)$. 7. $p_7(i)$ are biologically relevant parameters related specifically to the functional emergence of life and of self-consciousness, for example $ p_7(1)$ might characterise the possibility of supra-molecular chemistry and $ p_7(2)$ that of living cells. This builds on the complexity allowed by $ p_6(i)$ and relates again to the parameter set $p_1(i)$ It is important to note that these parameters will describe the set of possibilities we are able to characterise on the basis of our accumulated scientific experience. The limits of our understanding are relevant here, in the relation between what we conceive of as this space of possibilities, and what it really is. There may be universes which we believe are possible on the basis of what we know of physics, that may in fact not be possible. There may also be universes which we conceive of as being impossible for one reason or another, that turn out to be possible. And it is very likely that we simply may not be able to imagine or envisage all the possibilities. However this is by no means a statement that “all that can occur” is arbitrary. On the contrary, specifying the set of possible parameters determines a uniform high-level structure that is obeyed by all universes in $\mathcal{M}$. We see, then, that a possibility space $\mathcal{M\,}$ is the set of universes (one-parameter sets of states $\mathcal{S}$) obeying the dynamics characterised by a parameter space $\mathcal{P}$, which may be considered to be the union of all allowed parameters $p_j(i)$ for all $i,j\,\,$as briefly discussed above: $$\mathcal{M}=\{\mathcal{S},\mathcal{P}\},~~\mathcal{P}=\cup _{i,j}\,p_j(i).$$ Because the parameters $\mathcal{P}$ determine the dynamics, the set of paths in $\mathcal{S}$ characterising individual universes $m$ are determined by this prescription. In some particular envisaged ensemble, some of these parameters (‘class parameters’) may be fixed across the ensemble, thus defining a class of universes considered, while others (‘member parameters’) will vary across the ensemble, defining the individual members of that class. Thus $$\mathcal{P=P}_{class}\mathcal{\cup ~P}_{member}.$$ As we consider more generic ensembles, class parameters will be allowed to vary and so will become member parameters. In an ensemble in which all that is possible happens, all parameters will be member parameters; however that is so hard to handle that we usually analyse sub-spaces characterised by particular class parameters. Describing the Geometry of Possible Universes --------------------------------------------- Cosmological models are characterised by a preferred timelike vector field $ u:~u^{a}u_{a}=-1,$ usually the fluid flow vector (Ellis 1971a), but sometimes chosen for other reasons, e.g. to fit local symmetries. To describe a cosmological spacetime locally we must give a description of its (generally inhomogeneous and anisotropic) geometry via suitable parameters $p_{5}(i)$. This description may be usefully given in terms of a tetrad basis as follows (see Ellis and van Elst 1999, Wainwright and Ellis 1996, Uggla, *et al.* 2003): *Feature 1*: a set of local coordinates $\mathcal{X}$ $=\{x^i\}$ must be chosen in each chart of a global atlas. This will in particular have a time coordinate $t\,$ which will be used to characterise evolution of the universe; this should be chosen in as uniform as possible a way across all the universes considered, for example it may be based on surfaces of constant Hubble parameter $H$ for the preferred vector field $u$. *Feature 2*: in each chart, to determine the geometry we must be given the components $\mathcal{E}=[e_{\;a}^i(x^j)]$ of an orthonormal tetrad with the fluid flow vector chosen as the timelike tetrad vector ($a,b,c..$ are tetrad indices; four of these components can be set to zero by suitable choice of coordinates). Together the coordinates and the tetrad form the reference frame $$\mathcal{P}_{\mathrm{frame}}\equiv \{\mathcal{X},\mathcal{E\}}.$$ The metric tensor is then $$ds^2=g_{ij}\,(x^k)dx^idx^j=\eta _{ab}\,e_{\;i}^a(x^k)\,e_{\;j}^b(x^l)dx^idx^j$$ where $\eta _{ab}$ is the Minkowski metric:  $$\eta _{ab}=e_a.e_b=diag(-1,+1,+1,+1)$$ (because the tetrad is orthonormal) and $e_{\;j}^b(x^l)$ are the inverse of $ e_{\;a}^i(x^j):$ $$e_{\;a}^i(x^j)e_i^{\;b}(x^j)=\delta _a^b.$$ Thus the metric is given by $$ds^2=-\left( e_{\;i}^0dx^i\right) ^2+\left( e_{\;i}^1dx^i\right) ^2+\left( e_{\;i}^2dx^i\right) ^2+\left( e_{\;i}^3dx^i\right) ^2 \label{metric2}$$ The basic geometric quantities used to determine the spacetime geometry are the rotation coefficients $\Gamma _{\;bc}^a$ of this tetrad, defined by $$\Gamma _{\;bc}^a=e_j^{\;a}e_{\;c;k}^je_{\;b}^k.$$ They may conveniently be given in terms of geometric quantities $$\mathcal{P}_{\mathrm{geometry}}\equiv \{\dot{u}_\alpha ,\theta ,\sigma _{\alpha \beta },\omega _{\alpha \beta },\Omega _\gamma ,a^\alpha ,n_{\alpha \beta }\}. \label{geom}$$ characterised as follows: $$\begin{aligned} \Gamma _{\alpha 00} &=&\dot{u}_{\alpha }, \\ \Gamma _{\alpha 0\beta } &=&\frac{1}{3}\theta +\sigma _{\alpha \beta }-\omega _{\alpha \beta }, \\ \Gamma _{\alpha \beta 0} &=&\epsilon _{\alpha \beta \gamma }\Omega ^{\gamma }, \\ \Gamma _{\alpha \beta \gamma } &=&a_{[\alpha }\delta _{\beta ]\gamma }+\epsilon _{\gamma \delta \lbrack \alpha }n_{\;\beta ]}^{\delta }+\frac{1}{ 2 }\epsilon _{\alpha \beta \delta }n_{\;\delta }^{\gamma },\end{aligned}$$ where $\dot{u}_{\alpha }$ is the acceleration of the fluid flow congruence, $ \theta $ is its expansion, $\sigma _{\alpha \beta }\,=\sigma _{(\alpha \beta )}$ is its shear $(\sigma _{\;b}^{b}=0),$ and $\omega _{\alpha \beta }=\omega _{\lbrack \alpha \beta ]}$ its vorticity, while $n_{\alpha \beta }=n_{(\alpha \beta )}$ and $a_{\alpha }$ determine the spatial rotation coefficients (see Wainwright and Ellis 1996, Ellis and van Elst 1999). Greek indices (with range $1-3)$ indicate that all these quantities are orthogonal to $u^{a}.$ They are spacetime fields, although in particular high-symmetry cases they may be independent of many or of all the coordinates. The Jacobi identities, Bianchi identities, and Einstein field equations can all be written out in terms of these quantities, as can the components $E_{\alpha \beta },$ $H_{\alpha \beta }$ of the Weyl tensor (see Ellis and van Elst 1999). Except in the special cases of isotropic spacetimes and locally rotationally symmetric spacetimes (see Ellis 1967, van Elst and Ellis 1996), the basis tetrad can be chosen in an invariant way so that three of these quantities vanish and all the rest are scalar invariants. Thus the geometry is determined by the 36 spacetime functions in the combined set ($\mathcal{E},\mathcal{P}_{\mathrm{geometry}}$) with some chosen specification of coordinates $\mathcal{X}$, with the metric then determined by (\[metric2\]). For detailed dynamical studies it is often useful to rescale the variables in terms of the expansion (see Wainwright and Ellis 1995, Uggla et al 2003 for details). Note that the same universe may occur several times over in this space; the *equivalence problem* is determining when such multiple representations occur. We do not recommend going to a quotient space where each universe occurs only once, as for example in the dynamical studies of Fischer and Marsden (1979), for the cost of doing so is to destroy the manifold structure of the space of spacetimes. It is far better to allow multiple representations of the same universe (for example several representations of the same Bianchi I universe occur in the Kasner ring in the space of Bianchi models, see Wainwright and Ellis 1996) both to keep the manifold structure intact and because then the dynamical structure becomes clearer. *Feature 3*: To determine the global structure, we need a set of composition functions relating different charts in the atlas where they overlap, thus determining the global topology of the universe. Together these are the parameters $p_5(i)$ needed to distinguish model states. A particular model will be represented as a path through those states. The nature of that evolution will be determined by the matter present. Describing the Physics of Possible Universes -------------------------------------------- *Feature 4*: To determine the matter stress-energy tensor we must specify the quantities $$\mathcal{P}_{\mathrm{matter}}\equiv \{\mu ,q_\alpha ,p,\pi _{ab},\Phi _A\} \label{mat}$$ for all matter components present, where $\mu $ is the energy density, $ q_\alpha $ is the momentum flux density, $p$ is the pressure, $\pi _{ab}=\pi _{(ab)}$ the anisotropic pressure $(\pi _{\;b}^b=0),$ and $\Phi _A$ ($ A=1..A_{\max })$ is some set of internal variables sufficient to make the matter dynamics deterministic when suitable equations of state are added (for example these might include the temperature, the entropy, the velocity $ v^i$ of matter relative to the reference frame, some scalar fields and their time derivatives, or a particle distribution function)$.$ These are parameters $p_4(i)\,$ for each kind of matter characterised by $p_3(i).$ Some of these dynamical quantities may vanish (for example, in the case of a ‘perfect fluid’, $q_\alpha =0,$ $\pi _{ab}=0)\,$and some of those that do not vanish will be related to others by the equations of state (for example, in the case of a barotropic fluid, $p=p(\mu ))$ and dynamic equations (for example the Klein Gordon equation for a scalar field). These equations of state can be used to reduce the number of variables in $ \mathcal{P}_{\mathrm{matter}}$; when they are not used in this way, they must be explicitly stated in a separate parameter space $\mathcal{\ P}_{ \mathrm{eos}}$ in $p_3(i).$ In broad terms $$\begin{aligned} \mathcal{P}_{\mathrm{eos}} &\equiv &\{q_\alpha =q_\alpha (\mu ,\Phi _A),\;~~p=p(\mu ,\Phi _A),\; \nonumber \\ &{}&\;\pi _{ab}=\pi _{ab}(\mu ,\Phi _A),\;\dot{\Phi}_A=\dot{\Phi}_A(\Phi _A)\}.\end{aligned}$$ Given this information the equations become determinate and we can obtain the dynamical evolution of the models in the state space; see for example Wainwright and Ellis (1996), Hewitt et al (2002), Horwood et al (2002) for the case of Bianchi models (characterised by all the variables defined above depending on the time only) and Uggla et al (2003), Lim et al (2003) for the generic case. *Feature 5*: However more general features may vary: the gravitational constant, the cosmological constant, and so on; and even the dimensions of spacetime or the kinds of forces in operation. These are the parameters $\mathcal{P}_{\mathrm{physics}}$ comprising $p_1(i)$ and $p_2(i).$ What complicates this issue is that some or many of these features may be emergent properties, resulting for example from broken symmetries occurring as the universe evolves. Thus they may come into being rather than being given as initial conditions that then hold for all time. Initially one might think that considering all possible physics simply involves choices of coupling constants and perhaps letting some fundamental constant vary. But the issue is more fundamental than that. Taking seriously the concept of including *all* possibilities in the ensembles, the space of physical parameters $\mathcal{P}_{\mathrm{physics}}$ used to describe $\mathcal{M}$, the parameters $p_{2}(i)$ might for example include a parameter $p_{\mathrm{grav}}(i)$ such that: for $i=1$ there is no gravity; for $i=2$ there is Newtonian gravity; for $i=3$ general relativity is the correct theory at all energies – there is no quantum gravity regime; for $ i=4$ loop quantum gravity is the correct quantum gravity theory; for $i=5$ a particular version of superstring theory or M-theory is the correct theory. Choices such as these will arise for all the laws and parameters of physics. In some universes there will be a fundamental unification of physics expressible in a basic “theory of everything”, in others this will not be so. Some universes will be realised as branes in a higher dimensional spacetime, others will not. The Anthropic subset -------------------- We are interested in the subset of universes that allow intelligent life to exist. That means we need a function on the set of possible universes that describes the probability that life may evolve. An adaptation of the Drake equation (Drake and Shostak) gives for the probability of intelligent life in any particular universe $m$ in an ensemble, $$P_{life}(m)=F*\Pi \label{life1}$$ where the existence of a habitat for life is expressed by the product $$\Pi =P_{gal}*R*f_S*f_p*n_e \label{life4}$$ and the coming into existence of life, given such a habitat, is expressed by the product $$F=f_l*f_i. \label{life5}$$ Here $P_{gal}\,$is the probability of galaxies forming in the universe considered, $R$ is the average rate of star formation in galaxies, $f_S$ is the fraction of these stars that can provide a suitable environment for life (they are ‘Sun-like’), $f_p$ is the fraction of stars that are surrounded by planetary systems, $n_e$ is the mean number of planets in each such system that are suitable habitats for life (they are ‘Earth-like’), $f_l$ is the fraction of such planets on which life actually originates, and $f_i$ represents the fraction of those planets on which there is life where self-conscious beings develop. The anthropic subset of a possibility space is that set of universes for which $P_{life}(m)>0.$ The probabilities {$P_{gal},R,f_S,f_p,n_e,f_l,f_i$) are functions of the physical and cosmological parameters characterised above, so there will be many different representations of this parameter set depending on the degree to which we try to represent such interrelations. The astrophysical issues expressed in the product $\Pi $ are the easier ones to investigate. We can in principle make a cut between those consistent with the eventual emergence of life and those incompatible with it by considering each of the factors in $\Pi \,\,$in turn, taking into account their dependence on the parameters $ p_1(i)$ to $p_5(i),$ and only considering the next factor if all the previous ones are non-zero (an approach that fits in naturally with Bayesian statistics and the successive allocation of relevant priors). $\,$In this way we can assign bio-friendly intervals to the possibility space $\mathcal{M}$. If $\ \Pi \,$ is non-zero we can move on to considering similarly whether $F$ is non-zero, based on the parameters $p_6(i)$ to $p_7(i)$ determining if true complexity is possible, which in turn depend on the physics parameters $p_1(i)$ in a crucial way that is not fully understood. It will be impossible at any stage to characterise that set of the multiverse in which *all* the conditions *necessary* for the emergence of self-conscious life and its maintenance have been met, for we do not know what those conditions are (for example, we do not know if there are forms of life possible that are not based on carbon and organic chemistry). Nevertheless it is clear that life demands unique combinations of many different parameter values that must be realised simultaneously. When we look at these combinations, they will span a very small subset of the whole parameter space (Davies 2003, Tegmark 2003). If we wish to deal with specifically human life, we need to make the space $ \mathcal{M}$ large enough to deal with all relevant parameters for this case, where free will arises. This raises substantial extra complications, discussed in the companion (more philosophical) paper. Parameter space revisited ------------------------- It is now clear that some of the parameters discussed above are dependent on other ones, so that while we can write down a more or less complete set at varying levels of detail they will in general not be an independent set. There is a considerable challenge here: to find an independent set. *Inter alia* this involves solving both the initial value problem for general relativity and the way that galactic and planetary formation depend on fundamental physics constants (which for example determine radiation transfer properties in stars and in proto-planetary gas clouds), as well as relations there may be between the fundamental constants and the way the emergent complexity of life depends on them. We are a long way from understanding all these issues. This means we can provide necessary sets of parameter values but cannot guarantee completeness or independence. The Set of Realised Universes ============================= We have now characterised the set of possible universes. But in any given ensemble, they may not all be realised, and some may be realised many times. The purpose of this section is to set up a formalism making clear which of the *possible* universes (characterised above) occur in a specific *realised* ensemble. A distribution function describing an ensemble of realised universes -------------------------------------------------------------------- In order to select from $\mathcal{M}$ a set of realised universes we need to define on $\mathcal{M}$ a distribution function $f(m)$ specifying how many times each type of possible universe $m$ in $\mathcal{M}$ is realised. The function $f(m)$ expresses the contingency in any actualisation – the fact that not every possible universe has to be realised, and that any actual universe does not have to be realised as a matter of necessity. Things could have been different! Thus, $f(m)$ describes the *ensemble of universes* or * multiverse* envisaged as being realised out of the set of possibilities. If these realisations were determined by the laws of necessity alone, they would simply be the set of possibilities described by $ \mathcal{M}$. In general they include only a subset of possible universes, and multiple realisations of some of them. This is the way in which chance or contingency is realised in the ensemble[^2]. The class of models considered is determined by all the parameters held constant (‘class parameters’). Considering the varying parameters for the class (‘member parameters’), some will take only discrete values, but for each one allowed to take continuous values we need a volume element of the possibility space $M$ characterised by parameter increments $dp_j(i)$ in all such varying parameters $p_j(i)$. The volume element will be given by a product $$\pi =\Pi _{i,j}\,m_{ij}(m)\,dp_{j}(i) \label{measure}$$ where the product $\Pi _{i,j}$ runs over all continuously varying member parameters $i,j$ in the possibility space, and the $m_{ij}$ weight the contributions of the different parameter increments relative to each other. These weights depend on the parameters $p_{j}(i)$ characterising the universe $m$. The number of galaxies corresponding to the set of parameter increments $dp_{j}(i)$ will be $dN$ given by $$dN=f(m)\pi \label{dist1}$$ for continuous parameters; for discrete parameters, we add in the contribution from all allowed parameter values. The total number of galaxies in the ensemble will be given by $$N=\int f(m)\pi \label{dist2}$$ (which will often diverge), where the integral ranges over all allowed values of the member parameters and we take it to include all relevant discrete summations. The probable value of any specific quality $p(m)$ defined on the set of galaxies will be given by $$P=\int p(m)f(m)\pi \label{prob}$$ Such integrals over the space of possibilities give numbers, averages, and probabilities. Hence, a (realised) ensemble $E$ of galaxies is described by a possibility space $\mathcal{M}$, a measure $\pi $ on $\mathcal{M}$, and a distribution function $f(m)\,$on $\mathcal{M}:$ $$E=\{\mathcal{M},\pi ,f(m)\}. \label{ensemble}$$ The distribution function $f(m)$ might be discrete (e. g., there are 3 copies of universe $m_1$ and 4 copies of universe $m_2$, with no copies of any other possible universe), or continuous (e.g. characterised by a given distribution of densities$\Omega _i$). In many cases a distribution function will exclude many possible universes from the realisation it specifies. Now it is conceivable that all possibilities are realised – that all universes in $\mathcal{M}$ exist at least once. This would mean that the distribution function $$f(m)\neq 0\mathrm{~for~all~}m\in \mathcal{M}.$$ But there are an infinite number of distribution functions which would fulfil this condition, and so a really existing ‘ensemble of all possible universes’ is not unique. In such ensembles, all possible values of each distinguishing parameter would be predicted to exist in different members of the multiverse in all possible combinations with all other parameters at least once, but they may occur many times. One of the problems is that this often means that the integrals associated with such distribution functions would diverge, preventing the calculation of probabilities from such models (see our treatment of the FLRW case below). From this consideration we have the second key issue: **Issue 2:** What determines $f(m)$? What is the meta-cause that delimits the set of realisations out of the set of possibilities? The answer to this question has to be different from the answer to *Issue 1*, precisely because here we are describing the contingency of selection of a subset of possibilities from the set of all possibilities, determination of the latter being what is considered in *Issue 1*. Again, the meta-question is whether this can be answered scientifically. Measures and Probabilities -------------------------- It is clear that $f(m)$ will enable us to derive numbers and probabilities relative to the realisation it defines only if we also have determined a unique measure $\pi $ on the ensemble, characterised by a specific choice of the $\,$weights $m_{ij}(m)$ in (\[measure\]), where these weights will depend on the $p_{j}(i)$. There are three issues here. First, what may seem a naturalmeasure for $\mathcal{\ M}$ in one set of coordinates will not be natural in another set of coordinates. Hence the concept of a measure is not unique, as is illustrated below in the FLRW case. This is aggravated by the fact that the parameter space will often contain completely different kinds of quantities (density parameters and the values of the gravitational constant and the cosmological constant, for example), and assigning the weights entails somehow assigning a relative weighting between these quite different kinds of quantities. Second, it is possible that we might be able to assign probabilities $\chi (m)$ to points of $\mathcal{M}$ from some kind of physical argument, and then predict $f(m)$ from these, following the usual line of argument for determining entropy in a gas. However, we then have to determine some reason why $\chi (m)$ is what it is and how it then leads to $f(m)$. In the entropy case, we assume equal probability in each phase space volume; why should that hold for an ensemble of universes? Realising such probabilities seems to imply a causal mechanism relating the created members of the multiverse to one another so they are not in fact causally disjoint, otherwise, there is no reason why any probability law (Gaussian normal, for example) should be obeyed. We will return to this point later. Finally, the relevant integrals may diverge. In that case, assigning mean values or averages for physical quantities in an ensemble of universes is problematic (see Kirchner and Ellis 2003 and references therein). The Anthropic subset -------------------- The expression (\[life1\]) can be used in conjunction with the distribution function $f(m)$ of galaxies to determine the probability of life arising in the whole ensemble: $$P_{life}(E)=\int f(m)*P_{gal}*R*f_S*f_p*n_e*f_l*f_i*\pi \label{life3}$$ (which is a particular case of (\[prob\]) based on (\[life1\])). An anthropic ensemble is one for which $P_{life}(E)>0.$ If the distribution function derives from a probability function, we may combine the probability functions to get an overall anthropic probability function- for an example see Weinberg et al discussed below, where it is assumed that $P_{gal}\,$ is the only relevant parameter for the existence of life. This is equivalent to assuming that $R*f_S*f_p*n_e*f_l*f_i=1.$ This assumption might be acceptable in our physically realised universe, but there is no reason to believe it would hold generally in an ensemble because these parameters will depend on other ensemble parameters, which will vary. Problems With Infinity ---------------------- When speaking of multiverses or ensembles of universes – possible or realised – the issue of infinity often crops up. Researchers often envision an *infinite* set of universes, in which all possibilities are realised. Can there really be an infinite set of really existing universes? We suggest that, on the basis of well-known philosophical arguments, the answer is No. There is no conceptual problem with an infinite set – countable or uncountable – of *possible* or *conceivable* universes. However, as stressed by David Hilbert (1964), it can be argued that a *really existing* infinite set is not possible. As he points out, following many others, the  existence of the actually infinite inevitably leads to well-recognised unresolvable contradictions in set theory, and thus in definitions and deductive foundations of mathematics itself (Hilbert, pp. 141-142). His basic position therefore is that “Just as operations with the infinitely small were replaced by operations with the finite which yielded exactly the same results . . ., so in general must deductive methods based on the infinite be replaced by finite procedures which yield exactly the same results. (p. 135) He concludes, “Our principle result is that the infinite is nowhere to be found in reality. It neither exists in nature nor provides a legitimate basis for rational thought . . . The role that remains for the infinite to play is solely that of an idea . . . which transcends all experience and which completes the concrete as a totality . . .” (Hilbert, p. 151). Others (see Spitzer 2000 and Stoeger 2003 and references therein) have further pointed out that realised infinite sets are not constructible – there is no procedure one can in principal implement to complete such a set – they are simply incompletable. But, if that is the case, then “infinity” cannot be arrived at, or realised. On the contrary, the concept itself implies its inability to be realised! This is precisely why a realised past infinity in time is not considered possible from this standpoint – since it involves an infinite set of completed events or moments. There is no way of constructing such a realised set, or actualising it. Thus, it is important to recognise that infinity is not an actual number we can ever specify or reach – it is simply the code-word for it continues without end. Whenever infinities emerge in physics – such as in the case of singularities – we can be reasonably sure, as is usually recognised, that there has been a breakdown in our models. An achieved infinity in any physical parameter (temperature, density, spatial curvature) is almost certainly *not* a possible outcome of any physical process – simply because it means traversing in actuality an interval of values which never ends. We assume space extends forever in Euclidean geometry and in many cosmological models, but we can never prove that any realised 3-space in the real universe continues in this way - it is an untestable concept, and the real spatial geometry of the universe is almost certainly not Euclidean. Thus Euclidean space is an abstraction that is probably not realised in physical practice. In the physical universe spatial infinities can be avoided by compact spatial sections, either resultant from positive spatial curvature or from choice of compact topologies in universes that have zero or negative spatial curvature, (for example FLRW flat and open universes can have finite rather than infinite spatial sections). Future infinite time is never realised: rather the situation is that whatever time we reach, there is always more time available. Much the same applies to claims of a past infinity of time:there may be unbounded time available in the past in principle, but in what sense can it be attained in practice? The arguments against an infinite past time are strong – it’s simply not constructible in terms of events or instants of time, besides being conceptually indefinite.[^3] The same problem of a realised infinity may be true in terms of the supposed ensembles of universes. It is difficult enough conceiving of an ensemble of many ‘really existing’ universes that are totally causally disjoint from our own, and how that could come into being, particularly given two important features. Firstly, specifying the geometry of a generic universe requires an infinite amount of information because the quantities in $\mathcal{P}_{ \mathrm{\ geometry}}$ are fields on spacetime, in general requiring specification at each point (or equivalently, an infinite number of Fourier coefficients) - they will almost always not be algorithmically compressible. This greatly aggravates all the problems regarding infinity and the ensemble. Only in highly symmetric cases, like the FLRW solutions, does this data reduce to a finite number of parameters. One can suggest that a statistical description would suffice, where a finite set of numbers describe the statistics of the solution, rather than giving a full description. Whether this suffices to adequately describe an ensemble where ‘all that can happen, happens’ is a moot point. We suggest not, for the simple reason that there is no guarantee that all possible models will obey any known statistical description. That assumption is a major restriction on what is assumed to be possible. Secondly, many universes in the ensemble may themselves have infinite spatial extent and contain an infinite amount of matter, with the paradoxical conclusions that entails (Ellis and Brundrit 1979). To conceive of physical creation of an infinite set of universes (most requiring an infinite amount of information for their prescription, and many of which will themselves be spatially infinite) is at least an order of magnitude more difficult than specifying an existent infinitude of finitely specifiable objects. The phrase ‘everything that can exist, exists’ implies such an infinitude, but glosses over all the profound difficulties implied. One should note here particularly that problems arise in this context in terms of the continuum assigned by classical theories to physical quantities and indeed to spacetime itself. Suppose for example that we identify corresponding times in the models in an ensemble and then assume that *all* values of the density parameter occur at each spatial point at that time. Because of the real number continuum, this is an uncountably infinite set of models – and genuine existence of such an uncountable infinitude is highly problematic. But on the other hand, if the set of realised models is either finite or countably infinite, then almost all possible models are not realised – the ensemble represents a set of measure zero in the set of possible universes. Either way the situation is distinctly uncomfortable. However we might try to argue around this by a discretization argument: maybe differences in some parameter of less than say $10^{-10}$ are unobservable, so we can replace the continuum version by a discretised one, and perhaps some such discretisation is forced on us by quantum theory. If this is the intention, then that should be made explicit. That solves the ‘ultraviolet divergence’ associated with the small-scale continuum, but not the ‘infrared divergence’ associated with supposed infinite distances, infinite times, and infinite values of parameters describing cosmologies. Ensembles of FLRW Universes =========================== Having established the broad set of issues concerning multiverses that we believe need to be addressed, we shall for the remainder of this paper limit ourselves to the FLRW sector $\mathcal{M}_{\mathrm{FLRW}}$ of the ensemble of all possible universes $\mathcal{M}$ in order to illustrate these issues[^4]. We assume the family considered is filled with matter components characterised by a $ \gamma $-law equation of state, and mainly restrict our attention to their cosmological parameters, although full consideration of anthropic issues would be characterised by including all the other parameters. Our descriptive treatment will consider FLRW universe domains (whether a true multiverse or separate domains in a single spacetime) as distinct but with common physical characteristics. Properties of FLRW models ------------------------- FLRW models are homogeneous and isotropic models described by the metric $$\mathrm{d}s^2=-\mathrm{d}t^2+a^2(t)\left( \frac{{}r^2}{1-kr^2}+r^2\mathrm{d} \Omega ^2\right) , \label{metric}$$ where $\mathrm{d}\Omega ^2=\mathrm{d}\vartheta ^2+\sin ^2(\vartheta )\mathrm{ \ \ \ \ \ d}\varphi ^2$ denotes the line element on the two-dimensional unit sphere, $a(t)$ is the scale-factor, and $$k=\left\{ \begin{array}{ll} 1 & \mbox{for closed models} \\ 0 & \mbox{for flat models} \\ -1 & \mbox{for open models} \end{array} \right.$$ is the normalised curvature. The FLRW model is completely determined by $k$ and the scale-factor $a(t)$, which incorporates the time-evolution and is obtained from the Einstein-Field equations together with the matter description.[^5] Assuming gravity is described by the Einstein field equations, the evolution of FLRW models is described by the Friedmann equation $$H^2(\Omega -1)=\frac k{a^2}, \label{Fried}$$ where $H \equiv \dot{a}/a$ (a dot denotes differentiation with respect to proper time) is the Hubble parameter and $\Omega $ the density parameter. We restrict our discussion to models with only a cosmological constant $\Lambda $ and one matter component which obeys a $\gamma $-law equation of state, i.e., its pressure $p$ and density $\rho $ are related by $p=(\gamma -1)\rho $, where $\gamma $ is constant. This specification of parameters $p_3(i)$ includes in particular the case of dust ($\gamma =1$) and radiation ($\gamma =4/3$). The total density parameter is $$\Omega =\Omega _m + \Omega _\Lambda . \label{omega}$$ where the matter density parameter is $\Omega_m \equiv \frac{\kappa \rho }{ 3H^2}$ and the vacuum-energy density parameter is $\Omega _\Lambda \equiv \frac \Lambda {3H^2}$ (representing a cosmological constant). These form the parameters $p_4(i).$ The second time derivative of the scale factor is determined by the Raychaudhuri equation $$2q=(3\gamma -2)\Omega _m-\Omega _\Lambda , \label{ray}$$ where $q\equiv -\frac{\ddot{a}}{aH^2}$ is the dimensionless deceleration parameter. The matter evolution is given by the energy-conservation equation $$\dot{\rho}=-3\gamma \rho H \label{cons}$$ or equivalently by $$\dot{\Omega}_m=\Omega _mH(q+1-3\gamma ). \label{omega_dot}$$ Besides the normalised curvature $k$ there are two constants of motion, $ \chi \equiv \kappa \rho a^{3\gamma }/3$ and the cosmological constant $ \Lambda $. Given these parameters, the dynamical evolution is determined from the implied initial conditions: {$a(t_0),\Omega _{i0},\gamma _i,k\}\Rightarrow a(t)$. Parametrising FLRW models ------------------------- In order to define a FLRW ensemble we need a set of independent parameters which uniquely identify all possible models. We want to consider all possible FLRW models with the same physical laws as in our universe, but possibly different coupling constants. There is then one set of parameters $ p_2(i)$ which defines the gravitational “physics” of the model in terms of the coupling constants – for simplicity let us only consider the gravitational constant $G$ here – and further sets $p_5(i),$ $p_4(i)$ which identify the geometry and matter content of the actual model, and which are related to each other via the Einstein field equations. Among the various options there are two particularly useful parametrisations. Ehlers and Rindler (1989) developed a parametrisation in terms of the observable density parameters (they also include a radiation component) and the Hubble parameter. With $\Omega _k \equiv \frac k{H^2a^2}$ the Friedmann equation becomes $$\Omega_m + \Omega _\Lambda -1=\Omega _k. \label{uli-friedmann}$$ The curvature parameter $\Omega _k$ determines $k = sgn(\Omega _k).\,$For $k \neq 0$ the scale-factor, and hence the metric (\[metric\]), is determined by $$a^2(t)=\frac k{H^2\Omega _k}=\frac k{H^2(\Omega_m + \Omega_\Lambda -1)},$$ while for $k=0$ its value is unimportant because of scale-invariance in that case. Hence any *state* is completely described by $\Omega_m, \Omega _\Lambda $, and $H$. In order to parametrise the models rather than the states, we need to select one particular time $t_0$ for each model at which we take the above parameters as representative parameters $\Omega_{m 0}, \Omega_{\Lambda 0}$, and $H_0$ for this model.[^6] We note that this time $t_0$ can be model dependent because not all models will reach the age $t_0$. All big-bang FLRW models start with an infinite positive Hubble parameter whose absolute value reaches or approaches asymptotically a minimum value $ H_{\mathrm{min}}$. Hence we could define the time $t_{0}$ as the time when the model first takes a certain value $H_{0}(p_{I})>H_{\mathrm{min}}(p_{I})$ , where $p_{I}$ represents the model parameters. One particular choice of $ H_{0}(p_{I})$ is given by $$H_{0}(p_{I})\equiv \exp (H_{\mathrm{min}}(p_{I}))$$ On the other hand, by setting $H_{0}(p_{I})=\mathrm{const}$ and excluding all models which never reach this value one finds easily a parametrisation of all models which reach this Hubble value during their evolution. While above choice of parameters give a convenient parametrisation in terms of observables which covers closed, flat, and open models, it is disturbing that for each model an arbitrary time has to be chosen. This also leads to a technical difficulty, because the parameters $\{H_{0},\Omega _{m0},\Omega_{\Lambda 0}\}$ are subject to the constraint $ H_{0}=H_{0}(p_{I}) $. For these reasons it is often convenient to use a set of parameters which are comoving in the state-space, i.e., parameters which are constants of motion. As mentioned above, for open and closed models such a set is given by the matter constant $\chi$, the cosmological constant $\Lambda$, and the normalised curvature constant $k$. For flat models one can rescale the scale factor, which allows us to set $\chi=1$. These parameters are related to the observational quantities by (for $k=\pm1$ ) $$\chi =\frac{\Omega _{m}H^{2}}{(H^{2}|1-\Omega _{m}-\Omega _{\Lambda }|)^{3\gamma /2}}\quad \mbox{ and }\quad \Lambda =3H^{2}\Omega _{\Lambda }.$$ The evolution of these models through state space is illustrated here in terms of two different parametrisations of the state space, see Figures 1a and 1b. For a detailed investigation of these evolutions for models with non-interacting matter and radiation, see Ehlers and Rindler (1989). The possibility space --------------------- The structures defined so far are the uniform structures across the class of models in this possibility space, characterised both by laws of physics (in particular General Relativity) and by a restricted class of geometries. It is clear that universes in a multiverse should be able to differ in at least some properties from each other. We have just characterised the geometrical possibilities we are considering. The next question is, which physical laws and parameters can vary within the ensemble, and which values can they take? For this simplified discussion let us just assume that only the gravitational constant $G$ and the cosmological constant $\Lambda $ (which also qualifies as a model parameter) are variables, with the ranges[^7] $G\in [0,\infty )$ and $\Lambda \in (-\infty ,\infty )$. However, if we consider all that is possible  within this restricted class of FLRW models, maybe we should consider $G\in (-\infty ,\infty ).$ There is still considerable uncertainty as to the nature of an ensemble even within this restricted context. Whatever is chosen here defines the set of possibilities that can arise. The measure ----------- For a complete probabilistic description of an ensemble we need not only a distribution function $P$, but also a measure $\pi$ for the parameter space (see Section 3.1). The information entropy $$S\equiv -\int \mathrm{d}xP(x)\log \left( \frac{P(x)}{\mu (x)}\right) \label{uli-entropy}$$ is then maximised for the probability distribution equal to this measure, representing the state of minimal knowledge. Without knowledge of the creation mechanism it is impossible to determine this measure with certainty. Nevertheless, we might ask what our best guess for such a measure should be in a state of minimal information, where only a certain set of independent parameters, describing the ensemble, and their ranges are known. The only known method for constructing such a measure is Jaynes’ principle. Its application to FLRW models with $\gamma$-law equation of state has been discussed in Kirchner and Ellis (2003). One identifies a set of transformations $x^{\prime }(x)$ in possibility space which leaves the mathematical structure invariant, and demands that the measure is invariant under these transformations. The two most important cases are given by parameters which can take all real values, and those that take on all positive real values. In the first case, if $z$ is a valid parameter value, so is $z^{\prime }=z+\alpha $ for all real $\alpha $. According to Jaynes’ principle the measure should obey $\mathrm{\pi =}\mu (z)\mathrm{d}z=\mu (z^{\prime }) \mathrm{d}z^{\prime }$ and hence $\mu (z)=\mu (z^{\prime })$, i.e., the measure is constant. If on the other hand a parameter $u$ only takes non-zero positive values, we can generate another valid parameter value by $u^{\prime }=\lambda u$, where $\lambda \in R^{+}$. Demanding invariance of the measure yields $\mathrm{\pi }\propto \frac 1u \mathrm{d}u$. These transformations are not unique, and hence one could find many different measures. Nevertheless, in the state of minimum information we don’t know what the natural parametrisation is for the possibility space and different measures correspond to different guesses. Surprisingly Jaynes’ principle is “relatively invariant” under simple parametrisation changes. For example, introducing a new parametrisation for a positive quantity $G$ by $G = m^{n}$ for positive $m$, or by $G = \exp (\lambda )$ for real $ \lambda $, will give the same measure. There are two important points to note. Firstly the measure is derived from the chosen set of parameters. Generally a different choice of parameters yields a different minimum-information measure, predicting another maximum-entropy distribution function. Let us consider the example of an ensemble of dust-FLRW models. The different open and closed models are most conveniently parametrised by the constants of motion, which are given by the cosmological constant $\Lambda $ and $\chi \equiv a\rho ^{3\gamma }$, where $ \rho $ is the energy density. This leads to the minimum-information measure (Kirchner and Ellis 2003). $$\mathrm{\pi }\propto \frac{\mathrm{d}\Lambda \mathrm{d}\chi }{\sqrt{\chi }}. \label{uli-measure-1}$$ Considering dust models ($\gamma =1$) and the subset of all big-bang models which reach a certain Hubble parameter $H_0$ at a time $t_0$ during their evolution this measure becomes $$\mathrm{\pi }\propto \sqrt{\frac{\Omega _{m0}}{|\Omega _0-1|^{3/2}}}\left| \frac 1{\Omega _{m0}}-\frac{3/2}{\Omega _0-1}\right| \mathrm{d}\Omega _{m0} \mathrm{d}\Omega _{\Lambda 0},$$ with $\Omega _{\Lambda 0}\leq 1+\Omega _{m0}/2$. On the other hand, as mentioned above, there is a convenient parametrisation for this particular subset of models (Ehlers and Rindler 1989) in terms of the observables $ \Omega _{m0}$ and $\Omega _{\Lambda 0}$ (in Ehlers and Rindler (1989) an additional radiation component was also included). Using this parametrisation yields the minimum information measure $$\mathrm{\pi }\propto \frac 1{\sqrt{\Omega _{m0}}}\mathrm{d}\Omega _{m0} \mathrm{d}\Omega _{\Lambda 0},$$ which is clearly different from the above result. Secondly, the measure is in general non-normalisable and hence there is no normalisable maximum-entropy distribution. Without additional information we are not able to calculate certain probabilities. Since it seems questionable whether there will ever be additional information about the ensemble of universes available, one has to accept that certain questions will have no well defined probabilities. It should be mentioned that we encounter similar problems when we want to find a probability measure for physical parameters like the gravitational constant $G$. Let us assume that $G$ can take any non-zero positive value. Jaynes’ principle then suggests the probability measure $\mathrm{\pi } _G\propto \frac{\mathrm{d}G}G$. On the other hand, if we decide to use [^8] $m=\sinh (G)$ as our parameter, then we find the different measure $$\mathrm{\pi =}\frac{\mathrm{d}m}m=\frac{\cosh (G)}{\sinh (G)}\mathrm{d}G.$$ Distribution Functions on $\mathcal{M}_\mathrm{FLRW}$ ----------------------------------------------------- Now, having properly parametrised $\mathcal{M}_{\mathrm{FLRW}}$ and defined a measure on it, we can represent particular multiverses by giving distribution functions over the parameter-space (as discussed in Section 3.4). Given a distribution function $f$ it determines the number of universes in a small parameter-interval by $$\mathrm{d}N=f(p_I)\mathrm{\pi },$$ which is invariant under a change of parametrisation. Hence it is the combination of measure and distribution function which is of importance. While distribution functions can be parametrised by any set of coordinates over the possibility space, we need different distribution functions for different possibility spaces. For example, if universes in a multiverse *must* have a common value for the gravitational constant $G$ then distribution functions must not depend on $G$. It is clear that a particular distribution function can be expressed in any set of coordinates. Obviously there is a vast set of possible distribution functions. We want to examine some particular examples. Firstly, one could have a distribution function which is constant over the parameter-space. The actual ensemble then really depends on the measure and is the maximum-entropy distribution (which maximises (\[uli-entropy\])[^9]). If we choose the observational quantities $H,\Omega _m,\Omega _\Lambda $ to represent the model and allow for different values of the gravitational coupling constant $G$ this would be $$f(H,\Omega _m,\Omega _\Lambda ,G)=\mathrm{const.}$$ for all allowed values of the stated parameters. On the other hand, if we choose the constants of motion as coordinates in possibility space $$f(k,\chi ,\Lambda ,G)=\mathrm{const.}$$ The probability $P_{\mathcal{A}}$ to find a universe in a certain parameter-region $\mathcal{A}$ is given by $$P_{\mathcal{A}}=\frac{\int_{\mathcal{A}}f(k,\chi ,\Lambda ,G)\mathrm{\pi }.}{ \int f(k,\chi ,\Lambda ,G)\mathrm{\pi }.},$$ where the integral in the denominator extends over the whole possibility space. For many distribution functions, like for the above constant distribution function together with (\[uli-measure-1\]), this expression is not well defined. Let us assume that the measure is non-integrable, i.e., non-normalisable, and that we have a constant distribution function. If the set $\mathcal{A}$ does not include any point of non-integrability of the measure then $P_{\mathcal{A}}=0$, if it includes all points of non-integrability then $P_{\mathcal{A}}=1$, but if it includes only some of the non-integrabilities then $P_{\mathcal{A}}$ is not well defined. Of course, the above expression might be integrable for all sets $\mathcal{A} $ given a good distribution function. For instance the distribution function $$f(k, \chi,\Lambda )\propto \exp (-\chi -\Lambda ^{2})$$ together with the measure (\[uli-measure-1\]) is integrable everywhere. On the other hand $$f(k, \chi ,\Lambda )\propto \frac{\exp (-\Lambda ^{2})}{\sqrt{\chi +1}}$$ diverges for $\chi \rightarrow \infty $. A distribution function can also introduce an additional divergence, for instance $$f(k, \chi, \Lambda) \propto \frac{\exp (-\Lambda ^{2})}{\sqrt{\chi }}$$ is non-integrable at $\chi =0$ and $\chi \rightarrow \infty $. A multiverse might contain a finite or infinite countable number of universes. In these cases the distribution function contains Dirac $\delta$ -functions, e.g., $$f(k, \chi, \Lambda)= \left\{ \matrix{\ \sqrt{\chi}5\delta^2(\chi-3, \Lambda-.5) & \mbox{ for }k=1\cr 2\delta(\Lambda-.7)& \mbox{ for }k=0\cr \sqrt{\chi}3\delta^2(\chi-.1,\Lambda-2)& \mbox{ for }k=-1 } \right.$$ which represents a multiverse which contains $5$ copies of closed FLRW models with $\chi=3$ and $\Lambda=.5$, etc. The distribution function $$f(k, \chi, \Lambda)= \left\{ \matrix{\ \sqrt{\chi} \sum_{i=1}^\infty \delta(\chi-1/i, \Lambda-1) & \mbox{ for } k=1\cr 0 &\mbox{ for } k \ne 1 } \right.$$ represents an ensemble with a countably infinite number of universes – all are closed with $\Lambda=1$ and one for each $\chi=1/i$. Similarly one could imagine an ensemble of $10^7$ copies of our universe, which would be represented by the distribution function $$f(k,\chi, \Lambda) = 10^7 \sqrt{\chi} \delta^{k_0}_{k} \delta(\chi-\chi_0)\delta(\Lambda-\Lambda_0),$$ where $k_0, \chi_0, \Lambda_0$ represent the parameter values for our universe. This is very unlikely in terms of a generating mechanism, but for ensembles without generating mechanisms it is as likely as any other possibility. If a multiverse is “tested” by its prediction that our universe is a likely member, then such an ensemble should be the most satisfying one – but then we might just as well be happy with one copy, i.e., just our universe. Similar distribution functions determine the distribution of physical parameters like the gravitational constant $G$. For example with $G \in \mathbf{R}$ the minimum information measure is $\mathrm{d}\mu _G=\mathrm{d}G$ and $$f(G)=\exp (-G^2)$$ gives a Gaussian distribution around $G=0$. If, on the other hand $G\in (0,\infty )$ the measure is $\mathrm{d}G/G$ and $f(g)=G\exp (-G)$ would be an example of a distribution function. One can imagine various types of distributions, e. g., a Gaussian distribution in $G$ or in $H_{0}$, or in the other parameters. But, in order to establish these in a non-arbitrary way, we need a theory of how this particular ensemble is selected for from all the other possible ones. The problem of infinities again ------------------------------- Even within the restricted set of FLRW models, one of the most profound issues is the problem of realised infinities: if all that is possible in this restricted subset happens, we have multiple infinities of realised universes in the ensemble. First, there are an infinite number of possible spatial topologies in the negative curvature case, so an infinite number of ways that universes which are locally equivalent can differ globally. Second, even though the geometry is so simple, the uncountable continuum of numbers plays a devastating role locally: is it really conceivable that FLRW universes actually occur with *all* values independently of both the cosmological constant and the gravitational constant, and also all values of the Hubble constant at the instant when the density parameter takes the value 0.97? This gives 3 separate uncountably infinite aspects of the ensemble of universes that is supposed to exist. The problem would be allayed if spacetime is quantized at the Planck level, as suggested for example by loop quantum gravity. In that case one can argue that all physical quantities also are quantized, and the uncountable infinities of the real line get transmuted into finite numbers in any finite interval – a much better situation. We believe that this is a physically reasonable assumption to make, thus softening a major problem for many ensemble proposals. But the intervals are still infinite for many parameters in the possibility space. Reducing the uncountably infinite to countably infinite does not in the end resolve the problem of infinities in these ensembles. It is still an extraordinarily extravagant proposal. The anthropic subset -------------------- We can identify those FLRW universes in which the emergence and sustenance of life is possible at a broad level[^10] – the necessary cosmological conditions have been fulfilled allowing existence of galaxies, stars, and planets if the universe is perturbed, so allowing a non-zero factor $\Pi =$ $ P_{gal}*R*f_S*f_p*n_e\,$ as discussed above. These are indicated in the Figures above (anthropic universes are those intersecting the regions labelled “life allowing”). The fraction of these that will actually be life-bearing depends on the fulfilment of a large number of other conditions represented by the factor $F=f_l*f_i,\,$ which will also vary across a generic ensemble, and the above assumes this factor is non-zero.$.$ On the origin of ensembles ========================== Ensembles have been envisaged both as resulting from a single causal process, and as simply consisting of discrete entities. We discuss these two cases in turn, and then show that they are ultimately not distinguishable from each other. Processes Naturally Producing Ensembles --------------------------------------- Over the past 15 or 20 years, many researchers investigating the very early universe have proposed processes at or near the Planck era which would generate a really existing ensemble of expanding universe domains, one of which is our own observable universe. In fact, their work has provided both the context and stimulus for our discussions in this paper. Each of these processes essentially selects a really existing ensemble through a generating process from a set of possible universes, and often lead to proposals for a natural definition of a probability distribution on the space of possible universes. Here we briefly describe some of these proposals, and comment on how they fit within the framework we have been discussing. Andrei Linde’s (1983, 1990) chaotic inflationary proposal (see also Linde (2003) and references therein) is one of the best known scenarios of this type. The scalar field (inflaton) in these scenarios drives inflation and leads to the generation of a large number of causally disconnected regions of the Universe. This process is capable of generating a really existing ensemble of expanding FLRW-like regions, one of which may be our own observable universe region, situated in a much larger universe that is inhomogeneous on the largest scales. No FLRW approximation is possible globally; rather there are many FLRW-like sub-domains of a single fractal universe. These domains can be very different from one another, and can be modelled locally by FLRW cosmologies with different parameters. Linde and others have applied a stochastic approach to inflation (Starobinsky 1986, Linde, *et al.* 1994, Vilenkin 1995, Garriga and Vilenkin 2001, Linde 2003), through which probability distributions can be derived from inflaton potentials along with the usual cosmological equations (the Friedmann equation and the Klein-Gordon equation for the inflaton) and the slow-roll approximation for the inflationary era. A detailed example of this approach, in which specific probability distributions are derived from a Langevin-type equation describing the stochastic behaviour of the inflaton over horizon-sized regions before inflation begins, is given in Linde and Mezhlumian (2003) and in Linde *et al.* (1994). The probability distributions determined in this way generally are functions of the inflaton potential. This kind of scenario suggests how overarching physics, or a law of laws(represented by the inflaton field and its potential), can lead to a really existing ensemble of many very different FLRW-like regions of a larger Universe. However these proposals rely on extrapolations of presently known physics to realms far beyond where its reliability is assured. They also employ inflaton potentials which as yet have no connection to the particle physics we know at lower energies. And these proposals are not directly observationally testable – we have no astronomical evidence the supposed other FLRW-like regions exist. Thus they remain theoretically based proposals rather than established fact. There additionally remains the difficult problem of infinities: eternal inflation with its continual reproduction of different inflating domains of the Universe is claimed to lead to an infinite number of universes of each particular type (Linde, private communication). How can one deal with these infinities in terms of distribution functions and an adequate measure? As we have pointed out above, there is a philosophical problem surrounding a realised infinite set of any kind. Finally, from the point of view of the ensemble of all possible universes often invoked in discussions of multiverses, all possible inflaton potentials should be considered, as well as all solutions to all those potentials. They should all exist in such a multiverse, which will include chaotic inflationary models which are stationary as well as those which are non-stationary. Many of these potentials may yield ensembles which are uninteresting as far as the emergence of life is concerned, but some will be bio-friendly. The price of this process for creating anthropically favourable universe regions is the multiplication of realised infinities, most of which will be uncountable (for example the parameters in any particular form of inflaton potential will take all possible values in an interval of real numbers). Probability distributions for the cosmological constant ------------------------------------------------------- Weinberg (2000) and Garriga and Vilenkin (2001) derive a probability distribution for the cosmological constant in the context of an ensemble of regions generated in the same inflationary sequence via the action of a given inflaton potential where the cosmological constant is given by the potential energy of a scalar-field. In multi-domain universes, where spatial variations in a scalar-field cause different regions to inflate at different rates, the cosmological constant should be distributed according to some probability distribution $P(\rho _\Lambda )$. During inflation the scalar field undergoes randomisation by quantum fluctuations, such that later on its values in different regions are distributed according to the length in field space (Garriga and Vilenkin 2002). This leads to a probability distribution (or distribution function – the probability distribution is just the normalised distribution function) of values of the vacuum-energy density $\rho _\Lambda $ in these regions given by $$P(\rho _\Lambda )\mathrm{d}\rho _\Lambda \propto \frac{\mathrm{d}\rho _\Lambda }{\mid V^{\prime }(\phi )\mid },$$ where $V(\phi )$ is the inflaton potential, and the prime signifies differentiation with respect to the inflaton $\phi $. It has been suggested (Vilenkin 1995, Weinberg 1997) that the way the probability distribution for existence of galaxies depends on the cosmological constant can be approximated by $$P_{gal}=N(\rho _\Lambda )P(\rho _\Lambda ) \label{dm-l-01}$$ where $N(\rho _\Lambda )$ is the fraction of baryons that form galaxies. The requirement of structure formation as a pre-requisite for life places strong anthropic constraints on the domains in which observers could exist; these constraints must be satisfied in the really existing universe. Let us first note that galaxy formation is only possible for a narrow range around $\rho _\Lambda =0$ (Weinberg 2000). It has been shown that anthropic restrictions demand $\rho _\Lambda \lesssim 10^{-28} \rm \frac g{cm^3}$ (Kallosh and Linde 2002, Garriga and Vilenkin 2002). Consequently the anthropic selection factor $N(\rho _\Lambda )$ is sharply peaked and vanishes for $ |\rho _\Lambda |>\rho _{\Lambda \max }$ for some $\rho _{\Lambda \max }$, which is of the same order of magnitude as the observed cosmological constant. In scalar-field models $P(\rho _\Lambda )$ is in direct relation to the *a priori* distribution of the scalar-field fluctuations and it has been argued (Weinberg 2000) that for a wide class of potentials the variations of $P(\rho _\Lambda )$ over the anthropically allowed range (where $N(\rho _\Lambda )\neq 0$) should be negligible. Nevertheless, as has been shown in (Vilenkin-Garriga 2002) this is not always the case, in particular for power-law potentials $V(\phi )=\phi ^n$ with $n>1$ one finds an integrable divergence at $\rho _\Lambda =0$. It is clear that a similar relation to (\[dm-l-01\]) should hold for multiverses in the wider sense. Nevertheless, one could imagine multiverses containing universes with and without scalar-field, or with different potentials. Hence we cannot link the distribution of the cosmological constant to that of the scalar-field in a unique way, and there is a vast choice for possible *a priori* probability distributions for the cosmological constant. Let us assume that the cosmological constant is a remnant of some underlying (unknown) theory and as such might be restricted to some domain of values. Depending on this domain one finds different possible minimum information measures, which result from Jaynes’ principle. If the domain is given by all real numbers then the (non-normalisable) measure will be constant. If on the other hand the domain is given by all positive real numbers then the minimum-information measure gives an (non-normalisable) *a priori* probability distribution proportional to $1/\rho _\Lambda $ (Kirchner and Ellis 2003). In this case the divergence is located inside the anthropically allowed region and is non-integrable. For this case the expectation value vanishes, i.e., $$\bar{\rho}_\Lambda =\frac{\int_0^{\rho_{\Lambda \max}}\rho _{\Lambda} \frac 1{\rho _\Lambda }\mathrm{d} \Lambda }{\int_0^{\rho_{\Lambda \max }} \frac 1{\rho _\Lambda }\mathrm{d}\Lambda }=0$$ and we fail to explain the observed non-zero value of the cosmological constant. An interesting alternative is given by allowing the cosmological constant to take values in the domain $R^{+}\cup \{0\}$ (e.g., if the cosmological constant prediction is given by a quadratic term). The minimum information-measure is then proportional to $1/\sqrt{\rho _\Lambda }$. Again there is a divergence in the anthropically allowed region, but this time it is integrable. The expectation value becomes $$\bar{\rho}_\Lambda =\frac{\int_0^{\rho _{\Lambda \max }}\rho _\Lambda \frac 1{ \sqrt{\rho _\Lambda }}\mathrm{d}\Lambda }{\int_0^{\rho _{\Lambda \max }}\frac 1{ \sqrt{\rho _\Lambda }}\mathrm{d}\Lambda }=\frac 13\rho _{\Lambda \max }.$$ The existence of regularities ----------------------------- Consider now a genuine multiverse. Why should there be any regularity at all in the properties of universes in such an ensemble, where the universes are completely disconnected from each other? If there are such regularities and specific resulting properties, this suggests a mechanism creating that family of universes, and hence a causal link to a higher domain which is the seat of processes leading to these regularities. This in turn means that the individual universes making up the ensemble are not actually independent of each other. They are, instead, products of a single process, as in the case of chaotic inflation. A common generating mechanism is clearly a causal connection, even if not situated in a single connected spacetime – and some such mechanism is needed if all the universes in an ensemble have the same class of properties, for example being governed by the same physical laws or meta-laws. The point then is that, as emphasized when we considered how one can describe ensembles, any multiverse with regular properties that we can characterise systematically is necessarily of this kind. If it did not have regularities of properties across the class of universes included in the ensemble, we could not even describe it, much less calculate any properties or even characterise a distribution function. Thus in the end the idea of a completely disconnected multiverse with regular properties but without a common causal mechanism of some kind is not viable. There must necessarily be some pre-realisation causal mechanism at work determining the properties of the universes in the ensemble. What are claimed to be totally disjoint universes must in some sense indeed be causally connected together, albeit in some pre-physics or meta-physical domain that is causally effective in determining the common properties of the multiverse. Related to this is the issue that we have emphasized above, namely where does the possibility space come from and where does the distribution function come from that characterises realised models? As emphasized above, we have to assume that some relevant meta-laws pre-exist. We now see that we need to explain also what particular meta-laws pre-exist. If we are to examine ‘all that might be, exists’, then we need to look at the ensemble of all such meta-laws and a distribution function on this set. We seem to face an infinite regress as we follow this logic to its conclusion, and it is not clear how to end it except by arbitrarily calling a stop to this process. But then we have not looked at all conceivable possibilities. Testability and Existence ========================= Finally, the issue of evidence and testing has already been briefly mentioned. This is at the heart of whether an ensemble or multiverse proposal should be regarded as physics or as metaphysics. Evidence and existence ---------------------- Given all the possibilities discussed here, which specific kind of ensemble is claimed to exist? Given a specific such claim, how can one show that this is the particular ensemble that exists rather than all the other possibilities? There is no direct evidence of existence of the claimed other universe regions, nor can there be any, for they lie beyond the visual horizon; most will even be beyond the particle horizon, so there is no causal connection with them; and in the case of a true multiverse, there is not even any possibility of any indirect causal connection of any kind - the universes are then completely disjoint and nothing that happens in any one of them is linked to what happens in any other one. What weight does a claim of such existence carry, in this context when no direct observational evidence can ever be available? The point is that there is not just an issue of showing a multiverse exists - if this is a scientific proposition one needs to be able to show which specific multiverse exists; but there is no observational way to do this. Indeed if you can’t show which particular one exists, it is doubtful you have shown any one exists. What does a claim for such existence mean in this context? These issues are discussed in more depth in the accompanying philosophical paper, where we consider the various ways one may claim entities exist even when there is no direct or even indirect evidence for such existence. One ends up in deep philosophical waters. That is unavoidable if one is to seriously argue the claim for existence of a multiverse. Even the concept of what ‘existence’ might mean in this context needs careful consideration. Observations and Physics ------------------------ The one way one might make a reasonable claim for existence of a multiverse would be if one could show its existence was a more or less inevitable consequence of well-established physical laws and processes. Indeed, this is essentially the claim that is made in the case of chaotic inflation. However the problem is that the proposed underlying physics has not been tested, and indeed may be untestable. There is no evidence that the postulated physics is true in this universe, much less in some pre-existing metaspace that might generate a multiverse. Thus belief in the validity of the claimed physics that could lead to such consequences is just that, a belief - it is based on unproved extrapolation of established physics to vastly beyond where it has been tested. The issue is not just that the inflaton is not identified and its potential untested by any observational means - it is also that, for example, we are assuming quantum field theory remains valid far beyond the domain where it has been tested, and we have faith in that extreme extrapolation despite all the unsolved problems at the foundation of quantum theory, the divergences of quantum field theory, and the failure of that theory to provide a satisfactory resolution of the cosmological constant problem. Observations and probabilities ------------------------------ The ‘doomsday argument’ has led to a substantial literature on relating existence of universe models to evidence, based on analysis of probabilities, often using a model of choosing a ball randomly from an urn, and of associated selection effects (see e.g. Bostrom 2002). However usually these models either in effect assume an ensemble exists, or else are content to deal with potentially existing ensembles rather than actually existing ones (see e.g. Olum 2002). That does not deal with the case at hand. One would have to extend those arguments to trying to decide, on the basis of a single ball drawn from the urn, as to whether there was one ball in the urn or an infinite number. It is not clear to us that the statistical arguments used in those papers leads to a useful conclusion in this singular case, which is the case of interest for the argument in this paper. In any case, in the end those papers all deal just with observational probabilities, which are never conclusive. Indeed the whole reason for the anthropic literature is precisely the fact that biophilic universes are clearly highly improbable within the set of all possible universes (see e.g. the use of Anthropic arguments as regards the value of $\Lambda $ referred to in Section 5.2). We are working in a context where large improbabilities are the order of the day. Indeed that is why multiverse concepts were introduced in the first place - to try to introduce some form of scientific explanation into a context where the probabilities of existence of specific universe models preferred by observation are known to be very small. Observations and disproof ------------------------- Despite the gloomy prognosis given above, there are some specific cases where the existence of a chaotic inflation (multi-domain) type scenario can be disproved. These are when we live in a ‘small universe’ where we have already seen right round the universe (Ellis and Schreiber 1986, Lachieze-Ray and Luminet 1995) for then the universe closes up on itself in a single FLRW-like domain and so no further such domains that are causally connected to us in a single connected spacetime can exist. This ‘small universe’ situation is observationally testable, and indeed it has been suggested that the CBR power spectrum might already be giving us evidence that this is indeed so, because of its lack of power on the largest angular scales (Luminet et al, 2003). This proposal can be tested in the future by searching for identical circles in the CMB sky. That would disprove the usual chaotic inflationary scenario, but not a true multiverse proposal, for that cannot be shown to be false by any observation. Neither can it be shown to be true. Conclusion ========== The introduction of the multiverse or ensemble idea is a fundamental change in the nature of cosmology, because it aims to challenge one of the most basic aspects of standard cosmology, namely the uniqueness of the universe (see Ellis 1991, 1999 and references therein). However previous discussions have not made clear what is required in order to define a multiverse, although some specific physical calculations have been given based on restricted low-dimensional multiverses. The aim of this paper is to make clear what is needed in order to properly define a multiverse, and then examine some of the consequences that flow from this. Our fundamental starting point is the recognition that there is an important distinction to be made between possible universes and realised universes, and our main conclusion is that a really existing ensemble or multiverse is not *a priori* unique, nor uniquely defined. It must somehow be selected for. We have pointed out a clear distinction between an ensemble of possible universes $\mathcal{M}$,and an ensemble of really existing universes, which is envisioned as generated by the given primordial process or action of an overarching cosmic principle. These effectively select a really existing multiverse from the possibilities in $\mathcal{M}$, and, as such, effectively define a distribution function over $\mathcal{M}$. Thus, there is a definite causal connection, or law of laws , relating all the universes in these multiverses. It is this really existing ensemble of universes, *not* the ensemble of all possible universes, which provides the basis for anthropic arguments. Anthropic universes lie in a small subset of $\mathcal{M}$, whose characteristics we understand to some extent. It is very likely that the simultaneous realisation of *all* the conditions for life will pick out only a very small sector of the parameter space of all possibilities: anthropic universes are fine-tuned. The fine-tuning problem is very controversial. Two counter-attacks maintain that there is no fine-tuning problem, so it is not necessary to construct solutions to it by employing the multiverse idea. The first promotes the view that whatever happens will always be unlikely (any hand of cards is as unlikely as any other). Thus, since it is just an example of chance, there is nothing special about a universe that admits life. The counter response is that the existence of life is quite unlike anything else in the physical world – its coming into being is not just like choosing one out of numerous essentially identical hands of cards. It is like being transformed into an entirely different higher level game, and so does indeed require explanation. The second counter-attack argues that inflation explains the current state of the universe, making its apparently unlikely state probable. However, this move is only partially successful, since very anisotropic or inhomogeneous models may never inflate. The counter response is that this does not matter: however small the chances are, if it works just once then that is sufficient to give a model close enough to the standard FLRW cosmological models to be friendly to life. But this does not account for the rest of the coincidences enabling life, involving particle masses and the values of the fundamental constants. Perhaps progress in quantum cosmology will in the future lead to some unique theory of creation and existence that will guide the discussion. At present, uniqueness eludes us. Among those universes in which the necessary cosmic conditions for life have been fulfilled is the subset of almost-FLRW universes which are possible models of our own observable universe, given the precision of the observational data we have at present. It is, however, abundantly clear that really existing multiverses which can be defined as candidates for the one to which our universe belongs are *not* unique, and neither their properties nor their existence is directly testable. The only way in which arguments for the existence of one particular kind of multiverse would be scientifically acceptable is if, for instance, there would emerge evidence (either direct or indirect) for the existence of specific inflaton potential which would generate one particular kind of ensemble of expanding universe domains. Despite these problems, the idea of a multiverse is probably here to stay with us - it is an important concept that needs exploration and elucidation. Does the idea that ‘all that can exist, exists’ in the ensemble context provide an explanation for the anthropic puzzles? Yes it does do so. The issue of fine tuning is the statement that the biophilic set of universes is a very small subset of the set of possible universes; but if all that can exist exists then there are universe models occupying this biophilic subspace. However there are the following problems: (i) the issue of realised infinities discussed above, (ii) the problem of our inability to describe such ensembles because we don’t know what all the possibilities are, so our solution is in terms of a category we cannot fully describe, and (iii) the multiverse idea is not testable or provable in the usual scientific sense; existence of the hypothesized ensemble remains a matter of faith rather than proof. Furthermore in the end,  it simply represents a regress of causation. Ultimate questions remain: Why this multiverse with these properties rather than others? What endows these with existence and with this particular type of overall order? What are the ultimate boundaries of possibility – what makes something possible, even though it may never be realised? In our view these questions - Issues 1 and 2 discussed in this paper - cannot be answered scientifically because of the lack of any possibility of verification of any proposed underlying theory. They will of necessity have to be argued in philosophical terms. The concept of a multiverse raises many fascinating issues that have not yet been adequately explored. The discussion given here on how they can be described will be useful in furthering this endeavour. Acknowledgements {#acknowledgements .unnumbered} ================ We thank A. Linde, A. Lewis, A. Malcolm, and J.P. Uzan for helpful comments and references related to this work and an anonymous referee of an earlier version for comments. GFRE and UK acknowledge financial support from the University of Cape Town and the NRF (South Africa). [99]{} Bostrom, N. 2002, *Anthropic Bias* (Routledge). Barrow, J D and Tipler, F J 1986, *The Cosmic Anthropological Principle* (Oxford University Press). Davies, P. 2003, Multiverse or Design? Reflections on a ‘Third Way’, unpublished manuscript and talk given at the Centre for Theology and the Natural Sciences, Berkeley, California, March 22, 2003. Drake, F. and Shostak, S. *Astrobiology: The Search for Life in Space* (Cambridge University Press). Ehlers, J. and Rindler, W. 1989, ‘A phase space representation of Friedmann-Lemaître universes containing both dust and radiation and the inevitability of a big bang’. *Mon Not Roy Ast Soc* **238**, 503-521. Ellis, G F R 1967, “The dynamics of pressure-free matter in general relativity”, *Journ Math Phys* 8, 1171-1194. Ellis, G F R 1971, “Relativistic Cosmology”. In *    General Relativity and Cosmology, Proc Int School of Physics “Enrico Fermi”* (Varenna), Course XLVII. Ed. R K Sachs (Academic Press), 104-179. Ellis, G F R 1991, “Major Themes in the relation between Philosophy and Cosmology”. *Mem Ital Ast Soc* 62, 553-605. Ellis, G F R 1999, “Before the Beginning: Emerging Questions and Uncertainties”. In *Toward a New Millenium in Galaxy Morphology*, Ed. D Block, I Puerari, A Stockton and D Ferreira (Kluwer, Dordrecht, 2000). *Astrophysics and Space Science* 269-279: 693-720. Ellis, G. F. R. and Brundrit, G. B. 1979, Life in the infinite universe. *Qu Journ Roy Ast Soc* **20**, 37-41. Ellis, G F R and Bruni, M. 1989, A covariant and gauge-free approach to density fluctuations in cosmology. *Phys Rev* D40, 1804-1818. Ellis, G.F.R. and MacCallum, M A H 1969, A class of homogeneous cosmological models. *Comm Math Phys* 12, 108-141. Ellis, G.F.R. and Maartens, R. 2003. “The emergent universe: Inflationary cosmology with no singularity”. gr-qc/0211082. Ellis, G.F.R., Murugan, J. and Tsagas, C. G. 2003. “The Emergent Universe: An Explicit Construction”. gr-qc/0307112 Ellis, G.F.R, and Schrieber, G. 1986. *Phys Lett* A115, 97. Ellis, G.F.R, and van Elst, H 1999, “Cosmological Models” (Cargese Lectures 1998). In *Theoretical and Observational Cosmology*. Ed. M. Lachieze-Ray (Kluwer, Nato Series C: Mathematical and Physical Sciences, Vol 541, 1999), 1-116. \[gr-qc/9812046\]. van Elst, H. and Ellis, G.F.R. 1996, The covariant approach to LRS perfect fluid spacetime geometries. *Class. Quant. Grav.* 13:1099-1127 Fischer, A.E. and Marsden, J.E. 1979. “The initial value problem and the dynamical formulation of general relativity”. In *General Relativity: An Einstein Centenary Survey*, Ed. S. W. Hawking and W. Israel (Cambridge University Press). Garriga J., Vilenkin A. 2000, ‘On likely values of the cosmological constant’, arXiv/astro-ph/9908115. Garriga, J., and Vilenkin, A., 2001, ‘Many worlds in one’, *Phys Rev* D64, 043511 (arXiv: gr-qc/0102010); gr-qc/0102090. Garriga, J., and Vilenkin, A. 2002, arXiv: astro-ph/0210358, arXiv:astro-ph/0210358 Guth, A., and Pi, S. Y. 1981, *Phys. Rev. Lett.* 49, 1110. Hewitt, C. G., Horwood,  J. T. and Wainwright, J. 2003, “Asymptotic dynamics of the exceptional Bianchi cosmologies”, *Class. Quant. Grav.* **20** 1743-1756 \[gr-qc/0211071\]. Hilbert, D. 1964, On the Infinite , in *Philosophy of Mathematics*, edited by Paul Benacerraf and Hilary Putnam, Englewood Cliff, N. J. Prentice Hall, pp. 134-151. Horwood, J. T., Hancock, M. J., The, D., and Wainwright, J. 2003. “Late-time asymptotic dynamics of Bianchi VIII cosmologies”, *Class. Quant. Grav.* **20**  1757-1778 \[gr-qc/0210031\]. Kallosh R, Linde A.D. 2002, ‘M-theory, cosmological constant and anthropic principle’, arXiv/hep-th/0208157. Kirchner, U., and Ellis, G. F. R. 2003, A probability measure for FLRW models *Class. Quantum Grav.* **20** 1199-1213. Lachieze-ray, M., and Luminet, J.P. 1995. *Phys Rep* **254**, 135 \[gr-qc/9605010\]. Lim, W.C., van Elst, H., Uggla, C., and Wainwright, J. 2003, “Asymptotic isotropization in inhomogeneous cosmology”, gr-qc/0306118. Linde, A. D. 1983, *Physics Letters* 129B, 177. Linde, A. D. 1990, *Particle Physics and Inflationary Cosmology*, Harwood Academic Publishers, Chur, Switzerland. Linde, A. D. 2003, Inflation, Quantum Cosmology and the Anthropic Principle, to appear in *Science and Ultimate Reality: From Quantum to Cosmos*, honouring John Wheeler’s 90th birthday, J. D. Barrow, P. C. W. Davies and C. L. Harper, editors, Cambridge University Press (2003) (arXiv:hep-th: 0211048 v2). Linde, A. D., Linde, D. A., and Mezhlumian, A. 1994, *Phys. Rev*. D49, 1783. Linde, A. D., and Mezhlumian, A. 1993, Stationary Universe, arXiv:gr-qc/9304015. Luminet, J.-P., Weeks, J.R., Riazuelo, A., Lehoucq, R., and Uzan, J.-P. 2003, “Dodecahedral space topology as an explanation for weak wide-angle temperature correlations in the cosmic microwave background”. *Nature*, xxx yyy. Olum, K. 2002, The doomsday argument and the number of possible observers, *Phil Q* 52, 164 (gr-qc/0009081). Rees, M. J. 2001, *Just Six Numbers: The Deep Forces that Shape the Universe* (Basic Books). Rees, M.J. 2001 *Our Cosmic Habitat*, (Princeton University Press). M J Rees 2001, ‘Concluding Perspective’, astro-ph/0101268. Sciama, D. 1993, Is the universe unique?  In *Die Kosmologie der Gegenwart*, Ed G Borner and J Ehlers (Serie Piper). Smolin, L. 1999, *The Life of the Universe* (Oxford University Press)*.* Spitzer, R. J. 2000, Definitions of Real Time and Ultimate Reality, *Ultimate Reality and Meaning* **23** (No. 3), 260-276. Starobinsky, A. A. 1986, Stochastic De Sitter (Inflationary) Stage in the Early Universe, in *Current Topics in Field Theory, Quantum Gravity and Strings*, eds. H. J. de Vega and N. Sanchez (Springer, Heidelberg), p. 107. Stoeger, W. R. 2003, What is ‘the Universe’ that Cosmology Studies, to be published in the Ian Barbour Festschrift. Tegmark, M. 1998, Is the Theory of Everything Merely the Ultimate Ensemble Theory? *Annal Phys* 270, 1-51 (arXiv.org/gr-qc/9704009). Tegmark, M. 2003, Parallel Universes , astro-ph/0302131 and * Scientific American* (May 2003), 41-51. Uggla, C., van Elst, H., Wainwright, J. and Ellis, G. F. R. 2003, ‘The past attractor in inhomogeneous cosmology’. gr-qc/0304002. Vilenkin, A. 1995, *Phys. Rev. Lett*. 74, 846. Wainwright, J. and Ellis, G. F. R. (Eds) (1996). *The dynamical systems approach to cosmology*. Cambridge University Press. Weinberg, S. 1997, in *Critical Dialogs in Cosmology* , edited by Neil Turok, World Scientific, Singapore. Weinberg, S. 2000, The Cosmological Constant Problems’, astro-ph/0005265. ‘A Priori Probability Distribution of the Cosmological Constant’, *Phys. Rev.* D61, 103505 (2000) (astro-ph/0002387). [^1]: Connected implies Locally causally connected, that is all universe domains are connected by $C^{0}$ timelike lines which allow any number of reversals in their direction of time, as in Feynman’s approach to electrodynamics. Thus it is a union of regions that are causally connected to each other, and transcends particle and event horizons; for examples all points in de Sitter space time are connected to each other by such lines. [^2]: It has been suggested to us that in mathematics terms it does not make sense to distinguish identical copies of the same object: they should be identified with each other because they are essentially the same. But we are here dealing with physics rather than mathematics, and with real existence rather than possible existence, and then multiple copies must be allowed (for example all electrons are identical to each other; physics would be very different if there were only one electron in existence). [^3]: One way out would be, as quite a bit of work in quantum cosmology seems to indicate, to have time originating or emerging from the quantum-gravity dominated primordial substrate only “later.” In other words, there would have been a “time” or an epoch before time as such emerged. Past time would then be finite, as seems to be demanded by philosophical arguments, and yet the timeless primordial state could have lasted “forever,” whatever that would mean. This possibility avoids the problem of constructibility. [^4]: Many discussions implicitly suggest that this is the whole possibility space, as they only consider FLRW models as possibilities. However these clearly form a very small subspace of all geometrical possibilities. [^5]: The way the tetrad description given above relates to FLRW universes is described in detail in Ellis and MacCallum (1968); the standard coordinates given here are more convenient if one discusses only the FLRW models. [^6]: Hence, in general different models correspond to the same values of $ \Omega_m $ and $\Omega_\Lambda$, depending on the value of $H$. [^7]: It is worth noting that when $G=0$, we *do* obtain FLRW solutions to the Einstein-Field equations: the Milne universe which is effectively empty – no gravity effective. [^8]: However, it should be noted that any power ($m=G^n,m\in {^+},n\neq 0$) and logarithmic relationship ($m=\ln (G),m\in $) leads to the same measure $ \mathrm{d}\mu _{G}$. [^9]: In (\[uli-entropy\]) $P(x) \mathrm{d}x$ corresponds to $f(p_I)\pi$. [^10]: More accurately, perturbations of these models can allow life – the exact FLRW models themselves cannot do so.

ArXiv

--- abstract: | Magnetic ordering in TbMnO$_{3}$ was revisited by a neutron diffraction study. In addition to the previously reported *A*-type and *G*-type modulated structures with a propagation vector $(0,q_\mathrm{Mn},0)$,[@quezel77; @blasco00] we found *C*-type and *F*-type orderings of the Mn moments. All the components appear below $T_N^\mathrm{Mn} = 46$ K, and the *G*, *C*, and *F*-type components are enhanced below $T_\mathrm{lock} = 28$ K, where $q_\mathrm{Mn}$ ordering locks at its low temperature value. The locking of the propagation vector also yields the squaring up of the spin arrangement. The magnetic moments of Tb ions show a quasi long-range ordering below $T_N^\mathrm{Tb}=7$ K. It drastically promotes the development of the *G*, *C*, and *F*-type components while suppressing the *A*-type components. author: - 'R. Kajimoto' - 'H. Yoshizawa' - 'H. Shintani' - 'T. Kimura' - 'Y. Tokura' title: ' Neutron diffraction study of TbMnO$_3$: Magnetic structure revisited\' --- Perovskite manganese oxides $R$MnO$_3$ ($R = \mbox{trivalent rare earth ion}$) are famous as the parent materials of the colossal magnetoresistive manganites. The Mn ions in $R$MnO$_{3}$ become trivalent having the $t_{2g}^3 e_{g}^1$ configuration. This electronic configuration induces a degree of freedom of the $e_g$ orbital, which produces many novel properties. In the most well studied material, LaMnO$_{3}$, the $e_g$ orbitals form staggered ordering of $d_{3x^2-r^2}$ and $d_{3y^2-r^2}$ orbitals, which induces the layered-type (so-called *A*-type) antiferromagnetic ordering of Mn spins. In contrast, when the ionic radius of $R$ is substantially small, $R$MnO$_{3}$ shows a modulated spin ordering with a propagation vector $(0,q_\mathrm{Mn},0)$ (in the orthorhombic *Pbnm* cell). [@quezel77; @blasco00; @munoz01; @brinks01; @munoz02] The emergence of the incommensurate spin structure is explained as a competition between the nearest-neighbor (NN) spin interactions and the next-nearest-neighbor (NNN) interactions caused by the combination of the GdFeO$_{3}$-type distortion and the $d_{3x^2-r^2}/d_{3y^2-r^2}$ orbital ordering (We will refer to this model as the NN-NNN model.).[@kimura_prb] TbMnO$_{3}$ is one of the latter series of $R$MnO$_{3}$. It shows an incommensurate sinusoidal spin ordering with $q_\mathrm{Mn}\sim 0.295$ below $T_N \sim 41$ K with spins oriented along the \[010\] direction. $q_\mathrm{Mn}$ decreases as temperature is lowered, then locked at $q_\mathrm{Mn} = 0.28$ below $T_\mathrm{lock} \sim 30$ K.[@quezel77] Very recently, it was found that the spontaneous electric polarization $P$ parallel to the $c$ axis appears below $T_\mathrm{lock}$.[@kimura_nature] Moreover, the magnitude or the direction of $P$ can be drastically changed by applying a magnetic field.[@kimura_nature] The magnetic control of the ferroelectric polarization will provide an attractive possibility of new magnetoelectric devices. These recent discoveries in the $R$MnO$_{3}$ series motivated us to reinvestigate the magnetic properties of these materials in detail. Here, we report the result of a neutron diffraction study on the elastic properties of the magnetic ordering in TbMnO$_{3}$. The study of the inelastic properties by neutron scattering is underway, and will be published elsewhere.[@kaji_unpub] A single crystal of TbMnO$_{3}$ was grown by the floating zone method. A detailed procedure of the sample preparation was described elsewhere.[@kimura_nature] The neutron diffraction experiments were performed using the triple axis spectrometer GPTAS installed at the JRR-3M research reactor in JAERI, Tokai, Japan. A neutron wave length of $k_i = 2.57$ [Å]{}$^{-1}$ was selected by the 002 reflection of a pyrolytic graphite (PG) monochromator. The horizontal collimation of 40$^{\prime}$-40$^{\prime}$-40$^{\prime}$-blank (from monochromator to sample) was used for most of the measurements, while the 20$^{\prime}$-20$^{\prime}$-20$^{\prime}$-blank collimation was utilized for the investigation of temperature dependence of the wave vector of the magnetic ordering. Two PG filters were placed before the monochromator and after the sample to suppress contaminations of higher-order harmonics. The crystal structure has the *Pbnm* symmetry, and the measurements were performed in the $(0kl)$ scattering plane ($a^* = 1.08$ [Å]{}$^{-1}$ and $c^* = 0.853$ [Å]{}$^{-1}$). The sample was mounted in an aluminum can filled with helium gas, and was attached to the closed-cycle helium gas refrigerator. All the temperature dependence data were collected upon heating to avoid uncertainty due to the hysteresis. First, we surveyed magnetic reflections in the $(0kl)$ scattering plane with the relaxed collimation (40$^{\prime}$-40$^{\prime}$-40$^{\prime}$-blank). Figure \[kscan\_profs\] shows neutron scattering profiles along the $(0,k,0)$ and $(0,k,1)$ lines at 4 K, 10 K, and 55 K. At 4 K, surprisingly many superlattice reflections due to the magnetic moments of both Mn and Tb ions are observed at every position with $k\approx n/7$ ($n=\mbox{integers}$). At 10 K, on the other hand, only the Mn peaks are observed with broad peaks coming from the short-range order of the Tb moments. ![(Color online) Neutron diffraction profiles along (a) the $(0,k,0)$ line and (b) the $(0,k,1)$ line at 4, 10, and 55 K. Vertical lines indicate the positions of the Mn ($k\approx \mbox{integer} \pm 2/7$ and 6/7) and Tb ($k\approx \mbox{integer} \pm 3/7$ and 6/7) superlattice peaks. “Al” denotes diffractions from the aluminum sample cell.[]{data-label="kscan_profs"}](line_scans2.eps){width="0.95\hsize"} The magnetic reflections due to the Mn ordering were observed at $(h,k \pm q_\mathrm{Mn},l)$ with $q_\mathrm{Mn}=0.28$, where solid vertical lines are drawn in Fig. \[kscan\_profs\]. The observed Mn peaks can be classified into five groups depending on the values of $h$, $k$, and $l$ as *A*-type ($h+k=\mbox{even}$ and $l=\mbox{odd}$), *G*-type ($h+k=\mbox{odd}$ and $l=\mbox{odd}$), *C*-type ($h+k=\mbox{odd}$ and $l=\mbox{even}$), and *F*-type ($h+k=\mbox{even}$ and $l=\mbox{even}$). Quezel *et al.* observed strong *A*-type reflections and weak *G*-type peaks, but they did not observe any *C*-type and *F*-type reflections.[@quezel77] They concluded that the *A*-type peaks come from a sinusoidally modulated spin ordering with spins parallel to the $b$ axis. In contrast to Quezel’s result, we observed clear peaks of both *C*-type and *F*-type in addition to the *A*-type and *G*-type peaks. Since the small number of the observed magnetic peaks prevents us to unambiguously determine the real space spin configuration, we just performed a preliminary analysis of the magnetic structure: We calculated the intensities for the sinusoidal and the helical spin configurations assuming the spins lie on the $ab$, $bc$, or $ca$ planes, and compared them with the observed intensities at 10 K. The intensities of the *A*-type peaks are best described by the sinusoidal model with spins parallel to the $b$ axis as reported by Quezel *et al*. Analysis of the *G*-type peaks gave the similar result. For the *C*-type and *F*-type peaks, however, we failed to satisfactorily reproduce the observed intensities, though the existence of the peaks on the $(0,k,0)$ line indicates the spin arrangements have substantial $x$ or $z$ components. ![(a)-(d) Temperature dependences of the intensities of the Mn superlattice peaks of (a) *A*-type, (b) *G*-type, (c) *F*-type, and (d) *C*-type. (e) Temperature dependence of the intensity of the Mn third harmonics at $(0,0.835,1)=(0,3q_\mathrm{Mn},1)$.[]{data-label="Mn_Tdep"}](Mn_Tdep.eps) Figures \[Mn\_Tdep\](a)-(d) show temperature dependences of the scattering intensities of the Mn magnetic peaks. Each panel corresponds to the superlattice peak of (a) *A*-type: $(0,0.278,1)=(0,q_\mathrm{Mn},1)$, (b) *G*-type: $(0,0.721,1)=(0,1-q_\mathrm{Mn},1)$, (c) *F*-type: $(0,0.280,2)=(0,q_\mathrm{Mn},2)$, and (d) *C*-type: $(0,1.280,2)=(0,1+q_\mathrm{Mn},2)$. All the intensities start to increase at $T_N^\mathrm{Mn} = 46$ K, which is slightly higher than the reported transition temperatures.[@quezel77; @blasco00; @kimura_prb] There are two distinct anomalies in the temperature profiles below $T_N^\mathrm{Mn}$: with decreasing temperature, the slopes show upturns at $T_\mathrm{lock} = 28$ K, and then a drastic drop of the *A*-type peak (a) or increases of the others (b)-(d) below $T_N^\mathrm{Tb} = 7$ K. The former temperature coincides with the temperature where the wave vector of the Mn ordering locks at its low temperature value as described later. The latter anomaly is concomitant with the ordering of the Tb moments (see below). All the weak (*G*, *C*, and *F*) components show a similar temperature dependence, and they gradually develop below $T_N^\mathrm{Mn}$. Their temperature dependence is quite different from the sharp increase of the *A*-type component, suggesting that the weak components arise from some secondary effect. One of probable causes is the development of the ordering of the Tb moments, which is manifested by the steep increase of these components in compensation for the *A*-type component on the (quasi) long-range ordering of the Tb moments (Fig. \[Mn\_Tdep\]). In addition to the fundamental ($1q_\mathrm{Mn}$) peaks described above, we also observed the third harmonics ($3q_\mathrm{Mn}$) of each type of the Mn ordering at $(h,k \pm 3q_\mathrm{Mn},l)$ (dashed lines in Fig. \[kscan\_profs\]). Though we could not detect the third harmonics for the *F*-type ordering, it may be because of the weakness of the fundamental ($1q_\mathrm{Mn}$) *F*-type peaks. The existence of the higher harmonics of the Mn ordering means that the spin ordering is not an ideal sinusoidal wave. The ratio of the intensity of the $3q_\mathrm{Mn}$ peak to that of the $1q_\mathrm{Mn}$ peak is $\sim$10$^{-2}$. This value is much smaller than that for an ideal square wave (1/9), indicating the deviation of the spin arrangement from the sinusoidal wave is quite small. Nevertheless, the existence of the higher harmonics is interesting in that the NN-NNN model predicts the square ordering of the Ising-like spins.[@kimura_prb] In Fig. \[Mn\_Tdep\](e) is shown the temperature dependence of the intensity of the $3q_\mathrm{Mn}$ peak of the *A*-type ordering. It gradually develops below $T_\mathrm{lock}$, indicating that the sinusoidal ordering pattern of the Mn moments becomes distorted at $T < T_\mathrm{lock}$. The superlattice peaks due to the ordering of the Tb moments were observed at $(h,k \pm q_\mathrm{Tb},l)$ with $h,k,l=\mbox{integer}$ and $q_\mathrm{Tb}=0.42$ (dotted-and-dashed lines in Fig. \[kscan\_profs\]). The Tb peaks are observed in all zones similar to the Mn ordering. Though the previous study on a single crystal reported the Tb peaks are broad and less intense compared to the Mn peaks,[@quezel77] we observed quite sharp and much intense peaks at the lowest temperature. Figure \[Tb\_Tdep\](a) shows the temperature dependence of the Tb peak intensity at $(0,0.424,1)=(0,q_\mathrm{Tb},1)$. It shows a steep increase below $T_N^\mathrm{Tb} = 7$ K. In addition, superlattice peaks are observed at $k \pm 0.15$ in all zones (dashed lines in Fig. \[kscan\_profs\]). The intensity of these peaks shows the same temperature dependence as the Tb peaks \[Fig. \[Tb\_Tdep\](b)\], indicating that these peaks are related to the Tb ordering. Since the positions of the superlattices correspond to $2q_\mathrm{Tb}$, these peaks may be attributed to the second harmonics of the Tb ordering. The existence of the even harmonics means a space symmetry breaking.[@chatt94] On the other hand, a magnetic ordering can induce the second harmonics of lattice origin.[@pynn76] The exact origin of the peaks at $2q_\mathrm{Tb}$ remains an open question. ![Temperature dependences of the peak intensities due to the Tb ordering at (a) $(0,0.424,1) = (0,q_\mathrm{Tb},1)$ and (b) $(0,1.151,2) = (0,2-2q_\mathrm{Tb},2)$.[]{data-label="Tb_Tdep"}](Tb_Tdep.eps) Next, we measured temperature dependence of a profile of an *A*-type Mn peak to derive the change of the wave vector of the Mn spin ordering as a function of temperature. The NN-NNN model predicts that the wave number shows a step-wise temperature variation as locking at certain rational values.[@kimura_prb] In the previous diffraction measurements, however, only a smooth temperature dependence below $T_\mathrm{lock}$ was observed,[@quezel77; @kimura_prb] suggesting that the effect of the locking is, if it exists, relatively weak. In the present study, we collected the temperature dependence data at substantially fine intervals of temperature ($\Delta T \sim 0.5$ K) with the tight collimation (20$^{\prime}$-20$^{\prime}$-20$^{\prime}$-blank) to detect small anomalies in the wave vector of the Mn ordering. With this condition, the widths of the Mn peaks are still resolution limited for both the \[010\] and the \[001\] directions, while the Tb peaks have finite intrinsic widths. The correlation lengths of the Tb ordering, which is derived from the peak width of $(0,q_\mathrm{Tb},1)$ peak at 5 K by assuming a Gaussian line shape and deconvoluting the resolution, is $\sim$140 [Å]{} both along the \[010\] and \[001\] directions. We show in the inset of Fig. \[Mn\_prof\_Tdep\](a) the temperature variation of the profile of the Mn peak at $(0,q_\mathrm{Mn},1) \approx (0,0.28,1)$. One can see clearly the peak position shifts toward higher $Q$ as the temperature increases. By fitting these data to Gaussians (solid lines in the figure), we extract the temperature dependence of the peak height, the peak position (the wave number $q_\mathrm{Mn}$), and the peak width (full width at half-maximum, FWHM), which are summarized in Fig. \[Mn\_prof\_Tdep\]. With increasing temperature, the peak intensity shows an increase at $T < T_N^\mathrm{Tb}=7$ K, then decreases \[Fig. \[Mn\_prof\_Tdep\](a)\]. The slope of the temperature profile exhibits an inflection at $T_\mathrm{lock}=28$ K. Finally, the intensity vanishes at $T_N^\mathrm{Mn}=46$ K. For the wave number $q_\mathrm{Mn}$, it is nearly constant at $0.276~\mbox{r.l.u. (reciprocal lattice unit)} \approx 5/18$ (or 11/40) r.l.u. at low temperature with a slow decrease as a function of temperature \[Fig. \[Mn\_prof\_Tdep\](b)\]. It shows a steep rise above $\sim$34 K, but the upturn of the wave number starts at $T_\mathrm{lock}$ \[inset of Fig. \[Mn\_prof\_Tdep\](b)\]. It changes the slope of its temperature dependence at 38 K, then it increases monotonically without any locking behavior. For the peak width, it is resolution limited below $T_N^\mathrm{Mn}$, and diverges above $T_N^\mathrm{Mn}$ \[Fig. \[Mn\_prof\_Tdep\](c)\]. We also measured temperature dependence of the profile along the \[001\] direction, but did not find any anomalies. ![Temperature dependences of (a) the peak intensity, (b) the wave number, and (c) the peak width of the Mn $(0,q_\mathrm{Mn},1)$ peak. Inset of panel (a): Temperature dependence of the Mn peak profile at $(0,q_\mathrm{Mn},1)$ along the \[010\] direction.[]{data-label="Mn_prof_Tdep"}](0q1_kscan_Tdep.eps) According to the NN-NNN model, the wave number of the Mn ordering $q_\mathrm{Mn}$ decreases in a step-wise manner as lowering the temperature, and reaches zero in the ground state. However, the step-wise behavior is indistinct and $q_\mathrm{Mn}$ locks at the finite value below $T_\mathrm{lock}$ in the real material, although the observed temperature dependence of $q_\mathrm{Mn}$ semiquantitatively agrees with the model. This means some higher order effects than the NN and NNN spin interactions that stabilize the modulated structure must be considered. The coupling between the Mn moments and the Tb moments may play a role, but it fails to explain the fact that the similar incommensurate spin ordering and the locking behavior at the low temperature were also observed in orthorhombic YMnO$_{3}$, where the rare earth ion has no magnetic moment.[@munoz02] Further studies are needed to solve this problem, and the study of the dynamics will be helpful. To sum up the present study, we have investigated the ordering of the Mn and Tb moments in a single crystal of TbMnO$_{3}$ by neutron diffraction. The ordering process of the magnetic moments are summarized as follows: At $T_N^\mathrm{Mn} = 46$ K, the Mn moments shows a sinusoidally modulated *A*-type ordering with a propagation vector $(0,q_\mathrm{Mn},0)$, as reported by Quezel *et al*.[@quezel77] In addition, small *G*, *C*, and *F*-type with the same propagation vector coexist. $q_\mathrm{Mn}$ is $\sim$0.29 at $T_N^\mathrm{Mn}$, and smoothly decreases as the temperature is lowered showing a weak anomaly at 38 K. At $T_\mathrm{lock} = 28$ K, $q_\mathrm{Mn}$ locks at $0.276 \approx 5/18$ or 11/40. The locking of the wave vector amplifies the Mn moment, especially enhances the *G*, *C*, and *F*-type components, and squares the spin arrangements. The Tb moments order below $T_N^\mathrm{Tb} = 7$ K, which promotes the development of the additional Mn components. The authors thank S. Ishihara and K. Hirota for valuable discussions and critical reading of the manuscript. This work was supported by KAKENHI from the MEXT, Japan. [99]{} S. Quezel, F. Tcheou, J. Rossat-Mignod, G. Quezel, and E. Roudaut, Physica B **86-88**, 916 (1977). J. Blasco, C. Ritter, J. García, J. M. de Teresa, J. Pérez-Cacho, and M. R. Ibarra, Phys. Rev. B **62**, 5609 (2000). A. Munõz, M. T. Casáis, J. A. Alonso, M. J. Martínez-Lope, J. L. Martínez, M. T. Fernández-Díaz, Inorg. Chem. **40**, 1020 (2001). H. W. Brinks, J. Rodríguez-Carvajal, H. Fjellv[å]{}g, A. Kjekshus, and B. C. Hauback, Phys. Rev. B **63**, 094411 (2001). A. Muñoz, J. A. Alonso, M. T. Cassais, M. J. Martínez-Lope, J. L. Martínez, and M. T. Fernández-Díaz, J. Phys.: Condens. Matter **14**, 3285 (2002). T. Kimura, S. Ishihara, H. Shintani, T. Arima, K. T. Takahashi, K. Ishizaka, and Y. Tokura, Phys. Rev. B **68**, 060403(R) (2003). T. Kimura, T. Goto, H. Shintani, K. Ishizaka, T. Arima, and Y. Tokura, Nature (London) **426**, 55 (2003). R. Kajimoto, H. Mochizuki, H. Yoshizawa, H. Shintani, T. Kimura, and Y. Tokura, in preparation. T. Chattopadhyay, P. Burlet, J. Rossat-Mignod, H. Bartholin, C. Vettier, and O. Vogt, Phys. Rev. B **49**, 15096 (1994). R. Pynn, W. Press, S. M. Shapiro, S. A. Werner, Phys. Rev. B **13**, 295 (1976).

ArXiv

--- abstract: 'Rotation of polarization in an external gravitational field is one of the effects of general relativity that can serve as a basis for its precision tests. A careful analysis of reference frames is crucial for a proper evaluation of this effect. We introduce an operationally-motivated local reference frame that allows for a particularly simple description. We present a solution of null geodesics in Kerr space-time that is organized around a new expansion parameter, allowing a better control of the series, and use it to calculate the resulting polarization rotation. While this rotation depends on the reference-frame convention, we demonstrate a gauge-independent geometric phase for closed paths in general space-times.' author: - Aharon Brodutch - 'Tommaso F. Demarie' - 'Daniel R. Terno' title: Photon polarization and geometric phase in general relativity --- Introduction ============ Electromagnetic waves — visible light and other bands of the spectrum — are our prime source of information about the Universe [@pad1]. Since measurements of the light deflection near the Sun [@e1919] were made in 1919 wave propagation is used to test general relativity (GR). Two of the “classical tests" of GR, light deflection and time delay, can be understood in terms of geometric optics [@will]. The first post-eikonal approximation [@bw] allows to track the evolution of electric and magnetic fields along the light rays and thus discuss polarization. In this approximation we speak about photons with a null four-momentum $\ak$ and a transversal four-vector polarization $\af$. Both vectors are parallel-transported along the trajectory, which is a null geodesic [@mtw; @chandra]: $$\begin{aligned} & \ak\!\cdot\ak=0, \qquad \anab_\ak\ak=0, \label{momenone}\\ & \ak\!\cdot\af=0, \qquad \anab_\ak\af=0, \label{polone} \end{aligned}$$ where $\anab_\ak$ is a covariant derivative along $\ak$. In the last decades polarization has been yielding important astrophysical and cosmological data. Cosmic microwave background [@cmb], blazar flares [@fermi; @fermilat], astrophysical jets [@fermi; @honda], and searches for dark matter [@dark-lens] are just several examples where polarization conveys crucial information. Photons are commonly used as physical carriers of quantum information. The abstract unit of quantum information is a qubit (a quantum bit) [@qinfo], and two linearly independent polarizations encode the two basis states of a qubit. One of the branches of quantum information is quantum metrology, which aims to improve precision measurements by using explicitly quantum effects, such as entanglement [@macc]. Relativistic properties of the information carriers [@pt04; @alsing09] become important when quantum technology is used in precision tests of relativity [@qtest] or quantum information processing on the orbit [@milsat]. Gravity causes polarization to rotate. This effect is known as a gravimagnetic/Faraday/Rytov-Skrotskiĭ rotation [@skrot; @gravimag; @fayos; @kerr-farpol]. Helicity is invariant under rotations, but states of a definite helicity acquire phases $e^{\pm i\Delta\chi}$. Depending on the context we refer to $\Delta\chi$ either as a polarization rotation or as a phase. Once evaluated this phase can be encapsulated as a quantum gate [@qinfo] and incorporated into (quantum) communication protocols or metrology tasks. In the Schwarzschild space-time, as well as at the leading order of the post-Newtonian approximation, this phase is known to be zero [@skrot; @pleb; @god; @gravimag]. It is higher-order gravitational moments that are held responsible for the rotation of the polarization plane. In particular, the GR effects were shown to dramatically alter polarization of the X-ray radiation that is coming from the accretion disc of the (then presumed) black hole in Cyg X-1 [@bh:stark]. Numerous analytical and numerical studies of trajectories and polarizations in different models and astrophysical regimes were performed ([@gravimag; @fayos; @kerr-bray; @kerr-farpol; @lense; @burns; @kopeikin; @faraoni08], and references therein). The scenarios included fast-moving gravitating bodies, influence of gravitational lenses and propagation through gravitational waves. The results are often contradictory. Some of the contradictions result from genuine differences in superfluously similar physical situations [@fayos; @burns; @faraoni08]. On the other hand, polarization rotation is operationally meaningful only if the evolving polarization vector is compared with some standard polarization basis (two linear polarizations, right- and left-circular polarizations, etc.) at each point along the ray [@lpt1; @we-mach]. Setting up and aligning detectors requires alignment of local reference frames. Much work has been done recently on the role of reference frames in communications, especially in the context of quantum information [@Ref]. Partial knowledge of reference frames can lead to loss of communication capacity, and mistakes in identifying the information content of a physical system. The lack of definition for polarization standards and an *ad hoc* introduction of the angle adjustments is one reason for the variety of quoted values for the phase $\Delta\chi$. Even the consensual result of a zero phase in the Schwarzschild space-time should be qualified. Without specifying the appropriate reference frame it is either meaningless or wrong. In this article we investigate the role of local reference frames in defining $\Delta\chi$. The result is obviously gauge-dependent, and additional considerations should be used to fix the gauge. Communicating reference frames is sometimes a difficult procedure which may involve a high communication cost. On a curved background of GR it may require some knowledge of the metric at each point along the trajectory. Using the method presented by us in [@we-mach] we build a local reference frame in stationary space-times, and then fix the standard polarizations in a single construction using what we call the [*Newton gauge*]{}. Our construction does not require communication between the parties, gives a precise meaning to the idea that there is no polarization rotation in the Schwarzschild space-time and reproduces the absence of phase in the Minkowski space-time. From an operational point of view it allows us to set up detectors at any point in space based only on the local properties at that point. We illustrate the use of this gauge by studying polarization rotation in the Kerr space-time, and derive an explicit expression for $\Delta\chi$ in the scattering scenario that is described below. The calculations are based on the results of [@kerr-bray; @kerr-farpol; @chandra] with the added value of the Newton gauge. Careful bookkeeping is required with respect to the various coordinate systems used in different parts of the calculation, as well as different orders of the series expansion. To simplify the latter, we introduce a new expansion parameter. Closed paths result in a gauge-invariant gravity-induced phase. We discuss them in Section IV. The rest of the paper is organized as follows: First we discuss null trajectories in the Kerr space-time. In Section III we review the Wigner’s construction of setting standard polarizations in a local frame, set-up the Newton gauge and explicitly calculate the rotation in the scattering scenario, where the light is emitted and observed far from the gravitating body ($r_1,r_2\gg M,a$), but can pass close to it. Finally, we discuss gauge-dependent and gauge-invariant aspects of our results from mathematical and operational points of view (Section IV). Summary of the important facts about the Kerr space-time, as well as detailed calculations and special cases are presented in the appendices. We use $-+++$ signature, set $G=c=1$ and use Einstein’s summation convention in all dimensions. Three-dimensional vectors are written in boldface and the unit vectors are distinguished by carets, such as $\hbb$. Local tetrad components are written with carets on indices, such as $k^{\hat{\mu}}$, and the four-vector itself as $\ak=k^{\hat{\mu}}\are_{(\mu)}$, where $\are_{(\mu)}$ are vectors of a local orthonormal tetrad. Null geodesics ============== Null geodesics in the Kerr space-time ------------------------------------- The Kerr metric in the Boyer-Lyndquist coordinates is given by [@ll2] ds\^2=-(1-)dt\^2+dr\^2+\^2d\^2+(r\^2+a\^2+\^2)\^2d\^2-\^2dt d, where $M$ is the mass and $a=J/M$ the angular momentum per unit mass of the gravitating body. We use the standard notations \^2=r\^2+a\^2\^2, =r\^2-2Mr+a\^2. Thanks to the three conserved quantities — the energy $E=-k_0$, the $z$-component of the angular momentum $L$ and the constant $\eta$ [@carter; @chandra] — the geodesic equations in Kerr space-time are integrable in quadratures. When dealing with photons it is convenient to set the energy to unity and to re-scale other quantities [@chandra]. Explicit form and the asymptotic expansions of $D=L/E$ and $\eta$ are given in Appendix \[kerrdetail\]. Fixing the energy and using the null vector condition leaves us with only two components of the four-momentum $\ak$. In specifying the initial data we usually take $k^\theta_1$ and $k^\phi_1$ as independent. We discuss polarization from the point of view of static observers that are at rest in the “absolute" space $t=\mathrm{const}$. At every point (outside the ergosphere, if the model represents a black hole) we introduce a chronometric orthonormal frame [@fronov; @mtw] (Appendix \[kerrdetail\]), which will be used to express the initial conditions and observed quantities. The Hamilton-Jacobi equation for a null geodesic separates and the trajectories can be deduced from it [@chandra]. We label their initial and final points as $(r_1,\theta_1,\phi_1)$ and $(r_2,\theta_2,\phi_2)$, respectively. The gauge convention that we adapt (Sec. \[secnewton\]) makes the knowledge of $\Delta\phi$ redundant for discussions of polarization rotation in the scattering scenario (For the calculation of $\phi$ see Appendix \[kerrdetail\]). In particular we are interested in the integrals :=\^r=\^, \[ert\] where $$\begin{aligned} & R=r^4+(a^2-D^2-\eta)r^2+2M\bigl(\eta+(D-a)^2\bigr)r-a^2\eta, \\ & \Theta=\eta+a^2\cos^2\!\theta-D^2\cot^2\theta.\end{aligned}$$ Null geodesics in Kerr space-time are classified according to the sign of the constant $\eta$. The scattering scenario corresponds to $\eta>0$, and typically $\eta\sim r_1^2$. Following [@kerr-bray] we also assume that the angle $\theta$ reaches its extremal value (either maximum or minimum) only once, $\theta_1\rightarrow\theta_{\max\!/\min}\rightarrow\theta_2$. Series expansions of the above integrals are most conveniently written with the help of a new constant \^2:=D\^2+. We use the parametrization $D=\Lambda \cos\alpha$, and $\eta=\Lambda^2\sin^2\!\alpha$, with $0\leq\alpha\leq \pi$. In this notation the minimal coordinate distance from the center is r\_=-M -+-+(\^[-2]{}). We perform the integration by the methods of [@kerr-bray; @chandra] [(Appendix B)]{}. In the scattering scenario $r_1, r_2\rightarrow\infty$, while the constants of motion are kept finite. Expansion in terms of $\Lambda$ gives =(\_1+\_2-(3(\_1+\_2)+\_1\_1 +\_2\_2))+(\^[-4]{})). \[rte\] Where $\cos\psi=\cos\theta/\mu_+$ and $\mu_+$ is defined in the appendix. To separate conceptual issues from the computational details we make two simplifying assumptions. First, the initial and final points are taken to lie in the asymptotically flat regions, $r_i\rightarrow\infty$, while $\eta$ and $D$ are kept finite. Second, in these regions the polar angle $\theta$ (nearly) reaches its asymptotic values $\theta_\mrin$ and $\theta_\mout$, respectively. The asymptotic form of momentum can be deduced from the equations of motion [@chandra]. In the chronometric tetrad basis the outgoing momentum tends to k\^(1,1,-,), where $\beta_\mout:=+\sqrt{\Theta(\theta_\mout)}$ and $s=\pm 1$. This expression is true in any scenario where an outgoing photon reaches the asymptotically flat region. The asymptotic expression for the incoming momentum is similar, k\^(1,-1,,), where $\beta_\mrin:=+\sqrt{\Theta(\theta_\mrin)}$, and $s=\pm 1$ correspond to $\theta_\mrin\rightarrow\theta_{\max}/\theta_{\min}\rightarrow\theta_\mout$, respectively. Equating the radial and angular expressions for $\eR$ we obtain \_=-\_+(\^[-2]{}), with the higher order terms and special cases described in Appendix \[kerrtraj\]. The plus sign corresponds to the trajectory in which $\theta$ first decreases with $r$ and reaches $\theta_{\min}$ before increasing to $\theta_\mout$. Null geodesics in 1+3 formalism ------------------------------- In stationary space-times a tetrad of a static observer is naturally related to Landau-Lifshitz 1+3 formalism [@ll2]. Static observers follow the congruence of time-like Killing vectors that defines a projection from the space-time manifold $\cal{M}$ onto a three-dimensional space $\Sigma_3$, $\pi:\mathcal{M}\rightarrow \Sigma_3$. In practice it is performed by dropping the time-like coordinate of an event, and vectors are projected by a push-forward map $\pi_*\ak=\bk$ in the same way. Contravariant vector components satisfy ()\^m(\_\*)\^m=()\^m=dx\^m/d, m=1,2,3. where $\lambda$ is the affine parameter. The metric $\sg$ on $\mathcal{M}$ can be written in terms of a three-dimensional scalar $h$, a vector $\mathbf{g}$, and a metric $\gamma$ on $\Sigma_3$ as ds\^2=-h(dx\^0-\_mdx\^m)\^2+dl\^2, where the the three-dimensional distance is given by $dl^2=\gamma_{mn}dx^mdx^n$. The metric components are \_[mn]{}=(\_[mn]{}-), and h=-\_[00]{}, \_[m]{}=-\_[0m]{}/\_[00]{}. The inner product of three-vectors will always refer to this metric, $\bk\!\cdot\!\bof=\gamma_{mn}k^mf^n$. Vector products and differential operators are defined as appropriate dual vectors [@ll2]. Finally, the spatial projection of a null geodesic has a length $l$ that is related to the affine parameter $\lambda$ as ()\^2==\^2=: k\^2. For a static observer the three spatial basis vectors of the local orthonormal tetrad are projected into an orthonormal triad, $\pi_*\are_{(m)}= \hbe_{(m)}$, $\hbe_{(m)}\!\cdot\!\hbe_{(n)}=\delta_{mn}$. We adapt a gauge in which polarization is orthogonal to the observer’s four-velocity, $\au\!\cdot\!\af=0$. In a three-dimensional form this condition reads as $\bk\!\cdot\!\bof=0$. In stationary space-times the evolution equations Eq. , can be reduced into a convenient three-dimensional form [@fayos]. Using the relations between four- and three-dimensional covariant derivatives, $\anab_\mu$ and $D_m$, respectively, the propagation equations and are brought to a three-dimensional form [@fayos; @gravimag], =+, =. \[3deq\] The angular velocity of rotation ${\mbox{\boldmath$\Omega$}}$ is =2-()-\_g, where $\bg$ and $h$ play the roles of a vector potential of a gravimagnetic field $\mathbf{B}_g$, =-k\_0 -k\_0\_g, and a scalar potential of a gravielectric field, \_g=-, respectively [@inertia]. Local reference frames and polarization rotation ================================================ To discuss how photon’s polarization rotates it is necessary to define the standard polarization directions along its trajectory. Only after physically defining two standard polarizations it is meaningful to talk about polarization rotation, and only in a particular gauge the Schwarzschild space-time induces a zero phase on an open trajectory. Wigner phase — polarization convention -------------------------------------- Transversality of electromagnetic waves makes the choice of two standard polarization directions momentum-dependent. On a curved background it also depends on the location. Wigner’s construction of the massless representation of the Poincare group [@litgroup; @wkt] is the basis for classification of states in quantum field theory. We use it at every space-time point to produce standard polarization vectors. The construction — part of the induced representation of the Poincaré group [@wkt] — consists of a choice of a standard reference momentum $\ak_S$ and two polarizations, a standard Lorentz transformation $L(\ak)$ that takes $\ak_S$ to an arbitrary momentum $\ak$, and a decomposition of an arbitrary Lorentz transformation $\Lambda$ in terms of the standard transformations and Wigner’s little group element $W$. The standard reference 3-momentum is directed along the $z$-axis of an arbitrarily chosen reference frame, with its $x$ and $y$ axes defining the two linear polarization vectors $\hbb_{1,2}^S$, respectively. Hence $\ak_S=(1,0,0,1)$, and imposing the polarization gauge relates 3- and 4- polarization vectors, $\arb_{1,2}=(0,\hbb_{1,2})$. The standard Lorentz transformation can be taken as [^1] L()=R()B\_z(k), where $B_z(k)$ is a pure boost along the $z$-axis that takes $\ak_S$ to $(k,0,0,k)$, and the standard rotation $R(\hbk)$ brings the $z$-axis to the desired direction $\hbk(\theta,\phi)$ first by rotating by $\theta$ around the $y$-axis and then by $\phi$ around the $z$-axis, R()=R\_z()R\_y(). The standard polarization vectors for an arbitrary momentum are defined as \_i()= R()\_i\^S, and a general real 4-vector of polarization can be written as =\_1()+\_2(). \[poldec\] Helicity is invariant under Lorentz transformations. The corresponding polarization vectors are \_()=(\_1()i \_2()). Under a Lorentz transformation $\Lambda$ a state of a definite helicity acquires a phase $e^{\pm i\xi(\Lambda,\bk)}$, which can be read-off from Wigner’s little group element W(,)=L\^[-1]{}()L()=R\_z()T(,). The little group element $W$ leaves the standard momentum $\ak_S$ invariant. Here $T(\alpha, \beta)$ form a subgroup which is isomorphic to the translations of a Euclidean plane and $R_z(\xi)$ is a rotation around the origin of that plane, which in this case is also a rotation around the $z$-axis. If the transformation in question is a pure rotation $\eR$, then the little group element is then a rotation $R_z(\xi)$, and the polarization three-vector is rotated by $\eR$ itself [@lpt1]. The phase $\chi$ has a simple geometric interpretation: it is the angle between, say, $\hbb_1(\eR\bk)$ and $\eR\hbb_1(\bk)$. It is zero if the standard polarizations of the new momentum are the same as the rotated standard polarizations of the old momentum. This is what happens if a rotation $R_2(\omega)$ is performed around the current $\hbb_2(\bk)$: the resulting phase $\Delta\chi\big(R_2(\omega)\big)$ is zero [@we-mach]. Indeed, if $\hbk=\hbk(\theta,\phi)$, then by setting $\hbk'=R_2(\omega)\hbk$ and using the decomposition R\_2()=R()R\_y()R\^[-1]{}(), we find that $\hbk'=\hbk'(\theta+\omega,\phi)$, and the little group element is W=R\^[-1]{}(’)R\_[2]{}() R()=R\_y\^[-1]{}(+)R\_y()R\_y()=, indicating the absence of rotation with respect to standard polarization basis. In a curved space-time one has to provide the standard $(xyz)$ directions at every point. The next section deals with this problem. Newton gauge {#secnewton} ------------ The three-dimensional propagation equations in stationary space-times result in a joint rotation of polarization and unit tangent vectors =, =. \[3dprot\] In the Schwarzschild space-time test particles move in the plane passing through the origin [@mtw; @chandra; @ll2], ${\mbox{\boldmath$\Omega$}}=-\bE_g\times\bk$ and the covariant time-time component of the metric is related to the Newtonian gravitational potential $\varphi$ as $h=\sg_{00}=1+2\varphi(r)=1-2M/r$. *Requiring* the resulting phase to be zero, as in [@skrot; @gravimag; @kopeikin], constrains a choice of local reference frames. We use the zero phase condition to make a physically motivated choice of standard polarizations that does not require references to a parallel transport or communication between the observers. Orient the local $z$-axis along the direction of the free-fall acceleration, $\hbz\|\bnab\varphi$ and the standard polarizations as \_2:=/||, \_1:=\_2, \[polarbasis\] where $\bf{w}$ is a free-fall acceleration in the frame of a static observer. Then ${\mbox{\boldmath$\Omega$}}=\Omega\hbb_2$, and the propagation induces no phase. In a general static space-time we take the $z$-axis along the local free fall direction as seen by a static observer. In the Kerr space-time its components are [@fronov] w\_==--+(r\^[-4]{}), w\_== a\^2 M +(r\^[-5]{}), w\_=0. \[wcompo\] This convention, that we will call the Newton gauge, is consistent: if we set $\hbz=-\hat{\mathbf{r}}$ in the flat space-time, then no phase is accrued as a result of the propagation. In addition to being defined by local operations, the Newton gauge has two further advantages. First, it does not rely on a weak field approximation to define the reference direction. Second, if the trajectory is closed or self-intersecting, the reference direction $\hbz$ is the same at the points of the intersection. Kerr space-time examples ------------------------ We impose the temporal gauge in local frames, hence f\^0 f\^t-=-+(r\^[-3]{}), \[mygauge\] where we used $f^{\hat{0}}=f^\mu e^{\hat{0}}_{(\mu)}$. This transversality reduces to a familiar three-dimensional expression $\ak\cdot\af\equiv\mathbf{k}\cdot\mathbf{f}= f^{\hat{1}}k^{\hat{1}}+f^{\hat{2}}k^{\hat{2}}+f^{\hat{3}}k^{\hat{3}}=0$. This also implies f\_t= e\_[()t]{} f\^=e\_[(0)t]{} f\^=0. In the Kerr space-time the Walker-Penrose quantity [@wp] $K_2+iK_1$, K\_1=r B- a A , K\_2= r A + a B , \[wp12\] where $$\begin{aligned} A =&(k^t f^r-k^r f^t)+a(k^r f^\phi-k^\phi f^r)\sin^2\!\theta, \\ B=&\big( (r^2+a^2)(k^\phi k^\theta-k^\theta f^\phi)-a(k^t f^\theta-k^\theta f^t)\big)\sin\theta,\end{aligned}$$ is conserved along null geodesics [@chandra]. Transversality and the gauge make the Walker-Penrose constants functions of only two polarization components, for example $f^{\hat{2}}$ and $f^{\hat{3}}$. For a generic outgoing null geodesic in the asymptotic regime the constants become K\_1=\_f\^\_-s\_f\^\_, K\_2=s\_f\^\_+\_f\^\_, where \_:=D\_-a\_, and the analogous expression gives the constant in terms of the initial data. Hence, f\^\_=0, f\^\_=-(-s\_K\_1+\_K\_2), f\^\_=-(-s\_K\_2-\_K\_1). To determine the polarization rotation $\chi$ we take the initial polarization to be, say, $\mathbf{f}_{\mrin}= \hbb_1^{\mrin}$, which often considerably simplifies the expression for $K_1$ and $K_2$. Since $K_1$ and $K_2$ are linear functions of polarization, expressing the above result in the basis $(\hbb_1^{\mout},\hbb_2^{\mout})$ the desired rotation angle $\chi$ is observed from ( [c]{} 1\ 0 ) T( [c]{} 1\ 0 )=( [c]{}\ ), where $T$ is an orthogonal matrix that is described below. We assume fixed $D$ and $\eta$. Using Eq.  we find that at the limit $r_\mrin\ra\infty$ the initial standard linear polarization directions are \_1\^=(0,s\_\_, D), \_2\^=(0,D,-s\_\_). Similarly, the final standard polarizations are \_1\^=(0,s\_\_, -D), \_2\^=(0,D,s\_\_). The rotation matrices $N_{\mrin}$ and $N_\mout$ from the polarization bases to $(\htheta,\hphi)$ basis are given in the Appendix C. Following [@kerr-farpol] using the asymptotic relationship between Walker-Penrose constants and the components of polarization one can introduce a transformation matrix $R$, ( [c]{} f\^\_\ f\^\_ )=R( [c]{} f\^\_\ f\^\_ ). The transformation matrix has the form R=( [cc]{} 1 & -x\ -x & -1 ), where the parameter $x$ is given by x=s, with $\gamma_\mrin:=D\csc\theta_\mrin-a\sin\theta_\mrin$. Finally, T=N\_R N\_\^[-1]{}. \[defT\] Expansion in the inverse powers of $\Lambda$ results in T\_[21]{}==-\_+(\^[-3]{}). \[mainf\] This is our main new result. Few special cases are of interest. If the initial propagation is parallel to the $z$-axis and the impact parameter equals $b$, then \^2=b\^2-a\^2, D=0, and the polarization is rotated by =++(\^[-4]{}), and the antiparallel initial direction gives the opposite sign. Motion in the equatorial plane corresponds to $\eta=0$ ($\sin\alpha=0$) and is qualitatively similar to the motion in Schwarzschild space-time. If the trajectory starts there but eventually moves outside, then the polarization is rotated by =s+(\^[-4]{}). In the case of initial propagation along the $z$-axis agrees with both [@skrot] and [@fayos], if we take into account the respective definitions of reference frames. On the other hand, in a generic setting our Eq.  differs by a power of $\Lambda$ from [@kerr-farpol] or [@gravimag] (both works predict $\chi\sim\Lambda^{-3}$ but disagree on the pre-factors). The origin of this disagreement is in the following. In addition to using a more transparent series expansion, we use a different polarization basis from the one in [@kerr-farpol]. In particular, we do not require the knowledge of $\phi_\mout-\phi_\mrin$. Moreover, [@gravimag] considered only a Machian effect (Sec. \[phasesec\]), which is dominated by the reference-frame term [@we-mach]. Geometric phase {#phasesec} =============== A more geometric perspective on polarization rotation is possible both in a static space-time, where we use the 1+3 formalism, as well as in arbitrary space-times. While the phase $\Delta\chi$ is gauge-dependent for an open trajectory, we will show that it is gauge-invariant on a closed path. Consider the differential equations for polarization rotation [@we-mach]. By setting $\af=\arb_1$ at the starting point and using the parallel transport equations Eqs.  and we arrive to the equation =\_(\_2)= \_\_2. In a static space-time projection of the polarization on $\Sigma_3$ results in =, hence the desired equation is =( \_2+)=+. \[defrot\] The first term corresponds to the original Machian effect that was postulated in [@god], but the observable quantity involves both the Machian and the reference-frame terms [@we-mach]. Now we discuss the geometric meaning of these equations. First consider a basis of 1-forms $(\sigma^1, \sigma^2, \sigma^3)$ that is dual to the orthonormal polarization basis $(\hbb_1, \hbb_2,\hbk)$ at every point of the trajectory. A matrix of connection 1-forms $\omega$ is written with the help of Ricci rotation coefficients $\omega^{\hat{\imath}}_{\hat{\jmath}\,\hat{l}}$ as $\omega^{\hat{\imath}}_{\ \hat{l}}=\omega^{\hat{\imath}}_{\hat{\jmath}\,\hat{l}}\sigma^j$ [@mtw; @fra], and a linear polarization is written as =f\^\_1+f\^\_2=\_1+\_2. Its covariant derivative equals to =k\^\_f\^c\_a+\^\_f\^c+(f\^)\_c, a,c=1,2. Taking into account the antisymmetry of the connections $\omega^{\hat{\imath}}_{\ \hat{l}}=\omega_{\hat{\imath} \hat{l}}=-\omega_{\hat{l} \hat{\imath}}$, we find =(-\_1 f\^+\_2 f\^)(-\^\_k)+ k(\^\_f\^+\^\_f\^). A comparison with Eq. leads to the identification $\Omega^1={\mbox{\boldmath$\Omega$}}\!\cdot\!\hbb_1=k\omega^{\hat{3}}_{\hat{3}\hat{2}}$, $\Omega^2=-k\omega^{\hat{3}}_{\hat{3}\hat{1}}$ and to an alternative equation for the polarization rotation, =+\^\_k. Given a trajectory with a tangent vector $\bk$ one can define a SO(2) line bundle with the connection $\bar{\omega}=\omega^{\hat{1}}_{\hat{3}\hat{2}}kd\lambda$, similarly to the usual treatment of geometric phase [@fra; @geom]. Freedom of choosing the polarization frame $(\hbb_1,\hbb_2)$ at every point of the trajectory is represented by a SO(2) rotation $R_\hbk\big(\psi(\lambda)\big)$. Under its action the connection transforms as $\omega\rightarrow R\omega R^{-1}+R^{-1}dR$ [@fra], so +\^\_k+. In a static space-time we can consider a closed trajectory in space. Then the resulting phase is gauge-invariant, since the last term above is a total differential and drops out upon the integration on a closed contour, =d+|. \[3dom\] We can formalize it with the help of a curvature 2-form is introduced as $\bar{\theta}:=d\bar{\omega}+\bar{\omega}\wedge \bar{\omega}.$ For the SO(2) bundle it reduces to $\bar{\theta}=d\bar{\omega}$, so by using Stokes’ theorem the reference-frame term can be rewritten as a surface integral of the bundle curvature as |=|. A more practical expression follows from our previous discussion: $ \Delta\chi=\arcsin\hbf_\mout\!\cdot\!\hbb_2$. Conservation of $K_1$ and $K_2$ in the Kerr space-time ensures that if a trajectory is closed as a result of the initial conditions, then $\hbf_\mout=\hbf_\mrin$ and $\Delta\chi=0$. The Newton gauge is designed to give a zero phase along any trajectory in the Schwarzschild space-time. As a result of the gauge invariance of Eq. , no gravitationally-induced phase is accrued along a closed trajectory in the Schwarzschild space-time, regardless of the gauge convention. In a general space-time we introduce an orthonormal tetrad such that $\ak=k\are_0+k\are_3$ at every point of the trajectory. We again impose a temporal gauge and set $\are_{1,2}=\arb_{1,2}$, where the local polarization basis is chosen according to some procedure. Then from Eq.  it follows that $$\begin{aligned} \anab_\ak \af &= k^\hmu \mathsf{e}_{(\nu)} \omega_{\hmu \hrho}^\hnu f^\hrho + \frac{d f^\hnu}{d\lambda}\mathsf{e}_{(\nu)} = \left( -f^{\hat{2}} \mathsf{e}_{(1)} + f^{\hat1} \mathsf{e}_{(2)} \right) \left( \frac{d\chi}{d\lambda} -(\omega^{\hat1}_{\hat0 \hat2}+\omega^{\hat{1}}_{\hat3\hat2}) k \right) \nonumber\\ &+\big(k(\omega^{\hat 0}_{\hat0\hat{c}}+\omega^{\hat0}_{\hat3\hat{c}})f^{\hat{c}}\big)\are_{(0)}+\big(k(\omega^{\hat 3}_{\hat0\hat{c}}+\omega^{\hat3}_{\hat3\hat{c}})f^{\hat{c}}\big)\are_{(3)}, \qquad c=1,2. $$ where we used again the antisymmetry of the connection. From the parallel transport condition $\anab_\ak \af = 0$, we can see that (\^[0]{}\_[0]{}+\^[0]{}\_[3]{})f\^=(\^[3]{}\_[0]{}+\^[3]{}\_[3]{})f\^=0. and =(\^[1]{}\_[0 2]{}+\^\_[32]{}) k [d]{}:=| We cannot have a closed trajectory in chronologically-protected space-time, so to obtain a gauge-invariant result we consider two future-directed trajectories that begin and end in the same space-time points. This layout is similar to the two arms of a Mach-Zender interferometer [@bw]. We also align the initial and final propagation directions of the beams and use the same rules to define the standard polarizations. Similarly to the previous case the curvature is just $\bar{\Theta}=d\bar{\Omega}$, and the Stokes theorem gives the phase as =\_[\_1]{}|-\_[\_2]{}|=\_|=|. Conclusions and outlook ======================= Rotation one ascribes to polarization as light propagates on a curved background depends on the gauge conventions that are used along the way. However, the closed-loop phase is gauge-independent (assuming we use the same reference frame for transmission and detection). Similarly to other instances of a geometric phase, the phase $\Delta\chi$ is given by the integral of the (bundle) curvature over the surface that is bounded by the trajectory. The Newton gauge provides a convenient local definition of the standard polarizations. It is motivated by its relative simplicity and path-independence of the reference frames. It is also the gauge in which the statement of a zero accrued phase along an arbitrary path in the Schwarzschild space-time is correct. Our results characterize the behavior of polarization qubits on curved backgrounds. Specific regimes and realistic scenarios, both on the Kerr background an beyond, are to be investigated farther. Of special interest are those scenarios that will be experimentally feasible in the near future, such as sending photons between satellites [@milsat]. Once the expression for $\Delta\chi$ is obtained, it can be represent it as the action of a quantum gate [@qinfo], similarly to the the special-relativistic scenarios [@pt04; @bat05]. This will allow us to use the full toolbox of quantum optics to design experiments. Understanding of the polarization phase in GR opens new possibilities for optics-based precision measurements [@we-mach], both classical and quantum. We thank P. Alsing, B.-L. Hu, A. Kempf, and N. Menicucci for discussions and comments. Research at Perimeter Institute is supported by the Government of Canada through Industry Canada and by the Province of Ontario through the Ministry of Research & Innovation. Some aspects of the Kerr space-time {#kerrdetail} =================================== Chronometric tetrad ------------------- We fix an orthonormal tetrad at every space-time point (outside the static limit) by demanding that $\are_0$ is the four-velocity of a static observer, and the vectors $\are_1$ and $\are_2$ are proportional to the tangent vectors $\pad_r$ and $\pad_\theta$. Their covariant components $e_{(\alpha)\mu}$ are \_[(0)]{}=(-,0,0,-)and \_[(1)]{}=(0,/,0,0), \_[(2)]{}=(0,0,,0), \_[(3)]{}=(0,0,0,) Setting $E=-k_0=1$ leads to k\^t=, and $k^r$ is expressed from the null condition $\ak^2=0$, (k\^r)\^2=. The spatial components of the momentum in the coordinate and tetrad bases are related to $k^{\hat{m}}=e^{(m)}_{~~\mu} k^\mu$ as $$\begin{aligned} k^{\hat{1}}&=k^r\rho/\sqrt{\Delta},\\ k^{\hat{2}}&=k^\theta\rho,\\ k^{\hat{3}}&=k^\phi\rho\sin\theta\sqrt\frac{\Delta}{\Delta-a^2\sin^2\!\theta},\end{aligned}$$ while $k^2=\bk^2=k^{\hat{m}}k_{\hat{m}}$ satisfies k\^2=1+=1++(r\^[-2]{}). Constants of motion ------------------- The re-scaled ($E=1$) $z$-component of the angular momentum is D=(k\^(a\^2+r\^2)(\^2-2Mr)-2 Mar(1-k\^a \^2)). The re-scaled Carter’s constant :=K-(D-a)\^2, where the constant $K$ is most conveniently expressed as K:=(a -D/)\^2+(\^2k\^)\^2. The asymptotic expressions when $r\rightarrow\infty$ and the momentum components are fixed are $$\begin{aligned} & D= k_{\hat{3}}r\sin\theta+\co(r^{-1}),\\ &\eta=\big(k_{\hat{3}}^2\cos^2\!\theta+k_{\hat{2}}^2\big)r^2 + a \cos^2\!\theta\big(a(k_{\hat{2}}^2+k_{\hat{3}}^2-1)-4Mk_{\hat{3}}\sin\theta\big)+\co(r^{-1}),\\ &\Lambda=r\sqrt{k_{\hat{2}}^2+k_{\hat{3}}^2}+\co(r^{-1}),\\ &\cos\alpha =\frac{k_{\hat{3}}\sin\theta}{\sqrt{k_{\hat{2}}^2+k_{\hat{3}}^2}}+\co(r^{-2}).\end{aligned}$$ Scattering with the impact parameter $b$ that measures the coordinate distance $r\cos\theta$ from the $z$-axis is most conveniently described with the help of a fiducial Cartesian system. Its axes are “parallel" to the fictitious Cartesian axes of the Byer-Lindquist coordinates, and in the asymptotic region it is just parallel to the global Cartesian grid. We introduce the momentum components $\hat{p}_i$ that are related to the spherical components $\hat{k}_i$ the the usual relations. In this case for the initial momentum parallel to the $z$-axis we have D=(r\_1\^[-3]{}), =+(r\_1\^[-2]{}). Some care is needed in treating $R(r)$ as a function of the initial conditions when $r_1\rightarrow\infty$. Introducing the constants $c_0$, $c_1$, $c_2$, we write it as \[Rinr1\] R(r)=:r\^4-c\_2\^2r\_1\^2 r\^2+c\_1 r\_1\^2 r+c\_0 r\_1\^2, where $$\begin{aligned} c_2^2 &:=(\Lambda^2-a^2)/r_1^2=(k_1^{\hat{2}})^2+(k_1^{\hat{3}})^2+\co(r_1^{-2}) \\ c_1 &:=2M(\Lambda^2+a^2-2\Lambda a\cos\alpha)/r_1=2 M (c_2^2-\frac{2ak_1^{\hat{2}}\sin\theta_1}{r_1})+\co(r_1^{-2}), \\ c_0 &:= -a^2\Lambda^2\sin^2\!\alpha=-a^2 \big((k_1^{\hat{2}})^2\cos^2\!\theta_1+(k_1^{\hat{3}})^2\big)+\co(r_1^{-2})\end{aligned}$$ Trajectories {#kerrtraj} ============ The $r$ integral {#secrinteg} ----------------- The integral over $r$ in Eq.  (\[ert\]) is split into three parts, :=\^r=\_-\_1-\_2, where \_:=2\_[r\_]{}\^, \_i:=\_[r\_i]{}\^. \[rpart\] We present a corrected expression for $\eR_i$ and evaluate the term $\eR_\infty$ to the fourth order in $\Lambda$. Starting from $\eR_\infty$ in using the method of [@kerr-bray] we factor out $(r-r_{\min})$ and substite $r=r_{\min}/x$, writing $R(r)$ as R=(r\_\^2+Ax\^2++)=:R(x), where the constants $A$, $B$, $C$ are determined by the polynomial division. Noting that $\tilde{ R}(1)=0$ we have r\_\^2=-A-- which gives R(x)=-A--+Ax\^2++ Rewriting the equation and inserting the values for $A,B$ and $C$ gives us R(r)=(\^2-a\^2)(1-x\^2)(1+(1+x\^2) -) Defining f(x):=, g(x):=-2, we finally get \_=2 \^\_[r\_]{}=\_0\^1 dx((1-x\^2)(1+f(x)+g(x)))\^[-]{}. Expansion the integrand up to the third order in $\Lambda$ gives Eq. (15). $$\begin{aligned} \eR_\infty&=\frac{1}{{\sqrt{\Lambda^2-a^2}}}\left(\pi+\frac{4M}{\Lambda}+ \frac{\pi}{\Lambda^2}\left(\frac{15}{4}M^2-\frac{3}{4}a^2\sin^2\!\alpha\right)-\frac{8 aM}{\Lambda^2}\cos\alpha \right.\nonumber \\ &+\left.\frac{M}{\Lambda^3}\left[\frac{128}{3}M^2+a^2(10 \cos^2\!\alpha-6\sin^2\!\alpha)+15\pi \label{rr} Ma\cos\alpha\right]\right)+\co(\Lambda^{-5}).\end{aligned}$$ \[rinfo3\] The second and third integrals can be performed as an approximation in powers of $r_1$, using Eq. (\[Rinr1\]). Since R(r)=(r\^4-c\_2\^2r\_1\^2r\^2)(1+), the leading terms in the expansion are \_[r\_i]{}\^=-(+)+(r\_1\^[-5]{}), with $r_i$ being either $r_1$ or $r_2$, where in the latter case we assume that $r_2\gtrsim r_1$. The $\theta$ integral --------------------- We calculate the integral $\int1/\sqrt{\Theta}d\theta$ using the method of [@chandra]. Noting that =+(a-D)\^2-(a-D)\^2, we perform a change of variables $\mu:=\cos\theta$ and obtain I\_==, where fro $\eta>0$ \_=a\^2(\_-\^2+\^2)(\_+\^2-\^2), 0\^2\_+\^2, and \_\^2:=((\^2-a\^2)). The calculation is performed using the approximations to the auxiliary integral \[ellipticf\] \_[\_\*]{}\^[\_+]{}=F(,k\^2), where $F(\psi,k^2)$ is the elliptic integral of the first kind, and k\^2:=\_+\^2/(\_-\^2+\_+\^2), :=\_\*/\_+, with $0\leq\psi\leq\pi$ and $\mu_+=+\sqrt{\mu_+^2}$. The asymptotic expansion in the powers of $\Lambda$ gives $$\begin{aligned} \mu_+^2 &=\sin^2\alpha+\frac{a^2}{4\Lambda^2}\sin^22\alpha+\co(\Lambda^{-4}) \\ \mu_-^2 &=\frac{\Lambda^2}{a^2}-\cos^2\alpha+\frac{a^2}{2\Lambda^2}\sin^22\alpha+\co(\Lambda^{-4})\end{aligned}$$ so $$\begin{aligned} & k^2 = \frac{ a^2 \sin^2\alpha}{\Lambda^2}+ \co(\Lambda^{-4}),\\ & \frac{1}{a\sqrt{\mu_+^2+\mu_-^2}}=\frac{1}{\sqrt{\Lambda^2-a^2}}\left( 1-\frac{a^2}{\Lambda^2}\sin^2\alpha+\co(\Lambda^{-4})\right).\end{aligned}$$ Trajectories of the type $\theta_1\rightarrow\theta_{\min}/\theta_{\max}\rightarrow\theta_2$ correspond to $\cos\theta_{\min}=\mu_+$, and $\cos\theta_{\max}=-\mu_+<0$, respectively. In both cases the integration leads to =, and the expansion in powers of $\Lambda$ leads to Eq. (\[rte\]). Using that at the zeroth order (the flat space-time) $\psi_1+\psi_2=\pi$, we expand $\psi_2=\pi-\psi_1+\delta_\psi$ as \_2=-\_1 +\_k. Equating the two expression for $\eR_\infty$ in the scattering scenario leads to $$\begin{aligned} \xi_1&=4M, \\ \xi_2&=15 M^2 \pi/4-8 Ma\cos\alpha, \\ \xi_3&= M(-6 a^2-128 M^2+45 Ma\pi\cos\alpha-24 a^2 \cos2\alpha +a^2(3 +\cos2\psi_1) \sin^2\alpha).\end{aligned}$$ Similarly, setting $\theta_2=\pi-\theta_1+\delta_\theta$, \_=\_k, and expanding the both sides of (-\_)=\_+(-\_), we find \_1=4M, In the special case of scattering with the initial momentum parallel (anti-parallel) to the $z$-axis the angular momentum is zero, so $\mu_+=1$ and $\theta_\mrin=\psi_\mrin=0,\pi$, and it is easy to see that $\theta_\mout=\delta_\psi,\pi-\delta_\psi$, respectively. The $\phi$ motion ----------------- The equation for $\phi$ is given by [@chandra] $$\begin{aligned} \phi= & \pm \left( D\int^r\!\frac{dr}{\sqrt{R}}+2Ma\int^r\!\frac{r dr}{\Delta \sqrt{R}}- a^2 D\int^r\!\frac{dr}{\Delta \sqrt{R}} \right) \nonumber \\ & \pm D\left(\int^\theta\!\frac{d\theta}{\sin^2\!\theta\sqrt{\Theta}}-\int^\theta\!\frac{d\theta}{\sqrt{\Theta}}\right).\end{aligned}$$ If $\theta\neq\mathrm{const}$ the first and the last terms cancel thanks to Eq. . For brevity we write the remaining terms as \_2-\_1=: \^rR\^(r) dr \^T()d, We again decompose the radial integral as $\eR^\phi=\eR^\phi_\infty-\eR^\phi_1-\eR^\phi_2$, where \^\_=2\_[r\_]{}\^R\^(r) dr, \^\_i=\^\_[r\_i]{} R\^(r) dr. Following the same procedure as in Sec. B.1 we find \^\_=(++(\^[-3]{})). Expanding $R^\phi(r)$ for $r\gg r_1$ we obtain $$\begin{aligned} R^\phi(r)&=\frac{2aM}{\sqrt{(r^2-c_2^2r_1^2)}\left (\sqrt{1+ \frac{c_1r_1^2}{r(r^2-c_2^2r_1^2)}+\frac{c_0r_1^2}{r^2(r^2-c_2^2r_1^2)}}\right)(r^2+a^2-2Mr)}+\\ \nonumber &\frac{a^2D}{r\sqrt{(r^2-c_2^2r_1^2)}\left (\sqrt{1+ \frac{c_1r_1^2}{r(r^2-c_2^2r_1^2)}+\frac{c_0r_1^2}{r^2(r^2-c_2^2r_1^2)}}\right)(r^2+a^2-2Mr)} \nonumber \\ &=\frac{a}{r^2\sqrt{r^2-c_2^2r_1^2}}\left[2M\left(1+\frac{2M}{r}\right)+\frac{aD}{r}\right]+\co(r^{-5}),\end{aligned}$$ where the coefficients $c_i$ are defined in Appendix A2. Hence \_[r\_i]{}\^R\^(r) dr= -(c\_2r\_1- r\_i\^2())+(r\_1\^[-5]{}). Integral $\int Td\theta$ leads to the elliptic integral of the third kind. A standard change of variables $\cos\theta=\mu=\mu_+\cos t$ and the identity \^2=(1-\_+\^2)(1+p\^2t) where p:==\^2(1+)+(\^[-4]{}) lead to \_[\_]{}\^[\_\*]{} T()d= \_0\^[\_\*]{}, with $\cos\psi_*=\cos\theta_*/\mu_+$ as in Appendix B2. Hence \_[\_]{}\^[\_\*]{} T()d= (-p,\_\*,k\^2), where $\Pi(q,\psi,k^2)$ is the elliptic integral of the third kind. Expanding the pre-factor in the powers of $\Lambda$ gives =+(\^[-4]{}), while (-p,,0)=(-())=:\_0(p,) and (-p,,k\^2)=\_0(p,)+(-\_0(p,))+(k\^4). Taking into account that =||(1-\^2)+(\^[-4]{}), we get for the scattering scenario \_-\_=+\_i, where the first two terms of the series are \_1=, and $$\begin{aligned} \Phi_2 &= \frac{ \xi_2 (3 + \cos\!2\alpha) \sec^3\!\alpha + 4 \sec\alpha (\xi_1^2 \sin\!2\psi_1-\xi_2 \cos\!2\psi_1 ) \tan^2\alpha}{(1 + \sec^2\!\alpha - \cos 2\psi \tan^2\!\alpha)^2} +{\mbox{$1\over4$}} a^2 \cos \alpha+ 2Ma,\end{aligned}$$ where $\xi_i$ are given in Appendix B2. Newton gauge relationships {#gaugesec} ========================== In this Appendix we present a relationship between the components of polarization in the chronometric tetrad and the Newton gauge basis. This relationship allows to calculate the Walker-Penrose conserved quantity from the operationally meaningful polarization information and in the asymptotic regime gives the matrices $N$ of Eq. . The transversally of polarization $\hbk\cdot\hbf=0$ ensures the linearity of the relationship between $(f^{\hat2},f^{\hat3})$ and $(f^x,f^y)$, where =f\^x\_x+f\^y\_y. Since ${\bf b}_y=\hat{\bf w}\times{\bf k}/|\hat{ \bf w}\times{\bf k}|$ and ${\bf b}_x=-{\bf b}_y\times\hat{\bf k}$, by using Eq.  we find $$\begin{aligned} &{\bf b}_y=\frac{1}{\cal{N}}\left(w_\htw{k}_\hth,-w_\htw{k}_\hth,-w_\htw{k}_\ho+w_\ho{k}_\htw\right),\\ &{\bf b}_x=\frac{1}{{\cal{N}}k}\left(w_\htw{k}_\ho {k}_\htw-w_\ho\big(k_\hth^2+{k}_\htw^2\big),\quad w_\htw \big({k}_\htw^2+{ k}_\ho^2\big)-w_\ho { k}_\ho { k}_\htw,\quad -{k}_\hth(w_\ho{k}_\ho+w_\htw{k}_\hth)\right),\\ &{\cal N}^2=(w_\htw{k}_\hth)^2+(w_\htw{k}_\hth)^2+(-w_\htw{k}_\ho+a_\ho{k}_\htw)^2.\end{aligned}$$ Projecting $\hbf $ on these directions gives $$\begin{aligned} &f^y=\frac{1}{{\cal{N}}{k}_\ho}\left(f^\htw(-w_\ho k_\ho k_\hth-w_\htw k_\hth k_\htw)+f^\hth\big(w_\ho k_\ho k_\htw+a_\ho( k_\hth^2+ k_\ho^2)\big)\right),\\ &f^x=\frac{1}{{\cal{N}}{k}^\ho}\left(f^\htw(w_\htw k_\ho-w_\ho k_\htw)+f^\hth(-w_\ho k_\hth)\right).\end{aligned}$$ We define the transformation matrix $N$ as $$\left(\begin{array}{c} f^\htw\\f^\hth \end{array}\right)=N \left(\begin{array}{c} f^y\\f^x \end{array}\right).$$ In the asymptotic regime where $f^\ho\rightarrow0$ it becomes orthogonal. Taking $k^{\hat{m}}=(1,s\beta/r,D/r \sin\theta)$ we obtain in the leading order $1/r$ $$\begin{aligned} N=\frac{1}{\sqrt{\beta^2+D^2/\sin^2\!\theta}}\left( \begin{array}{cc} D/\sin\theta&s\beta\\ -s\beta&D/\sin\theta \end{array} \right).\end{aligned}$$ [99]{} T. Padmanabhan, *Theoretical Astrophysics*, vol. 1, (Cambridge University Press, Cambridge, 2000). A. Fowler, Observatory **42**, 241 (1919); F. W. Dyson, A. S. Eddington, C. R. Davidson, Phil. Trans. R. Soc. A, **220**, 291 (1920); D. Kennefick, Phys. Today **62** (3), 37 (2009). C. M. Will, *Theory and Experiment in Gravitational Physics*, (Cambridge University Press, 1993); C. M. Will, Am. J. Phys. **78**, 1240 (2010). M. Born and E. Wolf, *Principles of Optics* (Cambridge University Press, Cambridge, 1999). C. W. Misner, K. S. Thorn, J. A. Wheeler, *Gravitation*, (Freeman, San Francisco, 1973). S. Chandrasekhar, *The Mathematical Theory of Black Holes* (Oxford University Press, New York, 1998). P. D. Naselsky, D. I. Novikov, and I. D. Novikov, *The Physics of the Cosmic Microwave Background*, (Cambridge University Press, 2007). The Fermi-LAT collaboration and members of 3C 279 multi-band campaign, Nature, **463**, 919 (2010); P. Laurent, J. Rodriguez, J. Wilms, M. Cadolle Bel, K. Pottschmidt, and V. Grinberg, Science, **332**, 438 (2011). M. Honda and Y. Honda, Astrophys. J. **569**, L39 (2002). P. F. Michelson, W. B. Atwood, and S. Ritz, Rep. Prog. Phys. **73**, 074901 (2010). R. Massey, T. Kitching, and J. Richard, Rep. Prog. Phys. **73**, 086901 (2010). D. Bru[ß]{} and G. Leuchs, *Lectures on Quantum Information* (Wiley-VCH, Weinheim, 2007); M. A. Nielsen and I. L. Chuang, *Quantum Computation and Quantum Information*, (Cambridge University, Cambridge, 2000). M. Zwierz, Carlos A. Pérez-Delgado, and P. Kok, Phys. Rev. Lett. **105**, 180402 (2010); L. Maccone and V. Giovannetti, Nat. Phys. **7**, 376 (2011). A. Peres and D. R. Terno, Rev. Mod. Phys. **76**, 93 (2004). P.M. Alsing, G.J. Stephenson Jr, e-print arXive:0902.1399 (2009). C. Feiler , Space Sci. Rev. **148**, 123 (2009). R. Ursin (39 authors, 27 institutions), Europhysics News **40**, (3) (2009); e-print arXiv:0806.0945v1. G. V. Skrotskii, Soviet Phys. Doklady **2**, 226 (1957). M. Nouri-Zonoz, Phys. Rev. D **60**, 024013 (1999); M. Sereno, Phys. Rev. D **69**, 087501 (2004). F. Fayos and J. Llosa, [Gen. Rel. Grav.]{} **14**, 865 (1982). H. Ishihara, M. Takahashi, and A. Tomimatsu, Phys. Rev. D **38**, 472 (1988). J. Plebanski, Phys. Rev. **118**, 1396 (1960). B. B. Godfrey, [Phys. Rev. D]{} **1**, 2721 (1970). R. F. Stark and P. A. Connors, Nature **266**, 429 (1977); P. A. Connors, T. Piran, and R. F. Stark, Astrophys. J. **235**, 224 (1980). I. Bray, Phys. Rev. D **34**, 367 (1986). S. M. Kopeikin and G. Schäfer, Phys. Rev. D **60**, 124002 (1999); S. M. Kopeikin and B. Mashhoon, Phys. Rev. D **65**, 064025 (2002). P. P. Kronberg, C. C. Dyer, E. M. Burbidge, and V. T. Junkkarinen, Astr. J. **367**, L1 (1991); C. C. Dyer and E. G. Shaver, Astrophys. J. **390**, L5 (1992). C. R. Burns, C. C. Dyer, P. P. Kronberg, and H.-J. Röser, Astrophys. J. **613**, 672 (2004). V. Faraoni, New Astron. **13**, 178 (2008) and the references therein. N. H. Lindner, A. Peres, and D. R. Terno, [J. Phys. A]{} **36**, L449 (2003). S. D. Bartlett, T. [Rudolph]{}, and R. W. [Spekkens]{}, Rev. Mod. Phys. **79**, 555 (2007). A. Brodutch and D. R. Terno, Phys. Rev. A **84**, 121501(R) (2011). L. D. Landau and E. M. Lifshitz, *The Classical Theory of Fields* (Butterworth-Heinemann, Amsterdam, 1980). V. P. Frolov and I. D. Novikov, *Black Hole Physics* (Kluwer, Dordrecht, The Netherlands,1998). B. Carter, Phys. Rev. **174**, 1559 (1968). E. Wigner, Ann. Math. **40** 149 (1939); S. Weinberg, *The Quantum Theory of Fields* vol. 1 (Cambridge University Press, 1996). W.-K. Tung, [*G*roup Theory in Physics]{} (World Scientific, Singapore, 1985) I. Ciufolini and J. A. Wheeler, *Gravitation and Inertia*, (Princeton University Press, Princeton, 1995). T. Frankel, *The Geometry of Physics*, (Cambridge University Press, Cambridge, 1999). M. V. Berry, Nature **326**, 277 (1987); R. O[ł]{}kiewicz, Rep. Math. Phys. **33**, 325 (1993). M. Walker and R. Penrose, Comm. Math. Phys. **18**, 265 (1970). S. D. Bartlett and D. R. Terno, Phys. Rev. A [**71**]{}, 012302 (2005). [^1]: When it does not lead to confusion we use the same letter to label a four-dimensional object and its three-dimensional part. In particular, $R$ stands both for a Lorentz transformation which is a pure rotation, and for the corresponding three-dimensional rotation matrix itself.

ArXiv

--- abstract: 'The first example of a turbulent system where the failure of the hypothesis of small-scale isotropy restoration is detectable both in the ‘flattening’ of the inertial-range scaling exponent hierarchy, and in the behavior of odd-order dimensionless ratios, e.g., skewness and hyperskewness, is presented. Specifically, within the kinematic approximation in magnetohydrodynamical turbulence, we show that for compressible flows, the isotropic contribution to the scaling of magnetic correlation functions and the first anisotropic ones may become practically indistinguishable. Moreover, skewness factor now diverges as the Péclet number goes to infinity, a further indication of small-scale anisotropy.' author: - | N.V. Antonov,$^{1}$ J. Honkonen,$^{2}$ A. Mazzino$^{3}$ and P. Muratore-Ginanneschi$^{4}$\ \ $^4$ The Niels Bohr Institute, Blegdamsvej 17, DK-2100 Copenhagen, Denmark title: Manifestation of anisotropy persistence in the hierarchies of MHD scaling exponents --- =0.01cm [2]{} A wide interest has recently been devoted to the possible occurrence of small-scale isotropy restoration for scalar (see e.g., Ref. [@SS99] and references therein), Navier-Stokes [@BoOr96; @Arad99] and magnetohydrodynamical (MHD) turbulence [@LM99; @ABP99; @ALM99]. The scenario can be summarized as follows. In the presence of anisotropic large-scale injection mechanisms, the inertial-range statistics is characterized by an infinite hierarchy of scaling exponents; however the leading contribution to scaling comes from the isotropic component. From this point of view, one might argue that large-scale anisotropy does not affect inertial-range scaling properties. Actually, focussing on a larger set of observables, small-scale anisotropies become manifest. It turns out that the behavior of odd-order dimensionless ratios (e.g., skewness and hyperskewness) is completely different from the case of small-scale isotropy restoration. Such indicators go to zero down to the inertial range much slower than predicted by dimensional considerations [@P96] or, more dramatically, they diverge at the smallest scales [@ALM99] (see also [@A98; @A99]). The main aim of this Letter is to present a model of MHD turbulence where, by varying the degree of compressibility of the velocity field, anisotropic persistence is now detectable both from the ‘flattening’ of the hierarchy of inertial-range scaling exponents (the isotropic component and the first anisotropic ones may become practically indistinguishable) and the divergence of skewness factor with the Péclet number. We give the basic ideas and results; longer and more exhaustive technical discussions will be presented in a forthcoming paper. In the presence of a mean component $\bbox{B}^o$ (actually varying on a very large scale $L$) and for the compressible velocity field $\bbox{v}$, the kinematic MHD equations describing the evolution of the fluctuating (divergence-free) part $\bbox{B}$ of the magnetic field are [@Zeldo83]: $$\begin{aligned} \label{fp} &&\partial_t B_\alpha +\bbox{v}\cdot\bbox{\partial}\,B_\alpha =-(B_{\alpha}+B^o_{\alpha})\bbox{\partial}\cdot\bbox{v}+ \bbox{B}\cdot\bbox{\partial}\,v_{\alpha}+\nonumber\\ &&\qquad \bbox{B}^o\cdot \bbox{\partial}\,v_{\alpha}+ \kappa_{0}\,\partial^2 B_{\alpha},\qquad \alpha=1,\cdots ,d ,\end{aligned}$$ where $\kappa_0$ is the magnetic diffusivity. The field $\bbox{B}^o$ plays the same role as an external forcing driving the system and being also a source of anisotropy for the magnetic field statistics. Our choice for the velocity statistics generalizes that of the well-known kinematic Kazantsev–Kraichnan model for the compressible case: $\bbox{v}$ is a Gaussian process of zero average, homogeneous, isotropic and white in time. It is self-similar and defined by the two-point correlation function: $$\langle v_{\alpha}(t,\bbox{x}) v_{\beta}(t',\bbox{x}') \rangle = \delta (t-t')\, [ d^0_{\alpha\beta} - S_{\alpha\beta} (\bbox{x}- \bbox{x}') ], \label{2-point-v}$$ where $d^0_{\alpha\beta} = {\rm const} \,\, \delta _{\alpha\beta}$ and $S_{\alpha\beta}(\bbox{x}-\bbox{x}')$ is fixed by isotropy and scaling: $$S_{\alpha\beta}(\bbox{r})=r^{\xi}\left [{\cal X}\delta_{\alpha\beta} + {\cal Y}\frac{r_{\alpha}r_{\beta}}{r^2} \right ]\, , \qquad r\equiv \left |\bbox{x}-\bbox{x}' \right |\, , \label{eddydiff}$$ with the coefficients $${\cal X}=\frac{{\cal S}^2(d+\xi-1)-\xi {\cal C}^2}{(d+\xi)(d-1)\xi} \, ,\qquad {\cal Y}=\frac{d{\cal C}^2-{\cal S}^2}{(d+\xi)(d-1)}. \label{xy}$$ The degree of compressibility is thus controlled by the ratio $\wp\equiv {\cal C}^2/{\cal S}^2$, with ${\cal S}^2 \propto \langle (\partial_{i} v_{k} \partial_{i} v_{k}) \rangle $ and ${\cal C}^2\propto\langle(\bbox{\partial}\cdot\bbox{v})^2\rangle$. It satisfies the inequality $0\leq \wp \leq 1$; $\wp=0$ and 1 corresponding to the purely solenoidal and potential velocity fields, respectively. In the present Letter our attention will be focused on the inertial-range behavior of magnetic correlation functions, where power laws are expected in their decompositions on a set of orthonormal functions $P$ $$\langle B_{\parallel}^n (t, \bbox{x}) B_{\parallel}^q (t,\bbox{x}') \rangle ={\textstyle\sum_{j=0}^{\infty}} P_j(\cos\phi)\, r^{\zeta_j^{n,q}}, \label{scaling}$$ $B_{\parallel}$ being some component of $\bbox{B}$, e.g., its projection along the direction $\hat{\bbox{r}}\equiv\bbox{r}/r$ or $\hat{\bbox{B}}^o\equiv\bbox{B}^o/B^{o}$, and $\phi$ is the angle between $\bbox{r}$ and $\bbox{B}^o$. Notice that, due to the anisotropic injection mechanism, the inertial-range statistics is now characterized by an infinite hierarchy of exponents ($j$ denotes the $j$-th anisotropic sector), rather than just one exponent as in the isotropic case. If the Kolmogorov 1941 isotropization hypothesis holds, the contribution from the anisotropic sectors (i.e. $j\neq 0$ ) to the scaling of correlation functions should be negligible with respect to the isotropic component. Such picture indeed holds for even correlation functions and solenoidal velocity [@LM99; @ALM99], but it breaks down for compressibility strong enough.\ In order to prove this fact, consider first the pair correlation function, $C_{\alpha\beta}(t,\bbox{r}) \equiv\langle B_{\alpha} (t,\bbox{x}) B_{\beta}(t,\bbox{x}')\rangle$. Here, exploiting the zero-mode technique [@GK95; @CFKL95; @SS96; @V96], the complete set of scaling exponents $\zeta_j^{1,1}$ can be found nonperturbatively. We give the basic ideas of the strategy. The first key point is that exact, closed equation for $C_{\alpha\beta}$ can be found due to the time decorrelation of the velocity field: $$\begin{aligned} &&\partial_t\, C_{\alpha\beta}= S_{ij}\,\partial_i\partial_j \,C_{\alpha\beta}- \left (\partial_j\,S_{i\beta}\right ) \partial_i\,C_{\alpha j}-\nonumber\\ &&\qquad\left (\partial_jS_{\alpha i}\right )\left (\partial_i C_{j \beta}\right )+ \left (\partial_i\partial_j S_{\alpha\beta}\right ) \left ( C_{ij} + B_i^oB_j^o\right ) + \nonumber \\ &&\qquad 2\kappa_0\, \partial^2 C_{\alpha\beta}-B_\beta^oB_j^o \partial_i\partial_j S_{\alpha i}-B_\alpha^oB_j^o\partial_i\partial_j S_{\beta i}+\nonumber\\ &&\qquad B_\alpha^oB_\beta^o\partial_i\partial_j S_{i j}- C_{\alpha j}\,(\partial_i\partial_j \,S_{\beta i})+C_{\beta j}\, \partial_i\partial_j \,S_{\alpha \beta}+\nonumber\\ &&\qquad C_{\alpha \beta }\,\partial_i\partial_j\,S_{ i j}+2(\partial_i C_{\alpha \beta})\, \partial_j \,S_{i j} \label{eq-moto}\end{aligned}$$ (see [@ALM99] for the derivation in the incompressible case).\ $C_{\alpha\beta}$ can be projected on a vector basis including the large scale magnetic field $\bbox{B}^o$ [@LM99; @ALM99], or, alternatively, on the purely inertial range basis that span the irreducible representations of SO($d$) [@ALP99]. Deep in the inertial range, both descriptions yield a system of linear algebraic equations for the scaling law coefficients. The leading solutions are associated with the homogeneous solutions of such system, i.e. with zero modes. Their scaling exponents follow from the imposition that the determinant of the coefficients be zero. Schematically, the solutions can be expressed in the form: $$\zeta_j^{1,1}=\alpha+\sqrt{\beta+\gamma\sqrt{\delta}}\qquad (j\ \mbox{even}), \label{zero-modes}$$ where $\alpha$, $\beta$, $\gamma$ and $\delta$ are cumbersome functions of $\xi$, $d$ and $j$ and will not be reported here for the sake of brevity. Contributions to the scaling associated with odd $j$’s vanish due to the symmetry $\phi\mapsto -\phi$.\ In particular, the expression derived in Ref. [@RK97] in the isotropic case has been recovered here for $j=0$.\ For $j=0$ and $j=2$, the limit $\xi\to 0$ in Eq. (\[zero-modes\]) yields: $$\begin{aligned} &&\zeta_0^{1,1} = (-1+2\ \wp -d\ \wp)\ \xi +O(\xi^2), \label{small-xi1}\\ &&\zeta_2^{1,1} = \frac{2-\wp\ [4+d\ (d-2)\ (d+1)]}{(d-1)(d+2)}\xi +O(\xi^2), \label{small-xi2}\end{aligned}$$ while for large $d$ we have: $$\begin{aligned} &&\zeta_0^{1,1} =-\frac{\wp\ \xi}{1+\wp\ \xi}d -\frac{1-2 \wp\ (1-\xi)}{1+\wp\ \xi}\xi +O(1/d),\label{large-d1}\\ &&\zeta_2^{1,1} = -\frac{\wp\ \xi}{1+\wp\ \xi}d -\frac{\wp\ \xi(\xi-2)}{1+\wp\ \xi} +O(1/d) . \label{large-d2}\end{aligned}$$ In Fig. \[fig1\] we present the behavior of $\zeta_j^{1,1}$ obtained from the expression (\[zero-modes\]) for $\wp=0$ (thin lines) and $\wp=1$ (heavy lines). As one can see, for $\wp=0$ we have $\zeta_0^{1,1}<0$ but $\zeta_2^{1,1}>0$ and $\zeta_4^{1,1}>0$ so that, in particular, $(r/L)^{\zeta_0^{1,1}}\gg (r/L)^{\zeta_2^{1,1}}$ in the inertial range of scales. The situation changes for $\wp>\wp_{c}$ ($\wp_{c}\sim 0.1274$ for $j=2$ and $d=3$), when $\zeta_2^{1,1}$ becomes negative for all $0\le\xi\le 1$. This means that, for compressibility strong enough, $(r/L)^{\zeta_0^{1,1}}$ and $(r/L)^{\zeta_2^{1,1}}$ become very close (still, $\zeta^{1,1}_{j} < \zeta^{1,1}_{k} $ for $j<k$). As we can see from Eqs. (\[large-d1\]), (\[large-d2\]) the effect becomes dramatic for large dimensions and $\wp \neq 0$ when $\zeta_0^{1,1}\sim \zeta_2^{1,1}$. The contribution to the scaling of the correlation functions in Eq. (\[scaling\]) coming from the sector $j=2$ thus becomes less and less subleading as $\wp$ and/or $d$ increase. Further strong evidence of the crucial role of compressibility in the failure of the small-scale isotropy restoration can be obtained by looking at the higher-order correlation functions \[i.e. $n$ and/or $q>1$ in Eq. (\[scaling\])\] and, in particular, at dimensionless ratios of odd-order moments. Were isotropy restored at small scales, such ratios would go to zero for large Péclet number. The latter is defined as Pe $\equiv(L/\eta)^{1/\xi}$, where $L$ is the integral scale and $\eta\propto\kappa_0^{1/\xi}$ is the dissipation scale. Non-zero values of such indicators are thus the signature of anisotropy persistence. As we are going to show, in contrast to the incompressible case, for $\wp$ large enough the skewness factor now diverges as $\mbox{Pe}\to\infty$. To be more specific, the leading contribution in the expression (\[scaling\]) can be written as: $$\langle B_{\parallel}^n (t, \bbox{x}) B_{\parallel}^q (t,\bbox{x}') \rangle \propto (r/\eta)^{\alpha_0^{n,q}} \, (r/L)^{\beta_0^{n,q}} \propto r^{\zeta_0^{n,q}}\, , \label{final}$$ where we have defined $\zeta_0^{n,q}\equiv \alpha_0^{n,q}+\beta_0^{n,q}$. The expressions for $\zeta_0^{n,q}$ have been obtained up to first order in $\xi$ by means of the field theoretic renormalization group and operator product expansion. A detailed presentation of these techniques for the case $\wp=0$ can be found in [@ALM99] (see also Refs. [@A98; @A99; @AAV98] for the scalar case and [@UFN] for general review); below we confine ourselves to only the necessary information. The key role is played by the scaling dimensions $\Delta[n,j]$ of the $j$-th rank composite operators $ B_{\alpha_{1}}(x)\cdots B_{\alpha_{j}}(x) \bigl[B_{\alpha}(x)B_{\alpha}(x)\bigr]^{l}$ ($n\equiv 2l+j$ is the total number of $\bbox{B}$’s). The first order in $\xi$ (one-loop approximation) yields $$\Delta[n,j]=\frac{\xi}{2(d-1)(d+2)} \, \left\{ n(n-1) \left[2-\wp(d^{3}-d^{2}-2d+4)\right] - (n-j)(n+j+d-2)(d+1-2\wp) \right\} +O(\xi^{2}). \label{Dnp}$$ [2]{} The exponents in Eq. (\[scaling\]) are related to the dimensions (\[Dnp\]) as follows (see Ref. [@ALM99]): $$\zeta^{n,q}_{j} = \Delta[n+q,j] - \Delta[n,j_{n}] - \Delta[q,j_{q}],$$ where $j_{n} = 0 $ for $n$ even and 1 for $n$ odd. We thus conclude that the inequality $\partial \Delta[n,j] / \partial j \ge 0$, which follows from Eq. (\[Dnp\]) for all $\wp$ and $d\ge2$, generalizes the hierarchy discussed above to the higher-order functions. It becomes flatter and flatter as $\wp$ grows ($\partial^{2} \Delta[n,j] / \partial j \partial\wp \le 0$), while for $d\to\infty$ (and $\wp \neq 0$) the effect becomes even stronger: in the leading $O(d)$, expressions (\[Dnp\]) and (\[zeta\]) are now independent of $j$. The leading term of Eq. (\[scaling\]) in the inertial range is given by the contribution with the minimal $j$: $$\zeta_0^{n,q} = \cases{ -\xi{\displaystyle \frac{nq (1+\wp d^{2}-2\wp)+ (d+1-2\wp)}{(d+2)}} & \cr {\displaystyle -\frac{\xi nq }{(d+2)}} (1+\wp d^{2}-2\wp), & \cr} \label{zeta}$$ where the first holds if both $n$ and $q$ are odd, and the second otherwise. For $n=q=1$, expression (\[small-xi1\]) is recovered. Knowing the exponents $\alpha_0^{n,q}$ and $\beta_0^{n,q}$, dimensionless ratios of the form $R_{2n+1}(r) \equiv \langle B_{||}^{2n} (\bbox{x}) B_{||} (\bbox{x}') \rangle/\langle B_{||} (\bbox{x}) B_{||} (\bbox{x}')\rangle ^{(2n+1)/2}$ can be constructed and, as in Ref. [@P96], evaluated at the dissipative scale \[i.e. $r=\eta$ in Eq. (\[final\])\]. When doing this, the explicit dependence on Pe appears and the final $O(\xi)$ expressions read: $$R_{2n+1}(\mbox{Pe}) \propto \mbox{Pe}^{\sigma_{2n+1}}, \qquad n=1,2,\cdots, \label{skew}$$ with the exponents $$\sigma_{2n+1}(\wp )\equiv \frac{2 n^{2} (1-2\wp+\wp d^{2})}{(d+2)} -\frac{ (1-2\wp +\wp d)}{2}. \label{expon}$$ It is easy to verify from expression (\[expon\]) that for $d\geq 2$ and $n\geq 1$ we have $\partial\sigma_{2n+1}(\wp)/\partial \wp>0$. Negative values of $\sigma_{2n+1}(0)$ may thus become positive due to $\wp$. In particular, for $d=3$ we obtain $\sigma_{3}(\wp)=(23\wp -1)/10$, which becomes positive for $\wp > 1/23$. It then follows that $R_{3}\to\infty$ as $\mbox{Pe}\to\infty$, the footprint of the persistence of the small-scale anisotropy. Some remarks on the limit of large space dimensions are worth. One immediately realizes from Eq. (\[zeta\]) that, for $d\to\infty$ the scaling exponents reduce to $\zeta_0^{n,q}=-\xi nq d\wp$, that means $\zeta_0^{n,q}=n\zeta_0^{1,q}=q\zeta_0^{n,1}$, and thus the vanishing of intermittency for $d\to\infty$. Nevertheless, in this limit $\zeta_0^{1,1}=\zeta_2^{1,1}$ and, at the dissipative scale $\eta$, $\sigma_{2n+1}(\wp)=\wp d(2n^2-1/2)>0$, two clear signatures of small-scale anisotropy persistence. The latter may thus occur also in the absence of intermittency. Let us now examine the possible mechanisms at the origin of the small-scale anisotropy persistence in our problem. They can be easily grasped in two dimensions; our previous results being valid for all $d\geq2$, it is reasonable to expect that the mechanisms we are going to show hold also for higher $d$’s. Let us start from the incompressible case assuming, without loss of generality, that $\hat{\bbox{B}}^o$ is oriented along the $y$-axis (i.e. $\hat{\bbox{B}}^o \equiv \bbox{e}_y$). As we shall see, all our considerations will hold, [*a fortiori*]{}, in the compressible case. It can be shown that the magnetic field can be represented by the (scalar) magnetic flux function in the form $\bbox{B}=\bbox{\partial}\psi \times\bbox{e}_z$, where $\psi$ satisfies the passive scalar equation forced by a large scale gradient $\bbox{G}\equiv \bbox{B}^o\times \bbox{e}_z$. One finds $$\partial_t \Psi +\bbox{v}\cdot\bbox{\partial}\,\Psi = \kappa_{0}\,\partial^2 \Psi \label{duedim}$$ where $\Psi(t,\bbox{x})\equiv \psi+ \bbox{G}\cdot \bbox{x} $. Notice that, fort $d=2$, expression (\[zero-modes\]) recovers the results of Ref. [@CFKL95] obtained directly for the scalar problem. An interesting feature recently recognized for the passive scalar problem, for both synthetic advection by $\delta$-correlated velocity [@FGLP97] and Navier–Stokes advection [@P94; @HS94; @CLMV99; @CLMV99bis], is related to the formation of “cliffs”, i.e. very steep scalar gradients within very short distances. The region where such gradients develop are separated by “plateaus” where the scalar depends smoothly on the position (i.e. regions of gradient expulsion).\ The emergence of this ubiquitous pattern is explained by considering the action of the velocity derivative matrix $\partial_{\alpha}v_{\beta}$. As emphasized in Refs. [@P94; @HS94], scalar gradients are weak in the elliptic regions of the velocity field (in two dimensions they correspond to purely imaginary eigenvalues of the velocity derivative matrix) where its ‘rotational’ character inhibits the formation of strong scalar gradients. On the contrary, in the hyperbolic regions the flow is almost ‘irrotational’: the alignment of scalar gradients with the direction of the eigenvector corresponding to the most negative eigenvalue of the velocity derivative matrix (roughly, the direction corresponding to compression along one direction) is not discouraged, and actually observed [@P94], and strong scalar gradients develop. When the scalar is forced by isotropic injection mechanisms, no preferential direction arises and the intense gradients are randomly oriented. The final result is a small scale isotropic statistics. The situation changes in the case of non-isotropic injection mechanisms like the one encountered here. In this case strong gradients are oriented along $\bbox{G}$ and, as very recently shown in Refs. [@CLMV99; @CLMV99bis], small-scale isotropy is consequently not restored for the scalar field. Exploiting the relation between the magnetic field and the magnetic flux function, we can conclude that extreme magnetic fluctuations have a tendency to occur preferentially along the direction $\hat{\bbox{B}}^o$, the origin of the observed small-scale anisotropy. In the compressible case both eigenvalues can be negative Compression may thus occur in both directions, enhancing the formation of fronts in the magnetic flux function. In conclusion, we presented a simple model of MHD turbulence where, by varying the degree of compressibility of the velocity field, the persistence of anisotropy is detectable both from the hierarchy of inertial range exponents and from the divergence of skewness factor with the Péclet number. Although our results were obtained on the base of a specific model, they seem to be rather general: a similar behavior is also observed for a passive scalar advected by a compressible velocity field [@A99] (however, the flattening is less pronounced for the latter). The results presented here give the first evidence of a turbulent system displaying such twofold behavior. We are deeply grateful to L.Ts. Adzhemyan, L. Biferale, A. Celani, A. Lanotte, and M. Vergassola for discussions throughout the present work. N.V.A. acknowledges the hospitality of the University of Helsinki. N.V.A. was supported by the Academy of Finland, GRACENAS Grant No. 97-0-14.1-30, and RFFI Grant No. 99-02-16783. A.M. was partially supported by the INFM PA project No. GEPAIGG01. P.M.G. was partially supported by grant ERB4001GT962476 from EU. [99]{} B.I. Shraiman and E. Siggia, [*Scalar Turbulence*]{} (to appear in Nature). V. Borue and S.A. Orszag, J. Fluid Mech. [**306**]{}, 293 (1996). I. Arad, L. Biferale, I. Mazzitelli, and I. Procaccia, Phys. Rev. Lett. [**82**]{}, 5040 (1999). A. Lanotte and A. Mazzino, Phys. Rev. E [**60**]{}, R3483 (1999). I. Arad, L. Biferale, and I. Procaccia, Phys. Rev. E, [**61**]{}, 2654 (2000). N.V. Antonov, A. Lanotte, and A. Mazzino, E-print nlin.CD/0001039 (to appear in Phys. Rev. E). A. Pumir, Europhys. Lett. [**34**]{}, 25 (1996); [**37**]{}, 529 (1997); Phys. Rev. E [**57**]{}, 2914 (1998). N.V. Antonov, Phys. Rev. E [**60**]{}, 6691 (1999). N.V. Antonov, E-print chao-dyn/9907018 (to appear in Physica D); N.V. Antonov and J. Honkonen, (in preparation). Ya.B. Zeldovich, A.A. Ruzmaikin, and D.D. Sokoloff, [*Magnetic fields in Astrophysics*]{} (Gordon & Breach, New York, 1983). K. Gawȩdzki and A. Kupiainen, Phys. Rev. Lett. [**75**]{}, 3834 (1995); D. Bernard, K. Gawȩdzki, and A. Kupiainen, Phys. Rev. E [**54**]{}, 2564 (1996). M. Chertkov, G. Falkovich, I. Kolokolov, and V. Lebedev, Phys. Rev. E [**52**]{}, 4924 (1995); M. Chertkov and G. Falkovich, Phys. Rev. Lett. [**76**]{}, 2706 (1996). B.I. Shraiman and E.D. Siggia, Phys. Rev. Lett. [**77**]{}, 2463, (1996). M. Vergassola, Phys. Rev. E [**53**]{}, R3021 (1996). I. Arad, V.S. L’vov, and I. Procaccia, Phys. Rev. E. [**59**]{}, 6753 (1999). I. Rogachevskii and N. Kleeorin, Phys. Rev. E [**56**]{}, 417 (1997). L.Ts. Adzhemyan, N.V. Antonov, and A.N. Vasil’ev, Phys. Rev. E [**58**]{}, 1823 (1998); Theor. Math. Phys. [**120**]{}, 1074 (1999). L.Ts. Adzhemyan, N.V. Antonov, and A.N. Vasil’ev, Usp. Fiz. Nauk, [**166**]{}, 1257 (1996) \[Phys. Usp. [**39**]{}, 1193 (1996)\]. A.L. Fairhall, B. Galanti, V.S. L’vov, and I. Procaccia, Phys. Rev. Lett. [**79**]{}, 4166 (1997). A. Pumir, Phys. Fluids [**6**]{}, 2118 (1994). M. Holzer and E.D. Siggia, Phys. Fluids [**6**]{}, 1820 (1994); Erratum, Phys. Fluids [**7**]{}, 1519 (1995) A. Celani, A. Lanotte, A. Mazzino, and M. Vergassola, Phys. Rev. Lett. [**84**]{}, 2385 (2000). A. Celani, A. Lanotte, A. Mazzino, and M. Vergassola, [*Fronts in passive scalar turbulence*]{} (in preparation).

ArXiv

--- abstract: 'With an auxiliary weak external magnetic field, we reexamine the fundamental thermodynamic function, Gibbs free energy $F(T, h)$, to study the phase transitions in the classical spin lattice models. A cross derivative, i.e. the second-order partial derivative of $F(T, h)$ with respect to both temperature and field, is calculated to precisely locate the critical temperature, which also reveals the nature of a transition. The strategy is efficient and universal, as exemplified by the 5-state clock model, 2-dimensional (2D) and 3D Ising models, and the $XY$ model, no matter a transition is trivial or exotic with complex excitations. More importantly, other conjugate pairs could also be integrated into a similar cross derivative if necessary, which would greatly enrich our vision and means to investigate phase transitions both theoretically and experimentally.' author: - 'Y. Chen' - 'K. Ji' - 'Z. Y. Xie' - 'J. F. Yu' bibliography: - 'classical\_spin\_models.bib' title: 'Cross derivative: a universal and efficient method for phase transitions in classical spin models' --- Phases of matter and phase transitions have always been the hot topics in the statistical and the condensed matter physics. For decades, Landau’s symmetry-breaking theory was believed fully qualified to identify and describe different phases and phase transitions in-between. As a seminal illustration of the spontaneous symmetry breaking, the classical Ising model ($Z_2$ symmetry) on the square lattice, undergoes a typical order-disorder phase transition. Its opposite extreme, the continuous $XY$ model ($U(1)$ symmetry), involves the exotic topological vortices excitation and a phase transition without symmetry broken, i.e. the so-called Kosterlitz-Thouless (KT)[@KT1; @KT2] transition beyond Landau’s theory. Both types of transitions can be easily probed by the magnetic susceptibility, which reflects the system’s response to an external magnetic field and behaves distinctively across the critical point. One natural question is how the universality class evolves with the symmetry of the models, which arouses intensive interest in the intermediate $q$-state clock model with a finite $q$. As well known, when $q$ is no bigger than 4, it has one unique second-order phase transition; otherwise, there are two separate transitions sandwiching a critical KT phase with quasi-long-range order. So far, major debates focus on $q$ near 5, about the nature and the precise locations of the transitions. Usually, Monte Carlo (MC), one of the principal methods for many-body problems, calculates the helicity modulus[@Lapilli; @Baek1; @Baek3; @Kumano; @Okabe] to characterize the KT transition. As in the continuous $XY$ model on the square lattice, it jumps abruptly from finite to zero at the critical point[@Minnhagen]. For the $5$-state clock model, at the upper transition point, it behaves similarly to the $XY$ case. However, as to the lower one, inconsistent conclusions about the transition type were claimed by different groups with MC studies[@Kumano; @Baek3]. By proposing an extended universality class theory with MC simulations, Lapilli et al.[@Lapilli] even declared both transitions are not KT-type when $q\le6$, which is supported by Ref. \[\] from a Fish zero analysis for $q=6$ case. Another powerful method, the renormalization group (RG) predicted two KT transitions early[@Kadanoff], and a recent density matrix renormalization group (DMRG) study[@Christophe] favored this assertion by calculating the helicity modulus with relatively small system sizes. The tensor network states, generalized from DMRG to higher dimensional strongly correlated systems, have developed rapidly and been widely used to investigate both the classical and the quantum systems. Among those, the tensor renormalizaton group method based on the higher-order singular value decomposition (abbreviated as HOTRG)[@Xie2], has been successfully applied to study the 3D Ising model[@Xie2], the Potts model[@MPQin; @WShun], and the continuous $XY$ model[@JFYu]. Actually, it has been also utilized to study the 5-state clock model, where the magnetic susceptibility properly describes the upper phase transition, but does not work well for the lower one[@Chenyong], as also presented in Fig. \[fe\](b) below. Therefore, a gauge invariant factor from the fixed point tensor of the RG flow, proposed in Ref. \[\], was adopted to measure the degeneracy of each phase, which also precisely estimates the critical points of the 6-state clock model[@Jing]. Nevertheless, some important information is still missing, e.g., the reason why the magnetic susceptibility loses its efficacy for the lower transition in this model, and the nature of the transitions in particular, although many researches claimed both are KT transition. A duality analysis by the conformal field theory (CFT) deemed that two transitions are KT-type, but they still have some differences[@Elitzur; @Matsuo]. Recently, a universal entropy predicted by CFT on a Klein bottle[@TuPRL; @TuPRB] has different values at these two critical points[@TuPrivate], which is believed valuable to distinguish different CFTs. Here, we intend to detect and clarify the nature/mechanism of the classical phase transitions, within a unified frame, by reexamining the fundamental thermodynamic function, Gibbs free energy $F$, which is intrinsically a signpost of the universal entropy increase of a spontaneous change[@Atkins], and contains information about the phase transitions. However, the free energy and its temperature derivatives, i.e. the internal energy and the specific heat, are analytically continuous without any singularity in the KT transition. So, besides the temperature, we propose an auxiliary parameter, a weak external magnetic field, which interacts with spin degree of freedom and competes with thermal excitations, thus provides us a convenient tool to investigate the dynamical behavior of the system. By detailed analyses on the cross derivative of $F(T, h)$ with respect to both temperature and field, we can easily identify and precisely locate the transition points. Moreover, since the free energy is fundamental, this idea is readily applied to any classical spin system, like Ising or $XY$ model with trivial or exotic transitions. In other words, it is universal. First, we demonstrate this idea explicitly by the ferromagnetic $5$-state clock model with an in plane magnetic field, whose Hamiltonian is written as $$H = - J \sum_{\left< ij \right>}\cos(\theta_{i}-\theta_{j})-h\sum_{i}\cos\theta_{i},$$ where $\left< ij \right>$ means summing over all nearest neighbors. $\theta_i$ is the spin angle on lattice site $i$, selected from $\theta=2\pi k/q$, $k = 0, 1, 2, \ldots, q-1$. $J$ is the nearest coupling. $h$ is the applied field in unit of $J/\mu$, and $\mu$ is the magnetic moment of each spin. Both $J$ and $\mu$ are set as 1 for convenience.  (a) Gibbs free energy $F$ (blue blank square) of 5-state clock model with a magnetic field $h=4.0 \times 10^{-5}$; comparison of $-\partial F/\partial h$ (black empty circle) and $\boldsymbol{m}$ (red filled circle); (b) magnetic susceptibility $\partial \boldsymbol{m} / \partial h$ (blue blank square), cross derivative ${\partial}^2 F / \partial T \partial h$ (black blank circle), $-\partial \boldsymbol{m} / \partial T$ (red filled circle), and $-\partial{S} / \partial h$ (green cross).](fe.eps){width="48.00000%"} Here, we employ HOTRG method to compute the desired physical quantities. Details about the algorithms are in Refs. \[\]. Its accuracy, like other RG algorithms, is subject to the number of states kept during the RG process, labelled by the bond dimension $D$. Initially, it equals $q$, then expands exponentially along the RG process. Therefore, a truncation is necessary to ensure further steps sustainable. The free energy $F(T, h_1)$ with field $h_1$ is presented in Fig. \[fe\](a), wherein $-\partial F/ \partial h$ and the magnetization $\boldsymbol{m}$ are also shown. For comparison, the quantity $-\partial F/ \partial h$ is computed directly from $-[F(h_2)-F(h_1)]/(h_2-h_1)$ by using two close field strengths, and $\boldsymbol{m}$ is calculated by the impurity tensor algorithm[@Xie2; @JFYu; @Chenyong]. One can see they agree well with each other as should do. For this model, as discussed in Ref. \[\], the magnetic susceptibility can clearly identify the upper phase transition, but not be so convenient for the lower one. As shown in Fig. \[fe\](b) by blue blank squares, an exponential divergence clearly labels a phase transition near $T=1.0$. Meanwhile, a broad shoulder-shape structure emerges below, indicating something happens, but not as evident as the upper one. Instead, the cross derivative of the free energy with respect to both temperature and field, i.e. $\partial^2{F}/{\partial{T}\partial{h}}$, is able to characterize both transitions simultaneously. Clearly, as shown in Fig. \[fe\](b) by black blank circles, two separate sharp peaks show up. In particular, the upper one is coincident perfectly with the susceptibility curve, for both the position and the shape, although it decays exponentially from a much smaller peak other than divergence as in the magnetic susceptibility. The lower one, small but still obvious, locates near $T=0.90$. Here, a relatively small bond dimension $D=40$ is used just for illustration. As verified in Fig. \[fe\](a), $-\partial{F}/{\partial{h}}$ is just $\boldsymbol{m}$, then the cross derivative equals the temperature derivative of the magnetization $-\partial \boldsymbol{m} / \partial T$. As both shown in Fig. \[fe\](b), they match up well with each other. Similarly, one can choose function $-\partial S / \partial h$, as also presented in Fig. \[fe\](b), because $-\partial{F}/{\partial{T}}$ is just the thermodynamic entropy $S$. Additionally, the Maxwell relation[@Reichl] $\partial{S}/\partial{h}=\partial{\boldsymbol{m}}/\partial{T}$ is numerically verified by computing $S$ directly from the difference between Gibbs free energy and the internal energy, because both terms essentially spring from the cross derivative. For numerical simplicity and convenience, we adopt the notation $-\partial \boldsymbol{m} / \partial T$ hereafter, while keeping in mind its physical origin. Some may question the validity or the physical meaning of this cross derivative. One can imagine slicing the 3D curved surface $F(T, h)$ along $h$-axis, then performing the derivative $\partial{F}/{\partial{T}}$ for each $h$ slice, and observing its evolution along $h$-axis; or equivalently slicing $F(T, h)$ along $T$-axis and obtaining $\partial{F}/{\partial{h}}$, then investigating its evolution along $T$-axis. Thus, each captures the effects of both temperature and field, and the system dynamics can be easily deduced. This scheme may be elaborated by a formula $$\left( \frac{\partial}{\partial T}+\frac{\partial}{\partial h}\right)^2 F = \nabla ^2 F + 2\frac{\partial^2 F}{\partial T\partial h} ,\label{formu2}$$ where the left part in parentheses is a linear combination of two derivative operators in the two-dimensional orthogonal space expanded by temperature and field, and the Laplacian stands for the second-order derivatives with respect to each individual parameter, i.e. the specific heat and the magnetic susceptibility respectively, while neither is adequate to characterize the system dynamics comprehensively.  (a) Illustration of the peak positions of $-\partial \boldsymbol{m} / \partial T$ versus the magnetic fields for 5-state clock model with $D=40$, along with a power law fitting to extrapolate the transition temperature as $T_{c1}=0.9038$ and $T_{c2}=0.9557$ respectively; (b) The transition temperature versus the tensorial bond dimension $D$ to obtain the converged $T_c$ as 0.9063 and 0.9557, respectively.](q5_Tc.eps){width="48.00000%"} Similar to the procedure used in the continuous $XY$ model to locate the transition temperature[@JFYu], we vary the applied field, and obtain the peak positions of $-\partial \boldsymbol{m} / \partial T$, as presented in Fig. \[q5\_Tc\](a). To determine the critical points for a given $D$, an extrapolation to a zero field is performed by a power law fitting $T_p-T_c \sim h^x.$ As demonstrated in Fig. \[q5\_Tc\](a), the results with $D=40$ are obtained as $T_{c1}=0.9038$ and $T_{c2}=0.9557$. Likewise, we replicate the above process with different bond dimensions, and obtain the converged transition temperatures, i.e. $T_{c1}=0.9063$ and $T_{c2}=0.9557$, as shown in Fig. \[q5\_Tc\](b). Both agree well with the estimations from other researches[@Kumano; @Christophe; @Chenyong; @Borisen2; @Chatterjee]. Once obtaining the critical points, we can investigate the central charge $c$ as well as the critical exponent $\delta$ to determine the universality class of the transitions. According to the results of CFT[@Nightingale; @Afflect], we calculate the finite-size partition function on a torus, to obtain the central charge at two critical points and the sandwiched critical phase as $c=1.04$, which indicates that both transitions belong to the same $c=1$ CFT class. Meanwhile, the critical exponent $\delta$ is calculated, which signifies the change of the system magnetization with the applied magnetic field at the transition point as $m\sim h^{1/\delta}$. The results are $\delta_1=15.81$ and $\delta_2=15.77$ respectively, by using the bond dimension $D=70$. Both are consistent well with the theoretical value $\delta=15$ for the KT transition in the 2D $XY$ model[@KT2]. Combining $c$ and $\delta$ together, it probably implies two KT-type transitions. As also shown clearly in Fig. \[q5\_Tc\](a), the upper critical point shifts with the applied magnetic field, the stronger a field, the higher the transition temperature, similar to the $XY$ case[@JFYu], because more heat energy is needed to overcome the additional barrier introduced by the magnetic field. However, as seen in Fig. \[q5\_Tc\](a), the lower one moves oppositely, which seems to indicate a different scenario. Besides the vortex excitation, another typical topological excitation responsible for the melting of the magnetic order in magnetic systems is the domain wall[@Ortiz; @Einhorn; @Fertig; @Chatterjee], which probably plays an important role in this transition. To clarify the mechanism, we adopt the procedure of Refs. \[\] and the references therein, to investigate the influence of the vortices excitation on the phase transitions by introducing a parameter $\lambda$ to adjust the vortex core energy as $$\label{eqH5lamda} H = - J \sum_{\left< ij \right>} \cos(\theta_{i}-\theta_{j}) + \lambda\sum_{i'}\left |\omega_{i'}\right |,$$ where $\omega_{i'}=(\delta_{ba}-\delta_{cb}-\delta_{dc}-\delta_{ad})/5$, and $\delta_{ba}$ is $s_b-s_a$ wrapped in $[-1, 1]$. $s_a, s_b, s_c, s_d$ are spins on four vertexes of a square plaquette labelled by $i'$.  MC simulation of the Hamiltonian (Eq. \[eqH5lamda\]) with $L=128$ for different $\lambda$: (a) magnetization; (b) magnetic susceptibility; (c) $-\partial \boldsymbol{m} / \partial T$; (d) number densities of the domain walls ($\rho_d$) and the vortices ($\rho_v$), where $\rho_v$ is multiplied by 2 for a better view.](MC_all.eps){width="48.00000%"} By MC simulations about the above Hamiltonian (Eq. \[eqH5lamda\]) on a square lattice with $L=128$, we obtain the magnetization, the magnetic susceptibility and the deduced $-\partial \boldsymbol{m} / \partial T$ for different $\lambda$, as all shown in Fig. \[MC\_all\]. Increasing the vortex core energy to suppress its formation, a clear shift of the upper critical point can be seen from each curve. Again, $-\partial \boldsymbol{m} / \partial T$ looks much more convincing than the magnetic susceptibility for the lower transition. More importantly, as manifested in Fig. \[MC\_all\](b) and (c), this lower temperature phase transition is barely affected by the vortex suppression, which strongly suggests it is dominated by the domain wall excitation[@Ortiz; @Einhorn; @Fertig; @Chatterjee]. A more intuitive illustration is presented in Fig. \[MC\_all\](d), i.e. the number density of each excitation, by adopting the definition in Ref. \[\]. One can clearly observe that, near the lower transition point, the number density of the domain wall decreases negligibly, while the vortices are greatly suppressed even eliminated, when increasing $\lambda$. Furthermore, we calculate the aforementioned universal entropy $\ln{g}$ of CFT on a Klein bottle[@TuPRL] at the critical points, because CFT asserts the two transitions in this model are KT-type but with different $g$[@TuPrivate]. Our computation gives $g_1=3.30$ and $g_2=3.09$ respectively, both of which agree well with the CFT conclusion[@TuPrivate]. From the foregoing discussions, we can conclude that both transitions are indeed KT-type, but with subtle differences: the upper one is attributed to the unbinding of the vortices pairs, while the lower one is dominated by the domain wall excitation instead; and they belong to different CFTs. The differences may be closely related to why the magnetic susceptibility works fine for the upper transition but not so well for the lower one; and why the transition points shift oppositely with the external field as shown in Fig. \[q5\_Tc\](a). They may also be the reason why studies from different groups would give controversial estimations about the nature of the transitions.  $-\partial\boldsymbol{m} / \partial T$ and the power law fitting of its peak position varying with the applied field: (a) 2D $XY$ model with $D=40$; (b) 2D Ising model with $D=40$; (c) 3D Ising model with $D=10$.](combine.eps){width="48.00000%"} Briefly, with an auxiliary external magnetic field, the function $\partial \boldsymbol{m} / \partial T$ accurately reflects the interplay of the field and the temperature, and captures the implicit dynamics of the excitations in this model, hence correctly describes the phase transitions. While, the derivative of the free energy $F$ with respect to each single parameter, such as the specific heat or the magnetic susceptibility, is inadequate for lacking of information about the internal competition/interplay among those mingled complex excitations. The auxiliary magnetic field and the cross derivative provide us a convenient way to observe the response/dynamics of different excitations. To further check the universality of this idea, we apply it to the 2D $XY$, the 2D Ising, and the 3D Ising models separately. A sample of $-\partial\boldsymbol{m} / \partial T$ and the power law fitting of the peak position for each model are illustrated in Fig. \[combine\] as (a), (b), and (c) respectively. For the $XY$ case, the transition temperature is obtained at $T_c=0.8924(16)$, which is coincident to the previous estimation[@JFYu] from the magnetic susceptibility $T_c=0.8921(19)$ with same bond dimension $D=40$, and both conform to the results from other methods like MC[@Hasenbusch2; @Tomita] at $T_c=0.89294(8)$. For the 2D Ising case, the power law extrapolation yields the transition temperature at $T_c=2.26893(18)$, and a simultaneous prediction from the magnetic susceptibility (not shown in the figure) is $T_c=2.26904(22)$. They agree well with each other, and with the exact value $T_c=2/\ln{(\sqrt{2}+1)}\sim2.26919$, even using a relatively small bond dimension $D=40$. As to the 3D Ising case, the same procedure is carried out with the bond dimension $D=10$. The similar efficiency of the function $-\partial \boldsymbol{m} / \partial T$ is clearly demonstrated once again, from which the critical temperature is located at $T_c=4.5014(2)$. Also, the $T_c$ is determined at $4.5013(1)$ from the magnetic susceptibility. Both are consistent with the prediction at $T_c=4.5015$ by the HOTRG calculation with same $D$[@Xie2]. What’s more, we can also observe the singularity in the $-\partial \boldsymbol{m} / \partial T$ curve of the 2D/3D Ising model, indicating a second-order phase transition. It becomes more manifest and sharper if lowering the field down to zero, then a direct and accurate determination of the critical point can be obtained with no need for an extrapolation. These examples have verified the capability of our idea perfectly, and the proposed cross derivative $\partial^2{F}/{\partial{T}\partial{h}}$ seems more versatile and effective, no matter a transition is trivial or exotic, especially when multiple exotic excitations are involved and other quantities/methods are difficult to clarify. Also, we think this strategy is universal, as long as the free energy can be calculated accurately with a weak external magnetic field included. Experimentally, one can measure the system magnetization $\boldsymbol{m}(T, h)$, from which the phase transition information can be easily deduced. More importantly, the magnetic field and the magnetization in Gibbs free energy or the Hamiltonian are just one typical conjugate pair of generalized force and displacement[@Reichl]. Likewise, other conjugate pairs, if introduced into the Hamiltonian to regulate a system’s behavior, would play a similar role in investigating the phase transitions, e.g. the electric field and the polarization in an electronic system, which could be integrated into formula similarly. Thus, this idea will greatly enrich our vision and means to study the phase transitions both theoretically and experimentally. Considering its accuracy and simplicity, the idea we proposed in this work is efficient and universal to investigate the phase transitions in classical spin systems, trivial or complex, 2D or 3D. The predictions will be more accurate if the free energy or the physical quantities involved could be computed more precisely. We are grateful to Hong-Hao Tu, Fuxiang Li, and Yu-Chin Tzeng for valuable discussions and comments. Y. Chen thanks Mr. Yuan Si for helps on Monte Carlo simulations. This work was supported by the Shanghai Pujiang Program (No. 17PJ1407400), the National Natural Science Foundation of China (No. 11774420), the National R&D Program of China (No. 2016YFA0300503, No. 2017YFA0302900), and the Natural Science Foundation of Hunan Province (No. 851204035).

ArXiv

--- abstract: 'For the two-dimensional ideal electron gas with the Rashba spin-orbit interaction and in the presence of non-magnetic short-ranged potential impurities the spin-Hall conductivity $\sigma_{sH}$ is found by direct microscopic calculation. Within the semiclassical approximation $\hbar/\tau \ll \epsilon_F$ and $\Delta \ll \epsilon_F$ the value of $\sigma_{sH}$ [*is zero*]{} for arbitrary ratio of the spin-orbit splitting $\Delta$ and the inverse elastic scattering time $1/\tau$.' author: - 'Ol’ga V. Dimitrova' title: 'Vanishing Spin-Hall Conductivity in 2D disordered Rashba electron gas' --- Recently it has been proposed [@Murakami] that a [*dissipationless* ]{} spin current can be generated in response to an electric field in semiconductors with the spin-orbital interaction. For the case of an ideal two-dimensional (2D) electron gas with the Rashba coupling, Sinova et al. [@Sinova] have found a spin-Hall current of the transverse ($z$) spin component as a response to an in-plane electric field $E_{\nu}$, $j^z_\mu = \sigma_{sH} \epsilon_{\mu\nu} E_\nu$, with the “universal” (independent upon the spin-orbital band splitting $\Delta$) spin-Hall conductivity $\sigma_{sH}= e/8\pi\hbar$. In the presence of disorder and for the simplest standard model it was demonstrated a cancellation of the spin-Hall effect even in the case of arbitrary week disorder. [@Molenkamp2; @Shytov; @Khaetskii; @Schwab]. In this communication we provide a fully microscopic calculation of the spin-Hall conductivity for the generalized model of non-parabolic spectrum and arbitrary momentum dependence of the Rashba velocity (\[oneHam\]), within the semiclassical theory of disordered conductors, i.e. in the limit $\hbar/\tau\ll \epsilon_F$, $\Delta \ll \epsilon_F$. We show, that $\sigma_{sH} = 0$ [*independently*]{} upon the parameter $\tau\Delta/\hbar$ similarly to the case of the parabolic spectrum Ref. . 2D isotropic Rashba gas is an electron system with the broken inversion symmetry. In this case an electric field perpendicular to the plane could arise. It has no effect on the electron orbital motion but it couples to the electron spin via a relativistic spin-orbit interaction known as the Rashba term. The Hamiltonian of an electron consists of the kinetic energy term and the Rashba term: $$\label{oneHam} \hat{h}_{\alpha\beta}(\vec{p})=\epsilon(p)\delta_{\alpha\beta}+\alpha(p) \left( \sigma^x_{\alpha\beta} \hat{p}_{y}-\sigma^y_{\alpha\beta}\hat{p}_{x} \right) ,$$ where $\alpha(p)$ is the Rashba velocity, $\sigma^i$ ($i=x,y,z$) are the Pauli matrices and $\alpha,\beta$ are the spin indices. The Hamiltonian (\[oneHam\]) can be diagonalized by the unitary matrix: $$\label{Unitary} U(\vec{p})={1\over\sqrt{2}} \left( \begin{array}{cc} \displaystyle 1 & 1 \\ \displaystyle ie^{i\varphi_{\bf p}} & -ie^{i\varphi_{\bf p}} \end{array} \right),$$ where $\varphi_{\bf p}$ is the angle between the momentum $\vec{p}$ of the electron and the $x$-axis, with the eigenvalues: $$\label{eigenv} \epsilon_\lambda(p)=\epsilon(p) -\lambda p\alpha(p).$$ The eigenvalues $\lambda=\pm 1$ of the chirality operator and $\vec{p}$ constitute the quantum numbers of an electron state $(\vec{p},\lambda)$. Fermi circles of the Rashba gas with the different chiralities are split: $p_{F\pm}=p_F(1\pm\alpha(p_F)/v(p_F))$, where Fermi momentum $p_F$ solves the equation: $\epsilon(p_F)=\mu$, where $\mu$ is the chemical potential; $v(p)=d\epsilon(p)/dp$ is the band velocity of the electron. The electron velocity in the chiral state is $\partial \epsilon_{\lambda}(p)/ \partial p$. We assume $\alpha(p_F)\ll v(p_F)$, and neglect corrections of the order $\alpha/v$. The spin-orbital splitting is then: $\Delta=2p_F\alpha(p_F)$. The density of states on the two Fermi circles differs as: $\nu_{\pm}= \nu_F(1\pm \alpha_F/v_F)$, where $\nu_F=p_F/2\pi v(p_F)$. Contrary to the case of the parabolic spectrum and Rashba velocity independent on the momentum, for the generalized model (\[oneHam\]) the Fermi velocities are different on the two Fermi circles: $v_{F+}-v_{F-}=2\alpha_F\left(\frac{p_F}{mv_F}-1\right)- 2p_F\frac{d\alpha}{dp}\Big|_F$. In the following we use the units where $\hbar=1$. We consider the 2D ideal (non-interacting electrons) Rashba electron gas with the Hamiltonian: $$\label{mainHam} \hat{H}_R= \sum_{\vec{p}} a^{\dagger}_{\alpha}(\vec{p})\ \hat{h}_{\alpha\beta}(\vec{p})\ a_{\beta}(\vec{p}),$$ at zero temperature. $a^{\dagger}_{\alpha}(\vec{p})$ and $a_{\beta}(\vec{p})$ are the electron creation and annihilation operators. Electromagnetic vector potential $\vec{A}$ couples to the orbital motion of the electron according to the transformation: $\vec{p}\rightarrow \vec{p}- e\vec{A}/c$, in the Hamiltonian (\[oneHam\]). Variation of the Hamiltonian (\[mainHam\]) with respect to $\vec{A}$ gives the electric current operator: $\hat{J}_{\nu}=\sum_{\vec{p}} a^{\dagger}_{\alpha}(\vec{p}) (\hat{j}_{\nu})_{\alpha\beta} a_{\beta}(\vec{p})$, where the one-particle current operator reads (it is actually a velocity $\hat{j}_\nu=e\hat{v}_\nu$): $$\label{Jx} (\hat{j}_{\nu})_{\alpha\beta}(\vec{p})= e\left( v(p)\frac{p_{\nu}}{p}\delta_{\alpha\beta}+ \frac{d(p_\mu\alpha(p))}{dp_\nu} \epsilon^{zi\mu}\sigma^i_{\alpha\beta} \right),$$ with $\nu=x,y$ being the spatial index and $\epsilon^{zi\mu}$ is the 3D totally antisymmetric tensor. Under the non-uniform SU(2) electron spinor transformation: $a_\alpha(\vec{r})\mapsto U_{\alpha\beta}(\vec{r}) a_\beta(\vec{r})$, the Hamiltonian (\[mainHam\]) becomes dependent on the SU(2) “spin electromagnetic” vector potential $\hat{A}_\mu=A^0_\mu\sigma^0+A^i_\mu\sigma^i$, where $A^0_\mu$ coincides with the physical electromagnetic potential and $A^i_\mu= -i\,\textrm{Tr} (\sigma^iU^+\partial_\mu U)/2$. Although this latter potential is a pure gauge and has no physical consequences, variation of the Hamiltonian (\[mainHam\]) with respect to it defines the spin current of $i$-component of the spin along the direction $\mu$: $\hat{J}_{\mu}^i= \sum_{\vec{p}} a^{\dagger}_{\alpha}(\vec{p}) (\hat{j}_{\mu}^i)_{\alpha\beta} a_{\beta}(\vec{p})$, where the one-particle spin current operator reads: $$\label{Jyz} (\hat{j}_{\mu}^i)_{\alpha\beta}(\vec{p})= v(p)\frac{p_{\mu}}{p} \sigma^i_{\alpha\beta}+ \frac{d(p_\nu\alpha(p))}{dp_\mu} \epsilon^{zi\nu} \delta_{\alpha\beta}.$$ Our definition of the spin current (\[Jyz\]) coincides with the definition followed in Ref. : $\hat{J}_{\mu}^i= (\hat{v}_{\mu}\sigma^i+\sigma^i\hat{v}_{\mu})/2$, but differs from the definition followed in Refs. [@Murakami; @Sinova; @Molenkamp2] by a factor 2, which makes our value of the spin-Hall conductivity being twice as big as that in the literature. [@Murakami; @Sinova; @Molenkamp2] The interaction of electron with short-ranged non-magnetic impurities at positions $\vec{R}_i$, numerated by the index $i$, is described by the impurity Hamiltonian: $$\begin{aligned} \label{Himp} \hat{H}_{imp}=\sum_i\int u(\vec{r}-\vec{R}_i) a_\alpha^{\dagger}(\vec{r}) a_\alpha(\vec{r})d^2\vec{r},\end{aligned}$$ where $u(\vec{r})$ is a short-range impurity potential. We assume it to be sufficiently weak in order for the Born approximation to be valid. In this limit the impurity model (\[Himp\]) is equivalent to the model of the Gaussian random potential. We expand the electron Green function averaged over the realizations of the disorder potential purturbatively in power of the Hamiltonian (\[Himp\]) using the diagrammatic procedure [@AGD]. It is a sum of diagrams where a chain of electron bare Green functions is separated by impurity “crosses”. Two “crosses” are connected by the averaged impurity line: $n_{imp}u^2=1/2\pi\nu\tau$, where $n_{imp}$ is the density of impurities, and $\tau$ is the scattering mean free time. A “cross” does not change the electron spin and the electron frequency since the electron scattering off impurity is elastic. Therefore the impurity line carries zero frequency. Diagrams with crossings of two or more impurity lines are small as powers of the ratio $1/\epsilon_F \tau\ll 1$. The averaged Green function is a two by two matrix in the spin space and it is a solution of the Dyson equation: $$\label{Dyson} G_{\alpha\beta}^{-1}(\epsilon,\vec{p})- (\epsilon-\mu)\delta_{\alpha\beta}+ h_{\alpha\beta}(\vec{p})= -\frac{n_{imp}u^2}{V} \sum_{{\vec{p'}}} G_{\alpha\beta}(\epsilon,\vec{p'}).$$ It can be conveniently transformed into the chiral basis by the unitary matrix $U(\vec{p})$ (\[Unitary\]). The retarded and the advanced averaged Green functions are diagonal in the chiral basis: $G^{(R,A)}_{\lambda'\lambda}(\epsilon,\vec{p})= G^{(R,A)}_{\lambda}(\epsilon,\vec{p}) \delta_{\lambda'\lambda}$, and the solution to the Dyson Eq. (\[Dyson\]) reads [@AGD]: $$\label{Greenfunction} G^{R}_{\lambda}(\epsilon,\vec{p})= \frac{1}{\epsilon-\epsilon_\lambda(\vec{p}) +\mu + i/2\tau} \delta_{\lambda'\lambda},$$ where $\tau$ is explicitely independent of the chirality. The advanced Green function is a complex conjugate of the retarded one: $G^A_{\lambda}(\epsilon,\vec{p})= \{G^{R}_{\lambda}(\epsilon,\vec{p})\}^*$. The Green function in the spin basis is a non-diagonal two by two matrix: $$\label{GRA} G^{(R,A)}(\epsilon,\vec{p})=\left( \begin{array}{cc} G^{(R,A)}_{\uparrow\uparrow}(\epsilon,\vec{p}) & G^{(R,A)}_{\uparrow\downarrow}(\epsilon,\vec{p}) \\ \\ G^{(R,A)}_{\downarrow\uparrow}(\epsilon,\vec{p}) & G^{(R,A)}_{\downarrow\downarrow}(\epsilon,\vec{p}) \end{array} \right),$$ where (omitting for a moment the frequency and momentum notations): $$\begin{aligned} &&G^{(R,A)}_{\uparrow\uparrow}= G^{(R,A)}_{\downarrow\downarrow}= (G^{(R,A)}_+ + G^{(R,A)}_- )/2,\nonumber\\ &&G^{(R,A)}_{\uparrow\downarrow}= -ie^{-i\varphi_{\bf p}} (G^{(R,A)}_+ - G^{(R,A)}_- )/2, \nonumber\\ &&G^{(R,A)}_{\downarrow\uparrow}= ie^{i\varphi_{\bf p}} (G^{(R,A)}_+ - G^{(R,A)}_- )/2,\end{aligned}$$ with the chiral $G^{(R,A)}_{\pm}$ being defined in Eq. (\[Greenfunction\]). In order to calculate the current, induced in the electron system by an electric field, we use the Keldysh technique [@Keldysh]. Our result given by the Eq. (\[Nresponse\]) is well known but we derive it here for consistency. The averaged Keldysh Green function is a four by four matrix $\mathcal{G}(p,\epsilon)$ that can be conveniently factorized into a two by two Keldysh matrix whose elements are matrices in the spin space themselves: $$\begin{aligned} \label{G_Keldysh} &&\left( \begin{array}{cc} \mathcal{G}_{--} & \mathcal{G}_{-+} \\ \mathcal{G}_{+-} & \mathcal{G}_{++} \end{array} \right)= \left( \begin{array}{cc} 1-N(\epsilon) & -N(\epsilon)\\ 1-N(\epsilon) & -N(\epsilon) \end{array} \right) G^R(p,\epsilon)+ \nonumber\\ \nonumber \\ &&+\left(\begin{array}{cc} N(\epsilon) & N(\epsilon)\\ -1+N(\epsilon) &-1+N(\epsilon) \end{array} \right)G^A(p,\epsilon), \end{aligned}$$ where the electron distribution $N(\epsilon)$ is proportional to the unit matrix in the spin space. We choose the gauge for the uniform electric field $\vec{E}(t)=\vec{E}(\Omega) e^{-i\Omega t}$ to be a time dependent vector potential $\vec{A}(t)=\vec{A}(\Omega) e^{-i\Omega t}$, where $\vec{A}(\Omega)= -ic\vec{E}(\Omega)/\Omega$. Using the Keldysh technique [@Keldysh] we average the spin current operator over the electron state perturbed by the electromagnetic Hamiltonian: $\hat{H}_{em}= -\frac{1}{c} \int d^2\vec{r} \hat{j}_{\nu}(\vec{r}) A_{\nu}(t)$, in the first order of the perturbation theory. The spin-Hall conductivity $\sigma_{sH}$ is then found from the relationship: $\langle\hat{j}_{\mu}^z(\Omega)\rangle= \epsilon_{\mu\nu} \sigma_{sH}(\Omega) E_{\nu}(\Omega)$, as: $$\label{response_Keldysh} \sigma_{sH}=\frac{-1}{V\Omega} \sum_{\vec{p}} \int \frac{d\epsilon}{2\pi} \textrm{Tr} \left[\hat{j}_y^z(\vec{p}) \mathcal{G}(\epsilon+\Omega,\vec{p})\tau^z \hat{j}_x(\vec{p}) \mathcal{G}(\epsilon,\vec{p}) \right]_{-+}$$ where $\tau^z$ is the four by four matrix given by the direct product of the Pauli matrix $\sigma^z$ in the Keldysh space and the unit matrix in the spin space, the current operators in Eq. (\[response\_Keldysh\]) are the direct product of matrices (\[Jx\], \[Jyz\]) and the unitary matrix in the Keldysh space. $\textrm{Tr}$ in Eq. (\[response\_Keldysh\]) operates only in the spin space whereas the indices of $-+$ element corresponds to the Keldysh space. Substituting the Green function Eq. (\[G\_Keldysh\]), we obtain: $$\begin{aligned} \label{Nresponse} \sigma_{sH}(\Omega)= \frac{1}{V\Omega}\sum_{\vec{p}}\int\frac{d\epsilon}{2\pi}\Big\langle \textrm{Tr} \Big[ \hat{j}_y^z(\vec{p})\Big(N(\epsilon+\Omega) \nonumber\\ {}\times \left(G^R(\epsilon+\Omega,\vec{p})- G^A(\epsilon+\Omega,\vec{p})\right) \hat{j}_x(\vec{p})G^A(\epsilon,\vec{p}) \nonumber\\ {}+G^R(\epsilon+\Omega,\vec{p})\hat{j}_x(\vec{p}) N(\epsilon) \left(G^R(\epsilon,\vec{p})-G^A(\epsilon,\vec{p})\right) \Big)\Big] \Big\rangle,\end{aligned}$$ where the brackets indicate averaging over the disorder. In the non-crossing approximation the average in Eq. (\[Nresponse\]) is given by the sum of the one-loop and the ladder diagrams shown in Figs. \[fish\], \[ladder\]. ![\[fish\] The spin-Hall conductivity is given by a one-loop diagram.](spin1.eps){width="26.00000%"} First, we calculate the one-loop diagram Fig. \[fish\] and denote this part of the spin-Hall conductivity as $\sigma_{sH}^0$. It corresponds to the Eq. (\[Nresponse\]) with all Green functions being substituted by the averaged Green functions (\[Greenfunction\]). The second line in Eq. (\[Nresponse\]) contains the imaginary part of the Green function $G^R(\epsilon+\Omega,\vec{p})$ and the corresponding integral is convergent, therefore we change $\epsilon$ to $\epsilon-\Omega$ in this second line. At $T=0$ the Fermi-Dirac distribution function reads: $N(\epsilon)= \theta(-\epsilon)$. We take the integral over $\epsilon$ and in the zero-frequency limit $\Omega\rightarrow 0$ we find: $$\begin{aligned} \label{spH} \sigma^0_{sH}=-\frac{1}{V}\sum_{\vec{p}} \frac{v(p)}{2\pi p} \Big[\frac{1}{2p\alpha(p)}\left(S(\zeta)-S(\xi)\right)+ \nonumber\\ +\frac{1}{2\tau}\frac{\zeta\xi+1/4\tau^2}{(\zeta^2+1/4\tau^2) (\xi^2+1/4\tau^2)}\Big],\end{aligned}$$ where $S(x)=\arctan(x/2\tau)$, $\zeta=\epsilon(p)-p\alpha(p)-\mu$ and $\xi=\epsilon(p)+ p\alpha(p)-\mu$. In the large volume limit we substitute $\frac{1}{V}\sum_{\vec{p}}\rightarrow \int \frac{d^2\vec{p}}{(2\pi)^2}$, and then evaluate the integral over $\vec{p}$ in (\[spH\]) in the limit of large Fermi circle $\mu\gg \Delta, 1/\tau$. This procedure, known as the semiclassical approximation, expresses the momentum $p$ in terms of the quasiparticle energy $\xi$: $$\label{semi} \int\frac{d^2\vec{p}}{(2\pi)^2}\approx \frac{p}{2\pi v(p)} \int_{-\infty}^{\infty} \left(1+R(p) \xi\right) d\xi \int_{0}^{2\pi}\frac{d\varphi}{2\pi},$$ where $R(p)=p^{-1}v^{-1}(p)-m^{-1}(p)v^{-2}(p)$. The result reads: $$\label{loopResult} \sigma^{0}_{sH}=\frac{e}{4\pi}\left(1-\frac{1}{1+(\Delta\tau)^2}\right),$$ with the first and the second terms corresponding to the first and the second terms of Eq. (\[spH\]), respectively. An important observation is that the result (\[loopResult\]) coincides exactly with the result obtained from those terms in Eq. (\[Nresponse\]) that contains one retarded and one advanced Green functions. These two terms are proportional to $dN(\epsilon)/d\epsilon= -\delta(\epsilon)$ and all integrals are explicitely confined to the vicinity of the Fermi circle. Therefore the spin-Hall conductivity unlike the usual Hall conductivity is determined by the quasi-particles around the Fermi circle and not by the entire Fermi disk. The ladder diagrams shown in Fig. \[ladder\] represent the vertex corrections to the current. Additional impurity lines improve the convergence of the integral in Eq. (\[Nresponse\]). As a consequence, vertex corrections to the terms in Eq. (\[Nresponse\]) with the two advanced or with the two retarded Green functions vanish as $\textrm{max} (1/\tau,\Delta)/ \epsilon_F\ll 1$. Therefore we consider only the vertex corrections to the terms with one advanced and one retarded Green’s functions as it is shown in Fig. \[ladder\]. For these diagrams the semiclassical approximation (\[semi\]) is valid. ![\[ladder\] The vertex correction to the spin-Hall conductivity is given by a sum of non-crossing ladder diagrams.](spinn2.eps){width="30.00000%"} The sum of ladder diagrams with $n=1, 2,...$ impurity lines is given by the expression: $$\label{response_ladder} \sigma_{sH}^{\text{lad}}=-\int \frac{d^2\vec{p}}{(2\pi)^2}\textrm{Tr} \left[\tilde{J}_y^z G^R(0,\vec{p}) j_x(\vec{p}) G^A(0,\vec{p})\right],$$ where the sum of $n=1..\infty$ vertex corrections to the current $\hat{j}_y^z(\vec{p})$ (with at least one impurity line) is denoted by the matrix $\tilde{J}_y^z$. In the spin basis and for short-ranged impurity potentials it does not depend on the electron momentum $\vec{p}$ and satisfies the transfer matrix equation: $$\label{vertex_cor} \tilde{J}_y^z=\frac{1}{2\pi\tau\nu}\int\frac{d^2\vec{p}}{(2\pi)^2} G^A(0,\vec{p})\left[j_y^z(\vec{p})+ \tilde{J}_y^z\right]G^R(0,\vec{p}),$$ where the Green functions $G^{(R,A)}$ are given by Eq. (\[GRA\]). The “full” spin current operator with all vertex corrections included: $\hat{j}_y^z(\vec{p})+\tilde{J}_y^z$, is represented diagrammatically in Fig. \[vertex\]. In the equations for the current operators (\[Jx\], \[Jyz\]) we expand the electron velocity $v(p)=v(p_F)+ \xi/(v(p_F)m(p_F))$, where $m^{-1}(p)=dv(p)/dp$, to the first order in the deviation from the Fermi circle: $\xi/\mu$, small in the semiclassical approximation. We also expand the spin-orbital splitting: $p\alpha(p)= \alpha(p_F)\left(p_F+(1+(p_F/\alpha_F)d\alpha/dp_F)\xi/v_F\right)$, in the Green functions. We then evaluate Eq. (\[vertex\_cor\]) in the semiclassical approximation (\[semi\]) neglecting odd powers of $\xi$: $$\begin{aligned} \label{vc} &&\left(\tilde{J}_y^z\right)_{\uparrow\uparrow}= \left\{ \left[2+(\Delta\tau)^2\right] \left(\tilde{J}_y^z\right)_{\uparrow\uparrow} + (\Delta\tau)^2\left(\tilde{J}_y^z\right)_{\downarrow\downarrow} \right\}B, \nonumber\\ &&\left(\tilde{J}_y^z\right)_{\downarrow\downarrow}= \left\{ (\Delta\tau)^2 \left(\tilde{J}_y^z\right)_{\uparrow\uparrow}+ \left[2+(\Delta\tau)^2\right] \left(\tilde{J}_y^z\right)_{\downarrow\downarrow} \right\}B, \nonumber\\ &&\left(\tilde{J}_y^z\right)_{\uparrow\downarrow}= \left\{ -iv(p_F)\Delta\tau+ \left[2+(\Delta\tau)^2\right] \left(\tilde{J}_y^z\right)_{\uparrow\downarrow} \right\}B, \nonumber\\ &&\left(\tilde{J}_y^z\right)_{\downarrow\uparrow}= \left\{ iv(p_F)\Delta\tau+ \left[2+(\Delta\tau)^2\right] \left(\tilde{J}_y^z\right)_{\downarrow\uparrow} \right\}B,\end{aligned}$$ where $B=\frac12 \left[1+(\Delta \tau)^2\right]^{-1}$. From the first two lines of Eqs. (\[vc\]) we find: $(\tilde{J}_y^z)_{\uparrow\uparrow}= (\tilde{J}_y^z)_{\downarrow\downarrow}$, whereas from the last two lines we find: $(\tilde{J}_y^z)_{\downarrow\uparrow}=iv(p_F)/(\Delta\tau)$ and $(\tilde{J}_y^z)_{\uparrow\downarrow}= -iv(p_F)/(\Delta\tau)$. The integrand of Eq. (\[response\_ladder\]) does not depend on the diagonal elements of the matrix $\tilde{J}_y^z$ and therefore we set them to zero: $\tilde{J}_y^z= \sigma^y\ v(p_F)/\Delta\tau$. As it was expected the vertex corrections are proportional to the scattering rate. Integrating Eq. (\[response\_ladder\]) in the semiclassical approximation (\[semi\]) over $\xi$, we finally find the ladder part of the spin Hall conductivity (Fig. \[ladder\]): $$\label{spinHallladder} \sigma_{sH}^{\text{lad}}= \frac{e}{4\pi}\left(-1+\frac{1}{1+(\Delta\tau)^2}\right).$$ Remarkably all derivatives of $\alpha(p)$ and $v(p)$ over $p$ have canceled out. ![\[vertex\] The vertex of the spin current with the vertex corrections taken into account: $j_y^z(\vec{p})+\tilde{J}_y^z$.](vertex.eps){width="33.00000%"} The spin-Hall conductivity is the sum of Eqs. (\[loopResult\], \[spinHallladder\]) and is zero: $$\label{SHC} \sigma_{sH}=\sigma_{sH}^{0}+\sigma_{sH}^{\text{lad}}= 0.$$ It explicitely does not depend on the impurity scattering time $\tau$. But, we observe a discontinuity between the spin-Hall conductivity in the clean system $\sigma_{sH}= e/8\pi\hbar$ and the spin-Hall conductivity Eq. (\[SHC\]) in the presence of the infinitely small amount of non-magnetic scatterers. As it was shown in Ref. , this discontinuity is related to the dissipation in the system, which gives rise to the dissipative part in the spin-Hall conductivity $\sigma_{sH}^D=-e/8\pi\hbar$, which cancels the reactive part $\sigma_{sH}^R=e/8\pi\hbar$. In an analogous calculation, for the generalized model (\[oneHam\]), we find the magneto-electric effect - magnetization induced by the electric field: $$\label{magnetization} \langle \hat{S}^\mu\rangle=\epsilon^{\mu\nu}\frac{e\Delta\tau}{2\pi v_F}E_\nu,$$ where $\hat{S}^\mu=1/2\ \sum_{\vec{p}} a^{\dagger}_{\alpha}(\vec{p})\ \hat{\sigma}^\mu_{\alpha\beta}a_{\beta}(\vec{p})$ is the total spin operator and the total magnetization $\langle \hat{S}^y\rangle$ is in agreement with Ref. . One should notice that the steady in-plane magnetization (\[magnetization\]) is a consequence of the zero spin-Hall effect (\[SHC\]). Non-zero spin Hall conductivity would result in a non-steady in-plane magnetization. This follows from the evolution equation of the total spin of the system $\hat{S}_{\nu}$ and the commutation relationship: $$\label{totalspin} -i\frac{d}{dt}\hat{S}^\mu=[\hat{H},\hat{S}^{\mu}]= i\alpha(p)\frac{p}{v(p)}\hat{J}_{\mu}^z,$$ where $\mu=x,y$; $\hat{J}_{\mu}^z$ is the total spin current operator, defined in (\[Jyz\]). Remarkably, Eq. (\[totalspin\]) is valid for an electron system with any non-magnetic disorder and any electron-electron interaction. Moreover, valid is a more general equation for the local spin density evolution: $$\label{localspin} \frac{\partial}{\partial t} \hat{s}^\mu(\vec{r})+ \frac{1}{2}\frac{\partial}{\partial x_\nu} \hat{j}^\mu_\nu+ \alpha m\left(\epsilon^{\mu yl}\hat{j}_x^l- \epsilon^{\mu xk}\hat{j}_y^k\right)=0,$$ where $\mu=x,y,z$; in derivation of Eq. (\[localspin\]) we considered the standard model of parabolic spectrum. If the last term were zero then the above Eq. (\[localspin\]) would be the spin conservation equation. But in the Rashba metal neither component of the spin is conserved. To conclude, we have extended the result of Inoue et al. Ref.  $\sigma_{sH}=0$ for the case of the arbitrary electron dispersion, arbitrary strength of disorder and arbitrary momentum dependence of the Rashba velocity $\alpha(p)$. Our result agrees with Ref. . The author thanks M. V. Feigel’man and E. I. Rashba for illuminating discussions. I am grateful to A. V. Shytov for pointing out a mistake in my previous calculation. This research was supported by Dynasty Foundation, Landau Scholarship-Juelich, Program “Quantum Macrophysics” of Russian Academy of Sciences, and the RFBR grant \# 04-02-16348. [99]{} S. Murakami, N. Nagaosa, and S. C. Zhang, Science **301**, 1348 (2003); S. Murakami, N. Nagaosa, and S. C. Zhang, cond-mat/0310005. J. Sinova, D. Culcer, Q. Niu, N. A. Sinitsyn, T. Jungwirth, and A. H. MacDonald, Phys. Rev. Lett. **92**, 126603 (2004). J. I. Inoue, G. E. W. Bauer, and L. W. Molenkamp, Phys. Rev. B [**70**]{}, 041303(R) (2004). E. G. Mishchenko, A. V. Shytov, and B. I. Halperin, cond-mat/0406730. Al. Khaetskii, cond-mat/0408136. R. Raimondi and P. Schwab, cond-mat/0408233. E. I. Rashba, Sov. Phys. - Solid State [**2**]{}, 1109 (1960). E. I. Rashba, cond-mat/0311110. A. A. Abrikosov, L. P. Gorkov and I. E. Dzyaloshinski, [*Methods of Quantum Field Theory in Statistical Physics*]{} (Dover, New York, 1975). L. V. Keldysh, Sov. Phys. Zh. Eksp. Teor. Fiz. **47**, 1515 (1964) \[Sov. Phys. JETP **20**, 1018 (1965)\]. A.G.Aronov, Yu. B. Lyanda-Geller, Pisma ZhETF **50**, 398 (1989) V. M. Edelstein, Solid State Communications, Vol. **73**, No. 3, pp. 233-235, 1990.

ArXiv

--- author: - Dimitar Mekerov title: '$P$-connection on Riemannian almost product manifolds' --- \[section\] \[thm\][Lemma]{} \[thm\][Proposition]{} \[thm\][Corollary]{} \[thm\][Problem]{} \[section\] \[section\] [**Abstract**\ In the present work[^1], we introduce a linear connection (preserving the almost product structure and the Riemannian metric) on Riemannian almost product manifolds. This connection, called $P$-connection, is an analogue of the first canonical connection of Lichnerowicz in the Hermitian geometry and the $B$-connection in the geometry of the almost complex manifolds with Norden metric. Particularly, we consider the $P$-connection on a the class of manifolds with nonintegrable almost product structure.\ **Key words:** Riemannian manifold, Riemannian metric, almost product structure, linear connection, parallel torsion.\ **2000 Mathematics Subject Classification:** 53C15, 53C25, 53B05.]{} Introduction ============ In [@Gan-Gri-Mih2] a linear connection, called $B$-connection, is introduced on almost complex manifolds with Norden (or anti-Hermitian) metric. This connection (preserving the almost complex structure and the Norden metric) is an analogue of the first canonical connection of Lichnerowicz in the Hermitian geometry ([@Gau], [@Li], [@Ya]). In [@Mek3] the $B$-connection is considered on a class of almost complex manifolds with Norden metric and nonintegrable almost complex structure. This is the class ${\mathcal{W}}_3$ of the quasi-Kähler manifolds with Norden metric. In the present work, we introduce a linear connection (preserving the almost product structure and the Riemannian metric) on Riemannian almost product manifolds. This connection, called $P$-connection, is an analogue of the first canonical connection of Lichnerowicz in the Hermitian geometry and the $B$-connection in the geometry of the almost complex manifolds with Norden metric. Particularly, we consider the $P$-connection on the manifolds of the class ${\mathcal{W}}_3$ from the classification in [@Sta-Gri]. The systematic development of the theory of Riemannian almost product manifolds was started by K. Yano [@Ya]. In [@Nav] A. M. Naveira gives a classification of these manifolds with respect to the covariant differentiation of the almost product structure. Having in mind the results in [@Nav], M. Staikova and K. Gribachev give in [@Sta-Gri] a classification of the Riemannian almost product manifolds with zero trace of the almost product structure. Preliminaries ============= Let $(M,P,g)$ be a *Riemannian almost product manifold*, [i.e. ]{} a differentiable manifold $M$ with a tensor field $P$ of type $(1,1)$ and a Riemannian metric $g$ such that $$\label{Pg} P^2x=x,\quad g(Px,Py)=g(x,y)$$ for arbitrary $x$, $y$ of the algebra ${\mathfrak{X}}(M)$ of the smooth vector fields on $M$. Obviously $g(Px,y)=g(x,Py)$. Further $x,y,z,w$ will stand for arbitrary elements of ${\mathfrak{X}}(M)$. In this work we consider Riemannian almost product manifolds with ${{\rm tr}}{P}=0$. In this case $(M,P,g)$ is an even-dimensional manifold. The classification in [@Sta-Gri] of Riemannian almost product manifolds is made with respect to the tensor field $F$ of type (0,3), defined by $$\label{2} F(x,y,z)=g\left(\left(\nabla_x P\right)y,z\right),$$ where $\nabla$ is the Levi-Civita connection of $g$. The tensor $F$ has the following properties: $$\label{3} \begin{array}{l} F(x,y,z)=F(x,z,y)=-F(x,Py,Pz),\\[6pt] F(x,y,Pz)=-F(x,Py,z). \end{array}$$ The basic classes of the classification in [@Sta-Gri] are ${\mathcal{W}}_1$, ${\mathcal{W}}_2$ and ${\mathcal{W}}_3$. Their intersection is the class ${\mathcal{W}}_0$ of the *Riemannian $P$-manifolds*, determined by the condition $F(x,y,z)=0$ or equivalently ${\nabla}P=0$. In the classification there are include the classes ${\mathcal{W}}_1\oplus{\mathcal{W}}_2$, ${\mathcal{W}}_1\oplus{\mathcal{W}}_3$, ${\mathcal{W}}_2\oplus{\mathcal{W}}_3$ and the class ${\mathcal{W}}_1\oplus{\mathcal{W}}_2\oplus{\mathcal{W}}_3$ of all Riemannian almost product manifolds. In the present work we consider manifolds from the class ${\mathcal{W}}_3$. This class is determined by the condition $$\label{sigma} \mathop{{\mathfrak{S}}}_{x,y,z} F(x,y,z)=0,$$ where $\mathop{{\mathfrak{S}}}_{x,y,z}$ is the cyclic sum by $x, y, z$. This is the only class of the basic classes ${\mathcal{W}}_1$, ${\mathcal{W}}_2$ and ${\mathcal{W}}_3$, where each manifold (which is not Riemannian $P$-manifold) has a nonintegrable almost product structure $P$. This means that in ${\mathcal{W}}_3$ the Nijenhuis tensor $N$, determined by $$\label{4'} N(x,y)=\left(\nabla_x P\right)Py-\left(\nabla_{Px} P\right)y +\left(\nabla_y P\right)Px-\left(\nabla_{Py} P\right)x,$$ is non-zero. Further, manifolds of the class ${\mathcal{W}}_3$ we call *Riemannian ${\mathcal{W}}_3$-manifolds*. As it is known the curvature tensor field $R$ of a Riemannian manifold with metric $g$ is determined by $ R(x,y)z=\nabla_x \nabla_y z - \nabla_y \nabla_x z - \nabla_{[x,y]}z $ and the corresponding tensor field of type $(0,4)$ is defined as follows $ R(x,y,z,w)=g(R(x,y)z,w). $ Let $(M,P,g)$ be a Riemannian almost product manifold and $\{e_i\}$ be a basis of the tangent space $T_pM$ at a point $p\in M$. Let the components of the inverse matrix of $g$ with respect to $\{e_i\}$ be $g^{ij}$. If $\rho$ and $\tau$ are the Ricci tensor and the scalar curvature, then $\rho^*$ and $\tau^*$, defined by $\rho^*(y,z)=g^{ij}R(e_i,y,z,Pe_j)$ and $\tau^*=g^{ij}\rho^*(e_i,e_j)$, are called an *associated Ricci tensor* and an *associated scalar curvature*, respectively. We will use also the trace $\tau^{**}=g^{ij}g^{ks}R(e_i,e_k,Pe_s,Pe_j)$. The *square norm* of $\nabla P$ is defined by $$\label{5} {{\left\Vert\nabla P\right\Vert ^2}}=g^{ij}g^{ks}g\left(\left(\nabla_{e_i}P\right)e_k,\left(\nabla_{e_j}P\right)e_s\right).$$ Obviously ${{\left\Vert\nabla P\right\Vert ^2}}=0$ iff $(M,P,g)$ is a Riemannian $P$-manifold. In [@Mek1] it is proved that if $(M,P,g)$ is a Riemannian ${\mathcal{W}}_3$-manifold then $$\label{6} {{\left\Vert\nabla P\right\Vert ^2}}=-2g^{ij}g^{ks}g\left(\left(\nabla_{e_i}P\right)e_k,\left(\nabla_{e_s}P\right)e_j\right) =2\left(\tau-\tau^{**}\right).$$ A tensor $L$ of type (0,4) with the properties$$\begin{gathered} L(x,y,z,w)=-L(y,x,z,w)=-L(x,y,w,z), \label{2.11'}\medskip\\[4pt] \mathop{{\mathfrak{S}}} \limits_{x,y,z} L(x,y,z,w)=0 \quad \textit{(the first Bianchi identity)}\label{2.12}\medskip\end{gathered}$$ is called a *curvature-like tensor*. Moreover, if the curvature-like tensor $L$ has the property $$\label{2.13} L(x,y,Pz,Pw)=L(x,y,z,w),$$ we call it a *Riemannian $P$-tensor*. If the curvature tensor $R$ on a Riemannian ${\mathcal{W}}_3$-manifold $(M,P,g)$ is a Riemannian $P$-tensor, [i.e. ]{} $R(x,y,Pz,Pw)=R(x,y,z,w)$, then $\tau^{**}=\tau$. Therefore ${{\left\Vert\nabla P\right\Vert ^2}}=0$, [i.e. ]{}$(M,P,g)$ is a Riemannian $P$-manifold. $P$-connection ============== A linear connection ${\nabla}'$ on a Riemannian almost product manifold $(M,P,g)$ preserving $P$ and $g$, [i.e. ]{}${\nabla}'P={\nabla}'g=0$, is called a *natural connection* [@Ga-Mi]. The natural connection ${\nabla}'$ on a Riemannian almost product manifold $(M,P,g)$ determined by $$\label{2.1} {\nabla}'_x y=\nabla_x y -\frac{1}{2}\bigl(\nabla_x P\bigr)Py,$$ is called a *$P$-connection*. Let $T$ be a torsion tensor of the $P$-connection ${\nabla}'$ determined on $(M,P,g)$ by . Because of the symmetry of $\nabla$, from we have $T(x,y)=-\frac{1}{2}\{\bigl(\nabla_x P\bigr)Py- \bigl(\nabla_y P\bigr)Px\}$. Then, having in mind , we obtain $$T(x,y,z)=g(T(x,y),z)=-\frac{1}{2}\{F(x,Py,z)-F(y,Px,z)\}.$$ Hence and we have $$\label{2.3} \mathop{{\mathfrak{S}}} \limits_{x,y,z} T(x,y,Pz)=0.$$ Let $Q$ be the tensor field determined by $$\label{2.4} Q(y,z)=-\frac{1}{2}\bigl(\nabla_y P\bigr)Pz.$$ Having in mind , for the corresponding (0,3)-tensor field we have $$\label{2.5} Q(y,z,w)=-\frac{1}{2}F(y,Pz,w).$$ Because of the properties , implies $Q(y,z,w)=-Q(y,w,z)$. Let $R'$ be the curvature tensor of the $P$-connection ${\nabla}'$. Then, according to and we have [@Ko-No] $$\label{2.6} \begin{array}{l} R'(x,y,z,w)=R(x,y,z,w)+\left(\nabla_x Q\right)(y,z,w)-\left(\nabla_y Q\right)(x,z,w) \\[4pt] \phantom{R'(x,y,z,w)=} +Q\left(x,Q(y,z),w\right)-Q\left(y,Q(x,z),w\right). \end{array}$$ After a covariant differentiation of , a substitution in , a use of , , and some calculations, from we obtain $$\begin{array}{l} R'(x,y,z,w)=R(x,y,z,w)+\frac{1}{2}\left[\left(\nabla_x F\right)(y,z,Pw)-\left(\nabla_y F\right)(x,z,Pw)\right] \\[4pt] \phantom{R'(x,y,z,w)=} +\frac{1}{4}\left[g\bigl(\left(\nabla_y P\right)z,\left(\nabla_x P\right)w\bigr) -g\bigl(\left(\nabla_x P\right)z,\left(\nabla_y P\right)w\bigr)\right]. \end{array}$$ The last equality, having in mind the Ricci identity for Riemannian almost product manifolds $$\bigl(\nabla_x F\bigr)(y,z,w)-\bigl(\nabla_y F\bigr)(x,z,w)=R(x,y,Pz,w) - R(x,y,z,Pw),$$ implies $$\label{2.7} \begin{array}{l} R'(x,y,z,w)\\[4pt] =\frac{1}{4}\left\{2R(x,y,z,w)+2R(x,y,Pz,Pw)+K(x,y,z,w)\right\}, \end{array}$$ where $K$ is the tensor determined by $$\label{2.8} K(x,y,z,w)=-g\bigl(\left(\nabla_x P\right)z,\left(\nabla_y P\right)w\bigr) +g\bigl(\left(\nabla_y P\right)z,\left(\nabla_x P\right)w\bigr).$$ In this way, the following theorem is valid. \[thm2.2\] The curvature tensor $R'$ of the $P$-connection ${\nabla}'$ on a Riemannian almost product manifold $(M,P,g)$ has the form .$\Box$ From is follows immediately that the property and are valid for $R'$. Therefore, the property for $R'$ is a necessary and sufficient condition $R'$ to be a Riemannian $P$-tensor. Since $R$ satisfies , then from we obtain immediately the following \[thm2.3\] The curvature tensor $R'$ of the $P$-connection ${\nabla}'$ on a Riemannian ${\mathcal{W}}_3$-manifold $(M,P,g)$ is a Riemannian $P$-tensor iff $$\label{2.9} 2\mathop{{\mathfrak{S}}} \limits_{x,y,z} R(x,y,Pz,Pw)=-\mathop{{\mathfrak{S}}} \limits_{x,y,z} K(x,y,z,w).$$$\Box$ Let the following condition be valid for the Riemannian almost product manifold $(M,P,g)$: $$\label{2.10} \mathop{{\mathfrak{S}}} \limits_{x,y,z} R(x,y,Pz,Pw)=0.$$ We say that the condition characterizes a class ${\mathcal{L}}_2$ of the Riemannian almost product manifolds. The equality implies immediately the properties and for $P$. Then, according to and , we obtain the following \[thm2.4\] Let $(M,P,g)$ belongs to the class ${\mathcal{L}}_2$. Then the curvature tensor $R'$ of the $P$-connection ${\nabla}'$ is a Riemannian $P$-tensor iff the tensor $P$ determined by is a Riemannian $P$-tensor, too.$\Box$ Having in mind , the last theorem implies the following \[cor2.5\] Let the curvature tensor $R'$ of the $P$-connection ${\nabla}'$ be a Riemannian $P$-tensor on $(M,P,g)\in{\mathcal{L}}_2$. Then the tensor $H$, determined by $$\label{2.11} H(x,y,z,w)=R(x,y,z,w)+R(x,y,Pz,Pw)$$ is a Riemannian $P$-tensor, too.$\Box$ Curvature properties of the $P$-connection in ${\mathcal{W}}_3\cap{\mathcal{L}}_2$ ================================================================================== Let us consider the manifold $(M,P,g)\in{\mathcal{W}}_3\cap{\mathcal{L}}_2$ with a Riemannian $P$-tensor of curvature $R'$ of the $P$-connection ${\nabla}'$. Then, according to [Theorem \[thm2.4\]]{} and [Corollary \[cor2.5\]]{}, the tensors $K$ and $H$, determined by and , respectively, are also Riemannian $P$-tensors. Let $\rho'$ and $\rho(K)$ be the Ricci tensors for $R'$ and $K$, respectively. Then we obtain immediately from $$\label{3.1} \rho(y,z)+\rho^*(y,Pz)=2\rho'(y,z)-\frac{1}{2}\rho(K)(y,z).$$ From we have $$\label{3.2} \tau+\tau^{**}=2\tau'-\frac{1}{2}\tau(K),$$ where $\tau'$ and $\tau(K)$ are the scalar curvatures for $R'$ and $K$, respectively. It is known from [@Mek1], that ${{\left\Vert\nabla P\right\Vert ^2}}=2(\tau-\tau^{**})$. Then implies $$\label{3.3} \tau=\tau'-\frac{1}{4}\left(\tau(K)-{{\left\Vert\nabla P\right\Vert ^2}}\right).$$ From we obtain $$\rho(K)(y,z)=-g^{ij}g\bigl(\left(\nabla_{e_i} P\right)z,\left(\nabla_y P\right)e_j\bigr),$$ from where $$\tau(K)=g^{ij}g^{ks}g\bigl(\left(\nabla_{e_i} P\right)e_s,\left(\nabla_{e_k} P\right)e_j\bigr).$$ Hence, applying , we get $$\label{3.4} \tau(K)=\frac{1}{2}{{\left\Vert\nabla P\right\Vert ^2}}.$$ From and it follows $$\label{3.5} \tau=\tau'+\frac{1}{8}{{\left\Vert\nabla P\right\Vert ^2}}.$$ The equalities , , and implies the following \[prop3.1\] Let the curvature tensor $R'$ of the $P$-connection ${\nabla}'$ be a Riemannian $P$-tensor on $(M,P,g)\in{\mathcal{W}}_3\cap{\mathcal{L}}_2$. Then $$\label{3.6'} {{\left\Vert\nabla P\right\Vert ^2}}=-8(\tau'-\tau)=\frac{8}{3}(\tau'-\tau^{**})=2\tau(K).$$$\Box$ \[cor4.2\] Let the curvature tensor $R'$ of the $P$-connection ${\nabla}'$ be a Riemannian $P$-tensor on $(M,P,g)\in{\mathcal{W}}_3\cap{\mathcal{L}}_2$. Then the following assertions are equivalent: 1. $(M,P,g)$ a Riemannian $P$-manifold; 2. $\tau'=\tau$; 3. $\tau'=\tau^{**}$; 4. $\tau(K)=0$.$\Box$ Let the considered manifold with a Riemannian $P$-tensor of curvature $R'$ of the $P$-connection ${\nabla}'$ in ${\mathcal{W}}_3\cap{\mathcal{L}}_2$ be 4-dimensional. Since $H$ is a Riemannian $P$-tensor, then according to [@Sta], we have $$\label{3.6} H=\nu(H)(\pi_1-\pi_2)+\nu^*(H)\pi_3,$$ where $\nu(H)=\frac{\tau(H)}{8}$, $\nu^*(H)=\frac{\tau^*(H)}{8}$, $\tau(H)$ and $\tau^*(H)$ are the scalar curvature of $H$ and its associated one, and $$\begin{array}{l} \pi_1(x,y,z,w)=g(y,z)g(x,w)-g(x,z)g(y,w),\\[4pt] \pi_2(x,y,z,w)=g(y,Pz)g(x,Pw)-g(x,Pz)g(y,Pw),\\[4pt] \pi_3(x,y,z,w)=g(y,z)g(x,Pw)-g(x,z)g(y,Pw),\\[4pt] \phantom{\pi_2(x,y,z,w)=} +g(y,Pz)g(x,w)-g(x,Pz)g(y,w). \end{array}$$ From and we obtain $$\label{4.8} \tau(H)=\frac{4\tau'-\tau(K)}{2},\qquad \tau^*(H)=\frac{4\tau'^*-\tau^*(K)}{2},$$ where $\tau'^*$ and $\tau^*(K)$ are the associated scalar curvature to $\tau'$ and $\tau(K)$, respectively. We apply to and thus we obtain the following \[prop3.2\] Let the curvature tensor $R'$ of the $P$-connection ${\nabla}'$ be a Riemannian $P$-tensor on a 4-dimensional manifold $(M,P,g)\in{\mathcal{W}}_3\cap{\mathcal{L}}_2$. Then $$H=\frac{4\tau'-\tau(K)}{16}(\pi_1-\pi_2)+\frac{4\tau'^*-\tau^*(K)}{16}\pi_3.$$$\Box$ Let ${\mathcal{L}}_1$ is the subclass of ${\mathcal{L}}_2$ determined by $$\label{4.9} R(x,y,Pz,Pw)=R(x,y,z,w).$$ The equalities and imply the following \[prop3.3\] Let $(M,P,g)\in{\mathcal{W}}_3\cap{\mathcal{L}}_1$. Then $$R=R'-\frac{1}{4}K.$$$\Box$ \[cor3.4\] Let $(M,P,g)\in{\mathcal{W}}_3\cap{\mathcal{L}}_1$. Then $$\tau=\tau'-\frac{1}{4}\tau(K),\qquad \tau^*=\tau'^*-\frac{1}{4}\tau^*(K).$$$\Box$ \[cor3.5\] Let $(M,P,g)\in{\mathcal{W}}_3\cap{\mathcal{L}}_1$ and $\dim M=4$. Then $$\tau=\frac{1}{2}\tau(H),\qquad \tau^*=\frac{1}{2}\tau^*(H).$$$\Box$ [99]{} <span style="font-variant:small-caps;">Ganchev</span> G., K. <span style="font-variant:small-caps;">Gribachev</span>, V. <span style="font-variant:small-caps;">Mihova</span>. Publ. Inst. Math. (Beograd) (N.S.), **42**, 1987, 107–121. <span style="font-variant:small-caps;">Gauduchon</span> P. Bollettino U.M.I., **11**B, 1997, 257–288. <span style="font-variant:small-caps;">Lichnerowicz</span> A. Théorie globale des connexions et des goupes d’homotopie, Roma, Edizioni Cremonese, 1962. <span style="font-variant:small-caps;">Yano</span> K. Differential geometry on complex and almost complex spaces, Pure and Applied Math. vol. 49, New York, Pergamon Press Book, 1965. <span style="font-variant:small-caps;">Mekerov</span> D. C. R. Acad. Bulgare Sci., **61**, 2008, 1105–1110. <span style="font-variant:small-caps;">Staikova</span> M., K. <span style="font-variant:small-caps;">Gribachev</span>. Serdica Math. P., **18**, 1992, 150–161. <span style="font-variant:small-caps;">Naveira</span> A. Rend. Mat., **3**, 1983, 577–592. <span style="font-variant:small-caps;">Mekerov</span> D. P. Geom., **89**, 2008, No 1-2, 119–129. <span style="font-variant:small-caps;">Ganchev</span> G., V. <span style="font-variant:small-caps;">Mihova</span>. Ann. Univ. Sofia Fac. Math. Inform., **81**, 1987, 195–206. <span style="font-variant:small-caps;">Kobayshi</span> S., K. <span style="font-variant:small-caps;">Nomizu</span>. Foundations of differential geometry, vol. 1, 2, New York, Intersc. Publ., 1963, 1969. <span style="font-variant:small-caps;">Staikova</span> M. On the differential geometry of Riemannian almost product manifolds, Ph.D. Thesis, Sofia, 1993. *Dimitar Mekerov\ Department of Geometry\ Faculty of Mathematics and Informatics\ Paisii Hilendarski University of Plovdiv\ 236 Bulgaria Blvd.\ 4003 Plovdiv, Bulgaria\ e-mail: mircho@uni-plovdiv.bg* [^1]: Partially supported by project RS09-FMI-003 of the Scientific Research Fund, Paisii Hilendarski University of Plovdiv, Bulgaria

ArXiv

--- abstract: 'Recently, several data analytic techniques based on connection graph laplacian (CGL) ideas have appeared in the literature. At this point, the properties of these methods are starting to be understood in the setting where the data is observed without noise. We study the impact of additive noise on these methods, and show that they are remarkably robust. As a by-product of our analysis, we propose modifications of the standard algorithms that increase their robustness to noise. We illustrate our results in numerical simulations.' author: - 'Noureddine El Karoui [^1] and Hau-tieng Wu [^2]' bibliography: - 'research.bib' - 'noisyManifold.bib' date: 'May 18th, 2014' title: Connection graph Laplacian methods can be made robust to noise --- Introduction ============ In the last few years, several interesting variants of kernel-based spectral methods have arisen in the applied mathematics literature. These ideas appeared in connection with new types of data, where pairs of objects or measurements of interest have a relationship that is “blurred" by the action of a nuisance parameter. More specifically, we can find this type of data in a wide range of problems, for instance in the class averaging algorithm for the cryo-electron microscope (cryo-EM) problem [@singer_zhao_shkolnisky_hadani:2011; @Zhao_Singer:2013], in a modern light source imaging technique known as ptychography [@Marchesini_Tu_Wu:2014], in graph realization problems [@Cucuringu_Lipman_Singer:2012; @Cucuringu_Singer_Cowburn:2012], in vectored PageRank [@Chung_Zhao_Kempton:2013], in multi-channels image processing [@Batard_Sochen:2014], etc... Before we give further details about the cryo-EM problem, let us present the main building blocks of the methods we will study. They depend on the following three components: 1. an undirected graph ${\mathtt{G}}=({\mathtt{V}},{\mathtt{E}})$ which describes all observations. The observations are the vertices of the graph ${\mathtt{G}}$, denoted as $\{V_i\}_{i=1}^n$. 2. an [*affinity function*]{} $w:{\mathtt{E}}\to {\mathbb{R}}_+$, satisfying $w_{i,j}=w_{j,i}$, which describes how close two observations are ($i$ and $j$ index our observations). One common choice of $w_{i,j}=w(V_i,V_j)$ is of the form $w_{i,j}=\exp(-m(V_i,V_j)^2/\eps)$, where $m(x,y)$ is a metric measuring how far $x$ and $y$ are. 3. a [*connection function*]{} ${r}:{\mathtt{E}}\to \mathsf{G}$, where $\mathsf{G}$ is a Lie group, which describes how two samples are related. In its application to the cryo-EM problem, $r_{i,j}$’s can be thought of estimates of our nuisance parameters, which are orthogonal matrices. These three components form [*the connection graph*]{} associated with the data, which is denoted as $({\mathtt{G}}, w,{r})$. They can be either given to the data analyst or have to be estimated from the data, depending on the application. This fact leads to different connection graph models and their associated noise models. For example, in the cryo-EM problem, all components of the connection graph $({\mathtt{G}}, w,{r})$ are determined from the given projection images, where each vertex represents an image [@ElKaroui_Wu:2013 Appendix A]; in the ptychography problem [@Marchesini_Tu_Wu:2014], ${\mathtt{G}}$ is given by the experimenter, ${r}$ is established from the experimental setup, and $w$ is built up from the diffractive images collected in the experiment. Depending on applications, different metrics, deformations or connections among pairs of observations are considered, or estimated from the dataset, to present the local information among data (see, for example, [@Batard_Sochen:2012; @Boyer_Lipman_StClair_Puente_Patel_Funkhouser_Jernvall_Daubechies:2011; @Al-Aifari_Daubechies_Lipman:2013; @Collins_Zomorodian_Carlsson_Guibas:2004; @Wang_Huang_Guibas:2013; @Sun_Ovsjanikov_Guibas:2009; @Talmon_Cohen_Gannot_Coifman:2013; @Memoli_Sapiro:2005]). In this paper, since we focus on the connection graph Laplacian (CGL), we take the Lie Group $\mathsf{G}=O(k)$[^3], where $k\in{\mathbb{N}}$, and assume that ${r}$ satisfies ${r}_{i,j}={r}^{-1}_{j,i}$. Our primary focus in this paper is on $k=1$ and $k=2$. We now give more specifics about one of the problems motivating our investigation. **Cryo-EM problem** In the cryo-EM problem, the experimenter collects 2-dimensional projection images of a 3-dimensional macro-molecular object of interest, and the goal is to reconstruct the 3-dimensional geometric structure of the macro-molecular object from these projection images. Mathematically, the collected images $\mathcal{X}_{\text{cryoEM}}:=\{I_i\}_{i=1}^N\in {\mathbb{R}}^{m^2}$ can be modeled as the X-ray transform of the potential of the macro-molecular object of interest, denoted as $\psi:{\mathbb{R}}^3\to {\mathbb{R}}_+$. More precisely, in the setting that is usually studied, we have $I_i=X_\psi(R_i)$, where $R_i\in SO(3)$, $SO(3)$ is the 3-dimensional special orthogonal group, $X_\psi$ is the X-ray transform of $\psi$. The X-ray transform $X_\psi(R_i)$ is a function from $\mathbb{R}^2$ to $\mathbb{R}_+$ and hence can be treated by the data analyst as an image. We refer the reader to [@ElKaroui_Wu:2013 Appendix A] for precise mathematical details. (For the rest of the discussion, we write $R_i=[R_i^1\,\, R_i^2\,\, R_i^3]$ in the canonical basis, where $R_i^k$ are three dimensional unit vectors.) The experimental process produces data with high level of noise. Therefore, to solve this inverse problem, it is a common consensus to preprocess the images to increase the signal-to-noise ratio (SNR) before sending them to the cryo-EM data analytic pipeline. An efficient way to do so is to estimate the projection directions of these images, i.e $R_i^3$. This direction plays a particular role in the X-ray transform, which is different from the other two directions. If $R_i^3$’s were known, we would cluster the images according to these vectors and for instance take the mean of all properly rotationally aligned images to increase the SNR of the projection images - more on this below - in a cluster as a starting point for data-analysis. With these “improved" images, we can proceed to estimate $R_i$ for the $i$-th image by applying, for example, the common line algorithm [@Hadani_Singer:2011b], so that the 3-D image can be reconstructed by the inverse X-ray transform[@Epstein:2007]. We note that $R_i^3$ is a unit vector in $\mathbb{R}^3$ and hence lives on the standard sphere $S^2$. Conceptually, the problem is rendered difficult by the fact that the $X$-ray transform $X_\psi(R_i)$ is equivariant under the action of rotations that leave $R_i^3$ unchanged. In other words, if $r_\theta$ is an in-plane rotation, i.e a rotation that leaves $R_i^3$ unchanged but rotates $R_i^1$ and $R_i^2$ by an angle $\theta$, the image $X_\psi(r_\theta R_i)$ is $X_\psi(R_i)$ rotated by the angle $\theta$. In other words, $X_\psi(r_\theta R_i)=r_2(\theta)X_\psi(R_i)$, where $r_2(\theta)$ stands for the 2-dimensional rotation by the angle $\theta$. These in-plane rotations are clearly nuisance parameters if we want to evaluate the projection direction $R_i^3$. To measure the distance between $R_i^3$ and $R_j^3$, we hence use a rotationally invariant distance, i.e $d_{i,j}^2=\inf_{\theta \in [0,2\pi]}\norm{P_i-r_2(\theta)P_j}_2^2$. In other words, we look at the Euclidian distance between our two X-ray transforms/images after we have “aligned" them as best as possible. We now think of $R^3_i$’s - the vectors we would like to estimate - as elements of the manifold $S^2$, equipped with a metric $\mathsf{g}_\psi$, which depends on the macro-molecular object of interest. It turns out that the Vector Diffusion Maps algorithm (VDM), which is based on CGL and which we study in this paper, is effective in producing a good approximation of $\mathsf{g}_\psi$ from the local information $d_{i,j}$’s and the rotations we obtain by aligning the various X-ray transforms. This in turns imply better clustering of the $R_i^3$’s and improvement in the data-analytic pipeline for cryo-EM problems [@singer_zhao_shkolnisky_hadani:2011; @Zhao_Singer:2013]. The point of this paper is to understand how the CGL algorithms perform when the input data is corrupted by noise. The relationship between this method and the connection concept in differential geometry is the following: the projection images $P_i$ form a graph, and we can define the affinity and connection among a pair of images so that the topological structure of the 2-dimensional sphere $(S^2,\mathsf{g}_\psi)$ is encoded in the graph. This amounts to using the local geometric information derived from our data to estimate the global information - including topology - of $(S^2,\mathsf{g}_\psi)$. **Impact of noise on these problems** What is missing from these considerations and the current literature is an understanding of how noise impact the procedures which are currently used and have mathematical backing in the noise-free context. The aim of our paper is to shed light on the issue of the impact of noise on these interesting and practically useful procedures. We will be concerned in this paper with the impact of adding noise on the observations - collected for instance in the way described above. Note that additive noise may have impact in all three building blocks of the connection graph associated with the data. First, it might make the graph noisy. For example, in the cryo-EM problem, the standard algorithm builds up the graph from a given noisy data set $\{P_i\}_{i=1}^n=\{I_i+\xi_i\}_{i=1}^n$ - $\xi_i$ is our additive noise - using the nearest neighbors determined by a pre-assigned metric, that is, we add an edge between two vertices when they are close enough in that metric. Then, clearly, the existence of the noise $\xi_i$ will likely create a different nearest neighbor graph from the the one that would be built up from the (clean) projection images $\{I_i\}_{i=1}^n$. As we will see in this paper, in some applications, it might be beneficial to consider a complete graph instead of a nearest neighbor graph. The second noise source is how $w$ and ${r}$ are provided or determined from the samples. For example, in the cryo-EM problem, although $\{P_i\}$ are points located in a high dimensional Euclidean space, we determine the affinity and connection between two images by evaluating their rotationally invariant distance. It is clear that when $P_i$ is noisy, the derived affinity and connection will be noisy and likely quite different from the affinity and connection we would compute from the clean dataset $\{I_i\}_{i=1}^n$. On the other hand, in the ptychography problem, the connection is directly determined from the experimental setup, so that it is a noise-free even when our observations are corrupted by additive noise. In summary, corrupting the observations by additive noise might impact the following elements of the data analysis: 1. which scheme and metric we choose to construct the graph; 2. how we build up the affinity function; 3. how we build up the connection function. **More details on CGL methods**\ At a high-level, connection graph Laplacian (CGL) methods create a block matrix from the connection graph. The spectral properties of this matrix are then used to estimate properties of the intrinsic structure from which we posit the data is drawn from. This in turns lead to good estimation methods for, for instance, geodesic distance on the manifold, if the underlying intrinsic structure is a manifold. We refer the reader to Appendix \[Appendix:Section:CGL\] and references [@singer_wu:2012; @singer_wu:2013; @Bandeira_Singer_Spielman:2013; @Chung_Kempton:2013; @Chung_Zhao_Kempton:2013] for more information. Given a $n\times n$ matrix $W$, with scalar entries denoted by $w_{i,j}$ and a ${nk\times nk}$ block matrix $G$ with $k\times k$ block entries denoted by $G_{i,j}$, we define a $nk\times nk$ matrix $S$ with $(i,j)$-block entries $${S}_{i,j}=w_{i,j} G_{i,j}$$ and a $nk\times nk$ block diagonal matrix $D$ with $(i,i)$-block entries $${D}_{i,i}=\sum_{j\neq i} w_{i,j} \id_k,$$ which is assumed to be invertible. Let us call $$\begin{aligned} \label{definition:LWG} L(W,G):={D}^{-1} {S}\,\,\,\mbox{ and }\,\, L_0(W,G):=L(W\circ 1_{i\neq j}, G). \end{aligned}$$ In other words, $L_0(W,G)$ is the matrix $L(W,G)$ computed from the weight matrix $W$ where the diagonal weights have been replaced by $0$. Suppose we are given a connection graph $({\mathtt{G}},w,{r})$, and construct the $n\times n$ [*affinity matrix*]{} $W$ so that $w_{i,j}=w(i,j)$ and the [*connection matrix*]{} $G$, the ${nk\times nk}$ block matrix with $k\times k$ block entries $G_{i,j}={r}(i,j)$, [*the CGL associated with the connection graph $({\mathtt{G}},w,{r})$*]{} is defined as ${\id}_{nk}-L(W,G)$ and [*the modified CGL associated with the connection graph $({\mathtt{G}},w,{r})$*]{} is defined as ${\id}_{nk}-L_0(W,G)$. We note that under our assumptions on ${r}$, i.e $r_{i,j}=r_{j,i}^{-1}=r_{i,j}^*$ the connection matrix $G$ is Hermitian. We are interested in the large eigenvalues of $L(W,G)$ (or, equivalently, the small eigenvalues of the CGL ${\id}_{nk}-L(W,G)$), as well as the corresponding eigenvectors. In the case where the data is not corrupted by noise, the CGL’s asymptotic properties have been studied in [@singer_wu:2012; @singer_wu:2013], when the underlying intrinsic structure is a manifold. Its so-called synchronization properties have been studied in [@Bandeira_Singer_Spielman:2013; @Chung_Zhao_Kempton:2013]. The aim of our study is to understand the impact of additive noise on CGL algorithms. Two main results are Proposition \[prop:controlApproxdijNoisy\], which explains the effect of noise on the affinity, and Theorem \[thm:consistencyRotations\], which explains the effect of noise on the connection. These two results lead to suggestions for modifying the standard CGL algorithms: the methods are more robust when we use a complete graph than when we use a nearest-neighbor graph, the latter being commonly used in practice. One should also use the matrix $L_0(W,G)$ instead of ${L(W,G)}$ to make the method more robust to noise. After we suggest these modifications, our main result is Proposition \[prop:connectionGraphLaplacianApproxModifS\], which shows that even when the signal-to-noise-ratio (SNR) is very small, i.e going to 0 asymptotically, our modifications of the standard algorithm will approximately yield the same spectral results as if we had been working on the CGL matrix computed from noiseless data. We develop in Section 2 a theory for the impact of noise on CGL algorithms and show that our proposed modifications to the standard algorithms render them more robust to noise. We present in Section 3 some numerical results. **Notation:** Here is a set of notations we use repeatedly. ${\cal T}$ is a set of linear transforms. $\id_k$ stands for the $k\times k$ identity matrix. If $v\in\mathbb{R}^n$, $D(\{v\})$ is a $nk\times nk$ block diagonal matrix with the $(i,i)$-th block equal to $v_i \id_k$. We denote by $A\circ B$ the Hadamard, i.e entry-wise, product of the matrices $A$ and $B$. $\opnorm{M}$ is the largest singular value (a.k.a operator norm) of the matrix $M$. $\norm{M}_F$ is its Frobenius norm. Theory ====== Our aim in this section is to develop a theory that explains the behavior of CGL algorithms in the presence of noise. In particular, it will apply to algorithms of the cryo-EM type. We give in Subsection \[subsec:generalApproxResMatrices\] approximation results that apply to general CGL problems. In Subsection \[subsec:impactNoiseVDMProcedure\], we study in details the impact of noise on both the affinity and the connection used in the computation of the CGL when using the rotationally invariant distance (this is particularly relevant for the cryo-EM problem). We put these results together for a detailed study of CGL algorithms in Subsection \[subsec:consequencesForVDM\]. We also propose in Subsection \[subsec:consequencesForVDM\] modifications to the standard algorithms. General approximation results {#subsec:generalApproxResMatrices} ----------------------------- We first present a result that applies generally to CGL algorithms. \[lemma:approxBoundsLaplacian\] Suppose $W$ and $\widetilde{W}$ are $n\times n$ matrices, with scalar entries denoted by $w_{i,j}$ and $\tilde{w}_{i,j}$ and $G$ and $\widetilde{G}$ are $nd\times nd$ block matrices, with $d\times d$ blocks denoted by $G_{i,j}$ and $\tilde{G}_{i,j}$. Suppose that $$\sup_{i,j} |\tilde{w}_{i,j}-w_{i,j}|\leq \eps \;, \text{ and } \sup_{i,j} \norm{\widetilde{G}_{i,j}-G_{i,j}}_F \leq \eta\;,$$ where $\eps,\eta\geq 0$. Suppose furthermore that there exists $C>0$ such that $0\leq w_{i,j}\leq C$, $\sup_{i,j} \norm{G_{i,j}}_F\leq C$ and $\sup_{i,j} \norm{\widetilde{G}_{i,j}}_F\leq C$. Then, if $\inf_i \sum_{j\neq i} w_{i,j}/n>\gamma$ and $\gamma>\eps$, we have $$\opnorm{L(W,G)-L(\widetilde{W},\widetilde{G})}\leq \frac{1}{\gamma} C(\eta+\eps)+\frac{\eps}{\gamma (\gamma-\eps)} C^2\;.$$ The proof of this lemma is given in Appendix \[app:sec:ApproxResLaplMatrix\]. This lemma says that if we can approximate the matrix $W$ well entrywise and each of the individual matrices $G_{i,j}$ well, too, data analytic techniques working on the CGL matrix $L(\widetilde{W},\widetilde{G})$ will do essentially as well as those working on the corresponding matrix for $L(W,G)$ in the spectral sense. This result is useful because many methods rely on these connection graph ideas, with different input in terms of affinity and connection functions [@singer_zhao_shkolnisky_hadani:2011; @Zhao_Singer:2013; @Cucuringu_Lipman_Singer:2012; @Cucuringu_Singer_Cowburn:2012; @Boyer_Lipman_StClair_Puente_Patel_Funkhouser_Jernvall_Daubechies:2011; @Al-Aifari_Daubechies_Lipman:2013; @Chen_Lin_Chern:2013; @Collins_Zomorodian_Carlsson_Guibas:2004; @Sun_Ovsjanikov_Guibas:2009; @Talmon_Cohen_Gannot_Coifman:2013; @Memoli_Sapiro:2005]. However, it will often be the case that we can approximate $w_{i,j}$ - which we think of as measurements we would get if our signals were not corrupted by noise - only up to a constant. The following result shows that in certain situations, this will not affect dramatically the spectral properties of the CGL matrix. \[lemma:approxBoundsLaplacianMultiplicativeErrorAffinity\] We work under the same setup as in Lemma \[lemma:approxBoundsLaplacian\] and with the same notations. However, we now assume that $$\exists \{f_i\}_{i=1}^n\;, f_i>0 : \sup_{i,j} \left|\frac{\tilde{w}_{i,j}}{f_i}-w_{i,j}\right|\leq \eps \;, \text{ and } \sup_{i,j} \norm{\widetilde{G}_{i,j}-G_{i,j}}_F \leq \eta\;.$$ Suppose furthermore that there exists $C>0$ such that $0\leq w_{i,j}\leq C$, $\sup_{i,j} \norm{G_{i,j}}_F\leq C$ and $\sup_{i,j} \norm{\widetilde{G}_{i,j}}_F\leq C$. Then, if $\inf_i \sum_{j\neq i} w_{i,j}/n>\gamma$ and $\gamma>\eps$, we have $$\opnorm{L(W,G)-L(\widetilde{W},\widetilde{G})}\leq \frac{1}{\gamma} C(\eta+\eps)+\frac{\eps}{\gamma (\gamma-\eps)} C^2\;.$$ We note that quite remarkably, there are essentially no conditions on $f_i$’s: in particular, $\widetilde{w}_{i,j}$ and $w_{i,j}$ could be of completely different magnitudes. The previous lemma also shows that, for the purpose of understanding the large eigenvalues and eigenvectors of $L(W,G)$, we do not need to estimate $f_i$’s: we can simply use $L(\widetilde{W},\widetilde{G})$, i.e just work with the noisy data. Let us call $\widetilde{W_f}$ the matrix with scalar entries $\tilde{w}_{i,j}/f_i$. We note simply that $$L(\widetilde{W_f},\widetilde{G})=L(\widetilde{W},\widetilde{G})\;.$$ The assumptions of Lemma \[lemma:approxBoundsLaplacian\] apply to $(\widetilde{W_f},\widetilde{G})$ and hence we have $$\opnorm{L(W,G)-L(\widetilde{W_f},\widetilde{G})}\leq \frac{1}{\gamma} C(\eta+\eps)+\frac{\eps}{\gamma (\gamma-\eps)} C^2\;.$$ But since $L(\widetilde{W_f},\widetilde{G})=L(\widetilde{W},\widetilde{G})$, we also have $$\opnorm{L(W,G)-L(\widetilde{W},\widetilde{G})}\leq \frac{1}{\gamma} C(\eta+\eps)+\frac{\eps}{\gamma (\gamma-\eps)} C^2\;.$$ In some situations that will be of interest to us below, it is however, not the case that we can find $f_i$’s such that $$\exists \{f_i\}_{i=1}^n\;, f_i>0 :\, \sup_{i,j} \left|\frac{\tilde{w}_{i,j}}{f_i}-w_{i,j}\right|\leq \eps\;.$$ Rather, this approximation is possible only when $i\neq j$, yielding the condition $$\forall i, \exists f_i>0 :\, \sup_{i\neq j} \left|\frac{\tilde{w}_{i,j}}{f_i}-w_{i,j}\right|\leq \eps\;.$$ This apparently minor difference turns out to have significant consequences, both practical and theoretical. We propose in the following lemma to modify the standard way of the computing the CGL matrix to handle this more general case. \[lemma:approxCGLMatGeneralModif\] We work under the same setup as in Lemma \[lemma:approxBoundsLaplacian\] and with the same notations. We now assume that multiplicative approximations of the weights is possible only on the off-diagonal elements of our weight matrix: $$\exists \{f_i\}_{i=1}^n\;, f_i>0 : \sup_{i\neq j} \left|\frac{\tilde{w}_{i,j}}{f_i}-w_{i,j}\right|\leq \eps \;, \text{ and } \sup_{i,j} \norm{\widetilde{G}_{i,j}-G_{i,j}}_F \leq \eta\;.$$ Suppose furthermore that there exists $C>0$ such that $0\leq w_{i,j}\leq C$, $\sup_{i,j} \norm{G_{i,j}}_F\leq C$, and $\sup_{i,j} \norm{\widetilde{G}_{i,j}}_F\leq C$. Then, if $\inf_i \sum_{j\neq i} w_{i,j}/n>\gamma$ and $\gamma>\eps$, we have $$\opnorm{L_0(W,G)-L_0(\widetilde{W},\widetilde{G})}\leq \frac{1}{\gamma} C(\eta+\eps)+\frac{\eps}{\gamma (\gamma-\eps)} C^2\;,$$ and $$\opnorm{L(W,G)-L_0(\widetilde{W},\widetilde{G})}\leq \frac{1}{\gamma} C(\eta+\eps)+\frac{\eps}{\gamma (\gamma-\eps)} C^2+\frac{C^2}{n\gamma }\;.$$ **Comment:** Concretely, this lemma means that if we do not include the block diagonal terms in the computation of the CGL obtained from our “noisy data”, i.e $(\widetilde{W},\widetilde{G})$, we will get a matrix that is very close in spectral norm to the CGL computed from the “clean data”, i.e $(W,G)$. The significance of this result lies in the fact that recent work in applied mathematics has proposed to use the large eigenvalues and eigenvectors of $L(W,G)$ for various data analytic tasks, such as the estimation of local geodesic distances when the data is thought to be sampled from an unknown manifold. What our result shows is that even when $f_i$ are arbitrarily large, which we can think of as the situation where the signal to noise ratio in $\widetilde{W}$ is basically 0, working with $L_0(\widetilde{W},\widetilde{G})$ will allow us to harness the power of these recently developed tools. Naturally, working with $(\widetilde{W},\widetilde{G})$ is a much more realistic assumption than working with $(W,G)$ since we expect all our measurements to be somewhat noisy whereas results based on $(W,G)$ essentially assume that there is no measurement error in the dataset. Recall also that the computation of the $D$ matrix does not involve the diagonal weights. Therefore, $$L(W,G)=L_0(W,G)+D^{-1} \Delta(\{w_{i,i}\},G_{i,i})\;,$$ where $\Delta(\{w_{i,i}\},G_{i,i})$ is the block diagonal matrix with $(i,i)$ block diagonal $w_{i,i}G_{i,i}$, and $D^{-1} \Delta(\{w_{i,i}\},G_{i,i})$ is a block diagonal matrix with $(i,i)$ block diagonal $$\frac{w_{i,i}}{\sum_{j\neq i} w_{i,j}} G_{i,i}\;.$$ This implies that $$\opnorm{L(W,G)-L_0(W,G)}\leq \sup_i \frac{w_{i,i}}{\sum_{j\neq i} w_{i,j}} \opnorm{G_{i,i}}\;.$$ Our assumptions imply that $\sum_{j\neq i} w_{i,j}>\gamma n$, $\sup_{i}w_{i,i}\leq C$ and $\opnorm{G_{i,i}}\leq \norm{G_{i,i}}_F\leq C$. Hence, $$\opnorm{L(W,G)-L_0(W,G)}\leq \frac{C^2}{n\gamma}\;.$$ We still have $L(\widetilde{W_f},\widetilde{G})=L(\widetilde{W},\widetilde{G})$ and of course $$L_0(\widetilde{W_f},\widetilde{G})=L_0(\widetilde{W},\widetilde{G})\;.$$ Note further that the assumptions of Lemma \[lemma:approxBoundsLaplacianMultiplicativeErrorAffinity\] apply to the matrices $(W\circ 1_{i\neq j}, G)$ and $(\widetilde{W}\circ 1_{i\neq j},\widetilde{G})$. Indeed, the off-diagonal conditions on the weights are the assumptions we are working under. The diagonal conditions on the weights in Lemma \[lemma:approxBoundsLaplacianMultiplicativeErrorAffinity\] are trivially satisfied here since both $W\circ 1_{i\neq j}$ and $\widetilde{W}\circ 1_{i\neq j}$ have diagonal entries equal to 0. Hence, Lemma \[lemma:approxBoundsLaplacianMultiplicativeErrorAffinity\] gives $$\opnorm{L_0(W,G)-L_0(\widetilde{W},\widetilde{G})}\leq \frac{1}{\gamma} C(\eta+\eps)+\frac{\eps}{\gamma (\gamma-\eps)} C^2\;.$$ Using Weyl’s inequality (see [@bhatia97]) and our bound on $\opnorm{L(W,G)-L_0(W,G)}$, we therefore get $$\opnorm{L(W,G)-L_0(\widetilde{W},\widetilde{G})}\leq \frac{1}{\gamma} C(\eta+\eps)+\frac{\eps}{\gamma (\gamma-\eps)} C^2+\frac{C^2}{n\gamma }\;.$$ We now turn to the analysis of a specific algorithm, the class averaging algorithm in the cryo-EM problem, with a broadly accepted model of noise contamination to demonstrate the robustness of CGL-like algorithms. Impact of noise on the rotationally invariant distance {#subsec:impactNoiseVDMProcedure} ------------------------------------------------------ We assume that we observe noisy versions of the $k$-dimensional images/objects, $k\geq 2$, we are interested in. If the images in the - unobserved - clean dataset are called $\{{S}_i\}_{i=1}^n$, we observe $${I}_i={S}_i+ N_i\;.$$ Here $\{N_i\}_{i=1}^n$ are pure-noise images/objects. Naturally, after discretization, the images/objects we consider are just data vectors of dimension ${p}$ – we view ${S}_i$ and $N_i$ as vectors in $\mathbb{R}^{p}$. In other words, for a $k$-dim image, we sample $p$ points from the domain ${\mathbb{R}}^k$, which is denoted as $\mathfrak{X}:=\{\mathsf{x}_i\}_{i=1}^p\subset {\mathbb{R}}^k$ and called the [*sampling grid*]{}, and the image is discretized accordingly on these points. We also assume that the random variables $N_i$’s, $i=1,\ldots,n$, are independent. ### Distance measurement between pairs of images We start from a general definition. Take a set of linear transforms ${{{\cal T}}^{(k)}}\subset O(k)$. Consider the following measurement between two objects/images, $d_{ij}\geq 0$, with $$d_{ij}^2=\inf_{\mathtt{O}\in {{{\cal T}}^{(k)}}}\norm{{I}_i-\mathtt{O}\circ{I}_j}^2_2\;,$$ where $\circ$ means that the transform is acting on the pixels. For example, in the continuous setup where $I_j$ is replaced by $f_j \in L^2({\mathbb{R}}^k)$, given $\mathtt{O}\in SO(k)$, we have $\mathtt{O}\circ f_j(x):=f_j(\mathtt{O}^{-1} x)$ for all $x\in{\mathbb{R}}^k$. When ${{{\cal T}}^{(k)}}=SO(k)$, $d_{ij}$ is called [*the rotationally invariant distance (RID)*]{}. In the discrete setup of interest in this paper, we assume that $\mathfrak{X}=\mathtt{O}^{-1}\mathfrak{X}$ for all $\mathtt{O}\in {{{\cal T}}^{(k)}}$; that is, the linear transform is [*exact*]{} (with respect to the grid $\mathfrak{X}$), in that it maps the sampling grid onto itself. For concreteness, here is an example of sampling grid and associated exact linear transforms. Let $k=2$ and take the sampling grid to be the polar coordinates grid. Since we are in dimension 2, we pick $m$ rays of length $1$ at angles $2\pi k/m$, $k=0,\ldots,m-1$ and have $l$ equally spaced points on each of those rays. We consider $I_i$ to be the discretization of the function $f_i \in L^2(\mathbb{R}^2)$ which is compactly supported inside the unit disk, at the polar coordinate grid. The set ${{\cal T}}^{(2)}$ consisting of elements of $SO(2)$ with angles $\theta_k=2\pi\frac{k}{m}$, where $k=1,\ldots,m$, is thus exact and associated to the polar coordinate grid. The discretization and notation merit further discussion. As a linear transform of the domain ${\mathbb{R}}^k$, $\mathtt{O}\in {{{\cal T}}^{(k)}}$ can be represented by a $k\times k$ matrix. On the other hand, in the discretized setup we consider here, we can map ${{{\cal T}}^{(k)}}$ to a set ${{\cal T}}$ of $p\times p$ matrices ${\mathsf{O}}$ which acts on the discretized images ${I}_j$. These images are viewed as a set of $p$-dim vectors, denoted as ${I}_j^\vee$, and ${\mathsf{O}}$ acts on a “flattened” or “vectorized” (i.e 1-dimensional) version of the $k$-dimensional object of interest. Note that to each transform $\mathtt{O}$ there corresponds a unique $p\times p$ matrix ${\mathsf{O}}$. In the following, we will use $\mathtt{O}$ to denote the transform acting on the pixels, and use ${\mathsf{O}}$ to mean its companion matrix acting on the vectorized version of the object we are interested in. A simple but very important observation is that $$(\mathtt{O}\circ I_i)^\vee=O I_i^\vee\;.$$ In other words, we will have $\inf_{\mathtt{O}\in {{{\cal T}}^{(k)}}}\norm{{I}_i-\mathtt{O}\circ{I}_j}=\inf_{{\mathsf{O}}\in {{\cal T}}}\norm{{I}^\vee_i-{\mathsf{O}}{I}^\vee_j}$. To simplify the notation, when it is clear from the context, we will use ${I}_j$ to mean both the discretized object of interest and its vectorized version. In what follows, we assume that ${{{\cal T}}}$ always contains $\id_p$. We study the impact of noise on $d_{ij}$ through a uniform approximation argument. Let us call for ${\mathsf{O}}\in {{\cal T}}$, $$\begin{aligned} {d_{ij,\text{noisy}}}^2({\mathsf{O}}):=\norm{{I}_i^\vee-{\mathsf{O}}{I}_j^\vee}^2 \;, \text{ and } {d_{ij,\text{clean}}}^2({\mathsf{O}}):=\norm{{S}_i^\vee-{\mathsf{O}}{S}_j^\vee}^2\;.\end{aligned}$$ Essentially we will show that, when ${{\cal T}}$ contains only orthogonal matrices and is not “too large", $$\sup_{{\mathsf{O}}\in {{\cal T}}}\sup_{i\neq j} |{d_{ij,\text{noisy}}}^2({\mathsf{O}})-{d_{ij,\text{clean}}}^2({\mathsf{O}})-f(i,j)|=\lo_P(1)\;,$$ where $f(i,j)$ does not depend on ${\mathsf{O}}$. Our approximations will in fact be much more precise than this. But we will be able to conclude that in these circumstances, $$\sup_{i\neq j}\left|\inf_{{\mathsf{O}}\in {{\cal T}}}{d_{ij,\text{noisy}}}^2({\mathsf{O}})-\inf_{{\mathsf{O}}\in {{\cal T}}}{d_{ij,\text{clean}}}^2({\mathsf{O}})-f(i,j)\right|=\lo_P(1)\;.$$ We have the following theorem for any given set of transforms ${{\cal T}}$. \[thm:controlApproxUnifIndicesAndTransfo\] Suppose that for $1\leq i\leq n$, $N_i$ are independent, with $N_i^\vee\sim {\cal N}(0,\Sigma_i)$. Call $\mathsf{t_p}:=\sup_i \sup_{{\mathsf{O}}\in {{\cal T}}}\sqrt{\trace{({\mathsf{O}}\Sigma_i{\mathsf{O}}\trsp)^2}}$ and $\mathsf{s}_p:=\sup_{1\leq i \leq n}\sup_{{\mathsf{O}}\in {{\cal T}}}\sqrt{\opnorm{{\mathsf{O}}\Sigma_i{\mathsf{O}}\trsp}}$. Then, we have $$\begin{aligned} &\sup_{{\mathsf{O}}\in {{\cal T}}}\sup_{i\neq j} |{d_{ij,\text{noisy}}}^2({\mathsf{O}})-{d_{ij,\text{clean}}}^2({\mathsf{O}})-\trace{\Sigma_i+{\mathsf{O}}\Sigma_j{\mathsf{O}}\trsp}| \notag\\ &={\gO}_P\left( \sqrt{\log [\card{{{\cal T}}} n^2]} \Big(\mathsf{t_p}+\mathsf{s}_p \sup_{i,{\mathsf{O}}\in {{\cal T}}} \norm{{\mathsf{O}}{S}_i{\mathsf{O}}\trsp}\Big)+\log [\card{{{\cal T}}} n^2]\mathsf{s}_p^2\right)\;.\end{aligned}$$ We first note that $N_i^\vee-{\mathsf{O}}N_j^\vee\sim {\cal N}(0,\Sigma_i+{\mathsf{O}}\Sigma_j {\mathsf{O}}\trsp)$. Applying Lemma \[lemma:controlSupQuadFormsLaurentMassart\] to $\norm{N_i^\vee-{\mathsf{O}}N_j^\vee}^2$ with $Q_i=\id$, we get $$\begin{aligned} &\sup_{{\mathsf{O}}\in{{\cal T}}} \sup_{i\neq j} \left|\norm{N_i^\vee-{\mathsf{O}}N_j^\vee}^2-\trace{\Sigma_i+{\mathsf{O}}\Sigma_j{\mathsf{O}}\trsp}\right|=\\ &{\gO}_P\left(\sqrt{\log [\card{{{\cal T}}} n^2]} \sup_{i,j,{\mathsf{O}}} \left[\sqrt{\trace{(\Sigma_i+{\mathsf{O}}\Sigma_j {\mathsf{O}}\trsp)^2}}\right]+\log [\card{{{\cal T}}} n^2] \sup_{i,j,{\mathsf{O}}} \opnorm{(\Sigma_i+{\mathsf{O}}\Sigma_j {\mathsf{O}}\trsp)}\right)\;.\end{aligned}$$ Of course, using the fact that for positive semi-definite matrices, $(A+B)^2\preceq 2(A^2+B^2)$ in the positive-semidefinite order, we have $$\trace{(\Sigma_i+{\mathsf{O}}\Sigma_j {\mathsf{O}}\trsp)^2}\leq 2\trace{\Sigma_i^2+[{\mathsf{O}}\Sigma_j{\mathsf{O}}\trsp]^2} \text\;.$$ Hence, $$\begin{aligned} &\sup_{{\mathsf{O}}\in{{\cal T}}} \sup_{i\neq j} \left|\norm{N_i^\vee-{\mathsf{O}}N_j^\vee}^2-\trace{\Sigma_i+{\mathsf{O}}\Sigma_j{\mathsf{O}}\trsp}\right|=\\ &\quad{\gO}_P\left(\sqrt{\log [\card{{{\cal T}}} n^2]} \sup_{i} \sup_{{\mathsf{O}}\in{{\cal T}}}\sqrt{\trace{\left[{\mathsf{O}}\Sigma_i{\mathsf{O}}\trsp\right]^2}}+\log [\card{{{\cal T}}} n^2] \sup_i\sup_{{\mathsf{O}}\in{{\cal T}}} \opnorm{{\mathsf{O}}\Sigma_i{\mathsf{O}}\trsp}\right)\;.\end{aligned}$$ We also note that $$({S}_i^\vee-{\mathsf{O}}{S}_j^\vee)\trsp (N_i^\vee-{\mathsf{O}}N_j^\vee)\sim {\cal N}(0,\gamma_{i,j,{\mathsf{O}}}^2)\;,$$ where $$\gamma_{i,j,{\mathsf{O}}}^2=({S}_i^\vee-{\mathsf{O}}{S}_j^\vee)\trsp (\Sigma_i-{\mathsf{O}}\Sigma_j {\mathsf{O}}\trsp)({S}_i^\vee-{\mathsf{O}}{S}_j^\vee)\;.$$ We note that $$\gamma_{i,j,{\mathsf{O}}}^2\leq 2\left(\norm{{S}_i^\vee}^2+\norm{{\mathsf{O}}{S}_j^\vee{\mathsf{O}}\trsp}^2\right)(\opnorm{\Sigma_i}+\opnorm{{\mathsf{O}}\Sigma_j{\mathsf{O}}\trsp})\leq 8\sup_{i,{\mathsf{O}}\in {{\cal T}}} \opnorm{{\mathsf{O}}\Sigma_i{\mathsf{O}}\trsp} \sup_{i,{\mathsf{O}}\in {{\cal T}}} \norm{{\mathsf{O}}{S}_i^\vee{\mathsf{O}}\trsp}^2\;.$$ Recall also that it is well known that if $Z_1,\ldots, Z_N$ are ${\cal N}(0,\gamma_i^2)$ random variables, $$\sup_{1\leq k \leq N} |Z_k|={\gO}_P(\sqrt{\log N} \sup_k \gamma_k)\;.$$ This result can be obtained by a simple union bound argument. In our case, it means that $$\sup_{{\mathsf{O}}\in {{\cal T}}} \sup_{i\neq j}|({S}_i^\vee-{\mathsf{O}}{S}_j^\vee)\trsp (N_i^\vee-{\mathsf{O}}N_j^\vee)|={\gO}_P\left( \sqrt{\log [\card{{{\cal T}}} n^2]} \sup_{i,{\mathsf{O}}\in {{\cal T}}} \sqrt{\opnorm{{\mathsf{O}}\Sigma_i{\mathsf{O}}\trsp}} \sup_{i,{\mathsf{O}}\in {{\cal T}}} \norm{{\mathsf{O}}{S}_i^\vee{\mathsf{O}}\trsp} \right)\;.$$ In light of the previous theorem, we have the following proposition. \[prop:controlApproxdijNoisy\] Suppose that for all $1\leq i \leq n$ and ${\mathsf{O}}\in {{\cal T}}$, $\opnorm{{\mathsf{O}}\Sigma_i{\mathsf{O}}\trsp}\leq \sigma^2_p$, $\sqrt{\trace{[{\mathsf{O}}\Sigma_i{\mathsf{O}}\trsp]^2}/p}\leq s_p^2$, and $\norm{{\mathsf{O}}{S}_i^\vee{\mathsf{O}}\trsp}\leq K$, where $K$ is a constant independent of $p$. Then, $$\begin{aligned} \sup_{{\mathsf{O}}\in {{\cal T}}}\sup_{i\neq j} |{d_{ij,\text{noisy}}}^2({\mathsf{O}})-{d_{ij,\text{clean}}}^2({\mathsf{O}})-\trace{\Sigma_i+{\mathsf{O}}\Sigma_j{\mathsf{O}}\trsp}|={\gO}_P(u_{n,p})\;.\label{definition:unp}\end{aligned}$$ where $u_{n,p}:=\sqrt{\log [\card{{{\cal T}}} n^2]} (\sqrt{p}s^2_p+K\sigma_p)+\log [\card{{{\cal T}}} n^2]\sigma_p^2$.\ It follows that, if $\sqrt{\log [\card{{{\cal T}}} n^2]} \max(\sqrt{p}s^2_p,\sigma_p)\tendsto 0$, and ${{\cal T}}$ contains only orthogonal matrices, $$\begin{aligned} \sup_{{\mathsf{O}}\in {{\cal T}}}\sup_{i\neq j} |{d_{ij,\text{noisy}}}^2({\mathsf{O}})-{d_{ij,\text{clean}}}^2({\mathsf{O}})-\trace{\Sigma_i+\Sigma_j}|={\gO}_P(u_{n,p})\;=\lo_P(1)\;.\end{aligned}$$ Furthermore, in this case, $$\sup_{i\neq j}\left|d_{ij,\text{noisy}}^2-d_{ij,\text{clean}}^2-\trace{\Sigma_i+\Sigma_j}\right|=\lo_P(1)\;,$$ where $$\begin{aligned} d_{ij,\text{noisy}}^2:=\inf_{{\mathsf{O}}\in{{\cal T}}} \norm{{I}_i^\vee-{\mathsf{O}}{I}_j^\vee}^2\;,\quad d_{ij,\text{clean}}^2:=\inf_{{\mathsf{O}}\in{{\cal T}}} \norm{{S}_i^\vee-{\mathsf{O}}{S}_j^\vee}^2\;.\end{aligned}$$ In light of the previous proposition, the following set of assumptions is natural: #### Assumption G1 : $\forall i, {\mathsf{O}}\in {{\cal T}}$, $\opnorm{{\mathsf{O}}\Sigma_i{\mathsf{O}}\trsp}\leq \sigma^2_p$, $\sqrt{\trace{[{\mathsf{O}}\Sigma_i{\mathsf{O}}\trsp]^2}/p}\leq s_p^2$, and $\norm{{\mathsf{O}}{S}_i^\vee{\mathsf{O}}\trsp}\leq K$, where $K$ is a constant independent of $p$. Furthermore, $\sqrt{\log [\card{{{\cal T}}} n^2]} \max(\sqrt{p}s^2_p,\sigma_p)\tendsto 0$ and hence $u_{n,p}\tendsto 0$. We refer the reader to Proposition \[prop:cardinalSetOfDiscretizedRotations\] on page for a bound on $\card{{{\cal T}}}$ that is relevant to the class averaging algorithm in the cryo-EM problem. The first two statements are immediate consequences of Theorem \[thm:controlApproxUnifIndicesAndTransfo\]. For the second one, we use the fact that since ${\mathsf{O}}\in {{\cal T}}$ is orthogonal, $\trace{{\mathsf{O}}\Sigma_j{\mathsf{O}}\trsp}=\trace{\Sigma_j}$. Now, if $F$ and $G$ are two functions, we clearly have $|\inf F(x)-\inf G(x)|\leq \sup |G(x)-F(x)|$. Indeed, $\forall x$, $F(x)\leq G(x)+\sup |G(x)-F(x)|$. Hence, for all $x$, $$\inf_x F(x)\leq F(x)\leq G(x)+\sup |G(x)-F(x)|\;,$$ and we conclude by taking $\inf$ in the right-hand side. The inequality is proved similarly in the other direction. The results of Theorem \[thm:controlApproxUnifIndicesAndTransfo\] therefore show that $$\sup_{i\neq j}\left|d_{ij,\text{noisy}}^2-d_{ij,\text{clean}}^2-\trace{\Sigma_i+\Sigma_j}\right|={\gO}_P(\sqrt{\log [\card{{{\cal T}}} n^2]} (\sqrt{p}s^2_p+K\sigma_p)+\log [\card{{{\cal T}}} n^2]\sigma_p^2)$$ and we get the announced conclusions under our assumptions. We now present two examples to show that our assumptions are quite weak and prove that the algorithms we are studying can tolerate very large amount of noise. #### Magnitude of noise: First example Assume that $N_i^\vee\sim p^{-(1/4+\eps)}{\cal N}(0,\id_p)$, where $\eps>0$. In this case, $\norm{N_i^\vee}\sim p^{1/4-\eps}\gg \sup_i \norm{S_i^\vee}$ if $\eps<1/4$. In other words, the norm of the error vector is much larger than the norm of the signal vector. Indeed, asymptotically, the signal to noise ratio $\norm{S_i^\vee}/\norm{N_i^\vee}$ is 0. Furthermore, $\sigma_p=p^{-(1/4+\eps)}$ and $\sqrt{p}s_p^2=p^{-2\eps}$. Hence, if $\card{{{\cal T}}}={\gO}(p^{\gamma})$ for some $\gamma$, our conditions translate into $\sqrt{\log(np)} \max(p^{-(1/4+\eps)},p^{-2\eps})\tendsto 0$. This is of course satisfied provided $n$ is subexponential in $p$. See Proposition \[prop:cardinalSetOfDiscretizedRotations\] for a natural example of ${{\cal T}}$ whose cardinal is polynomial in $p$. #### Magnitude of noise: Second example We now consider the case where $\Sigma_i$ has one eigenvalue equal to $p^{-\eps}$ and all the others are equal to $p^{-(1/2+\eta)}$, $\eps, \eta>0$. In other words, the noise is much larger in one direction than in all the others. In this case, $\sigma_p^2=p^{-\eps}$ and $\trace{\Sigma_i^2}=p^{-2\eps}+(p-1)*p^{-(1+2\eta)}\leq p^{-2\eps}+p^{-2\eta}$. So if once again, $\card{{{\cal T}}}={\gO}(p^\gamma)$, our conditions translate into $\sqrt{\log(np)} \max(p^{-\eps}+p^{-\eta},p^{-\eps/2})\tendsto 0$. This example would also work if the number of eigenvalues equal to $p^{-\eps}$ were $\lo(p^{2\eps}/[\log (np)])$, provided $\sqrt{\log(np)}\max(p^{-\eta},p^{-\eps/2})\tendsto 0$. #### Comment on the conditions on the signal in Assumption G1 At first glance, it might look like the condition $\sup_{i,{\mathsf{O}}\in {{\cal T}}}\norm{{\mathsf{O}}S_i^\vee {\mathsf{O}}\trsp}\leq K$ is very restrictive due to the fact that, after discretization, $S_i$ has $p$ pixels. However, it is typically the case that if we start from a function in $L^2(\mathbb{R}^k)$, the discretized and vectorized image $S_i^\vee$ is normalized by the number of pixels $p$, so that $\norm{S_i^\vee}$ is roughly equal to the $L^2$-norm of the corresponding function. Hence, our condition $\sup_{i,{\mathsf{O}}\in {{\cal T}}}\norm{{\mathsf{O}}S_i^\vee {\mathsf{O}}\trsp}\leq K$ is very reasonable. ### The case of “exact rotations” We now focus on the most interesting case for our problem, namely the situation where $\mathtt{O}$ leaves our sampling grid invariant. We call ${{{\cal T}}^{(k)}}_{\text{exact}} \subset SO(k)$ the corresponding matrices $\mathtt{O}$ and ${{\cal T}}_{\text{exact}}$ the companion $p\times p$ matrices. We note that ${{{\cal T}}^{(k)}}_{\text{exact}}$ depends on $p$, but since this is evident, we do not index ${{{\cal T}}^{(k)}}_{\text{exact}}$ by $p$ to avoid cumbersome notations. From the standpoint of statistical applications, our focus in this paper is mostly on the case $k=1$ (which corresponds to “standard" kernel methods commonly used in statistical learning) and $k=2$. We show in Proposition \[prop:cardinalSetOfDiscretizedRotations\] that if ${\mathsf{O}}\in {{\cal T}}_{\text{exact}}$, ${\mathsf{O}}$ is an orthogonal $p\times p$ matrix. Furthermore, we show in Proposition \[prop:cardinalSetOfDiscretizedRotations\] that $\card{{{\cal T}}_{\text{exact}}}$ is polynomial in $p$. We therefore have the following proposition. \[prop:caseOfExactRotations\] Let $$\begin{aligned} d_{ij,\text{noisy}}^2:=\inf_{{\mathtt{O}}\in{{{\cal T}}^{(k)}}_\text{exact}} \norm{{I}_i-{\mathtt{O}}\circ {I}_j}^2\;,\quad d_{ij,\text{clean}}^2:=\inf_{{\mathtt{O}}\in{{{\cal T}}^{(k)}}_\text{exact}} \norm{{S}_i-{\mathtt{O}}\circ {S}_j}^2\;.\end{aligned}$$ Suppose $N_i$ are independent with $N_i^\vee\sim {\cal N}(0,\Sigma_i)$. When Assumption G1 holds with ${{\cal T}}_\text{exact}$ being the set of companion matrices of ${{{\cal T}}^{(k)}}_\text{exact}$, we have $$\sup_{i\neq j}\left|d_{ij,\text{noisy}}^2-d_{ij,\text{clean}}^2-\trace{\Sigma_i+\Sigma_j}\right|=\lo_P(1)\;,$$ and $$\sup_{{\mathtt{O}}\in {{{\cal T}}^{(k)}}_\text{exact} }\sup_{i\neq j} |{d_{ij,\text{noisy}}}^2({\mathtt{O}})-{d_{ij,\text{clean}}}^2({\mathtt{O}})-\trace{\Sigma_i+\Sigma_j}|={\gO}_P(u_{n,p})\;=\lo_P(1)\;.$$ ### On the transform ${\mathtt{O}}^*_{ij,\text{noisy}}$ We now use the notations $${d_{ij,\text{noisy}}}({\mathtt{O}})=\norm{I_i-{\mathtt{O}}\circ I_j}\;, \text{ and } {d_{ij,\text{clean}}}({\mathtt{O}})=\norm{S_i-{\mathtt{O}}\circ S_j}\;.$$ Naturally, the study of $$\label{eq:defOptimalTransform} {\mathtt{O}}^*_{ij,\text{noisy}}=\argmin_{{\mathtt{O}}\in {{{\cal T}}^{(k)}}_\text{exact}} {d_{ij,\text{noisy}}}({\mathtt{O}})$$ is more complicated than the study of $\inf_{{\mathtt{O}}\in {{{\cal T}}^{(k)}}_\text{exact}} {d_{ij,\text{noisy}}}({\mathtt{O}})$. We will assume that the clean images are nicely behaved when it comes to the ${d_{ij,\text{clean}}}({\mathtt{O}})$ minimization, in that rotations that are near minimizers of ${d_{ij,\text{clean}}}({\mathtt{O}})$ are close to one another. More formally, we assume the following. #### Assumption A0 : ${{{\cal T}}^{(k)}}_\text{exact}$ is a subset of $SO(k)$ and contains only exact rotations. Call ${\mathtt{O}}^*_{ij,\text{clean}}:=\argmin_{{\mathtt{O}}\in {{{\cal T}}^{(k)}}_\text{exact}} {d_{ij,\text{clean}}}^2({\mathtt{O}})$ and call ${{{\cal T}}^{(k)}}_{ij,\eps}:=\left\{{\mathtt{O}}\in {{{\cal T}}^{(k)}}_\text{exact}: \, {d_{ij,\text{clean}}}^2({\mathtt{O}})\leq {d_{ij,\text{clean}}}^2({\mathtt{O}}^*_{ij,\text{clean}})+\eps\right\}$. We assume that $$\exists\delta_{ij,p}>0: \, \forall \eps<\delta_{ij,p}\, \forall {\mathtt{O}}\in {{{\cal T}}^{(k)}}_{ij,\eps}\,,\; d({\mathtt{O}},{\mathtt{O}}^*_{ij,\text{clean}})\leq g_{ij,p}(\eps)\;,$$ for $d$ the canonical metric on the orthogonal group and some positive $g_{ij,p}(\eps)$. #### Assumption A1 : $\delta_{ij,p}$ can be chosen independently of $i,j$ and $p$. Furthermore, there exists a function $g$ such that $g(\eps)\tendsto 0$ as $\eps\tendsto 0$ and $g_{ij,p}(x)\leq g(x)$, if $x\leq \delta_{ij,p}\leq \delta$. We discuss the meaning of these assumptions after the statement and proof of the following theorem. \[thm:consistencyRotations\] Suppose that the assumptions underlying Theorem \[thm:controlApproxUnifIndicesAndTransfo\] hold and that Assumptions G1, A0 and A1 hold. Suppose further that ${{{\cal T}}^{(k)}}_\text{exact}$ is the set of exact rotations for our discretization. Then, for any $\eta$ given, where $0<\eta<1$, as $p$ and $n$ go to infinity, $$\label{eq:controlDistanceNoisyOptRotAndCleanOptRot} \sup_{i\neq j}d({\mathtt{O}}^*_{ij,\text{noisy}},{\mathtt{O}}^*_{ij,\text{clean}})={\gO}_P(g(u^{1-\eta}_{n,p}))\;,$$ where $u_{n,p}$ is defined in (\[definition:unp\]). (Under Assumption G1, $u_{n,p}\tendsto 0$ as $n$ and $p$ tend to infinity.) The informal meaning of this theorem is that under regularity assumptions on the set of clean images, the optimal rotation computed from the set of noisy images is close to the optimal rotation computed from the set of clean images. In other words, this step of the CGL procedure is robust to noise. Clearly, ${\mathtt{O}}^*_{ij,clean}$ is a minimizer of $L_{ij}({\mathtt{O}}):={d_{ij,\text{clean}}}^2({\mathtt{O}})+\trace{\Sigma_i+\Sigma_j}$, since the second term does not depend on ${\mathtt{O}}$. Naturally, if Assumptions A0 and A1 apply to ${d_{ij,\text{clean}}}^2({\mathtt{O}})$, they apply to ${d_{ij,\text{clean}}}^2({\mathtt{O}})+C$, for $C$ any constant. In particular, taking $C=\trace{\Sigma_i+\Sigma_j}$, we see that Assumptions A0 and A1 apply to the function $L_{ij}({\mathtt{O}})$. The approximation results of Proposition \[prop:caseOfExactRotations\] guarantee that, under Assumptions G1, A0 and A1, ${\mathtt{O}}^*_{ij,\text{noisy}}$ is a near minimizer of ${d_{ij,\text{clean}}}^2({\mathtt{O}})$. Indeed, we have by definition, $$\label{eq:almostTautologyTransformNoisy} {d_{ij,\text{noisy}}}^2({\mathtt{O}}^*_{ij,\text{noisy}})\leq {d_{ij,\text{noisy}}}^2({\mathtt{O}}^*_{ij,\text{clean}}).$$ But under assumption G1, Proposition \[prop:caseOfExactRotations\] and the fact that the elements of ${{\cal T}}_\text{exact}$ are orthogonal matrices imply that $$\begin{aligned} \label{eq:proof:22:dijLij} \forall {\mathtt{O}}\in {{\cal T}}, \forall i \neq j \; \; {d_{ij,\text{noisy}}}^2({\mathtt{O}})=L_{ij}({\mathtt{O}})+{\gO}_P(u_{n,p})\;.\end{aligned}$$ Hence, we can rephrase Equation as $$\begin{aligned} \label{eq:proof:22:LijLij} L_{ij}({\mathtt{O}}^*_{ij,\text{noisy}})\leq L_{ij}({\mathtt{O}}^*_{ij,\text{clean}})+{\gO}_P(u_{n,p})\;.\end{aligned}$$ Indeed, by plugging (\[eq:proof:22:dijLij\]) into (\[eq:almostTautologyTransformNoisy\]), we have $$\forall i\neq j, \; {d_{ij,\text{clean}}}^2({\mathtt{O}}^*_{ij,\text{noisy}})+\trace{\Sigma_i+\Sigma_j}\leq {d_{ij,\text{clean}}}^2({\mathtt{O}}^*_{ij,\text{clean}})+\trace{\Sigma_i+\Sigma_j}+{\gO}_P(u_{n,p})\;.$$ Now, by definition of ${d_{ij,\text{clean}}}^2$, we have $$\begin{aligned} \label{eq:proof:22:dijcleandijnoisy} {d_{ij,\text{clean}}}^2({\mathtt{O}}^*_{ij,\text{clean}})\leq {d_{ij,\text{clean}}}^2({\mathtt{O}}^*_{ij,\text{noisy}})\;.\end{aligned}$$ So by (\[eq:proof:22:LijLij\]) and (\[eq:proof:22:dijcleandijnoisy\]), we have shown that $$\begin{aligned} \forall i\neq j, \; {d_{ij,\text{clean}}}^2({\mathtt{O}}^*_{ij,\text{clean}})+\trace{\Sigma_i+\Sigma_j}&\leq {d_{ij,\text{clean}}}^2({\mathtt{O}}^*_{ij,\text{noisy}})+\trace{\Sigma_i+\Sigma_j} \;,\\ &\leq {d_{ij,\text{clean}}}^2({\mathtt{O}}^*_{ij,\text{clean}})+\trace{\Sigma_i+\Sigma_j}+{\gO}_P(u_{n,p})\;.\end{aligned}$$ This clearly implies that $$\begin{aligned} \forall i\neq j, \; {d_{ij,\text{clean}}}^2({\mathtt{O}}^*_{ij,\text{clean}})&\leq {d_{ij,\text{clean}}}^2({\mathtt{O}}^*_{ij,\text{noisy}})\leq {d_{ij,\text{clean}}}^2({\mathtt{O}}^*_{ij,\text{clean}})+{\gO}_P(u_{n,p})\;.\end{aligned}$$ Since $u_{n,p}\tendsto 0$ as $n$ and $p$ grow, this means that, for any given $\eta$, with $0<\eta<1$, with very high probability, $$\forall 1\leq i \neq j \leq n \;, {\mathtt{O}}^*_{ij,\text{noisy}}\in {{{\cal T}}^{(k)}}_{ij,u_{n,p}^{1-\eta}}\;.$$ We conclude, using Assumption A0, that with very high-probability, $$\forall 1\leq i \neq j \leq n \;, d({\mathtt{O}}^*_{ij,\text{noisy}},{\mathtt{O}}^*_{ij,\text{clean}})\leq g(u_{n,p}^{1-\eta})\;.$$ #### Interpretation of Assumptions A0-A1 Assumption A0 guarantees that all near minimizers of ${d_{ij,\text{clean}}}({\mathtt{O}})$ are close to one another and hence the optimum. Our uniform bounds in Proposition \[prop:caseOfExactRotations\] only guarantee that ${\mathtt{O}}^*_{ij,\text{noisy}}$ is a near minimizer of ${d_{ij,\text{clean}}}({\mathtt{O}})$ and nothing more. If ${d_{ij,\text{clean}}}({\mathtt{O}})$ had near minimizers that were far from the optimum ${\mathtt{O}}^*_{ij,\text{clean}}$, it could very well happen that ${\mathtt{O}}^*_{ij,\text{noisy}}$ end up being close to one of these near minimizers but far from ${\mathtt{O}}^*_{ij,\text{clean}}$, and we would not have the consistency result of Theorem \[thm:consistencyRotations\]. Hence, the robustness to noise of this part of the CGL algorithm is clearly tied to some regularity or “niceness” property for the set of clean images. In the cryo-EM problem, these assumptions reflect a fundamental property of a manifold dataset – [*its condition number*]{} [@Niyogi_Smale_Weinberger:2009]. Conceptually, the condition number reflects “how difficult it is to reconstruct the manifold” from a finite sample of points from the manifold. Precisely, it is the inverse of the reach of the manifold, which is defined to be the radius of the smallest normal bundle that is homotopic to the manifold. This also highlights the fact that even if we were to run the CGL algorithm on the clean dataset, without these assumptions, the results might not be stable and reliable since intrinsically distant points (i.e distant in the geodesic distance) might be identified as neighbors. #### About ${{{\cal T}}^{(k)}}_\text{exact}$ and extensions We are chiefly interested in this paper about 2-dimensional images and hence about the case $k=2$ (see the cryoEM example). It is then clear that when our polar coordinate grid is fine, ${{{\cal T}}^{(k)}}_\text{exact}$ is also a fine discretization of $SO(2)$ and contains many elements. (More details are given in Subsection \[subsec:cardExactRotations\].) The situation is more intricate when $k\geq 3$, but since it is a bit tangential to the main purpose of the current paper, we do not discuss it further here. We refer the interested reader to Subsection \[subsec:cardExactRotations\] for more details about the case $k\geq 3$. We also note that our arguments are not tied to using a standard polar coordinate grid for the discretization of the images. [For another sampling grid, we would possibly get another ${{{\cal T}}^{(k)}}_\text{exact}$]{}. Our arguments go through when : a) if ${\mathtt{O}}\in {{{\cal T}}^{(k)}}$, the operation ${\mathtt{O}}\circ $ maps our sampling grid of points onto itself; b) $\card{{{{\cal T}}^{(k)}}}$ grows polynomially in $p$. ### Extensions and different approaches At the gist of our arguments are strong concentration results for quadratic forms in Gaussian random variables. Naturally, our results extend to other types of random variables for which these concentration properties hold. We refer to [@ledoux2001] and [@nekCorrEllipD] for examples. A natural example in our context would be a situation where $N_i=\Sigma_i^{1/2} X_i$, and $X_i$ has i.i.d uniformly bounded entries. This is particularly relevant in the case where $\Sigma_i$ is diagonal for instance - the interpretation being then that the noise contamination is through the corruption of each individual pixel by independent random variables with possibly different standard deviations. The arguments in Lemma \[lemma:controlSupQuadForms\] handle this case, though the bound is slightly worse than the one in Lemma \[lemma:controlSupQuadFormsLaurentMassart\] when a few eigenvalues of $\Sigma_i$ are larger than most of the others. Indeed, the only thing that matters in this more general analysis is the largest eigenvalue of $\Sigma_i$, so that in the notation of Assumption **G1**, $\sqrt{p}s_p^2$ is replaced by $\sqrt{p}\sigma_p^2$. Hence, our approximation will require in this more general setting that $\sigma_p=\lo(p^{-1/4})$, whereas we have seen in the Gaussian case that we can tolerate a much larger largest eigenvalue. We also note that we could of course settle for weaker results on concentration of quadratic forms, which would apply to more distributions. For instance, using bounds on $\Exp{|\norm{N_i}^2-\Exp{\norm{N_i}^2}|^{k}}$ would change the dependence of results such as Proposition \[prop:controlApproxdijNoisy\] on $\card{{{\cal T}}}n^2$ from powers of logarithm to powers of $1/k$. This is in turn would mean that our results would become tolerant to lower levels of noise but apply to more noise distributions. Consequences for CGL algorithm and other kernel-based methods {#subsec:consequencesForVDM} ------------------------------------------------------------- ### Reminders and preliminaries Recall that in CGL methods performed with the rotationally invariance distance - henceforth RID - induced by $SO(k)$, we mostly care about the spectral properties - especially large eigenvalues and corresponding eigenvectors - of the CGL matrix $L(\widetilde{W},\widetilde{G})$, where $\widetilde{W}$ is a $n\times n$ matrix and $\widetilde{G}$ is a $nk\times nk$ block-matrix with $k\times k$ blocks defined through $$\widetilde{W}_{i,j}=\exp(-{d_{ij,\text{noisy}}}^2/\eps),\quad \widetilde{G}_{i,j}=\mathtt{O}^*_{ij,\text{noisy}}\;,$$ where $\mathtt{O}^*_{ij,\text{noisy}}$ is defined in Equation . The “good” properties of CGL stem from the fact that the matrix $L(W,G)$, the CGL matrix associated with the clean images, has “good" spectral properties. For example, when a manifold structure is assumed, the theoretical work of [@singer_wu:2012; @singer_wu:2013] relates the properties of $L(W,G)$ - the matrix obtained in the same manner as above when we replace ${d_{ij,\text{noisy}}}$ by ${d_{ij,\text{clean}}}$ and $\mathtt{O}^*_{ij,\text{noisy}}$ by $\mathtt{O}^*_{ij,\text{clean}}$ - to the geometric and topological properties of the manifold from which the data is sampled. The natural approximate “sparsity” of the spectrum of this kind of matrices is discussed in Section \[Appendix:Section:CGL\]. In practice, the data analyst has to work with $L(\widetilde{W},\widetilde{G})$. Hence, it could potentially be the case that $L(\widetilde{W},\widetilde{G})$ does not share many of the good properties of $L(W,G)$. Indeed, we explain below that this is in general the case and propose a modification to the standard algorithm to make the results of CGL methods more robust to noise. All these arguments suggest that it is natural to study the properties of the standard CGL algorithm applied to noisy data. We mention that CGL algorithms may apply beyond the case of the rotational invariance distance and $O(k)$ and we explain in Subsubsection \[subsubsec:CGLbeyondSOk\] how our results apply in this more general context. ### Modified CGL algorithm and rotationally invariant distance We now show that our modification to the standard algorithm is robust to noise. More precisely, we show that the modified CGL matrix $L_0(\widetilde{W},\widetilde{G})$ is spectrally close to the CGL matrix computed from the noise-free data, $L(W,G)$. \[prop:connectionGraphLaplacianApproxModifS\] Consider the modified CGL matrix $L_0(\widetilde{W},\widetilde{G})$ computed from the noisy data and the CGL matrix $L(W,G)$ computed from the noise-free data. Under Assumptions **G1** and **A0**-**A1**, we have, if $\trace{\Sigma_i}=\trace{\Sigma_j}=\trace{\Sigma}$ for all $(i,j)$, $$\opnorm{L_0(\widetilde{W},\widetilde{G})-L(W,G)}=\lo_P(1)\;,$$ provided there exists $\gamma>0$, independent of $n$ and $p$ such that $$\inf_i \sum_{j\neq i} \frac{\exp(-{d_{ij,\text{clean}}}^2/\eps)}{n}\geq \gamma>0\;.$$ Note that the previous result means that $L_0(\widetilde{W},\widetilde{G})$ and $L(W,G)$ are essentially spectrally equivalent: indeed we can use the Davis-Kahan theorem or Weyl’s inequality to relate eigenvectors and eigenvalues of $L_0(\widetilde{W},\widetilde{G})$ to those of $L(W,G)$ (see [@stewart90], [@bhatia97] or [@nekSparseMatrices] for a brief discussion putting all the needed results together). In particular, if the large eigenvalues of $L(W,G)$ are separated from the rest of the spectrum, the eigenvalues of $L_0(\widetilde{W},\widetilde{G})$ and corresponding eigenspaces will be close to those of $L(W,G)$. The proposition is a simple consequence of our previous results and Lemma \[lemma:approxCGLMatGeneralModif\] above. Indeed, in the notation of Lemma \[lemma:approxCGLMatGeneralModif\], we call $$w_{i,j}=\begin{cases} \exp(-{d_{ij,\text{clean}}}^2/\eps)& \text{ if } i\neq j \\ 1 &\text{ if } i=j \end{cases}\; \quad\text{and}\quad \tilde{w}_{i,j}=\begin{cases} \exp(-{d_{ij,\text{noisy}}}^2/\eps) &\text{ if } i\neq j \\ 1 &\text{ if } i=j \end{cases}\;.$$ Similarly, we call $$G_{i,j}=\begin{cases} \mathtt{O}^*_{ij,clean} &\text{ if } i\neq j\\ \id_d &\text{ if } i=j \end{cases} \quad\text{and}\quad \widetilde{G}_{i,j}=\begin{cases} \mathtt{O}^*_{ij,noisy} &\text{ if } i\neq j\\ \id_d &\text{ if } i=j \end{cases}\;.$$ Under Assumption G1, we know that, if $f_i=\exp(-2\trace{\Sigma}/\eps)$, $\sup_{i\neq j}|w_{i,j}-\tilde{w}_{i,j}/f_i|=\lo_P(1)$. Similarly, under Assumptions G1, A0 and A1, we know, using Theorem \[thm:consistencyRotations\] that $$\sup_{i,j}d(G_{i,j},\widetilde{G}_{i,j})=\lo_P(1)$$ and therefore, since $k$, the parameter of $SO(k)$, is held fixed in our asymptotics, $$\sup_{i,j}\norm{G_{i,j}-\widetilde{G}_{i,j}}_F=\lo_P(1)\;.$$ Since we assumed that $$\inf_i \sum_{j\neq i} \frac{\exp(-{d_{ij,\text{clean}}}^2/\eps)}{n}\geq \gamma>0\;,$$ i.e, in the notations of Lemma \[lemma:approxCGLMatGeneralModif\] $$\inf_i \frac{\sum_{j\neq i} w_{i,j}}{n}\geq \gamma>0\;,$$ where $\gamma$ is independent of $n$ and $p$, all the assumptions of Lemma \[lemma:approxCGLMatGeneralModif\] are satisfied when $n$ and $p$ are large enough, and we conclude that, in the notations of this lemma, $$\opnorm{L_0(W,G)-L_0(\widetilde{W},\widetilde{G})}=\lo_P(1)\;.$$ Furthermore, we have $0\leq w_{i,j},\tilde{w}_{i,j}\leq 1$, $\norm{G_{i,j}}_F\leq \sqrt{k}$ and $\norm{\widetilde{G}_{i,j}}_F\leq \sqrt{k}$, the latter two results coming from the fact that the columns of $G_{i,j}$ and $\widetilde{G}_{i,j}$ have unit norm. So we conclude that $$\opnorm{L(W,G)-L_0(\widetilde{W},\widetilde{G})}=\lo_P(1)\;.$$ #### Is the modification of the algorithm really needed? It is natural to ask what would have happened if we had not modified the standard algorithm, i.e if we had worked with $L(\widetilde{W},\widetilde{G})$ instead of $L_0(\widetilde{W},\widetilde{G})$. It is easy to see that $$L(\widetilde{W},\widetilde{G})=L_0(\widetilde{W},\widetilde{G})+\mathsf{D}$$ where $\mathsf{D}$ is a block diagonal matrix with $$\mathsf{D}(i,i)=\frac{\tilde{w}_{i,i}}{\sum_{j\neq i}\tilde{w}_{i,j}}\id_k=\frac{1}{\sum_{j\neq i}\tilde{w}_{i,j}}\id_k\;.$$ Under our assumptions, $$\opnorm{n\exp(-2\trace{\Sigma}/\eps) \mathsf{D}-D\left(\left\{\frac{\sum_{j\neq i} \exp(-{d_{ij,\text{clean}}}^2/\eps)}{n}\right\}_{i=1}^n\right)}=\lo_P(1)\;.$$ We also recall that under Assumption **G1**, $\trace{\Sigma}$ can be as large as $p^{1/2-\eta}$ - a very large number in our asymptotics. So in particular, if $n$ is polynomial in $p$, we have then $n^{-1}\exp(2\trace{\Sigma}/\eps)\tendsto \infty$. This implies that $$L(\widetilde{W},\widetilde{G})=L_0(\widetilde{W},\widetilde{G})+\mathsf{D}$$ is then dominated in spectral terms by $\mathsf{D}$. So it is clear that in the high-noise regime, if we had used the standard CGL algorithm, the spectrum of $L(\widetilde{W},\widetilde{G})$ would have mirrored that of $\mathsf{D}$ - which has little to do with the spectrum of $L(W,G)$, which we are trying to estimate - and the noise would have rendered the algorithm ineffective. By contrast, by using the modification we propose, we guarantee that even in the high-noise regime, the spectral properties of $L_0(\widetilde{W},\widetilde{G})$ mirror those of $L(W,G)$. We have hence made the CGL algorithm more robust to noise. **On the use of nearest neighbor graphs** In practice, variants of the CGL algorithms we have described use nearest neighbor information to replace $w_{i,j}$ by 0 if $w_{i,j}$ is not among the $k$ largest elements of $\{w_{i,j}\}_{j=1}^n$. In the high-noise setting, the nearest-neighbor information is typically not robust to noise, which is why we proposed to use all the $w_{i,j}$’s and avoid the nearest neighbor variant of the CGL algorithm, even though the latter probably makes more intuitive sense in the noise-free context. A systematic study of the difference between these two variants is postponed to future work. **Comparison with previous results in the literature** As far as we know, the study of the impact of high-dimensional additive noise on kernel methods was started in [@nekInfoPlusNoiseKernelMatrices10]. Compared to this paper, our extension is two-fold: 1) the noise level (i.e $\trace{\Sigma}$) that is studied in the current paper is much higher than what was studied in [@nekInfoPlusNoiseKernelMatrices10]. This is partly a result of the fact that the current paper focuses on the Gaussian kernel whereas [@nekInfoPlusNoiseKernelMatrices10] studied many more kernels. 2) [@nekInfoPlusNoiseKernelMatrices10] focused on standard kernel methods - based on the graph Laplacian, such as diffusion maps - where the connection information is not included in the data analysis. Incorporating this new element creates new difficulties. In other respects, we also refer to [@singer:2011] for another study of the influence of noise in a different setup. ### CGL beyond the rotational invariance distance {#subsubsec:CGLbeyondSOk} The previous analysis has been carried out for the RID and corresponding rotations for whom we studied the impact of additive noise in Subsection \[subsec:impactNoiseVDMProcedure\]. However, it is clear that our results apply much more broadly. We have the following proposition. \[prop:generalCGLApproxScheme\] Suppose we are given a collection $\mathsf{d}_{i,j,noisy}$ of (scalar-valued) dissimilarities between noisy versions of objects $i$ and $j$, $1\leq i,j\leq n$. Suppose objects $i$ and $j$ have (scalar-valued) dissimiliarity $\mathsf{d}_{i,j,clean}$. Consider the asymptotic regime where $n\tendsto \infty$ and suppose that there exists $\xi_n \in \mathbb{R}$ such that $$\sup_{i\neq j}|\mathsf{d}_{i,j,noisy}^2-\mathsf{d}^2_{i,j,clean}-\xi_n|=\lo_P(1)\;.$$ Call $\tilde{w}_{i,j}=\exp(-\mathsf{d}_{i,j,noisy}^2/\nu)$ and $w_{i,j}=\exp(-\mathsf{d}^2_{i,j,clean}/\nu)$ the corresponding affinities. $\nu$ is held fixed in our asymptotics, though the way affinities are computed may change with $n$. Suppose $\widetilde{G}_{i,j}$ is the connection between noisy versions of objects $i$ and $j$ and $G_{i,j}$ is the connection between the clean version of objects $i$ and $j$. Suppose that $w_{i,j}$, $G_{i,j}$ and $\widetilde{G}_{i,j}$ satisfy the assumptions of Lemma \[lemma:approxCGLMatGeneralModif\], with $\eps$ and $\eta$ possibly random but $\lo_P(1)$ and $\gamma$ bounded below as $n\tendsto \infty$. Then $$\opnorm{L(W,G)-L_0(\widetilde{W},\widetilde{G})}=\lo_P(1)\;.$$ This proposition is just a consequence of Lemma \[lemma:approxCGLMatGeneralModif\]. Indeed, the affinities are all bounded by 1. Furthermore, we can use $f_i=\exp(-\xi_n/\nu)$ and all the approximation results needed in Lemma \[lemma:approxCGLMatGeneralModif\] are true, so the result follows. ### A situation without robustness to noise So far, our work has been quite general and has shown that when the noise is Gaussian (or Gaussian-like) and its covariance $\Sigma_i$ is such that $\trace{\Sigma_i}=\trace{\Sigma_j}$ for all $i,j$, CGL algorithms can be made robust to noise. It has been recognized [@DiaconisFreedmanProjPursuit84; @nekCorrEllipD; @nekInfoPlusNoiseKernelMatrices10; @NEKRobustPaperPNAS2013Published] that to study the robustness of various statistical procedures in high-dimension, it is essential to move beyond the Gaussian-like situation and study for instance elliptical/scale mixture of Gaussian models. This largely due to the peculiar geometry of high-dimensional Gaussian and Gaussian-like vectors (see above references and [@HallMarronNeemanJRSSb05]). If we now write down a model for the noise where $N_i=\lambda_i Z_i$, where $Z_i$ are i.i.d ${\cal N}(0,\Sigma)$, $\lambda_i$’s are i.i.d with $\Exp{\lambda_i^2}=1$ and $\lambda_i \in \mathbb{R}$ is independent of $Z_i$, it is easy to modify our analysis (assuming for instance that $\lambda_i^2$ are bounded, though this condition could easily be relaxed) and to realize that our main approximation result in Proposition \[prop:controlApproxdijNoisy\] is replaced by $$\sup_{i\neq j}\left|d_{ij,\text{noisy}}^2-d_{ij,\text{clean}}^2-[\lambda_i^2+\lambda_j^2]\trace{\Sigma}\right|=\lo_P(1)\;.$$ In this situation, Theorem \[thm:consistencyRotations\] is still valid. However, Proposition \[prop:connectionGraphLaplacianApproxModifS\] is not valid anymore. The matrix $L_0(\widetilde{W},\widetilde{G})$ can be approximated by a matrix that depends both on the signal and the distribution of the $\lambda_i^2$’s. And there is no guarantee in general that this matrix will have approximately the same spectral properties as $L(W,G)$ or $L_0(W,G)$, the CGL matrix generated from the noise-free signals. This suggests that even our modification of the original CGL algorithm will not be robust to this “elliptical”-noise contamination. Numerical work ============== Although the robustness properties of CGL methods were not well studied in the past, these methods have been successfully applied to different problems; for example, [@singer_zhao_shkolnisky_hadani:2011; @singer_wu:2012; @Zhao_Singer:2013; @Marchesini_Tu_Wu:2014; @Cucuringu_Singer_Cowburn:2012; @Alexeev_Bandeira_Fickus_Mixon:2013]. In this section, we show simulated examples to illustrate the practical performance of our theoretical findings about CGL methods. We refer interested readers to the aforementioned papers for details and results of its applications. To demonstrate the main finding of this paper - that CGL methods are robust to high-levels of noise in the spectral sense - we take the noise to be a random Gaussian vector $Z\sim \mathcal{N}(0,cI_p/p^{\alpha})$, where $\alpha\leq 1$ and $c>0$. Note that the amount of noise, or the trace of the covariance matrix of $Z$, is $cp^{1-\alpha}$ and will blow up when $p\to \infty$ and $\alpha<1$. $1$-dim manifold {#Section:numerical:1dim} ---------------- Our first example is a dataset sampled from a low dimensional manifold, which is embedded in a high dimensional space. This dataset can be viewed as a collection of high dimensional points which is (locally) parametrized by only few parameters[^4], but in a nonlinear way. As a concrete example, we take the twisted bell-shaped simple and closed curve, denoted as ${\text{M}}$, embedded in the first $3$ axes of ${\mathbb{R}}^p$, where $p\gg 2$, via $\iota:[0,2\pi)\to {\mathbb{R}}^p$: $$\iota:\,t\mapsto [\,\cos(t),\,(1-0.8e^{-8\cos^2t})\cos(\pi(\cos(t)+1)/4),\,(1-0.8e^{-8\cos^2t})\sin(\pi(\cos(t)+1)/4),\,0,\ldots,0\,]\in{\mathbb{R}}^p\,.$$ ${\text{M}}$ is a 1-dim smooth manifold without boundary; that is, no matter how big $p$ is, locally the points on ${\text{M}}$ can be parametrized by only $1$ parameter. See Figure \[fig:numerical1:clean\] (A) for an illustration. We mention that one interesting dataset of this kind is the 2-D tomography from noisy projections taken at unknown random directions [@singer_wu:2013a]. For our numerical work, we independently sample $n$ points uniformly at random from $[0,2\pi)$. Due to the non-linear nature of $\iota$, it is equivalent to non-uniformly sampling $n$ points from ${\text{M}}$ independently. Denote the clean data as $\mathcal{Y}=\{y_i\}_{i=1}^n\subset {\text{M}}$. The data $\mathcal{X}=\{x_i\}_{i=1}^n$ we analyze is the clean data contaminated by noise, i.e $x_i=y_i+Z_i$, with $Z_i$ i.i.d with the same distribution as $Z$. We measure the [*signal-to-noise ratio*]{} of the dataset by the quantity $\text{snrdb}:=20\log\frac{\sqrt{{\mathbb{E}}X^TX}}{\sqrt{{\mathbb{E}}Z^TZ}}$. We take $n=p=1000$ and $\alpha=1/4$. Note that $\alpha=1/4$ is the critical value in our analysis beyond which our results do not apply. It corresponds to a high-noise level; for example, the snrdb will be $-9.25$ and $-18.73$ respectively when $c=0.25,\,0.4$. Then, we build up the connection graph by setting ${\mathtt{V}}:=\mathcal{X}$ and ${\mathtt{E}}:=\{(i,j);\,i,j\in {\mathtt{V}}\}$. Note that in practice, it is common to use a nearest-neighbor scheme to build up the graph, denoted as ${\mathtt{G}}^{\text{NN}}$, for the sake of computational efficiency. However, since the sets of nearest neighbors are not stable under the action of the noise, we also consider here the complete graph scheme, denoted as ${\mathtt{G}}$. Next we assign the weight function as $w:(i,j)\mapsto e^{-\|x_i-x_j\|^2_{{\mathbb{R}}^p}/m}$, where $m$ is the $25\%$ quantile of all Euclidean distances between pairs of $(x_i,x_j)\in{\mathtt{E}}$, and the connection function is defined to be a trivial one, that is, ${r}(i,j)=1$ for all $(x_i,x_j)\in{\mathtt{E}}$. With the connection graph, we build up the CGL matrix (in this $1$-dim manifold with the trivial connection, it is equivalent to the graph Laplacian (GL)) from ${\mathtt{G}}^{\text{NN}}$ and ${\mathtt{G}}$ for comparison, denoted as $L^{\text{NN}}(\widetilde{W},\widetilde{G})$ and $L(\widetilde{W},\widetilde{G})$ respectively (see (\[definition:LWG\])). We have seen in the analysis described earlier in the paper that, when $\alpha<1$, we have to remove the diagonal terms of the CGL matrix in order to preserve spectral properties. So, we also consider the matrix $L_0(\widetilde{W},\widetilde{G})$ for the comparison. We then evaluate the eigenvalues and eigenvectors of the above three different CGL’s. To simplify the notation, we use the same notations to denote the eigenvectors $u_1,u_2,u_3\ldots\in{\mathbb{R}}^n$ associated with the eigenvalues $1=\lambda_1>\lambda_2\geq \lambda_3\geq\ldots\geq 0$. We now show two sets of results to demonstrate the robustness of the CGL methods studied in this paper. : To achieve this, we may embed the sampled points into ${\mathbb{R}}^m$ by the [*truncated diffusion maps (tDM) with time $t>0$ and precision $\delta>0$*]{}: $$\Phi_{t,m}:\,x_i\mapsto (\lambda_2^t u_2(i),\, \lambda_3^t u_3(i),\,\ldots, \lambda_{m+1}^t u_{m+1}(i))\in {\mathbb{R}}^m,$$ where $\lambda_{m+1}>\delta$ and $\lambda_{m+2}\leq \delta$; that is, we map the $i$-th data point to ${\mathbb{R}}^m$ using the first $m$ non-trivial eigenvectors of the CGL. We choose $\delta=0.2$ in this simulation. The embedding results of $\mathcal{Y}$, $\Phi_{1,3}$, based on the above different CGL’s are shown in Figure \[fig:numerical1:clean\], and the results from $\mathcal{X}$ with $c=0.4$ are shown in Figure \[fig:numerical1:smallNoise\]. Ideally, we would expect to recover the “parametrization” of the dataset by the idea that the eigenvectors of the CGL represent a set of new coordinates for the data points, so the high dimensional dataset can be visualized in this new set of coordinates or its dimension can be reduced. In this specific example, we would expect to find a simple and closed curve out of the noisy dataset which represents the dataset in ${\mathbb{R}}^3$. Clearly when the dataset is clean, we succeed in the task no matter which CGL we use. However, if the dataset is noisy, at high-noise levels, the embedding might not be that meaningful if we use $L^{\text{NN}}(\widetilde{W},\widetilde{G})$ or $L(\widetilde{W},\widetilde{G})$. Indeed, as shown in Figure \[fig:numerical1:smallNoise\], with $L^{\text{NN}}(\widetilde{W},\widetilde{G})$ the structure of the dataset is barely recovered; with $L(\widetilde{W},\widetilde{G})$, even though we can get the simple closed curve[^5] back, there are several outliers which might deteriorate the interpretation. In this noisy case, we can only succeed in the task if we choose $L_0(\widetilde{W},\widetilde{G})$, as is discussed in this paper. ![ Clean samples from the twisted bell-shaped manifold. (A): the clean samples. Here we only plot the first $3$ axes of the high dimensional data $\mathcal{Y}$. The color of each point is a surrogate of the norm of each embedded point – blue means a relative small norm and dark red means a relative large norm; the scale above the figure refers to $\{\norm{x_i}_2\}_{i=1}^n$, i.e the norm of the data vectors in $\mathbb{R}^p$.(B): the results of the truncated diffusion maps (tDM), $\Phi_{1,3}$, when the connection graph is ${\mathtt{G}}^{\text{NN}}$ and the diagonal entries are not removed, where the number of nearest neighbors is chosen to be $100$; (C): the result of tDM, $\Phi_{1,3}$, when the connection graph is ${\mathtt{G}}$ and the diagonal entries are not removed; (D): the result of tDM, $\Phi_{1,3}$, when the connection graph is ${\mathtt{G}}$ and the diagonal entries are removed. Note that without surprise, the “parametrization” of the bell shaped manifold is recovered in (B), (C) and (D). For (B), (C), and (D), the scales above the figures refer to the norm of $\{\Phi_{1,3}(x_i)\}_{i=1}^n$; those vectors are of course 3-dimensional, which explains the difference in magnitude of our scales. []{data-label="fig:numerical1:clean"}](fig1_nn1000_pp1000_ll0.png){width="0.9\columnwidth"} ![Noisy samples from the twisted bell-shaped manifold with $\alpha=1/4$ and $c=0.4$. (A): the noisy samples. Note that we only plot the first $3$ axes of the data $\mathcal{X}$, so the “big” noise seems small, since 997 out of 1000 coordinates are not plotted. To emphasize the relationship among data points, the color of each point is a surrogate of the norm of each embedded point – blue means a relative small norm and dark red means a relative large norm; the scale above the figure refers to $\{\norm{x_i}\}_{i=1}^n$, i.e the norm of our 1000-dimensional vectors. (B): the results of the truncated diffusion maps (tDM), $\Phi_{1,3}$, when the connection graph is ${\mathtt{G}}^{\text{NN}}$ and the diagonal entries are not removed, where the number of nearest neighbors is chosen to be $100$. We can barely see the circle structure in the middle, and there are several big outliers; (C): the result of tDM, $\Phi_{1,3}$, when the connection graph is ${\mathtt{G}}$ and the diagonal entries are not removed. Note that when compared with (B), the embedding is better in the sense that the “parametrization”, the simple and close curve, is better recovered. But we can still observe several outliers; (D): the result of tDM, $\Phi_{1,3}$, when the connection graph is ${\mathtt{G}}$ and the diagonal entries are removed. Note that compared with (C), the embedding is yet better in the sense that the number of outliers is reduced and the parametrization of the manifold is recovered. Note that for (B), (C),(D), the scale above the figures refer to $\{\norm{\Phi_{1,3}(x_i)}\}_{i=1}^n$, which are 3-dimensional vectors. The different scales indicate the presence of outliers. Compare also with the scales in Figure \[fig:numerical1:clean\], (B), (C), (D).[]{data-label="fig:numerical1:smallNoise"}](fig2_nn1000_pp1000_ll10.png){width="0.9\columnwidth"} Estimating nearest neighbors of a given data point from a noisy dataset is not only important but also challenging in practice (for example, it is essential in the class averaging algorithm for the cryo-EM problem). This problem is directly related to local geodesic distance estimation when the dataset is modeled by the manifold. Their theoretical properties make diffusion maps and vector diffusion maps particularly well-suited for these tasks. To determine the neighbors, we need the notion of distance. In addition to the naive $L^2$ distance between points, we consider the [*diffusion distance*]{} between two points $x_i,x_j\in\mathcal{X}$ by $$d_{\text{DD}}(x_i,x_j):=\|\Phi_{t,m}(x_i)-\Phi_{t,m}(x_j)\|_{{\mathbb{R}}^m}.$$ Then, we determine the nearest neighbors of each data point based on these distances, where we choose $t=1$ and $\delta=0.2$ for the diffusion distance. More precisely, we first determine $10$ nearest neighbors of $x_i$, denoted as $x_{i_j}$, $j=1,\ldots,10$, from the noisy dataset $\mathcal{X}$, for all $i$. Then, since we know the ground truth, we may check the true relationship between $y_{i}$ and $y_{i_{j}}$, $j=1,\ldots,10$, i.e $d_{\text{DD}}(y_i,y_{i_j})$ for various CGL methods, or $\norm{y_i-y_{i_j}}$ if we use $L^2$ distance. Clearly, if the method preserves nearest neighbor information, at least approximately, the ranks of the $y_{i_j}$’s measured in terms of distances to $y_i$ should be small. To quantify the estimation accuracy, we collect the ranks of all estimated nearest neighbors, and plot the cumulative distribution results in Figure \[figure:numerical1:nnEst\]. In other words, if we call $R_{i_j}$ the rank of $y_{i_j}$ in terms of distance to $y_i$, we plot the cdf of $\{\{R_{i_j}\}_{j=1}^{10}\}_{i=1}^n$ for the various distances we use. (There are many other methods one could use to do these comparisons, such as using Kendall’s $\tau$ and variants (see [@comparingTopKLists2003]). The one we use here has the benefit of simplicity.) When the dataset is clean, all methods perform the same, as is predicted in Theorem \[thm:geod:DM\]. It is clear from the results that when the noise is large, the result based on the $L^2$ distance is much worse than the others. The performance based on the diffusion distance from $L^{\text{NN}}(\widetilde{W},\widetilde{G})$ is better when the noise level is not big, but still a non-negligible portion of error exists; the results based on $L(\widetilde{W},\widetilde{G})$ and $L_0(\widetilde{W},\widetilde{G})$ are much better, while the result based on $L_0(\widetilde{W},\widetilde{G})$ is the best. In conclusion, in addition to showing the robustness of CGL to noise, we have demonstrated the spectrally close relationship between $L_0(\widetilde{W},\widetilde{G})$ and $L(W,G)$, which is proved in Proposition \[prop:connectionGraphLaplacianApproxModifS\]. ![The result of nearest neighbors estimation. In all subfigures, the x-axis is the true rank of an estimated nearest neighbor and the y-axis is its cumulative distribution. To emphasize the difference, we only show the area ranging from $90\%$ to $100\%$ in the y-axis. The gray dashed (gray, black dashed and black respectively) curve is the cumulative distribution of the true ranks of the estimated nearest neighbors estimated from the ordinary Euclidean distance (diffusion distance based on $L^{\text{NN}}(\widetilde{W},\widetilde{G})$, $L(\widetilde{W},\widetilde{G})$ and $L_0(\widetilde{W},\widetilde{G})$ respectively). From left to right: clean samples from the bell shaped manifold, noisy samples with $\alpha=1/4$ and $c=0.25,0.4,0.5$ respectively. It is clear that when the noise is large, the result based on the $L^2$ distance is much worse than the others. The result based on $L^{\text{NN}}(\widetilde{W},\widetilde{G})$ is slightly better, but not that ideal, $L(\widetilde{W},\widetilde{G})$ is even better and $L_0(\widetilde{W},\widetilde{G})$ is the best. []{data-label="figure:numerical1:nnEst"}](fig3_nn1000_pp1000_ll0.png "fig:"){width="0.24\columnwidth"} ![The result of nearest neighbors estimation. In all subfigures, the x-axis is the true rank of an estimated nearest neighbor and the y-axis is its cumulative distribution. To emphasize the difference, we only show the area ranging from $90\%$ to $100\%$ in the y-axis. The gray dashed (gray, black dashed and black respectively) curve is the cumulative distribution of the true ranks of the estimated nearest neighbors estimated from the ordinary Euclidean distance (diffusion distance based on $L^{\text{NN}}(\widetilde{W},\widetilde{G})$, $L(\widetilde{W},\widetilde{G})$ and $L_0(\widetilde{W},\widetilde{G})$ respectively). From left to right: clean samples from the bell shaped manifold, noisy samples with $\alpha=1/4$ and $c=0.25,0.4,0.5$ respectively. It is clear that when the noise is large, the result based on the $L^2$ distance is much worse than the others. The result based on $L^{\text{NN}}(\widetilde{W},\widetilde{G})$ is slightly better, but not that ideal, $L(\widetilde{W},\widetilde{G})$ is even better and $L_0(\widetilde{W},\widetilde{G})$ is the best. []{data-label="figure:numerical1:nnEst"}](fig3_nn1000_pp1000_ll5.png "fig:"){width="0.24\columnwidth"} ![The result of nearest neighbors estimation. In all subfigures, the x-axis is the true rank of an estimated nearest neighbor and the y-axis is its cumulative distribution. To emphasize the difference, we only show the area ranging from $90\%$ to $100\%$ in the y-axis. The gray dashed (gray, black dashed and black respectively) curve is the cumulative distribution of the true ranks of the estimated nearest neighbors estimated from the ordinary Euclidean distance (diffusion distance based on $L^{\text{NN}}(\widetilde{W},\widetilde{G})$, $L(\widetilde{W},\widetilde{G})$ and $L_0(\widetilde{W},\widetilde{G})$ respectively). From left to right: clean samples from the bell shaped manifold, noisy samples with $\alpha=1/4$ and $c=0.25,0.4,0.5$ respectively. It is clear that when the noise is large, the result based on the $L^2$ distance is much worse than the others. The result based on $L^{\text{NN}}(\widetilde{W},\widetilde{G})$ is slightly better, but not that ideal, $L(\widetilde{W},\widetilde{G})$ is even better and $L_0(\widetilde{W},\widetilde{G})$ is the best. []{data-label="figure:numerical1:nnEst"}](fig3_nn1000_pp1000_ll8.png "fig:"){width="0.24\columnwidth"} ![The result of nearest neighbors estimation. In all subfigures, the x-axis is the true rank of an estimated nearest neighbor and the y-axis is its cumulative distribution. To emphasize the difference, we only show the area ranging from $90\%$ to $100\%$ in the y-axis. The gray dashed (gray, black dashed and black respectively) curve is the cumulative distribution of the true ranks of the estimated nearest neighbors estimated from the ordinary Euclidean distance (diffusion distance based on $L^{\text{NN}}(\widetilde{W},\widetilde{G})$, $L(\widetilde{W},\widetilde{G})$ and $L_0(\widetilde{W},\widetilde{G})$ respectively). From left to right: clean samples from the bell shaped manifold, noisy samples with $\alpha=1/4$ and $c=0.25,0.4,0.5$ respectively. It is clear that when the noise is large, the result based on the $L^2$ distance is much worse than the others. The result based on $L^{\text{NN}}(\widetilde{W},\widetilde{G})$ is slightly better, but not that ideal, $L(\widetilde{W},\widetilde{G})$ is even better and $L_0(\widetilde{W},\widetilde{G})$ is the best. []{data-label="figure:numerical1:nnEst"}](fig3_nn1000_pp1000_ll10.png "fig:"){width="0.24\columnwidth"} $2$-dim images {#Section:numerical:2dim} -------------- In Subsection \[Section:numerical:1dim\], we investigated numerically the influence of noise on CGL methods when the connection function is trivial. In this subsection, we discuss an example where the connection function plays an essential role in the analysis. We consider a dataset which contains randomly rotated versions of a set of objects, and the task is to align these objects in addition to classifying them. We encounter this kind of datasets and problems in, for example, image processing [@singer_zhao_shkolnisky_hadani:2011; @singer_wu:2012; @Zhao_Singer:2013], shape analysis [@Huang_Su_Guibas:2013], phase retrieval problems [@Marchesini_Tu_Wu:2014; @Alexeev_Bandeira_Fickus_Mixon:2013], etc. In [@singer_zhao_shkolnisky_hadani:2011; @singer_wu:2012; @Zhao_Singer:2013; @Marchesini_Tu_Wu:2014; @Alexeev_Bandeira_Fickus_Mixon:2013] and others, the CGL methods have been applied to solve the problem. To focus specifically on demonstrating the influence of noise on this problem, we work with 2-dimensional images observed in polar coordinates. If an image is defined in Cartesian coordinates, then in general a numerical rotation will introduce numerical artifacts and errors since resampling or interpolation procedures are then involved to compare two rotated images. These numerical issues are alleviated if we work with polar coordinates. To further minimize these numerical artifacts, we use surrogate images as our dataset – by a [*surrogate image*]{}, we mean a function defined on the circle $S^1$, which is discretized into $p$ equally spaced points. In other words, we consider images defined in polar coordinates, where we only have one sample on the radial axis. Now we discretize the $2\times 2$ rotational group, $SO(2)$, which is the same as the circle $S^1,$ into $p$ equally spaced points, that is, ${\mathcal T}^{(2)}:=\{e^{i2\pi k/p}\}_{k=1}^p$ - the sample $e^{i2\pi k/p}\in {\mathcal T}^{(2)}$ simply rotates vectors in $\mathbb{R}^2$ by an angle $2\pi k/p$. Note that since the surrogate images are defined on $p$ equally spaced points on $S^1$, the rotations in ${\mathcal T}^{(2)}$ act exactly on the images without introducing any numerical error. We choose $n_K$ different surrogate images, denoted as $\{f_i\}_{i=1}^{n_K}\subset {\mathbb{R}}^{p}$. Then we randomly and independently rotate each of them by $n_R$ angles; that is, for all $k=1,\ldots,n_K$ and $l=1,\ldots,n_R$, we have $S_i:=R_{k,l}\circ f_k$, where $R_{k,l}\in{\mathcal T}^{(2)}$, $R_{k,l}\circ f_k$ means rotating $f_k$ by $R_{k,l}$ and $i=(k-1)n_R+l$. We assume that $\argmin_{R\in \mathcal{T}^{(2)}}\|f_i-R\circ f_j\|>0$, for all $i,j=1,\ldots,n_K$; that is, the image $f_i$ is not a rotated version of another one $f_j$. In the end we get $n=n_Kn_R$ randomly rotated images $\{S_{i}\}_{i=1}^n\subset{\mathbb{R}}^{p}$. Denote by $\sigma$ the standard deviation of all pixels of all images in $\{f_i\}_{i=1}^{n_K}$. The data $\mathcal{X}=\{I_i\}_{i=1}^n$ we analyze is the clean data contaminated by the noise which is i.i.d. sampled from $Z$, that is, we have $I_i=S_i+Z_i$. We now build up the connection graph by setting ${\mathtt{V}}:=\{I_i\}_{i=1}^n$ and ${\mathtt{E}}:=\{(i,j);\,I_i,I_j\in {\mathtt{V}}\}$; that is, we take the complete graph scheme. Next we assign the weight function as $w:(i,j)\mapsto e^{-d^2_{\text{RID}}(I_i,I_j)/m}$, where $m$ is the $25\%$ quantile of all non-zero RID distances defined on ${\mathtt{E}}$, and the connection function as ${r}:(i,j)\mapsto \argmin_{R\in \mathcal{T}^{(2)}}\|I_i-R\circ I_j\|$. For comparison purposes, we also take the nearest neighbor scheme to construct the connection graph, denoted by $({\mathtt{G}}^{\text{NN}},w^{\text{NN}},{r}^{\text{NN}})$, where we choose $100$ nearest neighbors - as defined by the RID distance - to construct edges. When the images are noise-free, due to the connection function, we can recover $R_{k,l}$’s up to a rotation from the top eigenvector $v_1$ of different CGL’s built up from different connection graphs, $({\mathtt{G}},w,{r})$ or $({\mathtt{G}}^{\text{NN}},w^{\text{NN}},{r}^{\text{NN}})$, with or without removing the diagonal entries. To simplify the notation, we will use the same notation $v_1$ to denote the top eigenvector of the different CGL’s. Precisely, the estimated rotation is built up from $v_1$, denoted as $v\in{\mathbb{C}}^{n}$, by setting $v(i)=\frac{v_1(i)}{|v_1(i)|}$ when $|v_1(i)|>0$ and $v(i)=1$ when $|v_1(i)|=0$. (In a slight departure from the descriptions given earlier in the paper, the $r_{i,j}$’s are not $2\times 2$ matrices here, but complex numbers describing the corresponding rotations. Hence, $v_1$ is in ${\mathbb{C}}^n$. If we had used $2\times 2$ matrices, $v_1$ would have been in ${\mathbb{R}}^{2n}$ and we could have computed the vector $v$ by using pairs of entries of $v_1$.) To evaluate the performance of the estimated rotation when noise exists, we construct a complex vector $u\in{\mathbb{C}}^{n}$ whose $i$-th entry - where $i=(k-1)n_R+l$, $k=1,\ldots,n_K$ and $l=1,\ldots,n_R$ - is the complex form of $R_{k,l}$. We then evaluate the difference between the estimated rotation of the $l$-th object and the ground truth by observing the angle of $u(i)^*v(i)$. In other words, this quantity shows the discrepancy between the true rotation and the estimated rotation. To visualize this result, we plot the vector $z\in {\mathbb{R}}^{n}$ where $z(i)$ is the angle of the complex number $u(i)^*v(i)$. In Figure \[fig:numerical2\], the resulting $z$’s with $p=1000$, $n_K=5$, $n_R=200$, $\alpha=1/4$ and $c=6\sigma$ are illustrated. Note that since there are $5$ different images, we see a piecewise function with $5$ different values when the images are clean. When noise exists, we can see clearly the benefit of removing the diagonal entries (see Figure \[fig:numerical2\], (H)). We mention that depending on the problem, the affinity function and the connection function are constructed in different ways (see, for example [@singer_zhao_shkolnisky_hadani:2011; @singer_wu:2012; @Zhao_Singer:2013; @Huang_Su_Guibas:2013; @Marchesini_Tu_Wu:2014; @Alexeev_Bandeira_Fickus_Mixon:2013]). Also, the CGL is only one of several techniques we could use to analyze datasets described by the connection graph. We might consider other techniques, such as, semidefinite programming relaxation [@Wang_Singer:2013], nonlinear independent component analysis [@Talmon_Cohen_Gannot_Coifman:2013], and other methods, to obtain the information we are interested in, reorganize the data, etc... ![(A): a clean surrogate image. (B)-(D): alignment vectors $z$ computed from clean images; (E): a noisy surrogate image. (F)-(H): alignment vectors $z$ computed from noisy images with $c=6\sigma$. (A) and (E): the black curve is a clean surrogate image, and the gray curve is its noisy version; (B) and (F): the result from the CGL built up from ${\mathtt{G}}^{\text{NN}}$; (C) and (G): the result from the CGL built from ${\mathtt{G}}$ and the diagonal entries are not removed; (D) and (H): the result from the CGL built from ${\mathtt{G}}$ with the diagonal entries removed. It is clear that when the images are clean, all different CGL’s give equivalent results. But in the presence of noise, the CGL built up from ${\mathtt{G}}^{\text{NN}}$ is obviously worse.[]{data-label="fig:numerical2"}](fig4_nn1000_pp1000_ll0.png "fig:"){width="0.45\columnwidth"} ![(A): a clean surrogate image. (B)-(D): alignment vectors $z$ computed from clean images; (E): a noisy surrogate image. (F)-(H): alignment vectors $z$ computed from noisy images with $c=6\sigma$. (A) and (E): the black curve is a clean surrogate image, and the gray curve is its noisy version; (B) and (F): the result from the CGL built up from ${\mathtt{G}}^{\text{NN}}$; (C) and (G): the result from the CGL built from ${\mathtt{G}}$ and the diagonal entries are not removed; (D) and (H): the result from the CGL built from ${\mathtt{G}}$ with the diagonal entries removed. It is clear that when the images are clean, all different CGL’s give equivalent results. But in the presence of noise, the CGL built up from ${\mathtt{G}}^{\text{NN}}$ is obviously worse.[]{data-label="fig:numerical2"}](fig4_nn1000_pp1000_ll6.png "fig:"){width="0.45\columnwidth"} **<span style="font-variant:small-caps;">APPENDIX</span>**\ to “Connection graph Laplacian methods can be made robust to noise” Technical results ================= On quadratic forms {#subsec:ConcQuadForms} ------------------ \[lemma:controlSupQuadForms\] Suppose $Z_1,\ldots,Z_n$ are random vectors in $\mathbb{R}^p$, with $Z_i=\Sigma_i^{1/2} X_i$, where $X_i$ has mean 0 and covariance $\id_p$. We further assume that for every convex 1-Lipschitz function $f$, if $m_{f(X_i)}$ is a median of $f(X_i)$, $P(|f(X_i)-m_{f(X_i)}|>t)\leq 2 \exp(-c_it^2)$. $Z_i$’s are possibly dependent. Let $\{Q_i\}_{i=1}^n$ be $p\times p$ positive definite matrices. Call $\opnorm{Q_i}$ the largest eigenvalue of $Q_i$. Then we have $$\sup_{1\leq i \leq n}\left|\sqrt{Z_i\trsp Q_i Z_i}-\Exp{\sqrt{Z_i\trsp Q_i Z_i}}\right|={\gO}_P(\sup_i \sqrt{\opnorm{Q_i\Sigma_i/c_i}} \sqrt{\log n})\;.$$ This implies that, when $\sup_i \sqrt{\opnorm{Q_i\Sigma_i/c_i}} \sqrt{\log n}\tendsto 0$, $$\sup_{1\leq i \leq n}\left|Z_i\trsp Q_i Z_i-\trace{\Sigma_i Q_i}\right|={\gO}_P(\sup_i \sqrt{\opnorm{Q_i\Sigma_i/c_i}}\sqrt{\log n} \left[\sup_i \sqrt{\trace{\Sigma_i Q_i}}\vee 1\right])\;.$$ As explained in [@ledoux2001], the condition we require on $X_i$ is satisfied by many distributions. We refer also to [@nekCorrEllipD] for many examples. Here are two examples. The Gaussian distribution in dimension $p$ satisfies the previous assumptions with $c_i=1/2$, independently of the dimension. When $X_i$’s have independent coordinates supported on intervals of width at most $B_i$, $c_i$ is proportional to $1/B_i$. The map $X_i\rightarrow \sqrt{X_i\trsp\Sigma_i^{1/2}Q_i\Sigma_i^{1/2} X_i}$ is convex and $\sqrt{\opnorm{Q_i\Sigma_i}}=\sqrt{\opnorm{\Sigma_i^{1/2}Q_i\Sigma_i^{1/2}}}$-Lipschitz as a function of $X_i$. Indeed, it is a norm, which gives convexity. The Lipschitz-property comes from the triangle inequality. Hence, under our assumptions, since $Z_i=\Sigma_i^{1/2}X_i$, we have $$P\left(\left|\sqrt{Z_i\trsp Q_i Z_i}-\Exp{\sqrt{Z_i\trsp Q_i Z_i}}\right|>t\right)\leq 2\exp(-c_i t^2/(\opnorm{Q_i\Sigma_i}))\;.$$ By a simple union bound, we get $$P\left(\sup_{1\leq i \leq n}\left|\sqrt{Z_i\trsp Q_i Z_i}-\Exp{\sqrt{Z_i\trsp Q_i Z_i}}\right|>t\right)\leq 2\sum_{i=1}^n \exp(-c_i t^2/[\opnorm{Q_i\Sigma_i}])\leq 2 n \exp(- t^2/(\sup_i \opnorm{Q_i\Sigma_i/c_i}))$$ Taking $t_K=K\sqrt{\log(n)\opnorm{Q_i\Sigma_i/c_i}}$, for $K$ a constant, gives the first result. The second result follows from remarking that $|a^2-b^2|=|a-b||a+b|\leq |a-b|^2+2|b||a-b|$. When $\sup_i \sqrt{\opnorm{Q_i\Sigma_i/c_i}} \sqrt{\log n}\tendsto 0$, this gives immediately $$\sup_{1\leq i \leq n}\left|Z_i\trsp Q_i Z_i-\left[\Exp{\sqrt{Z_i\trsp Q_i Z_i}}\right]^2\right|={\gO}_P(\sup_i \sqrt{\opnorm{Q_i\Sigma_i/c_i}}\sqrt{\log n} \sup_i \left[\sqrt{\trace{\Sigma_i Q_i}}\vee 1\right])\;,$$ after we notice that $\trace{\Sigma_iQ_i}=\Exp{Z_i\trsp Q_i Z_i}\geq \left[\Exp{\sqrt{Z_i\trsp Q_i Z_i}}\right]^2$. Finally, using the variance bound in Proposition 1.9 of [@ledoux2001], we see that, $$\Exp{Z_i\trsp Q_i Z_i}-\left[\Exp{\sqrt{Z_i\trsp Q_i Z_i}}\right]^2\leq 2\opnorm{\Sigma_i Q_i/c_i}\;.$$ Under our assumption that $\sup_i \sqrt{\opnorm{Q_i\Sigma_i/c_i}} \sqrt{\log n}\tendsto 0$, we have $\sup_i \sqrt{\opnorm{Q_i\Sigma_i/c_i}}\tendsto 0$ and therefore $$\sup_i \opnorm{\Sigma_i Q_i/c_i} = \lo(\sup_i\sqrt{\opnorm{\Sigma_i Q_i/c_i}})\;.$$ This gives the second bound. In the case of the Gaussian distribution, the previous bounds can be improved, using an observation found in [@LaurentMassart2000]. \[lemma:controlSupQuadFormsLaurentMassart\] Suppose $Z_1,\ldots,Z_n$ are random vectors in $\mathbb{R}^p$, with $Z_i\sim{\cal N}(0,\Sigma_i)$. $Z_i$’s are possibly dependent. Let $\{Q_i\}_{i=1}^n$ be $p\times p$ positive definite matrices. Then we have, if $S_i=\Sigma_i^{1/2}Q_i\Sigma_i^{1/2}$ $$\sup_{1\leq i \leq n}\left|Z_i\trsp Q_i Z_i-\trace{S_i}\right|={\gO}_P\left(\sup_{1\leq i \leq n}\sqrt{\log(n)}\sqrt{\trace{S_i^2}}+\opnorm{S_i}\log(n)\right)\;.$$ By rotational invariance of the Gaussian distribution, we have $$W_i\triangleq Z_i\trsp Q_i Z_i-\trace{S_i}\equalInLaw \sum_{k=1}^p \lambda_k(S_i) (X_k^2-1)\;,$$ where $X_k$’s are i.i.d ${\cal N}(0,1)$. Using Lemma 1, p, 1325 in [@LaurentMassart2000], we see that $$P\left(\frac{|W_i|}{2}>\sqrt{\trace{S_i^2}}\sqrt{x}+\opnorm{S_i} x\right)\leq \exp(-x)\;.$$ Taking $x=K\log(n)$ in the previous inequality and a simple union bound gives the announced result. Proof of Lemma \[lemma:approxBoundsLaplacian\] {#app:sec:ApproxResLaplMatrix} ---------------------------------------------- We have $L(W,G)=D^{-1}S=(D/n)^{-1}(S/n)$. If we call $d_{i,i}=\sum_{j\neq i} w_{i,j}$ and $\tilde{d}_{i,i}=\sum_{j\neq i} \tilde{w}_{i,j}$, we see that $$|d_{i,i}/n-\tilde{d}_{i,i}/n|\leq \sup_{j\neq i} |w_{i,j}-\tilde{w}_{i,j}|\;.$$ Hence, $$\sup_{1\leq i \leq n} |d_{i,i}/n-\tilde{d}_{i,i}/n|\leq \sup_{1\leq i \leq n} \sup_{j\neq i} |w_{i,j}-\tilde{w}_{i,j}|\leq \sup_{i,j} |w_{i,j}-\tilde{w}_{i,j}|\leq \eps\;.$$ We conclude that $$\opnorm{D/n-\widetilde{D}/n}\leq \eps\;.$$ Under our assumptions, it is clear that $\opnorm{(D/n)^{-1}}\leq 1/\gamma$. The previous display also implies that $\opnorm{(\widetilde{D}/n)^{-1}}\leq 1/(\gamma-\eps)$. Furthermore, since $$(D/n)^{-1}-(\widetilde{D}/n)^{-1}=(D/n)^{-1}[D/n-\widetilde{D}/n](\widetilde{D}/n)^{-1}\;,$$ we see that $$\opnorm{(D/n)^{-1}-(\widetilde{D}/n)^{-1}}\leq \frac{\eps}{\gamma (\gamma-\eps)}\;.$$ Also, $$\norm{S/n-\widetilde{S}/n}_F^2\leq \sup_{i,j}\norm{S_{i,j}-\widetilde{S}_{i,j}}_F^2\;.$$ Naturally, since $S_{i,j}=w_{i,j}G_{i,j}$ and $\widetilde{S}_{i,j}=\widetilde{w}_{i,j}\widetilde{G}_{i,j}$, $$\norm{S_{i,j}-\widetilde{S}_{i,j}}_F^2\leq |w_{i,j}|^2 \norm{G_{i,j}-\widetilde{G}_{i,j}}_F^2+|w_{i,j}-\widetilde{w}_{i,j}|^2 \norm{\widetilde{G}_{i,j}}_F^2\leq C^2 (\eta^2+\eps^2)\;.$$ Hence, $$\opnorm{S/n-\widetilde{S}/n}\leq \norm{S/n-\widetilde{S}/n}_F\leq C (\eta+\eps).$$ We also note that $\norm{\widetilde{S}/n}_F\leq C^2$. So we can conclude that $$\opnorm{D^{-1}S-\widetilde{D}^{-1}\widetilde{S}}\leq \opnorm{D^{-1}(S-\widetilde{S})+(D^{-1}-\widetilde{D}^{-1})\widetilde{S}}\leq \frac{1}{\gamma} C(\eta+\eps)+\frac{\eps}{\gamma (\gamma-\eps)} C^2\;.$$ $\card{{{\cal T}}}$: an example when ${{{\cal T}}^{(k)}}_\text{exact}\subset SO(k)$ {#subsec:cardExactRotations} ----------------------------------------------------------------------------------- Naturally, when working with discretized images/objects with $p$ pixels/voxels, we need to also discretize $SO(k)$. In light of results like Proposition \[prop:controlApproxdijNoisy\], one natural question we have to deal with concerns the cardinality of the discretized set of transformations, ${{{\cal T}}^{(k)}}$, and the corresponding set for companion matrices, ${{\cal T}}$. The following proposition answers this question. The images/objects are discretized in polar coordinates. In other words, each point on our grid can be identified by its location on a ray emanating from the origin and reaching a point $p$ on the sphere of radius $r_0$ centered at the origin. The discretization of each ray does not have to be uniform. But this discretization is the same for all rays. In the case of $SO(2)$, this simply means that we discretize the circle of radius $r_0$, and our points lay on the corresponding radii. In this situation, it is natural to represent each point on our grid through $(r,\theta)$. To give a concrete example, we assume that $\theta \in \{2\pi\frac{k}{M}\}_{k=0}^{M-1}$ and $r\in \{\mathsf{r}_1,\ldots,\mathsf{r}_\alpha\}$ with $\mathsf{r}_\alpha=r_0$. The discretization of $SO(2)$ corresponds simply to rotations by an angle $\theta_k$, where $\theta_k=2\pi\frac{k}{M}$. These rotations clearly map our grid onto itself. We assume that our images or objects, after having been uniformly discretized in polar coordinates, fit in a $k$-dimensional cube. Then we assume that the rotation group is properly discretized so that each rotation is exact in the sense that it commutes with discretization. In other words, the rotation does not change the pixel values - pixels are simply swapped and pixel values are not averaged or aggregated in other ways. Note that when the image or object is discretized in Cartesian coordinates, the discretization and rotation will not commute and a distortion is inevitable. We postpone the study of such a phenomenon to future work. \[prop:cardinalSetOfDiscretizedRotations\] When ${{{\cal T}}^{(k)}}_\text{exact}$ is the discretized version of $SO(k)$ we just discussed, we have $$\card{{{{\cal T}}^{(k)}}_\text{exact}}=\card{{\cal T}_\text{exact}}=\gO(p^{k-1})\;.$$ Furthermore, the elements of ${\cal T}_\text{exact}$ are permutation matrices. In particular, they are orthogonal matrices. **Comment :** Proposition \[prop:cardinalSetOfDiscretizedRotations\] shows that in checking Assumption **G1**, we can assume that $\card{{{\cal T}}}$ is polynomial in $p$. This implies that Assumption G1 will be satisfied when $\max(\sigma_p,\sqrt{p}s_p^2) =\lo([\log(np)]^{-1/2})$. Hence the conditions we will have to check on $\sigma_p$ and $\sqrt{p}s_p^2$ will be quite unrestrictive and we will see that this implies that CGL algorithms are robust to considerable amount of additive noise. Our polar-coordinate discretization amounts to discretizing a sphere of radius $r_0$ in $\mathbb{R}^k$ with $M$ points and discretizing each ray linking a point on that sphere to the origin along $\alpha$ points. We naturally have the relationship $M\alpha=p$. Now elements of ${{{\cal T}}^{(k)}}_\text{exact}$ are orthogonal matrices with determinant 1, hence they can be characterized by their action on $k-1$ vectors in $\mathbb{R}^k$ which span a subspace of dimension $k-1$. Let us pick $k-1$ elements among our $M$ points on the sphere of radius $r_0$. We require that these $k-1$ elements span a subspace of dimension $k-1$ in $\mathbb{R}^k$. We call the corresponding vectors $v_1,\ldots,v_{k-1}$. Suppose now that ${\mathtt{O}}\in {{{\cal T}}^{(k)}}_\text{exact}$. Then, for each $i$, ${\mathtt{O}}v_i$ has to be one of the elements of our discretized sphere. Therefore, $$\card{{{{\cal T}}^{(k)}}_\text{exact}}\leq M^{k-1}=\left(\frac{p}{\alpha}\right)^{k-1}\leq p^{k-1}\;.$$ Now let ${\mathtt{O}}\in {{{\cal T}}^{(k)}}_\text{exact}$ and let $O$ be the companion matrix of ${\mathtt{O}}$. Note that if $I_i$ is our image and $I_i^\vee \in \mathbb{R}^p$ is its discretized version, $OI_i^\vee$ swaps the position of the entries of the vector $I_i^\vee$, since ${\mathtt{O}}$ maps our grid onto itself. Hence $O$ is a permutation matrix and it is therefore orthogonal. #### Another approach We note that another approach can be employed to generate a sampling grid $\mathfrak{X}$ and an associated set of exact transformations for $k\geq 3$. Take $m$ points in $R^k$ denoted as $X=\{x_i\}_{i=1}^m$. Take a finite subgroup, $T$ of $SO(k)$. Now consider the sampling grid $\mathfrak{X}\triangleq \{Rx_i;\, R\in T, x_i\in X\}$. Since $T$ is a subgroup, we know that the sampling grid $\mathfrak{X}$ is of finite size; that is, $|\mathfrak{X}|\leq m|T|$. It is also clear that $T$ is an exact set of transforms for $\mathfrak{X}$, by simply using the fact that $T$ is a group. Note that the standard polar coordinate grid in $\mathbb{R}^2$ described above can be viewed as an instance of the method we just discussed, with $T$ consisting of powers of the rotation by the angle $\frac{2\pi}{M}$. Note however that the classification of finite subgroups of $SO(k)$, for $k\geq 3$, imposes strong constraints on the sampling grids obtained by such a construction. Background on CGL methods {#Appendix:Section:CGL} ========================= In this section, we discuss the [*noise-free*]{} connection graph Laplacian (CGL) ${L(W,G)}$ defined in (\[definition:LWG\]), understand its asymptotical behavior under the assumption that the point clouds we collect are distributed on a manifold, and show that the CGL matrix built up under this assumption enjoy a sparsity property which allows the robustness result shown in this paper. In addition, we will discuss the fact that the CGL matrix can be viewed as a generalization of the graph Laplacian (GL) [@singer_wu:2013]. We will see that although GL and CGL share several similar properties but are fundamentally different. We would assume the background knowledge of differential geometry in the following discussion. For a reader who is not familiar with the subject, we refer him to [@doCarmo:1992; @gilkey:1974; @Bishop_Crittenden:2001; @berard2; @Berline_Getzler_Vergne:2004] for the topics we will encounter. We start from some notations. Denote ${\text{M}}$ to be a $d$-dimensional compact, connected and smooth Riemannian manifold embedded in ${\mathbb{R}}^p$ via ${\iota}$, where $d\leq p$. Denote the tangent bundle as $T{\text{M}}$. The tangent plane at $y\in{\text{M}}$ is denoted as $T_y {\text{M}}$. Introduce the metric $g$ on ${\text{M}}$ induced from the canonical metric of the ambient space ${\mathbb{R}}^p$. Denote $d(y,y')$ to be the geodesic distance between $y,y'\in{\text{M}}$. Denote by $\nabla$ the covariant derivative of the vector field, $\Delta_g$ the Laplace-Beltrami operator, $\nabla^2$ the connection Laplacian of the tangent bundle associated with the Levi-Civita connection, and by ${\mbox{Ric}}$ the Ricci curvature of $({\text{M}},g)$. We denote the spectrum of $\nabla^2$ (resp. $\Delta_g$) by $\{-\lambda_l\}_{l=0}^\infty$ (resp. $\{-\gamma_l\}_{l=0}^\infty$), where $0=\lambda_0\leq\lambda_1\leq \ldots$ (resp. $0=\gamma_0<\gamma_1\leq \ldots$), and the corresponding eigenspaces by $F_l:=\{{X}\in L^2(T{\text{M}}):~\nabla^2{X}=-\lambda_l {X}\}$ (resp. $E_l:=\{\phi\in L^2({\text{M}}):~\Delta_g\phi=-\gamma_l \phi\}$), $l=0,1,\ldots$. In general, while $\gamma_0=0$ always exists, $\lambda_0$ may not: a simple example is found considering $S^2$ with the standard metric. It is well known [@gilkey:1974] that $\dim(F_l)<\infty$, the eigen-vector-fields are smooth and form a basis for $L^2(T{\text{M}})$ (resp. $\dim(E_l)<\infty$, the eigenfunctions are smooth and form a basis for $L^2({\text{M}})$), that is, $L^2(T{\text{M}})=\overline{\oplus_{l\in{\mathbb{N}}\cup\{0\}} F_l}$ (resp. $L^2({\text{M}})=\overline{\oplus_{l\in{\mathbb{N}}\cup\{0\}} E_l}$), the completion of $\oplus_{l\in{\mathbb{N}}\cup\{0\}} F_l$ with relative to the measure induced by $g$. To simplify the statement, we assume that $\lambda_l$ (resp. $\gamma_l$) for each $l$ are simple and ${X}_l$ (resp. $\phi_l$) is a normalized basis of $F_l$ (resp. $E_l$). Denote $\mathcal{B}(F_k)$ (resp. $\mathcal{B}(E_k)$) the set of bases of $F_k$ (resp. $E_k$), which is identical to the orthogonal group $O(\dim(F_k))$ (resp. $O(\dim(E_k))$). Denote the set of the corresponding orthonormal bases of $L^2(T{\text{M}})$ by $\mathcal{B}(T{\text{M}},g)=\Pi^\infty_{k=1}\mathcal{B}(F_k)$ (resp. $\mathcal{B}({\text{M}},g)=\Pi^\infty_{k=1}\mathcal{B}(E_k)$). Given the collected data ${\mathcal{X}}=\{x_i\}_{i=1}^n\subset {\mathbb{R}}^p$, where $x_i$ are [*signal random vector*]{} i.i.d. sampled from a random vector $\mathtt{X}$. We assume a manifold structure inside the signal random vector; that is, we view $\mathtt{X}:\Omega \rightarrow {\mathbb{R}}^p$ as a measurable function with respect to the probability space $(\Omega,\mathcal{F},P)$, and assume that its range is ${\iota}({\text{M}})$. Note that we cannot define the probability density function (p.d.f.) of $\mathtt{X}$ on the ambient space ${\mathbb{R}}^p$ when $d<p$ since $\iota({\text{M}})$ is degenerate, but we can still discuss how we sample points from $\iota({\text{M}})$, which leads to [*the p.d.f. of $\mathtt{X}$ on ${\text{M}}$*]{}. Indeed, we employ the following definition which is based on the induced measure [@cheng_wu:2012 Section 4]. Denote $\tilde{\mathcal{B}}$ to be the Borel sigma algebra on ${\iota}({\text{M}})$, and $\tilde{P}_{\mathtt{X}}$ the probability measure of $\mathtt{X}$, defined on $ \tilde{\mathcal{B}}$, induced from $P$. Assume that $\tilde{P}_{\mathtt{X}}$ is absolutely continuous with respect to the volume density on ${\iota}({\text{M}})$ associated with $g$, that is, ${\textup{d}}\tilde{P}_{\mathtt{X}}(x)=f({\iota}^{-1}({x})){\iota}_*{\textup{d}}V(x)$, where $f:{\text{M}}\to {\mathbb{R}}$ and $x\in{\iota}({\text{M}})$. We interpret $f$ as [*the p.d.f. of $\mathtt{X}$ on ${\text{M}}$*]{}. To alleviate the notation, in the following we abuse the notation and will not distinguish between ${\iota}({\text{M}})$ and ${\text{M}}$. Pick up the kernel function $K(x)=e^{-x^2}$. Note that the kernel function can be more general, for example, $K\in C^2({\mathbb{R}})$, non-zero, non-negative and monotonic decreasing, but we focus ourselves on this kernel to make the explanation clear. Denote $ \mu^{(k)}_{l}:=\int_{\mathbb{R}^d}\|x\|^l K^{(k)}(\|x\|){\textup{d}}x$, where $k=0,1,2$, $l\in {\mathbb{N}}\cup\{0\}$, and $K^{(k)}$ means the $k$-th order derivative of $K$. We assume $\mu^{(0)}_0=1$ and $\sqrt{h}$ is small enough so that $\sqrt{h}$ is smaller than the reach [@Niyogi_Smale_Weinberger:2009] and the injectivity radius [@doCarmo:1992] of the manifold ${\text{M}}$, $\text{inj}({\text{M}})$. Connection Graph and Affinity Graph {#Appendix:B_2} ----------------------------------- An affinity graph, denoted as $({\mathtt{G}},w)$, is actually a special case of the connection graph in the sense that the connection function is not defined on the affinity graph. If we take a constant function ${r}_0:{\mathtt{E}}\to 1$, the affinity graph becomes a connection graph $({\mathtt{G}},w,{r}_0)$. We mention that in practice, if we decide to construct the connection graph from a given dataset, there are several different ways depending on the application and goal. For example, in addition to the examples discussed in the main context, in the geometric approach to the signal processing [@coifman_singer:2008; @Talmon_Cohen_Gannot_Coifman:2013], the affinity is defined to be the Mahalanobis distance reflecting the intrinsic property of the underlying state space; in the chair synchronization problem [@Huang_Su_Guibas:2013], the affinity between two chair meshes is defined based on their Hausdorff distance. The quality of the chosen affinity might influence the analysis result directly. Connection Graph Laplacian and its Applications ----------------------------------------------- Now we discuss the CGL. Consider the symmetric matrix ${L(W,G)}_s := {D}^{-1/2}{S}{D}^{-1/2}$, which is similar to ${L(W,G)}$. Since ${L(W,G)}_s$ is symmetric, it has a complete set of eigenvectors $v_{n,i}$, $i=1,\ldots,nd$ and its associated eigenvalues $\mu_{n,i}$, where the eigenvalues are bounded by $1$ [@singer_wu:2012]. We would order the eigenvalues in the decreasing order. Note that the eigenvectors of ${L(W,G)}_s$ is related to those of ${L(W,G)}$ via ${D}^{-1/2}$. First, note that ${L(W,G)}$ is an operator acting on $\boldsymbol{v}\in {\mathbb{R}}^{nd}$ by $$\begin{aligned} \label{Cmatrix} ({L(W,G)}\boldsymbol{v})[i] = \frac{\sum_{j: (i,j)\in E}w(i,j)r(i,j)\boldsymbol{v}[j]}{\sum_{k: (i,k)\in E} w(i,k)},\end{aligned}$$ where $\boldsymbol{v}$ can be viewed as a vector-valued function defined on ${\mathtt{V}}$ so that $\boldsymbol{v}[j]:=(\boldsymbol{v}((j-1)d+1),\ldots,\boldsymbol{v}(jd))\in{\mathbb{R}}^d$. We could interpret this formula as a generalized random walk on the affinity graph. Indeed, if we view the vector-valued function $\boldsymbol{v}$ as the status of a particle defined on the vertices, when we move from one vertex to the other one, the status is modified according to the relationship between vertices [*encoded*]{} in ${r}$. We mention that depending on the connection function, the structure ${L(W,G)}$ might be very different, which leads to different analysis results and conclusion. We will give a precise example regarding this statement later. Now we discuss an important property of the CGL – the synchronization, which has been studied in [@Chung_Zhao_Kempton:2013; @Bandeira_Singer_Spielman:2013] and applied to the following problems, for example, 1. a new imaging technique aiming to obtain the atomic scale resolution images of a macro-scale object called “ptychographic imaging problem” [@Marchesini_Tu_Wu:2014]; 2. a frame design called “polarization” for the phase retrieval problem [@Alexeev_Bandeira_Fickus_Mixon:2013]; 3. a spectral relaxation approach to solve the least squares solution of the rotational synchronization problem [@Wang_Singer:2013]; 4. graph realization problem by synchronization over the euclidean group [@Cucuringu_Lipman_Singer:2012; @Cucuringu_Singer_Cowburn:2012]. Here we give an intuition about this synchronization notion. Suppose there exists a vector-valued status $\boldsymbol{v}\in{\mathbb{R}}^{nd}$ of norm $1$ which is “synchronized” according to the encoded relationship ${r}$ in the sense that $\boldsymbol{v}[j]={r}(j,i)\boldsymbol{v}[i]$ for all $(i,j)\in {\mathtt{E}}$, then ${L(W,G)}\boldsymbol{v}[i]$ will be the same as $\boldsymbol{v}[i]$, and hence the functional associated with the eigenvalue problem $$\max_{\boldsymbol{v}\in{\mathbb{R}}^{nd};\,\|\boldsymbol{v}\|=1}\boldsymbol{v}^T{L(W,G)}\boldsymbol{v}$$ is maximized with the eigenvalue $1$, and its eigenvector is the synchronized vector $\boldsymbol{v}$. Thus, the top eigenvector of ${L(W,G)}$, when viewed as the vector-valued status on the vertex, is the “synchronized” status with respect to the connection function. We mention that the existence of the synchronized vector-valued function is equivalent to the notion of “consistency” studied in [@Chung_Zhao_Kempton:2013]. When the connection function is constant; that is, the connection matrix $G_0:=\boldsymbol{1}\boldsymbol{1}^T$, where $\boldsymbol{1}$ is a $n\times 1$ vector with all entries $1$, the GL is defined as ${L}:={\id}-L(W,G_0)$. Notice a natural interpretation of $L(W,G_0)$ – since the sum of each row of $L(W,G_0)$ is $1$, $L(W,G_0)$ is the [*transition matrix*]{} associated with a random walk on ${\mathtt{G}}$. To avoid confusion, the eigenvectors and eigenvalues of $L(W,G_0)_s$ are denoted as $u_{n,i}$ and $\nu_{n,i}$, where $i=1,\ldots,n$, and $0\leq\nu_{n,i}\leq 1$ are ordered in the decreasing order. As a special case of CGL, the GL has several applications which deserves discussion, for example 1. in the spectral clustering algorithm, we only need to find the first $k$ trivial eigenvectors [@VonLuxburg_Belkin_Bousquet:2008; @Lee_OveisGharan_Trevisan:2012]; 2. to evaluate the Cheeger ratio, we need the second eigenvalue [@Fan:1996]; 3. to visualize the high dimensional data, we need the first $3$ non-trivial eigenvectors; 4. in the cryo-EM problem, if we want to reconstruct the rotational position of each projection image, we need the first $9$ non-trivial eigenvectors [@Giannakis_Schwander_Ourmazd:2012]; 5. in the 2d random tomography problem [@singer_wu:2013a], only the first $2$ non-trivial eigenvectors are needed; 6. in the orientability detection problem [@singer_wu:2012], we need the first eigenvector. More algorithms depending on the eigenstructure of the GL can be found, to mention but a few, in [@Coifman_Maggioni:2006; @Szlam_Maggioni_Coifman:2008; @Jones_Maggioni_Schul:2008; @Sun_Ovsjanikov_Guibas:2009; @Ovsjanikov_Sun_Guibas:2008; @Rustamov:2007; @Reuter:2010; @Memoli:2011]. We comment that if the p.d.f $f$ is not uniform, then a specific normalization stated in [@coifman_lafon:2006] allows us to study the dynamics of the underlying dynamical system [@nadler_lafon_coifman:2005; @nadler_lafon_coifman:2006]. Asymptotical behavior of CGL ---------------------------- To better understand the CGL, we focus on the frame bundle and its associated tangent bundle [@singer_wu:2012] here to simplify the exploration. For CGL associated with a more general principal bundle structure, we refer the reader to [@singer_wu:2013]. \[Assumption:A\] 1. ${\text{M}}$ is a smooth and compact $d$-dim manifold. When the boundary $\partial{\text{M}}$ is not empty, it is assumed to be smooth, and we denote ${\text{M}}_\delta:=\{x\in{\text{M}}:\,d(x,\partial{\text{M}})\leq \sqrt{\delta}\}$; 2. The p.d.f. ${f}\in C^3({\text{M}})$ is uniformly bounded from below and above, that is, $0<p_m\leq {f}(x)\leq p_M<\infty$. However, to simplify the exploration, we assume here that ${f}$ is uniform, that is, ${f}$ is a constant function defined on ${\text{M}}$. When ${f}$ is non-uniform, its theoretical results can be found in [@coifman_lafon:2006; @singer_wu:2013]. Under Assumption \[Assumption:A\], we collect the data $\mathcal{X}$ independently and identically sampled from ${\text{M}}$ and build up the following graph. First define a graph ${\mathtt{G}}_{\text{M}}:=({\mathtt{V}},{\mathtt{E}})$ by taking ${\mathtt{V}}=\mathcal{X}$, and ${\mathtt{E}}=\{(x_i,x_j);\,x_i\in\mathcal{X}\}$. Then define the affinity function $w$ on ${\mathtt{E}}$ by $$w:\,(i,j)\mapsto K_h\left(x_i,x_j\right):=K\left( \|x_i-x_j \|^2_{\mathbb{R}^p}/h \right),$$ where $h>0$ is the chosen bandwidth. Note that we choose to use the Euclidean distance, instead of the geodesic distance, to build up $w$ since in practice we have only an access to the Euclidean distance (or other metric, depending on the application). Asymptotically this discrepancy will disappear. \[Assumption:B\] 1. For each point $x_i\in\mathcal{X}$, we also have a sample on the frame bundle $b(i)\in O({\text{M}})$ so that the $b(i)$ is the basis of the tangent space $T_{x_i}{\text{M}}$. In particular, we are given a group-valued function $b:\mathcal{X}\to O(d)$. With Assumption \[Assumption:B\], we define the [*connection function*]{} on ${\mathtt{E}}$ as $${r}:\,(i,j)\mapsto b(i)^{T}P_{x_i,x_j}b(j)\in O(d)$$ where $P_{x_i,x_j}$ presents the parallel transport of the vector field from $x_j$ to $x_i$. As a result, we have a connection graph $({\mathtt{G}}_{\text{M}}, w,{r})$. With $({\mathtt{G}}_{\text{M}}, w,{r})$, we build up the the CGL by $\id-L(W,G)$, where $W$ and $G$ are the weight matrix and connection matrix associated with $w$ and ${r}$. Under this framework, the GL is when we work with the trivial line bundle associated with ${\text{M}}$. The geometrical meaning of the connection function deserves some discussions. First, note that although all tangent planes $T_{x_i}{\text{M}}$ are isomorphic to ${\mathbb{R}}^d$ [@doCarmo:1992], but they are different in the sense that we cannot “compare” $T_{x_i}{\text{M}}$ and $T_{x_j}{\text{M}}$ directly. Precisely, it makes sense the say $u-v$ when $u,v\in{\mathbb{R}}^d$, but we can not evaluate $u_i-u_j$ when $u_i\in T_{x_i}{\text{M}}$ and $u_j\in T_{x_j}{\text{M}}$. To carry out the comparison between different tangent planes, we need a bit more work. Indeed, $b(i)\in O(d)$ is a basis of the tangent plane of $T_{x_i}{\text{M}}$, which practical meaning is mapping ${\mathbb{R}}^d$ isomorphically to $T_{x_i}{\text{M}}$. In other words, given a vector field $Y$, $b(i)^TY(x_i)$ evaluates its coordinate at $x_i$. The parallel transport $P_{x_i,x_j}$ is a geometrical generalization of the notion “translation” in the Euclidean space – it is an isometric map mapping $T_{x_j}{\text{M}}$ to $T_{x_i}{\text{M}}$. As a result, $r(i,j)\in O(d)$ is an isometric map from ${\mathbb{R}}^d$ to ${\mathbb{R}}^d$, and geometrically it maps the coordinate of a vector field at $x_j$, that is, $\boldsymbol{v}[j]$, to the vector field at $x_i$, that is, $b(j)\boldsymbol{v}[j]$, then parallelly transports $b(j)\boldsymbol{v}[j]$ to $x_i$, and then evaluate the coordinate of $P_{x_i,x_j}b(j)\boldsymbol{v}[j]$ with related to the basis $b(i)$. We emphasize that the connection function in the connection graph associated with the frame bundle encodes not only the geometry but also the topology of the manifold. In practice, this constraint may lead to a better understanding of the underlying data structure. For example, in the cryo-EM problem, this viewpoint leads to a better angular classification result. We now state the pointwise convergence and the spectral convergence of the ${L(W,G)}$. These theorems apply to the GL, while we replace the vector fields by the functions with the same regularity and the connection Laplacian operator by the Laplace-Beltrami operator and $b(i)=1$ (see [@singer_wu:2013] for details). \[thm:pointwise\_conv\_VDM\] Suppose Assumption \[Assumption:A\] and Assumption \[Assumption:B\] hold and ${X}\in C^4(T{\text{M}})$. For all $x_i\notin {\text{M}}_{\sqrt{h}}$ with high probability (w.h.p.) $$\begin{aligned} b(i)\big(({\id}-{L(W,G)}){\bar{X}}\big)[i]=h\frac{\mu_2}{2d}\nabla^2{X}(x_i)+O(h^{2})+O\left(\frac{1}{n^{1/2}h^{d/4-1/2}}\right)\nonumber\end{aligned}$$ where ${\bar{X}}\in{\mathbb{R}}^{nd}$ and ${\bar{X}}[i]=b(i)^{-1}{X}(x_i)$. For all $x_i\in{\text{M}}_{\sqrt{h}}$, we have w.h.p. $$\begin{aligned} b(i)\big(({\id}-{L(W,G)}){\bar{X}}\big)[i]=O(\sqrt{h})P_{x_i,x_0}\nabla_{\partial_d}{X}(x_0)+O(h)+O\left(\frac{1}{n^{1/2}h^{d/4-1/2}}\right),\nonumber\end{aligned}$$ where $x_0=\argmin_{y\in\partial{\text{M}}}d(x_i,y)$ and $\nabla_{\partial_d}$ is the derivative in the normal direction. To state the spectral convergence result, define an operator $T_{\text{C},h}:C(T{\text{M}})\to C(T{\text{M}})$: $$T_{\text{C},h}{X}(y):=\frac{\sum_{j: (i,j)\in E} K_h(y,x_j)P_{y,x_j}{X}(x_j)}{\sum_{j: (i,j)\in E} K_h(y,x_j)},$$ where ${X}\in C(T{\text{M}})$. To simplify the discussion, we assume that the eigenvalues of the heat kernel of the connection Laplacian $e^{t\nabla^2}$ are simple. When there exists an eigenvalue with multiplicity greater than $2$, the theorem can be proved using the projection operators onto the eigenspaces. We mention that in the special case GL, the point convergence theorem was first established in [@belkin_niyogi:2005] under the uniform sampling and boundary-free assumption, and then extended to a more general setup in [@hein_audibert_luxburg:2005; @coifman_lafon:2006; @singer:2006; @singer_wu:2013], and the spectral convergence was established in [@belkin_niyogi:2007; @VonLuxburg_Belkin_Bousquet:2008]. \[thm:spectralconvergence\_heatkernel\] Suppose Assumption \[Assumption:A\] and Assumption \[Assumption:B\] hold and fix $t>0$. Denote $\mu_{\text{C},t,h,i}$ to be the $i$-th eigenvalue of $T^{t/h}_{\text{C},h}$ with the associated eigenvector ${X}_{\text{C},t,h,i}$. Also denote $\mu_{t,i}>0$ to be the $i$-th eigenvalue of $e^{t\nabla^2}$ with the associated eigen-vector field ${X}_{t,i}$. We assume that both $\mu_{\text{C},t,h,i}$ and $\mu_{t,i}$ decrease as $i$ increases, respecting the multiplicity. Fix $i\in{\mathbb{N}}$. Then there exists a sequence $h_n\to 0$ such that $$\lim_{n\to \infty}\mu_{\text{C},t,h_n,i}=\mu_{t,i},\,\,\mbox{ and }\,\,\lim_{n\to \infty}\|{X}_{\text{C},t,h_n,i}-{X}_{t,i}\|_{L^2(T{\text{M}})}=0$$ in probability. With these Theorems, we are able to discuss why different connection functions lead to different analysis results. Consider $S^2$ embedded in ${\mathbb{R}}^3$ with the standard metric. If we define the connection function according to the Levi-Civita connection, then the top eigenvalue of $e^{t\nabla^2}$ is strictly less than $1$ due to the hairy-ball theorem [@gilkey:1974]. In other words, asymptotically we are not able to find a synchronized vector-valued status on it. On the other hand, if we take the trivial connection function, that is, ${r}(i,j)=I_2$, then asymptotically we obtain $\Delta_g$ acting on two independent functions. Since the dimension of the null space of $\Delta_g$ is the number of the connected components of the manifold, the top eigenvalue of the CGL with the trivial connection function is $1$; that is, a synchronized vector-valued status exists. See Figure \[S.B:fig:2\] for the result. The main reason leading to this difference is rooted in the connection theory, and we refer the interested reader to [@Bishop_Crittenden:2001]. ![The first $10$ eigenvalues of the ${\id}-{L(W,G)}$ with related to a non-trivial connection function determined from the Levi-Civita connection of the frame bundle of $S^2$ (left) and a trivial connection function (right). $1000$ points are sampled uniformly from $S^2$. Note that the eigenvalues on the right figure are the same as those of GL while the multiplicities of all the eigenvalues are $2$.[]{data-label="S.B:fig:2"}](consistency.png){width="0.6\columnwidth"} The “Sparsity” of the CGL ------------------------- We define the following “sparsity” condition. Fix $\gamma>0$. For a $n\times n$ matrix $Q$, we sort its eigenvalues $\nu_{Q,\ell}$, $\ell=1,\ldots,n$, so that $|\nu_{Q,1}|\geq |\nu_{Q,2}|\geq \ldots\geq |\nu_{Q,n}|$. Then $Q$ satisfies the [*$\gamma$-sparsity property*]{} if $$|\nu_{Q,\ell}|\leq e^{-C_Q\ell^{\gamma}}$$ for all $\ell$, where $C_Q>0$ depends on $Q$. We now claim that the CGL under the manifold assumption satisfies the $2/d$-sparsity property. Note that this theorem also holds for the GL. Asymptotically when $n\to\infty$, for $\ell\in{\mathbb{N}}$, $\mu_{n,\ell}\leq e^{-C_{\text{L}}\ell^{2/d}}$, where the constants $C_{\text{L}}>0$ depends $d$, the lower bound of the Ricci curvature $k$ and the diameter $D$. Note that the Weyl’s theorem [@Berline_Getzler_Vergne:2004] holds for the connection Laplacian, that is, $$\tilde{N}(\mu)\sim \frac{1}{(4\pi)^{d/2}\Gamma(d/2+1)}\mu^{d/2},$$ where $\tilde{N}(\mu)$ is the number of eigenvalues of $\nabla^2$ less than $\mu>0$, and we have the consequence that [@wu:2012] $$\begin{aligned} \lambda_j\geq c'(d,k,D)j^{2/d},\end{aligned}$$ where $j\in{\mathbb{N}}$ and $c'(d,k,D)$ is the universal constant depending only on $d$, the lower bound of the Ricci curvature $k$ and the diameter $D$. Note that since $\mu_{n,\ell}\to e^{-\lambda_\ell}$ in probability, we have $$\mu_{n,\ell}\leq e^{-c'(d,k,D)\ell^{2/d}}.$$ Hence, combined with Theorem \[thm:spectralconvergence\_heatkernel\], we conclude the claim with $C_{\text{L}}:=c'(d,k,D)$. Vector Diffusion Maps and Diffusion Maps {#subsection:Appendix:VDMDM} ---------------------------------------- In this subsection, we discuss a potential application of the CGL and GL – estimate the local geodesic distance. Fix $t>0$, we define the [*vector diffusion maps*]{} (VDM) $V_{t,n}:\mathcal{X}\to {\mathbb{R}}^{(nd)^2}$ by $$V_{t,n}: x_i \mapsto \left( (\mu_{n,l} \mu_{n,r})^{t} \langle v_{n,l}[i], v_{n,r}[i] \rangle\right)_{l,r=1}^{nd},$$ where $v_{n,l}[i]$ is a $d$-dim vector containing the $((i-1)d+1)$-th entry to the $(id)$-th entry of the eigenvector $v_{n,l}$. With this map, the Hilbert-Schmidt norm of the $(i,j)$-th block of ${L(W,G)}_s$ satisfies $$\|{L(W,G)}_s^{2t}(i,j)\|^2_{HS} = \langle V_{t,n}(x_i), V_{t,n}(x_j) \rangle,$$ that is, $\|{L(W,G)}_s^{2t}(i,j)\|^2_{HS}$ becomes an inner product for the finite dimensional Hilbert space. The reason we need to consider ${L(W,G)}_s^{2t}$ but not ${L(W,G)}_s^{t}$ is that all eigenvalues $\mu_{n,l}$ of ${L(W,G)}_s$ reside in the interval $[-1,1]$, and we can not guarantee the positivity of $\mu_{n,l}$ when $n$ is finite. We can then define the [*vector diffusion distance*]{} (VDD) to quantify the affinity between nodes $i$ and $j$: $$d_{\text{C},t,n}:=\| V_{t,n}(x_i)- V_{t,n}(x_j)\|^2.$$ The theoretical properties of VDM and VDD will be clear when $n\to \infty$. Fix $a\in \mathcal{B}(T{\text{M}},g)$ and $t>0$, define the following map embedding $x\in{\text{M}}$ to $\ell^2$: $$V^a_t:x \mapsto \left( \frac{1}{\sqrt{d}(4\pi)^{d/2}t^{(d+1)/2}} e^{-(\lambda_k + \lambda_l)t/2} \langle {X}_k(x), {X}_l(x) \rangle \right)_{k,l=1}^\infty.$$ With $V^a_t$, we define a new affinity between pairs of points by $$\begin{aligned} d_{\text{C},t}(x,y) := \|V^a_t(x)-V^a_t(y)\|_{\ell^2}.\end{aligned}$$ Due to Theorem \[thm:pointwise\_conv\_VDM\] and Theorem \[thm:spectralconvergence\_heatkernel\], the VDM (resp. VDD) is a discretization of $V^a_t$ (resp. $d_{\text{C},t}$), so we may abuse the notation and call $V^a_t$ VDM and $d_{\text{C},t}$ VDD. We have the following Theorem saying that locally $d_{\text{C},t}$ approximates the geodesic distance: \[thm:geod\] Take a Riemannian manifold $({\text{M}},g)$. For all $t>0$, the VDM $V^a_t$ is diffeomorphic. Furthermore, suppose $x,y\in {\text{M}}$ so that $x=\exp_yv$, where $v\in T_y {\text{M}}$. When $\|v\|^2\ll t\ll 1$ we have $$\begin{aligned} \label{thm:geod_approximation} d^2_{\text{C},t}(x,y)= \|v\|^2+O(t\|v\|^2).\end{aligned}$$ Although GL is a special case of CGL, with GL we may define a different embedding which has different features. Given $t>0$ and $0\leq \delta<1$, the [*diffusion maps (DM) with diffusion time $t$*]{}[^6] as $$\Phi_{t,n}:\,x_i \mapsto \left( \nu_{n,l}^tu_{n,l}(i) \right)_{l=2}^{n}.$$ One similar but different algorithm is the [*Laplacian eigenmaps*]{} [@belkin_niyogi:2003; @belkin_niyogi:2005], that is, $x_i$ is mapped to $\left(u_{n,l}(i) \right)_{l=2}^{m}$, which can be viewed as a special DM with diffusion time $t=0$ and $1<m\leq n$ is chosen by the user. Yet another similar quantity referred to as the [*global point signature*]{} proposed in [@Rustamov:2007], which maps $x_i$ to $\left( (-\ln \nu_{n,l} )^{-1/2} u_{n,l}(i) \right)_{l=2}^m$, where $m$ is chosen by the user. Another variation is the [*commute time embedding*]{} [@Qiu_Hancock:2007]. We mention in the Laplacian eigenmaps, global point signature and commute time embedding, the notion “diffusion” does not exist. Although these mappings are diffeomorphic to each other when $m=n$ via a linear transformation, asymptotically their behaviors are different. Furthermore, even if the connection function is trivial, the VDM and DM are different. With DM, we introduce a new metric between sampled points, which is referred to as [*diffusion distance*]{} (DD): $$\begin{aligned} d_{\textup{DM},t,n}(x_i,\,x_j) := \|\Phi_{t,n}(x_i)-\Phi_{t,n}(x_j)\|_{{\mathbb{R}}^{n-1}}.\end{aligned}$$ To study $d_{\textup{DM},t,n}$, we take $a\in \mathcal{B}({\text{M}},g)$ and $t>0$, and map $x\in {\text{M}}$ to the Hilbert space $\ell^2$ by [@berard_besson_gallot:1994] $$\Phi^a_t:x \mapsto \sqrt{\text{vol}({\text{M}})}\left( e^{-\gamma_\ell t}\phi_\ell(x) \right)_{\ell=1}^\infty,$$ With the map $\Phi^a_t$, we are able to define a new affinity between pairs of points: $$\begin{aligned} d_{\textup{DM},t}(x,y) := \|\Phi^a_t(x)-\Phi^a_t(y)\|_{\ell^2}.\end{aligned}$$ Due to Theorem \[thm:pointwise\_conv\_VDM\] and Theorem \[thm:spectralconvergence\_heatkernel\], the DM (resp. DD) is a discretization of $\Phi^a_t$ (resp. $d_{\text{DM},t}$), so we may abuse the notation and call $\Phi^a_t$ DM and $d_{\text{DM},t}$ DD. It has been shown that the DM satisfies the following “almost isometric” property [@singer_wu:2012]: \[thm:geod:DM\] Take a Riemannian manifold $({\text{M}},g)$. For all $t>0$, $\Phi^a_t$ is diffeomorphic. Furthermore, suppose $x,y\in {\text{M}}$ so that $x=\exp_yv$, where $v\in T_y {\text{M}}$. When $\|v\|^2\ll t\ll 1$ we have $$\begin{aligned} \label{thm:geod_approximation} d^2_{\textup{DM},t}(x,y)= \|v\|^2+O(t\|v\|^2).\end{aligned}$$ The above theorems, when combined with the above spectral convergence theorem, says that the VDD and DD provide an accurate estimation of the geodesic between two close points. While combined with the manifold sparsity property, we have the following practical fact – if we are allowed a positive small error when we estimate the geodesic distance, we do not need to recover the whole eigen-structure. Instead, the first few eigenvalues and eigenvectors are enough. We have the following statement shown in [@Portegies:2013; @Bates:2014]. Fix $\epsilon>0$ and $({\text{M}},g)$ is a $d$-dim manifold satisfying ${\mbox{Ric}}(g)\geq (d-1)kg,\,\mbox{vol}({\text{M}})\leq V,\,\text{inj}({\text{M}})\leq I$. Then there exists a $t_0=t_0(d,k,I,\epsilon)$ such that for all $0<t<t_0$, these is $N_E=N_E(d,k,I,V,\epsilon,t)$ so that if $N\geq N_E$, the [*truncated diffusion maps*]{} $$\begin{aligned} \Phi^a_{t,N}:\,x\mapsto\sqrt{\text{vol}({\text{M}})}\big(e^{-\gamma_\ell t}\phi_\ell(x)\big)_{\ell=1}^N\end{aligned}$$ is an embedding of ${\text{M}}$ into ${\mathbb{R}}^N$ and $$\begin{aligned} 1-\epsilon<\left|\frac{(2t)^{(n+2)/4}\sqrt{2}(4\pi)^{n/4}}{\sqrt{\text{vol}({\text{M}})}}{\textup{d}}\Phi_{N,t}|_x\right|<1+\epsilon.\end{aligned}$$ Before ending this section, we show an interesting example regarding the data visualization and embedding issue. Take the Trefoil knot ${\text{M}}$ embedded in ${\mathbb{R}}^3$ by $\iota(t)=[\sin(t)+2\sin(2t),\,\cos(t)-2\cos(2t),\,-\sin(3t)]$, where $t\in [0,2\pi)$. We refer to Figure \[fig5:Trefoil\] for an illustration. Note that the Trefoil knot is not homeomorphic to $S^1$. We sample $1000$ points uniformly from $[0,2\pi)$ independently; that is, we sample $1000$ points on ${\text{M}}$ non-uniformly. If we want to visualize the dataset, we may apply the tDM to embed ${\text{M}}$ to ${\mathbb{R}}^3$ (or ${\mathbb{R}}^2$). The result is shown in Figure \[fig5:Trefoil\]. The results deserve some discussion. Note that the tDM maps the Trefoil knot into a circle, which is not homeomorphic to the Trefoil knot; that is, the topology of the Trefoil knot is not preserved. Note that for the visualization purpose, we only choose the first $3$ (or $2$) eigenvectors, which leads to a map which deteriorate the topology. If we want to guarantee the preservation of the topology, we need the embedding theorem counting how many eigenvectors we need. This opens the following question, in particular when the dataset is noisy – how to balance between different data analysis results, for exampling, how to balance between preserving the topology information and data visualization? ![Left: the Trefoil knot; middle: the truncated DM with the first 2 non-trivial eigenvectors of the GL; right: the truncated DM with the first 3 non-trivial eigenvectors of the GL.[]{data-label="fig5:Trefoil"}](fig5Trefoil.png){width="0.5\columnwidth"} [^1]: UC Berkeley, Department of Statistics; **Contact:** nkaroui@berkeley.edu ; Support from NSF grant DMS-0847647 (CAREER) is gratefully acknowledged. [^2]: Stanford University, Department of Mathematics; **Contact:** hauwu@stanford.edu ; Support from AFOSR grant FA9550-09-1-0643 is gratefully acknowledged.**Keywords:** concentration of measure, random matrices, vector diffusion maps, spectral geometry, kernel methods;**AMS MSC 2010 Classification:** 60F99, 53A99 [^3]: We may also consider $U(k)$. But in this paper we focus on $O(k)$ to simplify the discussion. [^4]: By definition, although locally the manifold resembles Euclidean space near a point, globally it might not. Thus, in general we can only parametrize the manifold locally. This feature captures the possible nonlinear structure in the data. [^5]: The main idea behind tDM is embedding the dataset to a lower dimensional Euclidean space so that the structure underlying the data can be extracted. Please see Section \[subsection:Appendix:VDMDM\] for details. [^6]: In practice, we may consider [*the truncated diffusion maps (tDM) with diffusion time $t$ and accuracy $\delta$*]{}, which is defined as $\Phi_{t,n,m(\delta,t)}:\,x_i \mapsto \left( \nu_{n,l}^tu_{n,l}(i) \right)_{l=2}^{m(\delta,t)}$, where $m(\delta,t)\in{\mathbb{N}}$ such that $\lambda_{m(\delta,t)}^t>\delta\lambda_1^t$ and $\lambda_{m(\delta,t)+1}^t\leq\delta\lambda_1^t$ [@coifman_lafon:2006]

ArXiv

\ --------------------------- ------------------------------- ----------------------- ---------------------- Piotr Sapiezynski\*$^{1}$ Arkadiusz Stopczynski$^{1,2}$ Radu Gatej$^{3}$ Sune Lehmann$^{1,4}$ \[0em\] pisa@dtu.dk arks@mit.edu radu.gatej@econ.ku.dk sljo@dtu.dk --------------------------- ------------------------------- ----------------------- ---------------------- $^{1}$ Department of Applied Mathematics and Computer Science, Technical University of Denmark\ $^{2}$ Media Lab, Massachusetts Institute of Technology\ $^{3}$ Department of Economics, University of Copenhagen\ $^{4}$ Niels Bohr Institute, University of Copenhagen\ Abstract {#abstract .unnumbered} ======== We study six months of human mobility data, including WiFi and GPS traces recorded with high temporal resolution, and find that time series of WiFi scans contain a strong latent location signal. In fact, due to inherent stability and low entropy of human mobility, it is possible to assign location to WiFi access points based on a very small number of GPS samples and then use these access points as location beacons. Using just one GPS observation per day per person allows us to estimate the location of, and subsequently use, WiFi access points to account for 80% of mobility across a population. These results reveal a great opportunity for using ubiquitous WiFi routers for high-resolution outdoor positioning, but also significant privacy implications of such side-channel location tracking. Introduction {#introduction .unnumbered} ============ Due to the ubiquity of mobile devices, the collection of large-scale, longitudinal data about human mobility is now commonplace[[ ]{}]{}[@lazer2009life]. High-resolution mobility of individuals and entire social systems can be captured through a multitude of sensors available on modern smartphones, including GPS and sensing of nearby WiFi APs (access points or routers) and cell towers. Similarly, mobility data may be collected from systems designed to enable communication and connectivity, such as mobile phone networks or WiFi systems (e.g. at airports or on company campuses)[[ ]{}]{}[@lim2007real; @ferris2006gaussian]. Additionally, large companies such as Google, Apple, Microsoft, or Skyhook, combine WiFi access points with GPS data to improve positioning[[ ]{}]{}[@skyhook], a practice known as ‘wardriving’. While widely used, the exact utility and mechanics of wardriving are largely unknown, with only narrow and non-systematic studies reported in the literature[[ ]{}]{}[@rekimoto2007lifetag; @kawaguchi2009wifi]. As a consequence, it is generally not known how WiFi networks can be used for sensing mobility on a societal scale; this knowledge is proprietary to large companies. In the scientific realm, the mobility patterns of entire social systems are important for modeling spreading of epidemics on multiple scales: metropolitan networks[[ ]{}]{}[@eubank2004modelling; @sun2014efficient; @liang2013unraveling] and global air traffic networks[[ ]{}]{}[@colizza2007; @hufnagel2004forecast]; traffic forecasting[[ ]{}]{}[@kitamura2000micro]; understanding fundamental laws governing our lives, such as regularity[[ ]{}]{}[@gonzalez2008understanding], stability[[ ]{}]{}[@Lu18062012], and predictability[[ ]{}]{}[@song2010limits]. Predictability and stability of human mobility are also exploited by commercial applications such as intelligent assistants; for example Google Now[[ ]{}]{}[@google_now] is a mobile application, which learns users’ habits to, among other services, conveniently provide directions to the next inferred location. Mobility traces are highly unique and identify individuals with high accuracy[[ ]{}]{}[@de2013unique]. Sensitive features can be extracted from mobility data, including home and work locations, visited places, or personality traits[[ ]{}]{}[@de2013predicting]. Moreover, location data are considered the most sensitive of all the commonly discussed personal data collected from or via mobile phones[[ ]{}]{}[@staiano2014money]. Here, we show that a time sequence of WiFi access points is effectively equal to location data. Specifically, having collected both GPS and WiFi data with high temporal resolution (median of 5 minutes for GPS and 16 seconds for WiFi) in a large study[[ ]{}]{}[@10.1371/journal.pone.0095978], we use six months of data for 63 participants to model how lowering the rate of location sampling influences our ability to infer mobility. The study participants are students with heterogeneous mobility patterns. They all attend lectures on campus located outside of the city center, but live in dormitories and apartments scattered across the metro area at various distances from the university. By mapping the WiFi data, we are able to quantify details of WiFi-based location tracking, which are usually not available to the general public. We find that the geo-positioning inferred from WiFi access points (APs or routers) could boost efficacy in other data collection contexts, such as research studies. In addition, our findings have significant privacy implications, indicating that for practical purposes WiFi data should be considered location data. As we argue in the following sections, this finding is not recognized in current practices of data collection and handling. Results {#results .unnumbered} ======= Ubiquitously available WiFi access points can be used as location beacons, identifying locations based on BSSID (basic service set identifier, uniquely identifying every router) broadcast by APs. These locations are not intrinsically geographical, as the APs do not have geographical coordinates attached. However, since the placement of APs tends to remain fixed, mapping an AP to a location where it was seen once is sufficient to associate all the subsequent scans from the user device with geographical coordinates. See Supplementary Information for details on inferring the geographical locations of routers, as well as identifying (and discarding data from) mobile access points. WiFi networks are ubiquitous. In our population, 92% of all WiFi scans detect at least one access point, and 33% detect more than 10 APs, as shown in Figure \[fig:parishes\_map\]c. In densely-populated areas, an average of 25 APs are visible in every scan, with population density explaining 50% of the variance of the number of APs, as shown in Figure \[fig:parishes\_map\]b. WiFi scans containing at least one visible AP can be used for discovering the location of the user, with a typical spatial resolution on the order of tens of meters. ![**WiFi routers are located where people live.** **a:** Map of Greater Copenhagen Area, divided into parishes with color indicating average number of routers discovered per scan; rectangle overlay indicates the city center. **b:** The number of access points visible in each scan is correlated with the population density ($r^2=0.5$). Even in low population density areas there are several routers visible on average in each scan. Therefore, knowing the positions of only a subset of routers is enough for precise location sensing. **c:** Distribution of number of routers per scan. In our dataset 92% of scans contain at least one router.[]{data-label="fig:parishes_map"}](parish_map_panel.png){width="1\columnwidth"} We investigate three approaches to using access points as location beacons, all of which enable WiFi-based location tracking even with limited resources: (1) recovering APs’ locations from mobility traces collected during an initial training period (exploiting the long-term stability of human mobility), (2) recovering APs’ locations from randomly sampled GPS updates (exploiting low entropy of human mobility, see Supplementary Information for distinction between stability and low entropy), and (3) using only the most frequently observed APs for which location can be feasibly obtained from external databases. The task is to efficiently assign geographical coordinates (latitude and longitude) to particular APs, so they can be used as beacons for tracking user’s location. In the following sections, we refer to *time coverage* as the fraction of ten-minute timebins, in which at least one router with a known location is observed. Stability of human mobility allows for efficient WiFi-based positioning. {#stability-of-human-mobility-allows-for-efficient-wifi-based-positioning. .unnumbered} ------------------------------------------------------------------------ Human mobility has been shown to remain stable over long periods of time[[ ]{}]{}[@gonzalez2008understanding]. We find that participants in our study have stable routines, with locations visited in the first one, two, three, and four weeks of the study still visited frequently six months later. Learning the locations of routers seen during the first seven days (corresponding to ${\sim}$3.5% of the observations, shown in Figure \[fig:coverage\]a, left panel) provides APs’ locations throughout the rest of the experiment sufficient for recovering ${\sim}$55% of users mobility until the Christmas break around days 75-90. When the location of routers seen by each person is inferred using only this person’s data (the personal-only WiFi database case, shown using an orange line in Figure \[fig:coverage\]), the information expires with time: there is a stable decrease in time coverage after Christmas break. This decline is evident both when a week (Figure \[fig:coverage\]a, left panel) and four weeks (right panel) are used for training, with the time coverage dropping ${\sim}$18 percentage points between days 60 and 160. The histograms above each plot show the distribution of time coverage in selected points in time (at 7, 80, 190 days respectively). The distribution for day 190 reveals that the expiry of the personal database validity is driven by individuals who significantly altered routines, with 40% of participants spending only around 10% of time in locations they have visited in the first week. In contrast, when the inferred locations of routers are shared among people (the global database case, represented by a blue line) the information does not expire and shows no decreasing trend during the observation period. This implies that rather than moving to entirely new locations, people begin to visit places that are new to them, but familiar to other participants. The histograms of time coverage distribution in both panels of Figure \[fig:coverage\]a reveal that the individuals are heterogeneous in their mobility. The coverage in most cases is highly affected in the non-personal case (where the person does not collect their own location information, but data from others is used, marked using green in the figures), but 20% of participants retain a coverage of above 80% throughout the observation period, see Figure \[fig:coverage\]a, left panel. People living and working close to each other (like students in a dormitory) share a major part of their mobility and thus location of the APs they encounter can be estimated using data collected by others. The demonstrated stability of human mobility patterns over long periods has real-life privacy implications. Denying a mobile application access to location data, even after a short period, may not be enough to prevent it from tracking user’s mobility, as long as its access to WiFi scans is retained. Human mobility can be efficiently captured using infrequent location updates. {#human-mobility-can-be-efficiently-captured-using-infrequent-location-updates. .unnumbered} ----------------------------------------------------------------------------- Sampling location randomly across time (Figure \[fig:coverage\]b), rather than through the initial period (Figure \[fig:coverage\]a) provides a higher time coverage, which is retained throughout the observation. With around one sample per day per person on average, the location can be inferred 80% of the time in case of global lookup base and 70% in personal case (see Figure \[fig:coverage\]c, at training fraction of 0.007). The histograms in Figure \[fig:coverage\]b confirm that distribution of coverage in the non-personal case is bimodal within our population: mobility of some individuals can effectively be modeled using data from people around them, while patterns of others are so distinct they require using self-collected information. The single-mode distribution of coverage in the personal case and the fact that the distribution is unchanged between day 7 and day 190 show the lack of temporal decline when sampling happens throughout the observation period. The GPS sensor on a mobile device constitutes a major battery drain when active[[ ]{}]{}[@gps_consumption], whereas the WiFi frequently scans for networks by default. Our results show that GPS-based location sampling rate can be significantly reduced in order to save battery, while retaining high resolution location information through WiFi scanning. Our analyses also point to another scenario where WiFi time series can result in leaks of personal information. Infrequent location data can be obtained from a person’s (often public) tweets, Facebook updates, or other social networking check-ins and then matched with their WiFi records to track their mobility. Overall human mobility can be effectively captured by top WiFi access points. {#overall-human-mobility-can-be-effectively-captured-by-top-wifi-access-points. .unnumbered} ----------------------------------------------------------------------------- As previously suggested[[ ]{}]{}[@song2010limits], people’s mobility has low entropy and thus a few most prevalent routers can work effectively as proxies for their location. Figure \[fig:coverage\]d shows that inferring the location of just 20 top routers per person on average (which, given the median count of 22000 routers observed per person, corresponds to 0.1% of all routers seen) translates to knowing the location of individuals 90% of the time. Since our population consists of students, who attend classes in different lecture halls in various buildings across the campus, we expect that the number of access points necessary to describe mobility of persons with a fixed work location can be even lower. There are persons in our study, for whom just four access points correspond to 90% of time coverage (see Suppelementary Figure 4 for details). That the mobility of individuals in our sample overlaps is apparent in Figure \[fig:coverage\]d as the time coverage of three top routers in the personal case is the same as in the global coverage using the total of 80 routers (instead of 189 disjoint routers). As a consequence, a third party with access to records of WiFi scans and no access to location data, can effectively determine the location of each individual 90% of time by sending less than 20 queries to commercial services such as Google Geolocation API or Skyhook. Single-user analysis. {#single-user-analysis. .unnumbered} --------------------- To illustrate the ubiquity of WiFi access points and how effectively they can be used to infer mobility patterns, we present a small example dataset containing measured and inferred location information of one of the authors, collected over two days. During the 48 hours of observation, the researcher’s phone was scanning for WiFi with a median period of 44 seconds, measuring on average 19.8 unique devices per scan, recording 3822 unique access points. Only one scan during the 48 hours was empty, and one scan yielded 113 unique results. Figure \[fig:subsampling\]a shows the corresponding GPS trace collected with a median sampling period of 5 minutes. When dividing the 48 hours of the test period into 10 minute bins, a raw GPS trace provides location estimation in 89% of these bins. Four stop locations are marked with blue circles and include home, two offices, and a food market visited by the researcher. Figure \[fig:subsampling\]b shows the estimation of this trace based on the inferred locations of WiFi routers, see Supplementary Information for detailed information on the location inference. The four stop locations are clearly visible, but the transitions have lower temporal resolution and errors in location estimations. This method provides location information in 97% of temporal bins. Using WiFi increases overall coverage, but might introduce errors in location estimation of routers which were only observed shortly, for example during transition periods. Figure \[fig:subsampling\]c shows the estimation of this trace based on the locations of top 8 (0.2%) WiFi routers. The four important locations have been correctly identified, but information on transitions is lost. Information in 95% of temporal bins is available. Finally, Figure \[fig:subsampling\]d shows a graphical representation of how much time the researcher spends in any one of the top eight locations during the observation time. Note that the first four locations account for an overwhelming fraction of the 48 hours. Knowing the physical position of the top routers and having access to WiFi information reveals the location of the user for the majority of the timebins. The details of trajectories become lost as we decrease the number of routers we use to estimate locations. With too few routers might not be possible to determine which of possible routes the subject chose or how long she took to travel through each segment of the trip. On the other hand, the high temporal resolution of the scans allows for very precise discovery of arrival and departure times and of time spent in transit. Such information has important implications for security and privacy, as it can be used to discover night-watch schedules, find times when the occupants are not home, or efficiently check work time of the employees. ![**48 hours of location data of one of the authors, with the four visited locations visited marked in blue: home, two offices, and a food market.** Even though the author’s phone has sensed 3822 unique routers in this period, only a few are enough to describe the location more than 90% of time. [**a.**]{} traces recorded with GPS; [**b.**]{} traces reconstructed using all available data on WiFi routers locations - the transition traces are distorted, but all stop locations are visible and the location is known 97% of the time. [**c.**]{} with 8 top routers it is still possible to discover stop locations in which the author spent 95% of the time. In this scenario transitions are lost. [**d.**]{} timeseries showing when during 48 hours each of the top routers were seen. It can be assumed that AP 1 is home, as it’s seen every night, while AP 2 and AP 3 are offices, as they are seen during working hours. The last row shows the combined 95% of time coverage provided by the top 8 routers.[]{data-label="fig:subsampling"}](subsampling_curves.png){width="1\columnwidth"} Discussion {#discussion .unnumbered} ========== Our world is becoming progressively more enclosed in infrastructures supporting communication, mobility, payments, or advertising. Logs from mobile phone networks have originally been considered only for billing purposes and internal network optimization; today they constitute a global database of human mobility and communication networks[[ ]{}]{}[@gonzalez2008understanding]. Credit card records form high-resolution traces of our spending behaviors[[ ]{}]{}[@krumme2013predictability]. The omnipresent WiFi networks, intended primarily for communication, has now became a location tracking infrastructure, as described here. The pattern is clear: every new cell tower, merchant with credit card terminal, every new private or municipal WiFi network offer benefits to the connected society, but, at the same time, create opportunities and perils of unexpected tracking. Cities entirely covered by WiFi signal provide unprecedented connectivity to citizens and visitors alike; at the same time multiple parties have to incorporate this fact in their policies to limit privacy abuse of such infrastructure. Understanding and quantifying the dynamics of privacy and utility of infrastructures is crucial for building connected and free society. Since the creation of comprehensive databases containing geolocation for APs is primarily carried out by large companies[[ ]{}]{}[@skyhook], one might assume that WiFi based location tracking by ‘small players’, such as research studies or mobile applications, is not feasible. As we have shown above, however, APs can be very efficiently geolocated in a way that covers a large majority of individuals’ mobility patterns. In the results, we focused on outdoor positioning with spatial resolution corresponding to WiFi AP coverage: we assume that if at least one AP is discovered in a scan, we can assign the location of this AP to the user. This is a deliberately simple model, described in detail in Supplementary Information, but we consider the resulting spatial resolution sufficient for many aspects of research, such as studying human mobility patterns. The spatial resolution of dozens of meters is higher than for example CDR data[[ ]{}]{}[@gonzalez2008understanding], which describes the location with the accuracy of hundreds of meters to a few kilometers. Incorporating WiFi routers as location beacons can aid research by drastically increasing temporal resolution without additional cost in battery drain. Students live in multiple dormitories on and outside of campus, take multiple routes commuting to the university, frequent different places in the city, travel across the country and beyond. While the students spend most of their time within a few dozens of kilometers from their homes, they also make international and intercontinental trips (see Supplementary Figures 2 and 3 for details). Such long distance trips are not normally captured in studies based on telecom operator data. Our population is densely-connected and in this respect it is biased, in the same sense as any population of people working in the same location. We do simulate a scenario in which the individuals do not form a connected group by analyzing the results for personal-only database. We expect the obtained results to generalize outside of our study. Our findings connect to an ongoing debate about the privacy of personal data[[ ]{}]{}[@strandburg2014privacy]. Location data has been shown to be among the most sensitive categories of personal information[[ ]{}]{}[@staiano2014money]. Still, a record of WiFi scans is, in most contexts, not considered a location channel. In the Android ecosystem, which constitutes 85% of global smartphone market in Q2 2014[[ ]{}]{}[@android], the permission for applications to passively collect the results of WiFi scans is separate from the location permission; moreover, the *Wi-Fi connection information (ACCESS\_WIFI\_STATE)* permission is not considered ‘dangerous’ in the Android framework, whereas both high-accuracy and coarse location permissions are tagged as such[[ ]{}]{}[@android_permissions]. While it has been pointed out that Android WiFi permissions may allow for inference of sensitive personal information[[ ]{}]{}[@achara2014wifileaks], the effect has not been quantified through real-world data. Here we have shown that inferring location with high temporal resolution can be efficiently achieved using only a small percentage of the WiFi APs seen by a device. This makes it possible for any application to collect scanned access points, report them back, and inexpensively convert these access points into users’ locations. The impact is amplified by the fact that apps may passively obtain results of scans routinely performed by Android system every 15–60 seconds. Such routine scans are even run when the user disables WiFi. See Supplementary Information for additional analysis on data privacy in the Android ecosystem. Developers whose applications declare both location and WiFi permissions are able to use WiFi information to boost the temporal resolution of any collected location information. We have shown that even if the location permission is revoked by the user, or removed by the app developers, an initial collection of both GPS and WiFi data is sufficient to continue high-resolution tracking of the user mobility for subsequent months. Many top applications in the Play Store request *Wi-Fi connection information* but not explicit location permission. Examples from the top charts include prominent apps with more than 100 million users each, such as Candy Crush Saga, Pandora, and Angry Birds, among others. We are not suggesting that these or other applications collect WiFi data for location tracking. These apps, however, do have a *de facto* capability to track location, effectively circumventing Android permission model and general public understanding. Due to uniqueness of location traces, users can be easily identified across multiple datasets[[ ]{}]{}[@de2013unique]. Our results indicate that any application can use WiFi permission to link users to other public and private identities, using data from Twitter or Facebook (based on geo-tagged tweets and posts), CDR data, geo-tagged payment transactions; in fact any geo-tagged data set. Such cross-linking is another argument why WiFi scans should be considered a highly sensitive type of data. In our dataset, 92% of WiFi scans have at least one visible AP. Even in the most challenging scenario, when there are no globally shared locations and each individual frequents different places, top 20 WiFi access points per person can be efficiently converted into geolocations (using Google APIs or crowd-sourced data) and used as a stable location channel. These results should inform future thinking regarding the collection, use, and data security of WiFi scans. We recommend that WiFi records be treated as strictly as location data. Methods {#methods .unnumbered} ======= The dataset. {#the-dataset. .unnumbered} ------------ Out of the 130+ participants of the study[[ ]{}]{}[@10.1371/journal.pone.0095978], we selected 63 for which at least 50% of the expected data points are available. The methods of collection, anonymization, and storage of data were approved by the Danish Data Protection Agency, and complies both with local and EU regulations. Written informed consent was obtained via electronic means, where all invited participants read and digitally signed the form with their university credentials. The median period of WiFi scans for these users was 16 seconds, and the median period of GPS sampling was 10 minutes. The data spans a period of 200 days from October 1st, 2012 to April 27th, 2013. *Known routers* and *coverage* {#known-routers-and-coverage .unnumbered} ------------------------------ In the article we use a simple model of locating the WiFi routers. We consider an access point as *known* if it occurred in a WiFi scan within one second of a GPS location estimation. The shortcomings of this approach and possible remedies are described in more detail in Supplementary Information. We define *time coverage* as a fraction of ten-minute bins containing WiFi data in which at least one *known router* was scanned. For example, let us assume that the user has data in 100 out of 144 timebins during a day, and in 80 of these timebins there is a known router visible. Therefore, that user’s coverage for that day is 80%. The average time coverage for a day is the mean coverage of all users who had any WiFi information in that day. This way our results are independent from missing data caused by imperfections in data collection system deployed in the study. In Figure \[fig:coverage\] we present three different approaches to sampling, which we describe here in detail. **Initial-period sampling.** As presented in Figure \[fig:coverage\]a, we learn the location of the routers sequentially. With each GPS location estimation accompanied with a WiFi scan, we add the visible access points to the list of known routers. The learning curve can be observed for the first seven days (Figure \[fig:coverage\]a, left panel) or the first 28 days (Figure \[fig:coverage\]a, right panel). **Random subsampling.** In the random subsampling scenario we select a set fraction of available GPS location estimations, each paired with a WiFi scan. Each GPS estimation provides information on the position of all routers seen in the paired scan. This scenario can be realized after the data collection is finished, as the location estimations are used to locate the WiFi scans which happened both before and after said estimations. The results are presented in Figure \[fig:coverage\]b. **Top routers.** We select the top routers in a greedy fashion after the data collection is finished. We sort the routers in descending order by the number of user timebins they occur in. We choose the top one router, and then we select the routers which provide the biggest increase in the number of user timbebins covered. Due to high density of access points, each semantic place is described by presence of several routers, but location of only one of them has to be established to find the geographic position of the place. In this sampling method we do not rely on our own GPS data — top routers are found purely based on their occurrence in the WiFi scans, regardless of availability of GPS scans within the one second time delta. The results of such sampling are presented in Figure \[fig:coverage\]e. ![**The time coverage provided by the routers with known position depends on who collects the corresponding location data and when it happens.** In each subplot the orange line describes the scenario where each individual collects data about themselves and does not share it with others; the blue line corresponds to a system in which the location of routers discovered by one person is made known to other users; the green line presents a situation where each individual can use the common pool of known routers but does not discover access points herself. [**a. Stability of location.**]{} Learning the location of APs seen during the first seven (left panel) or 28 (right panel) days, leads to performance gradually decreasing with time in the personal case (orange line). The histograms of time coverage distribution for day 190 show that this decline is driven by a growing number of people who spend only ${\sim}$10% of time in the locations they visited in the beginning of the observation. The global approach (blue line) does not show this tendency, which indicates that people rotate between common locations rather than moving to entirely new places. [**b, c. Representativeness of randomly selected locations.**]{} Random subsampling with an average period of 24 hours (less than 1% of all available location estimations) is sufficient to find the most important locations in which people spend more than 80% of their time; using an average period of 4 hours (2.5% percent of all available location data) results in ${\sim}$85% coverage. The personal database does not expire since the location is sampled throughout the experiment, not only in the beginning. [**d. Limited number of important locations.**]{} Although querying commercial services for WiFi geolocation is costly, knowing the location of only the 20 most prevalent routers per person in the dataset results in an average coverage of ${\sim}$90%. Since people’s mobility overlaps, there is a benefit of using a global database rather than treating all mobility disjointly.[]{data-label="fig:coverage"}](coverage.png){width="1\columnwidth"} Data collection scenarios. {#data-collection-scenarios. .unnumbered} -------------------------- Each subplot in Figure \[fig:coverage\] contains series coming from three different simulated collection scenarios. In the **global** scenario, there is a pool of WiFi routers locations estimations coming from all users, and a router is considered known if at least one person has found its location. This scenario simulates the function of such services as for example mobile Google Maps. In the **personal** scenario each user can only use their own data, a router can be known to them only if they found its location themselves. It simulates collecting data in a disjoint society, where each person frequents different locations. Finally, in the **global with no personal data** scenario, each user can exploit estimations created by everybody else, but without contributing their own data. [ **Acknowledgements.** We thank Yves-Alexandre de Montjoye and Andrea Cuttone for useful discussions.\ **Funding.** The research project is supported by the Villum Foudation, as well as the University of Copenhagen through Programmes of Excellence for Interdisciplinary Research.\ **Contributions.** Designed the research - SL, RG, PS, AS; Analysed the data - PS, RG; wrote and reviewed the manuscript - PS, AS, RG, SL.\ **Additional Information.** The authors declare no competing financial interests. ]{} [10]{} url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} *et al.* . ** ****, (). , , & . ** ****, (). , & . In ** (, ). . . , & . In **, vol. , (). . In **, (, ). *et al.* . ** ****, (). , , & . ** **** (). , , & . ** **** (). , , , & . ** ****, (). <http://bit.ly/1w6k6bB>. , & . ** ****, (). , , & . ** ****, (). , & . ** ****, (). , & . ** (). <http://bit.ly/1yHpJiC>. , , & . ** ****, (). . . , , & . ** **** (). , , & . In **, (, ). *et al.* . ** (). *et al.* . ** ****, (). <http://bit.ly/1jWczTL>. . , , , *et al.* . ** **** (). *et al.* ** (, ). . . . . , , , *et al.* (). , , & . (, ). , , & . ** ****, (). . . Supplementary Information {#supplementary-information .unnumbered} ========================= Inferring location of routers. {#inferring-location-of-routers. .unnumbered} ------------------------------ In the article we use a deliberately simplistic model of locating the WiFi routers. We assume that if we find a WiFi scan and a GPS location estimation which happened within a one second time difference we can assume that all routers visible in the scan are at the geographical location indicated by the GPS reading. Due to effective outdoor range of WiFi routers of approximately 100 meters, this assumption introduces an obvious limitation of accuracy of location inference. Moreover, there are a number of mobile access points such as routers installed in public transportation or smartphones with hotspot capabilities. Such devices cannot effectively be used as location beacons and will introduce noise into location estimations unless identified and discarded. We propose and test the following method. For each GPS location estimation with timestamp $t_{GPS}$ we find WiFi scans performed by the same device at $t_{WiFi}$ so that $t_{GPS}-1s \leq t_{WiFi} \leq t_{GPS}+1s$ and select the one, for which $|t_{GPS}-t_{WiFi}|$ is the smallest. We then add the location estimation and its timestamp to the list of locations where each of the available WiFi access points was seen. For each device, we fit a density-based spatial clustering of applications with noise (DBSCAN) model[[ ]{}]{}[@dbscan] specifying 100 meters as the maximum distance parameter $\varepsilon$. If there are no clusters found, or the found clusters contain less than 95% of all locations associated with the said router we assume the router is mobile and to be discarded from further analysis. If only one cluster is detected and it contains at least 95% of all points, we assume the geometric median of these points is the physical location of the router. If there are more clusters found and they contain at least 95% of all points, we verify if these clusters are disjoint in time: if the timestamps of sightings do not overlap between those clusters, we assume the device is a static access point which has been moved to a different place during the experiment. Otherwise, we classify the access point as a mobile device and do not use it as a location proxy. In the proposed method we assume accuracy of tens of meters is satisfactory, and hence do not find a need to exploit the received signal strength information[[ ]{}]{}[@liu2007survey]. Arguably, with the sparse data that we operate on, employing received signal strength could lead to more confusion, as it can vary greatly for one location, depending on the position of the measuring smartphone, and presence of humans and other objects obstructing the signal. Supplementary Figure \[fig:wifi\_rssi\] shows timeseries of signal strengths received by a non-moving smartphone, which vary as much as 10dB, which corresponds to drastic differences in estimated distance to the source, as in free-space propagation model extending the distance $\sqrt{10} \approx 3.16$ times corresponds to 10dB loss in received signal strength. In the 200 days of observations, the participants have scanned 487216 unique routers, out of which 64983 were scanned within a second of a GPS estimation. As many as 57912 were only seen less than five times which we assumed to be the minimum number of sightings to be considered a cluster, which left only 7071 routers for further investigation. In 1760 cases there were no clusters found, or there was more than 5% noise. In 5267 cases there was only one cluster and less than 5% of noise. Out of 21 cases there were multiple clusters and less than 5% of noise, 9 revealed no time overlap between clusters. We verified our heuristic of determining which routers are mobile by classifying routers which are very likely mobile, as their networks are called AndroidAP (default SSID for a hotspot on Android smarphones), iPhone (default SSID for iPhones), Bedrebustur or Commutenet (names of networks on buses and trains in Copenhagen). Out of 340 such devices 323, or 95%, were identified as mobile, and 17 as fixed-location devices. All in all, out of 487216 unique APs we believe we managed to estimate the location of 5276, we identified 1771 as mobile, and did not have enough data to investigate 480169. Even though we only know the location of approximately 1% of all sensed routers, this knowledge is enough to estimate the location of users in 87% ten-minute timebins in the dataset. Long term stability and low entropy of human mobility. {#long-term-stability-and-low-entropy-of-human-mobility. .unnumbered} ------------------------------------------------------ Long-term stability in the context of human mobility means that individuals keep returning to the same locations over long time periods. Arguably, most people do not often move, change the work place, or find an entirely new set friends to visit. We use entropy in Shanon’s definition, as presented in equation (\[eq:entropy\]) $$\label{eq:entropy} H(X) = -\sum_{i}{P(x_i)\log{P(x_i)}},$$ where $X$ is the set of all possible locations, and $P(x_i)$ is the probability of a person being at location $i$. Therefore, the bigger the fraction of time a person spends in their top few places, the lower the entropy value of that person’s mobility. In this sense, long-term stability is necessary for the low entropy, and both contribute to the predictability of human mobility. ![Received signal strength can vary greatly even if the smartphone and the access points do not move.[]{data-label="fig:wifi_rssi"}](screenshot.png){width="1\columnwidth"} Mobility of the studied population. {#mobility-of-the-studied-population. .unnumbered} ----------------------------------- This article focuses on a population of students at a university. To show that their mobility is not constrained to the campus only, we present summary statistics about their mobility. Displacements in our dataset can be as big as 10000 km. Given such extreme statistics, the radius of gyration, while commonly used in literature to describe mobility on smaller scales [@gonzalez2008understanding], is not a suitable measure here. Instead, in Supplementary Figure \[fig:world\_map\] we show a qualitative overview in form of a heatmap of observed locations, as well as a distribution of time spent as a function of distance from home. For simplicity, we define the home location for each student as the location of the most prevalent access point in their data. We then calculate the median distance from home for each hour of the observation using their location data. For a more detailed view, we present the distribution for 48 randomly chosen students in Supplementary Figure \[fig:histograms\]. ![The article focuses on a population of students at a single university, but they are not constrained to the campus only. Our data captures human mobility at different scales: the participants spend most of their time at home (1), but they travel around the neighborhood (2), the city (3), to different cities in Denmark (4), different cities in Europe (5), and finally, other continents (6).[]{data-label="fig:world_map"}](world_map.png "fig:"){width=".875\columnwidth"}\ ![The article focuses on a population of students at a single university, but they are not constrained to the campus only. Our data captures human mobility at different scales: the participants spend most of their time at home (1), but they travel around the neighborhood (2), the city (3), to different cities in Denmark (4), different cities in Europe (5), and finally, other continents (6).[]{data-label="fig:world_map"}](general_population.png "fig:"){width="1\columnwidth"} ![Distribution of time spent at different distances from the inferred home location, presented for randomly selected 48 participants. In most cases, we see the home location as the most prevalent, and probably a “work” location as the next peak in the distribution.[]{data-label="fig:histograms"}](histograms.png){width="\columnwidth"} Time coverage of top routers. {#time-coverage-of-top-routers. .unnumbered} ----------------------------- In this section we present a more detailed view on time coverage of top routers selected separately for each person. Supplementary Figure \[fig:top\_locations\]A shows the fraction of time which participants spent near to one of their top 20 routers. It is worth noting, that while home location is immediately apparent, there seems to be no definite “work” location in our population. This can be attributed to the fact that the participants of the observation are students who attend classes in different buildings and lecture halls and do not have an equivalent of an office. Supplementary Figure \[fig:top\_locations\]B is an enriched version of Figure 2d from the main text of the article. It shows that even though 20 routers are needed on average to capture 90% of mobility, there are participants for whom just four routers suffice. ![A more detailed view of time coverage provided by top routers found through the greedy algorithm. A: there is a clear main location for a majority of participants, we therefore assume this to be the home location. B: even though 20 routers are needed on average to capture 90% of mobility, there are participants for whom just four routers suffice.[]{data-label="fig:top_locations"}](top_locations_illu.png){width="1\columnwidth"} Android Permissions. {#android-permissions. .unnumbered} -------------------- The scope of Android permission *ACCESS\_WIFI\_STATE* is described in the developer documentation as “allows applications to access information about Wi-Fi networks”[[ ]{}]{}[@android_permission_1]. This permission provides the requesting application with a list of all visible access points along with their MAC identifiers after each scan ordered by any application on the phone (via broadcast mechanism). Moreover, with this permission the applications can start in the background when the first WiFi scan results appear after the phone boots: the app’s BroadcastReceiver is called and the data can be collected without explicit *RECEIVE\_BOOT\_COMPLETED* permission. Requesting a WiFi scan requires the *CHANGE\_WIFI\_STATE* permission, marked as dangerous, but in most cases it is not necessary to request it: the Android OS by default performs WiFi scans in the intervals of tens of seconds, even when the WiFi is turned off; the setting to *disable* background scanning when WiFi is off is buried in the advanced settings. Application developers often use *ACCESS\_WIFI\_STATE* to obtain information whether the device is connected to the Internet via mobile or WiFi network. This information is useful, for example, to perform larger downloads only when the user in connected to a WiFi network and thus avoid using mobile data. This is an unnecessarily broad permission to use for this purpose, as the same information can be obtained with *ACCESS\_NETWORK\_STATE*, which provides all the necessary information without giving access to personal data of WiFi scans: ConnectivityManager cManager = (ConnectivityManager) getSystemService(Context.CONNECTIVITY_SERVICE); NetworkInfo mWifi = cManager.getNetworkInfo(ConnectivityManager.TYPE_WIFI); if (mWifi.isConnected()) { } //wifi is connected Since the *ACCESS\_WIFI\_STATE* together with *INTERNET* permission (for uploading the results) are effectively sufficient for high-resolution location tracking, we suggest the developers transition to using the correct permissions and APIs for determining connectivity and that accessing the result of WiFi scan requires at least the *ACCESS\_COARSE\_LOCATION* permission.

ArXiv

--- abstract: 'The XMM-OM instrument extends the spectral coverage of the XMM-Newton observatory into the ultraviolet and optical range. It provides imaging and time-resolved data on targets simultaneously with observations in the EPIC and RGS. It also has the ability to track stars in its field of view, thus providing an improved post-facto aspect solution for the spacecraft. An overview of the XMM-OM and its operation is given, together with current information on the performance of the instrument.' author: - 'K. O. Mason, A. Breeveld, R. Much, M. Carter, F. A. Cordova, M. S. Cropper, J. Fordham, H. Huckle, C. Ho, H. Kawakami, J. Kennea, T. Kennedy, J. Mittaz, D. Pandel, W. C. Priedhorsky, T. Sasseen, R. Shirey, P. Smith, J.-M. Vreux' title: 'The XMM-Newton Optical/UV Monitor Telescope' --- Introduction ============ The Optical/UV Monitor Telescope (XMM-OM) is a standalone instrument that is mounted on the mirror support platform of XMM-Newton (Jansen et al. 2001) alongside the X-ray mirror modules. It provides coverage between 170 nm and 650 nm of the central 17 arc minute square region of the X-ray field of view (FOV), permitting routine multiwavelength observations of XMM targets simultaneously in the X-ray and ultraviolet/optical bands. Because of the low sky background in space, XMM-OM is able to achieve impressive imaging sensitivity compared to a similar instrument on the ground, and can detect a $B=23.5$ magnitude A-type star in a 1000 s integration in “white” light (6 sigma). It is equipped with a set of broadband filters for colour discrimination. The instrument also has grisms for low-resolution spectroscopy, and an image expander (Magnifier) for improved spatial resolution of sources. Fast timing data can be obtained on sources of interest simultaneously with image data over a larger field. In the following sections we give an overview of the instrument followed by an account of its operation in orbit and the instrument characteristics. Instrument overview =================== The XMM-OM consists of a Telescope Module and a separate Digital Electronics Module, of which there are two identical units for redundancy (see Fig. 1). The Telescope Module contains the telescope optics and detectors, the detector processing electronics and power supply. There are two distinct detector chains, again for redundancy. The Digital Electronics Module houses the Instrument Control Unit, which handles communications with the spacecraft and commanding of the instrument, and the Data Processing Unit, which pre-processes the data from the instrument before it is telemetered to the ground. (8.8,10) (10,0) Optics ------ The XMM-OM uses a Ritchey Chrétien telescope design modified by field flattening optics built into the detector window. The f/2 primary mirror has a 0.3 m diameter and feeds a hyperboloid secondary which modifies the f-ratio to 12.7. A $45$[$^{\circ}$]{} flat mirror located behind the primary can be rotated to address one of the two redundant detector chains. In each chain there is a filter wheel and detector system. The filter wheel has 11 apertures, one of which is blanked off to serve as a shutter, preventing light from reaching the detector. Another seven filter locations house lenticular filters, six of which constitute a set of broad band filters for colour discrimination in the UV and optical between 180 nm and 580 nm (see Table 2 for a list of filters and their wavelength bands). The seventh is a “white light” filter which transmits light over the full range of the detector to give maximum sensitivity to point sources. The remaining filter positions contain two grisms, one optimised for the UV and the other for the optical range, and a 4 field expander (or Magnifier) to provide high spatial resolution in a 380–650 nm band of the central portion of the (FOV). Detector -------- The detector is a microchannelplate-intensified CCD (Fordham et al. 1992). Incoming photons are converted into photoelectrons in an S20 photocathode deposited on the inside of the detector window. The photoelectrons are proximity focussed onto a microchannelplate stack, which amplifies the signal by a factor of a million, before the resulting electrons are converted back into photons by a P46 phosphor screen. Light from the phosphor screen is passed through a fibre taper which compensates for the difference in physical size between the microchannelplate stack and the fast-scan CCD used to detect the photons. The resulting photon splash on the CCD covers several neighbouring CCD pixels (with a FWHM of approximately 1.1 CCD pixels, if fitted with a Gaussian). The splash is centroided, using a 33 CCD pixel subarray to yield the position of the incoming photon to a fraction of a CCD pixel (Kawakami et al. 1994). An active area of 256256 CCD pixels is used, and incoming photon events are centroided to 1/8th of a CCD pixel to yield 20482048 pixels on the sky, each 0.4765 arc seconds square. In this paper, to avoid confusion, while CCD pixels (256256 in FOV) will be referred to explicitly, a pixel refers to a centroided pixel (20482048 in FOV). As described later, images are normally taken with pixels binned 22 or at full sampling. The CCD is read out rapidly (every 11 ms if the full CCD format is being used) to maximise the coincidence threshold (see sect. 5.2). Telescope mechanical configuration ---------------------------------- The XMM-OM telescope module consists of a stray light baffle and a primary and secondary mirror assembly, followed by the detector module, detector processing electronics and telescope module power supply unit. The separation of the primary and secondary mirrors is critical to achieving the image quality of the telescope. The separation is maintained to a level of 2 by invar support rods that connect the secondary spider to the primary mirror mount. Heat generated by the detector electronics is transferred to the baffle by heat pipes spaced azimuthally around the telescope, and radiated into space. In this way the telescope module is maintained in an isothermal condition, at a similar temperature to the mirror support platform. This minimizes changes in the primary/secondary mirror separation due to thermal stresses in the invar rods. Fine focussing of the telescope is achieved through two sets of commandable heaters. One set of heaters is mounted on the invar support rods. When these heaters are activated, they cause the rods to expand, separating the primary and secondary mirrors. A second set of heaters on the secondary mirror support brings the secondary mirror closer to the primary when activated. The total range of fine focus adjustment available is $\pm10\mu$m. The filter wheel is powered by stepper motor, which drives the wheel in one direction only. The filters are arranged taking into account the need to distribute the more massive elements (grisms, Magnifier) uniformly across the wheel. Digital Electronics Module -------------------------- There are two identical Digital Electronics Modules (DEM) serving respectively the two redundant detector chains. These units are mounted on the mirror support platform, separate from the telescope module. Each DEM contains an Instrument Control Unit (ICU) and a Digital Processing Unit (DPU). The ICU commands the XMM-OM and handles communications between the XMM-OM and the spacecraft. The DPU is an image processing computer that digests the raw data from the instrument and applies a non-destructive compression algorithm before the data are telemetered to the ground via the ICU. The DPU supports two main science data collection modes, which can be used simultaneously. In Fast Mode, data from a small region of the detector are assembled into time bins. In Image Mode, data from a large region are extracted to create an image. These modes are described in more detail in the next sect. The DPU autonomously selects up to 10 guide stars from the full XMM-OM image and monitors their position in detector coordinates at intervals that are typically set in the range 10–20 seconds, referred to as a tracking frame. These data provide a record of the drift of the spacecraft during the observation accurate to $\sim 0.1$ arc second. The drift data are used within the DPU to correct Image Mode data for spacecraft drift (see sect. 5.5). Observing with XMM-OM ===================== Specifying windows ------------------ The full FOV of XMM-OM is a square 1717 arc minutes, covering the central portion of the X-ray FOV. Within this field the observer can define a number of data collection windows around targets or fields of interest. Up to five different Science Windows can be defined with the restriction that their boundaries may not overlap. However, one window can be completely contained within another. Because of constraints on the telemetry rate available, it is not possible to transmit the full data on every photon that XMM-OM detects. Instead a choice has to be made between image coverage and time resolution. Thus two types of Science Window can be defined, referred to as Image Mode and Fast Mode. A maximum of two of the five available science windows can be Fast Mode. *Image Mode emphasizes spatial coverage at the expense of timing information. Images can be taken at the full sampling of the instrument or binned by a factor of 2 or 4, to yield a resolution element on the sky of approximately 0.5, 1.0 or 2.0 arc seconds (a factor of four finer for the Magnifier). The maximum total size of the Science Windows is determined by the memory available in the DPU. A single Image Mode window binned by a factor of 22 can be up to 976960 detector pixels, which results in a 488480 binned pixel image being stored in the DPU. At full sampling (with no binning) the window can be up to 652652 pixels. Any drift in the pointing direction of the spacecraft is corrected in the image by tracking guide stars (section 5.5).* *Fast Mode emphasizes timing information at the expense of spatial coverage. The maximum total number of pixels that can be specified for a Fast Mode window is 512. Thus the maximum size of an approximately square window would be 2223 pixels ( = 506 total). Note that there is no binning within a Fast Mode window. The pixel locations of individual photons within the window are recorded and assigned a time tag, which has a user-specified integration time of between 100 ms and the tracking frame duration (10–20 s). No tracking correction is applied to Fast Mode data. This can be applied on the ground, from the drift history supplied by XMM-OM.* To simplify observation set-up, two standard observing sequences of five exposures have been created that together cover the whole XMM-OM FOV at one arc second sampling while at the same time monitoring a central target at full spatial sampling (0.5 arcsec). In the first variant, each of the five exposures contains an unbinned Image Mode window centred on the prime instrument’s boresight (the position of the main target), and a second Image Mode window, binned by 22 pixels, that is defined in each of the set of five exposures so as to form a mosaic of the entire field (see Fig. 2). The second variant is exactly the same as the first except that a Fast Mode window is added around the prime instrument’s boresight position. (8.8,8) (0,0) The length of an XMM-OM Image Mode only exposure can be set in the range 800–5000 s. However it should be noted that there is an approximately 300 second overhead associated with each individual exposure. The maximum length of an exposure that contains a single Fast Mode window is 4400 s, or 2200 s if there are two Fast Mode windows. At the time of writing a further mode is being commissioned which allows the full field to be imaged at 1 arcsec sampling in one go, at the expense of tracking information and correction. This is made possible by the impressive stability of the XMM-Newton spacecraft compared to pre-launch expectation. Window coordinates can be specified either in detector pixels, or in sky coordinates. To facilitate the latter, the XMM-OM performs a short V-band observation at the start of each pointing. The DPU compares the image with the positions of uploaded field stars to calibrate the absolute pointing of the OM. Filter selection ---------------- The XMM-OM filter wheel rotates in one direction only and, to conserve the total number of wheel rotations over the expected lifetime of XMM-Newton, the number of filter wheel rotations per pointing is limited to one (unless there are very strong scientific arguments for more). Thus filter observations have to be executed in a particular order during a given target pointing. The filter elements are listed in the order they occur in the filter wheel in Table 2. The instrument is slewed with the blocked filter in place, and thereafter a field acquisition exposure is performed in V. The same telescope focus setting is used for all the filters except for the Magnifier (see sect. 5.6.1), where the optimum focus is different (the image quality is the most sensitive to focus position when using the Magnifier). The XMM-OM instrument is optimised for the detection of faint sources. If the source count-rate is too high the response of the detector is non-linear. This “coincidence loss” occurs when the probability of more than one photon splash being detected on a given CCD pixel within the same CCD readout frame becomes significant. Coincidence loss is discussed in more detail in sect. 5.2. If a source is predicted to exceed the coincidence threshold for a given filter, then a different filter with lower throughput can be selected. Alternatively a grism can be selected which disperses the available light over many pixels. The XMM-OM detectors can also be permanently damaged by exposure to a source that is too bright, reducing both the quantum efficiency of the photocathode and the gain of the channelplates. This is a cumulative effect dependent on the total number of photons seen over the lifetime of the instrument at a particular location on the detector. The deterioration is therefore more severe for longer observations of a bright source. For this reason limits are imposed on the maximum brightness of stars in the FOV (see Table 1) and apply to any star in the FOV irrespective of whether it is within a science window or not. Even more stringent limits are applied to the central region of the detector that will usually contain the target of interest. In the event that there is a star in (or near to) the FOV that violates the brightness constraints, a different filter, which has lower photon throughput, can be selected. Also, if the bandpass is appropriate, the Magnifier can be used to exclude bright stars further than a few arc minutes from the field centre. The grisms (one optimised for the UV, the other for the optical) form a dispersed first order image on the detector, together with a zeroth order image that is displaced in the dispersion direction. The counts in the zeroth order image of field stars determines the brightness limits used for observing with the grisms. OM performance ============== (8.8,10) (0,10) The first light observation for XMM-OM took place on 2000 January 11. Since then the various engineering and science data taking modes of the instrument have been commissioned including full-field image engineering mode (not generally available for science observations because of the very large telemetry overhead required to transmit the data to the ground), and the Image and Fast science modes. The telescope focus has been optimised using the heater-based fine focus control, and the gain of the image intensifier has been optimised. The performance of the DPU in tracking image motion due to spacecraft drift has also been verified and distortion maps derived to relate XMM-OM detector coordinates to the sky. Photometric calibrations have been derived for all filter elements, but work continues on colour equations and to tie these more accurately into standard systems. Similarly, preliminary throughput and wavelength calibrations have been derived for the Grisms. To illustrate the capabilities of XMM-OM, we show in Figure 3 images of part of the Lockman Hole field in the White Light filter, and in five of the six colour filters (the remaining filter, UVM2, was not used during this observation). The images contains an $R=18.1$ magnitude AGN identified in the ROSAT observation of the field, and referred to as R32 by Schmidt et al.  (1998). The AGN is clearly UV bright. The XMM-OM detects approximately 12from the AGN in White Light, while the count rate in the colour filters ranges from a high of 2.9 in U, down to about 0.25 in the UV filter UVW2. To illustrate the spectral capability of XMM-OM, we show in Figure 4 the extracted spectrum of the DA white dwarf standard BPM16274. The Balmer absorption lines can be clearly discerned. (8.8,7) (0,0) Analysis issues =============== Throughput ---------- An initial estimate of the zero points of the various XMM-OM broadband filters (i.e. the magnitude which yields 1 count per second; Table 2.) was derived from calibration observations of two white dwarfs. - The OM response throughput was determined based on measurements of the spectrophotometric standards BPM16274 and LBB227. - Using the OM throughput model an OM exposure of Vega was simulated. - The zeropoints in U-, B-, V-filter were fixed in a way that the brightness of Vega matches the literature values. - In the UV-filters the brightness of Vega was set to 0.025 mag; the average [*U*]{} magnitude of Vega in the literature being the most appropriate for the UV filters, and to 0.03 mag (the average Vega [*V*]{} magnitude) for the broadband Magnifier and the White Light filters. The calculated OM zeropoints are written into the relevant XMM-CCF file, to be used by the XMM Science Analysis System (SAS) (Watson et al. 2001). Updates of the zeropoint definitions as well as more precise colour transformations (Royer et al. 2000) to the standard UBV system are expected once the results of a dedicated ground based photometric observation programme become available. In the framework of this programme several OM calibration fields are anchored to high quality secondary photometric standards deep fields established by ESO. Early results of this ground observations are expected in October 2000. From analysis of the Lockman Hole field, avoiding the central 1 arc minute of the FOV where the background is enhanced (section 5.6.4) the limiting magnitude after 1000 s is calculated to be 21.0 in [*V*]{}, 22.0 in [*B*]{} and 21.5 in [*U*]{} (6 sigma). The limiting magnitude for the White Light filter is very dependant on the spectral type of the star, because the bandpass is so broad. However, for an A0 star we estimate that the 6 sigma limiting magnitude above background is $\sim$23.5. Coincidence loss and deadtime ----------------------------- Coincidence loss is observed whenever the count rate is such that more than one photon arrives in the same place within a given readout frame. Losses become significant for a point source at a count rate of about $10$ (for 10% coincidence) when the full CCD chip is being readout (i.e. about 2.5 magnitudes brighter than the zero points listed in Table 2). A factor of approximately two improvement can be achieved by restricting the area of the CCD used, since this reduces the time required to readout the chip. The coincidence loss can be approximated by $$ph_{\rm in}= { \log{(1-{cts}_{\rm detected}*T)} \over T_{\rm ft} - T } \label{equation1}$$ where\ ------------------------ ---------------------------------------- $ph_{\rm in}$ infalling photon rate per second ${cts}_{\rm detected}$ measured count rate per second $T$ CCD frametime in units of seconds ${T}_{\rm ft}$ frametransfer time in units of seconds ------------------------ ---------------------------------------- Equation 1 applies strictly in the case of a perfectly point-like source. In practice a real stellar profile has wings, and the formula will break down at very high rates when coincidence among photons in the wings of the profile becomes significant. The CCD deadtime depends on the size and shape of the science window used but can be calculated accurately. The deadtime correction should be applied by the SAS after any coincidence loss corrections. Flat fields ----------- An LED can be used to illuminate the detector by backscatter of the photons from the blocked filter. These images are not completely flat due to the illumination pattern of the LED, the gross shape of which could be removed by comparing with sky flats. However, using the LED allows a large number of events to be collected in every pixel to give sufficiently high statistics for pixel to pixel sensitivity to be measured and the relative measurement of any variation of the detector response on a fine scale. The LED brightness is adjustable and is currently operated at a level that produces $3.25\exp{-3}$  per binned (22) pixel. So far, flat fields have been obtained to the level of 400 counts per binned (22) pixel allowing an accuracy of 5% in the sensitivity measurement. A CCF file in the SAS currently represents the accuracy of flat fields obtained before mid June, which is at the 10% level. Once sufficient flat fields have been obtained for a 2–3% sensitivity the relevant CCF file will be updated. Background ---------- The background count rate in the OM is dominated by the zodiacal light in the optical. In the far UV the intrinsic detector background becomes important. Images are regularly taken with the blocked filter and no LED illumination to measure the detector dark counts. The mean OM dark count rate is $2.56\exp{-4}$  per pixel. The variation in dark count rate across the detector is $\pm 9\%$ and shows mainly a radial dependence, being highest in an annulus at about 8 arcmin radius and lowest in the centre. When the spacecraft is pointing at a very bright star, the dark rate is noticeably increased (e.g. up to 65% higher for the $V=0$ star Capella) despite the blocked filter. Excluding those dark frames taken during Capella ($V=0$) and Zeta Puppis ($V=2$) observations, the counts per dark frame vary by only $\pm 7\%$ and show no trend of change with time. Tracking performance -------------------- The positions of selected guide stars in the XMM-OM FOV are measured each 10–20 second tracking frame, and an X-Y offset applied to image mode data obtained during the tracking frame before they are added to the master image in the DPU memory. The tracking offsets are computed in pixels irrespective of the binning parameter chosen. Using this “Shift and Add” technique, the final image is corrected on timescales greater than a few tens of seconds and on spatial scales down to $\sim 0.5$ pixels, for drift in the pointing direction of the spacecraft. The performance of tracking can be verified by comparing the PSFs of stars taken during Fast Mode (at high time resolution and with no tracking) with those data taken using Image Mode when tracking is enabled. This analysis has shown that XMM-OM tracking is performing as expected. Analysis of OM tracking histories show that the spacecraft drift is less than 0.5 pixels for approximately 75% of all frames taken, and therefore require no shift and add correction (a shift of one pixel will be made if the guide stars are calculated to have drifted more than $\pm 0.5$ pixels from their reference positions). Of the remaining, the corrections due to drift are rarely more than 2 pixels in any one direction. Tracking is turned off automatically when no suitable guide stars are found, which is usually due to poor statistics. This can occur in observations of very sparse fields (rare) or when using the UVM2 and UVW2 filters, where throughput is lower than in the optical bands. However, given the pointing stability of XMM-Newton and the intrinsically poorer resolution of the detector in the UV (section 5.6), this does not normally lead to any significant degradation in the PSF for non-magnified data. Image quality ------------- ### Point Spread Function After launch the measured PSFs in the V-filter had FWHM widths broader than expected from preflight measurements. The focus was therefore adjusted using the control heaters as discussed in section 2.3. Fig. 5 shows the gradual change in the PSF with the heater setting. As can be seen from the figure, the optimum setting for the Magnifier is clearly at $-100$% i.e. at the minimum separation of the primary and secondary mirrors, whereas for the V-filter it is above 70%. A value of 100% (maximum separation of mirrors) was chosen for subsequent measurements in all filters, except for the Magnifier where $-100$% is selected. To allow for the thermal settling time involved in a change of focus, twenty minutes of additional overhead time is inserted before and after a sequence of Magnifier exposures. (8.8,5.5) (0,0) The PSFs contain a contribution from the telescope optics and from the detector. They can be assumed to be radially symmetric in shape, with an approximately gaussian central peak and extended wings. The width of the PSF increases with photon energy because of the detector component, from 3.1 pixels (1.5 arcsecs) FWHM in the V band to $\sim 6$ pixels (3 arcsecs) pixels in the UV filters (see Fig. 6). (8.8,6) (0,0) ### Distortion The XMM-OM optics, filters and (primarily) the detector system result in a certain amount of image distortion. It is mainly in the form of barrel distortion, and if not corrected can result in shifts from the expected position of up to 20 arcsecs. By comparing the expected position with the measured position for a large number of stars in the FOV a distortion map has been derived. The preliminary V-filter analysis was performed on the LMC pointing and is based on 230 sources. A 3rd order polynomial was fitted to the deviations assuming that there is no error at the centre of the FOV (i.e. at address (1024.5,1024.5)). This polynomial can be used to correct source positions measured in other fields, and currently gives a positional RMS accuracy of 1.0 arcsec (1.9 pixels) in the V-filter (see Fig. 7; astrometry relative to stars of known position over restricted regions of the field can of course be more accurate than this). Using higher orders of the polynomial does not increase the accuracy and is detrimental particularly for sources at the edges of the FOV. Using functions other than polynomials has not yet been investigated, but may lead to an improvement to the correction for sources near the edges of the FOV. Distortion maps using the 3C273 field have been derived for the other filters, but are not yet to such high accuracy. Further work will either use fields with more sources in the FOV or combine data from several observations. The preliminary distortion maps have been entered into the appropriate CCF files and can be used in conjunction with the SAS. They are also used on board to automatically position windows on the detector that are specified in sky coordinates. This is important for small windows such as those used in Fast Mode. (8.8,6) (0,0) ### Modulo-8 pattern As discussed earlier, the XMM-OM detector functions by centroiding a photon splash to within a fraction (1/8th) of a physical CCD pixel. This calculation is performed in real-time by the detector electronics, and therefore has to be fast. It is done by means of a lookup table whose parameters are computed onboard once per revolution, based on a short image taken with the internal flood LED lamp, and periodically updated. The lookup table parameters are the mean values derived from a selected part of the active area of the detector (usually the central region). They do not take into account small variations in the shape of the photon splash over the detector face and as such are an approximation to the optimum value at a given location on the detector. The result of imperfections in the lookup table is that the size of the pixels is not equal on the sky. When displayed with a normal image display routine, therefore, uncorrected XMM-OM images can exhibit a faint modulation in the apparent background level repeating every eight pixels, corresponding to every physical CCD pixel (see Fig. 8). SAS tasks that, for example, search for sources in XMM-OM images take the variation in pixel size into account and compute the local 88 pattern post facto based on the measured image. Similarly the raw image can be resampled for display purposes. The SAS routine does not lose or gain counts, but resamples them according to the true pixel sizes. The detector centroiding process also breaks down if more than one photon splash overlaps on a given CCD frame. Thus an 88 pattern is often seen around bright stars (see Fig. 8a), or when two bright stars occur close together on an image. ### Scattered light (10,12.0) (0,0) (0,6.5) Artifacts can appear in XMM-OM images due to light being scattered within the detector. These have two causes: internal reflection of light within the detector window and reflection of off-axis starlight and background light from part of the detector housing. The first of these causes a faint, out of focus ghost image of a bright star displaced in the radial direction away from the primary image due the curvature of the detector window (Fig. 8a). The second effect is due to light reflecting off a chamfer in the detector window housing. Bright stars that happen to fall in a narrow annulus 12.1 to 13 arc minutes off axis shine on the reflective ring and form extended loops of emission radiating from the centre of the detector (Fig. 8b). Similarly there is an enhanced “ring” of emission near the centre of the detector due to diffuse background light falling on the ring (Fig. 8b). The reflectivity of the ring, and of the detector window, reduces with increasing photon energy. Therefore these features are less prominent when using the UV filters. Conclusion ========== The first stage of commissioning and calibrating XMM-OM has been completed. The instrument is fulfilling its role of extending the spectral coverage of XMM-Newton into the ultraviolet and optical band, allowing routine observations of targets simultaneously with EPIC and RGS. Specifically, the instrument has successfully been demonstrated to provide wide field simultaneous imaging with the X-ray camera, simultaneous timing studies, and boresight information to arcsecond accuracy. A number of results illustrating the scientific potential of XMM-OM are contained within this volume. We would like to thank all the people who have contributed to the instrument; in its building, testing and operation in orbit, as well as those who have analysed the calibration data. The author list only contains a small fraction of those people involved. XMM-OM was built by a consortium led by the Principal Investigator, Prof. K.O.Mason, and comprising, in the UK, the Mullard Space Science Laboratory and Department of Physics and Astronomy, University College London; in the USA, University of California Santa Barbara, Los Alamos National Laboratory, & Sandia National Laboratory; and in Belgium, the Centre Spatial Liege & the University of Liege. JMV acknowledges support from the SSTC-Belgium under contract P4/05 and by the PRODEX XMM-OM Project. The U.S. investigators acknowledge support from NASA contract NAS5-97119. The UK contribution was supported by the PPARC. Based on observations obtained with XMM-Newton, an ESA science mission with instruments and contributions directly funded by ESA Member States and the USA (NASA) Fordham, J. L. A., Bone, D. A., Norton, R. J., Read, P. D. 1992, Proc. ESA Symp. on Photon Counting Detectors for Space Instrumentation, ESA SP-356, 103 Jansen, F., Lumb, D., Altieri, B. et al. 2001, A&A, 365 (this issue) Kawakami, H., Bone, D., Fordham, J.L.A., Michel, R. 1994, Nucl. Instrum. & Meth. A 348, 707 Royer, P., Manfroid, J., Gosset, E., Vreux J.-M. 2000, A&AS, 145, 351. Schmidt, M., Hasinger, G., Gunn, J. et al. 1998, A&A, 329, 495-503 Watson, M. G., Augeres, J.-L., Ballet, J. et al. 2001, A&A, 365 (this issue)

ArXiv

--- author: - 'M. Gillon$^{1}$, B.-O. Demory$^{2,3}$, C. Lovis$^4$, D. Deming$^5$, D. Ehrenreich$^4$, G. Lo Curto$^6$, M. Mayor$^4$, F. Pepe$^4$, D. Queloz$^{3, 4}$, S. Seager$^7$, D. Ségransan$^4$, S. Udry$^4$' date: 'Received date / accepted date' title: 'The $Spitzer$ search for the transits of HARPS low-mass planets - II. Null results for 19 planets[^1]' --- Introduction ============ Starting from 2004 (Butler et al. 2004, Santos et al. 2004), exoplanet search projects have been detecting planets of a few to $\sim$20 Earth masses at an ever-increasing rate, revealing them to be very frequent around solar-like stars (e.g. Howard et al. 2012) where they tend to form compact multiple systems (Rowe et al. 2014). Based on their mass (or minimal mass for planets detected by radial velocity - RV), these objects are loosely classified as super-Earths ($M_p \le 10 M_\oplus$) and Neptunes ($M_p > 10 M_\oplus$). This division is based on the theoretical limit for the runaway accretion of H/He by a protoplanet, $\sim 10 M_\oplus$ (Rafikov 2006), and thus implicitly assumes that Neptunes are predominantly ice giants and that most super-Earths are massive terrestrial planets. However, the growing sample of transiting low-mass exoplanets with precise mass and radius measurements exhibit a wide diversity of densities that reveals a very heterogeneous population, making simplistic inferences hazardous when based on the comparison with solar system planets. A better understanding of this ubiquitous class of planets requires the thorough characterization of a significant sample, not only the precise measurements of their physical dimensions but also the exploration of their atmospheric composition to alleviate the strong degeneracies of composition models in this mass range (e.g. Seager et al. 2007, Valencia et al. 2013). Among the known transiting low-mass planets, GJ436b (Butler et al. 2004, Gillon et al. 2007) and GJ1214b (Charbonneau et al. 2009) are the most thoroughly characterized planets in the Neptune and super-Earth mass ranges, respectively, thanks to the small size ($\sim$ 0.45 and 0.2 $R_\odot$) and proximity (about a dozen of pc) of their M-dwarf host stars. First constraints on their atmospheric properties have indeed been obtained by several programs (e.g. Kreidberg et al. 2014, Knutson et al. 2014a, Ehrenreich et al. 2015). Extending this kind of detailed studies to Neptunes and super-Earths orbiting solar-type hosts requires the detection of such planets in transit in front of extremely bright and nearby stars. A straightforward method to achieve this goal is to search for the transits of the low-mass planets detected by RV surveys. Indeed, these surveys target the most nearby main-sequence stars, and they have now detected enough short-period low-mass planets to make it highly probable that a handful of them transit their parent stars, as demonstrated by the previous detections by the MOST space telescope of the transits of the super-Earths 55Cnce (Winn et al. 2011)[^2]. Thanks to the brightness of their host stars, the atmospheric characterization of these two planets has already started (e.g. Demory et al. 2012, Knutson et al. 2014b). More recently, the same approach enabled us to reveal with [*Spitzer*]{} the transiting configuration of the rocky planet HD219134b (Motalebi et al. 2015) which, at 6.5pc, is the nearest known transiting exoplanet. Searching for the transits of RV low-mass planets is one of the main objectives of the future European space mission CHEOPS (Broeg et al. 2013). However, CHEOPS is not due to launch before the end of 2017. To set it on its path, back in 2010 we set up an ambitious project using [*Spitzer*]{}/IRAC (Fazio et al. 2004) to search for the transits of the RV low-mass planets that have the highest geometric transit probabilities, focusing mainly on the shortest-period planets detected by the HARPS spectrograph (Mayor et al. 2003). Our [*Spitzer*]{} transit search was composed of a cryogenic program targeting HD40307b (ID 495, 27.5hr), and three so-called [*Warm*]{} (i.e. non cryogenic) programs (ID 60027, 90072, and 11180; 100hr, 300hr, and 9.5hr) that targeted 24 other RV low-mass planets. Its published results have so far been the non-detection of the transits of HD40307b (Gillon et al. 2010, hereafter G10) and GJ3634b (Bonfils et al. 2011), the detection and confirmation of the transits of 55Cnce (Demory et al. 2011, Gillon et al. 2012a), the confirmation of the transiting nature of HD97658b (Van Grootel et al. 2014), and the detection of a transit of HD219134b (Motalebi et al. 2015). Results for another planet will be presented in a forthcoming paper (Ségransan et al., in prep.) We report here null results for the 19 other targets of the project. We first present our targets and our determination of their transit ephemeris. In Sect. 3, we present our $Spitzer$ data and their reduction. Section 4 describes our data analysis. Its main results are presented in Sect. 5. Finally, we discuss our global results and give our conclusions in Sect. 6. Targets and transit ephemeris determination =========================================== Table 1 and Tables A.1 to A.4 list the 19 targets of this work. For each target, we performed an analysis of the available RVs, including new measurements for some of them gathered by HARPS[^3], to derive the most accurate transit ephemeris. This analysis was done with the adaptative Markov Chain Monte-Carlo (MCMC) algorithm described in G10 (see also Gillon et al. 2012a, 2012b, 2014). Our nominal model was based on a star and one or several planets on Keplerian orbits around their common center of mass. For some cases, we added a linear or quadratic trend to the model, based on the minimization of the Bayesian information criterion (BIC, Schwarz 1974) to elect our final model. We checked that planet-planet interactions had negligible influence on our solutions, using, for this purpose, the [*Systemic Console*]{} software (Meschiari et al. 2009). We also used this software to perform an initial optimization of parameters to check that no solution existed with a higher likelihood than the published one. During this initial stage of the analysis, the RV errors were assumed to be measurement uncertainties. Once this initial stage was completed, we measured the quadrature difference between the rms of the residuals and the mean error, and this ”jitter noise” (Wright 2005) was quadratically summed to the measurement uncertainties. We then performed a MCMC analysis to probe the posterior probability distribution functions (PDF) of the model and physical parameters Each MCMC analysis was composed of five Markov chains of 10$^5$ steps, the first 20% of each being considered as its burn-in phase and discarded. For each run, the convergence of the five Markov chains was checked using the statistical test of Gelman & Rubin (1992). The resulting posterior PDFs for the transit time and period were then used to schedule the [*Spitzer*]{} observations, with monitoring the $2-\sigma$ transit window as the goal, i.e. keeping the probability to miss a transit below 5%. Even if the resulting posterior PDFs for the orbital eccentricity of most target planets were compatible with zero, we did not assume a purely circular orbit for them to ensure the reliability of our derived transit ephemeris. Indeed, several examples of low-mass exoplanets with significantly eccentric few-days orbits (e.g. GJ436b, HAT-P-11b, HD215497b) remind us of the risks of systematically assuming circular orbits for close-in exoplanets based on tidal circularization arguments. An established theory for tidal dissipation mechanisms is still a long-term goal, and reaching it relies mostly on gathering new observational constraints. Once the [*Spitzer*]{} observations had been performed for a planet, we made a second MCMC analysis of the most up-to-date RV dataset, this time with the fitted transit time corresponding to the epoch covered by the [*Spitzer*]{} run. The purpose of this analysis was to derive the most accurate orbital parameters, transit ephemeris, and minimal masses for the planets under consideration. Indeed, for most of them HARPS gathered a significant number of RV measurements between the scheduling of the [*Spitzer*]{} observations and the final analysis of the [*Spitzer*]{} images. Tables A.1 to A.4 present the results obtained for each targeted planet from these MCMC analyses of the most up-to-date RV dataset. In addition to some basic parameters for the host stars, these tables give the origin of the RVs used as input in our MCMC analysis for each target, and provide the most relevant results of our MCMC analysis: the transit and occultation ephemeris, the minimal mass, the expected minimum transit depth corresponding to a pure iron planet (Seager et al. 2007), the equilibrium dayside temperature, the orbital parameters, and the expected duration for a central transit. The radius of the star was derived from the luminosity and effective temperature, taking bolometric corrections from Flower (1996). Tables A.1 to A.4 also present the median value and the 1-$\sigma$ errors for the prior transit probability. It was computed at each step of the MCMC with the following formula: $$P(tr) = \bigg(\frac{R_\ast}{a}\bigg) \bigg( \frac{1+e\sin\omega}{1-e^2}\bigg) \textrm{,}$$ where $R_\ast$ is the stellar radius, $a$ is the orbital semi-major axis, $e$ is the orbital eccentricity and $\omega$ is the argument of periastron. This probability estimate does not take into account that planets are more likely to be discovered by RV if their orbit is significantly inclined (e.g. Wisniewski et al. 2012). By performing Bayesian simulations that assume different prior PDFs for the planetary masses, Stevens & Gaudi (2013) have shown that this bias increases the transit probability of short-period low-mass RV planets like the ones considered here by only $\sim$20% in average. Furthermore, its actual estimation for a given planet depends strongly on the assumed prior PDFs for the planetary masses. We have thus neglected it in the context of this work. [*Warm Spitzer*]{} photometry ============================= Tables A.5 to A.11 provide a summary of the [*Spitzer*]{} observations. Since all our targets are very bright ($K$ between 2.96 and 6.90), all of them were observed in subarray mode (32x32 pixels windowing of the detector), the extremely fast Fowler sampling ($\sim$0.01s), which maximizes the duty cycle and signal-to-noise ratio (S/N). No dithering pattern was applied to the telescope (continuous staring). For each target, the exposure time was selected to maximize the S/N while staying in the linear regime of the detector, basing on the [*Warm Spitzer*]{} flux density estimator tool[^4] and on the instructions of the [*Warm Spitzer*]{} Observer Manual[^5]. The observations of program 60027 (Cycle 6) and 90072 (Cycle 9) were performed between 2009 Dec 14 and 2010 Sep 11, and between 2012 Dec 03 and 2014 May 14, respectively. For several of the Cycle 9 targets, we benefitted from the newly introduced PCRS peak-up mode (Ingalls et al. 2012). This mode provides enhanced accuracy in the position of the target on the detector, to a significant decrease of the so-called ‘pixel phase effect’ that is the most important source of correlated noise in high-S/N staring mode observation with IRAC InSb arrays (e.g. Knutson et al. 2008). For HD1461b, we supplemented our data with the IRAC 4.5$\mu$m observations presented by Kammer et al. (2014), as described in Sect. 5.2. On a practical level, each observation run was divided in one or several science astronomical observational requests (AOR) of 12hr at most, preceded by a short (20-30 min) AOR to allow the pointing of the telescope and the instrument to stabilize. The IDs of all AORs are given for each target in Tables A.5 to A.11. These tables also give the version of the [*Spitzer*]{} pipeline used to calibrate the corresponding images, the resulting files being called basic calibrated data (BCD) in the [*Spitzer*]{} nomenclature. Each subarray mode BCD is composed of a cube of 64 subarray images of 32$\times$32 pixels (pixel scale = 1.2 arc second). The following reduction strategy was used for all the [*Spitzer*]{} AOR. We first converted fluxes from the [*Spitzer*]{} units of specific intensity (MJy/sr) to photon counts, then aperture photometry was performed on each subarray image with the [IRAF/DAOPHOT]{}[^6] software (Stetson, 1987). For each AOR, we tested different aperture radii and background annuli, and selected the combination minimizing the white and red noises in the residuals of a short data fitting analysis. The center and width of the point-spread functions (PSF) were measured by fitting a 2D-Gaussian profile on each image. The $x-y$ distribution of the measurements was then studied, and measurements that had a visually discrepant position relative to the bulk of the data were then discarded. For each block of 64 subarray images, we then discarded the discrepant values for the measurements of flux, background, $x$- and $y$-positions using a 10-$\sigma$ median clipping for the four parameters, and the resulting values were averaged, the photometric errors being taken as the errors on the average flux measurements. Finally, a 50-$\sigma$ slipping median clipping was used on the resulting light curves to discard outliers (owing to, for example, cosmic hits). Global [*Warm Spitzer*]{} + RV data analysis ============================================ We analyzed the [*Spitzer*]{} photometric time-series supplemented by the RVs with our MCMC code. For each target, our model for the RVs was the same as the one presented in Sect. 2. The assumed photometric model consisted of the eclipse model of Mandel & Agol (2002) to represent the possible eclipses of the probed planets, multiplied for each light curve by a baseline model that aimed to represent the other astrophysical and instrumental effects at the source of photometric variations. We assumed a quadratic limb-darkening law for the stars. For each light curve that corresponded to a specific AOR, we based the selection of the baseline model on the minimization of the BIC. Tables A.5 to A.11 present the baseline function elected for each AOR. Following Gillon et al. (2014), the instrumental models included three types of low-order polynomials. The first one had as variables the $x$- and $y$-positions of the center of the PSF to represent the so-called pixel phase effect of the IRAC InSb arrays (e.g. Knutson et al. 2008). The second one had the PSF widths in the $x$- and/or the $y$-direction as variables, its inclusion in the baseline model strongly increasing the quality of the fit for all AORs (the so-called PSF breathing effect, see also Lanotte et al. 2014). The third function was a polynomial of the logarithm of time to represent a sharp increase of the detector response at the start of some AORs (the so-called ramp effect, Knutson et al. 2008). To improve the quality of the modeling of the pixel phase effect, especially the fitting of its highest frequency components,for most AORs we supplemented the $x$- and $y$-polynomial with the bi-linearly-interpolated sub-pixel sensitivity (BLISS) mapping method (Stevenson et al. 2012). The sampling of the position space was selected so that at least five measurements fall within the same sub-pixel. See Gillon et al. (2014) for more details. The jump parameters of the MCMC, i.e. the parameters randomly perturbed at each step of the Markov Chains, were as follows: - The stellar mass $M_\ast$, radius $R_\ast$, effective temperature $T_{eff}$, and metallicity \[Fe/H\]. For these four parameters, normal prior PDFs were assumed based on the values given in Tables A.1-4. - For the potential transiting planet, the parameter $b' = a \cos{i}/R_\ast$, where $a$ is the orbital semi-major axis and $i$ is the orbital inclination. $b'$ would correspond to the transit impact parameter in the case of a circular orbit. The step was rejected if $b' > a/R_\ast$. For the other planets of the system, $b'$ was fixed to 0. - The parameter $K_2 = K \sqrt{1-e^2} \textrm{ } P^{1/3}$ for all planets of the system, $K$ being the RV orbital semi-amplitude, $e$ the orbital eccentricity, and $P$ the orbital period. - The orbital period $P$ of each planet. - For each planet, the two parameters $\sqrt{e} \cos{\omega}$ and $\sqrt{e} \sin{\omega}$, $e$ being the orbital eccentricity and $\omega$ being the argument of periastron. - The planet/star area ratio $dF = (R_p/R_\ast)^2$ for the potential transiting planet. At each step of the MCMC, the planetary radius corresponding to a pure iron composition was computed under the formalism given by Seager et al. (2007), and if the planetary radius derived from $dF$ and $R_\ast$ was smaller, the step was rejected. A similar rejection was done for $R_p > 11 R_\oplus$, an implausibly large size for the low-mass planets considered here. The goal of these prior constraints on $R_p$ was to avoid fitting extremely shallow transits and ultra-grazing transits of unrealistically big planets in the correlated noise of the light curves to ensure an unbiased posterior transit probability. For the other planets of the multiple systems, $dF$ was fixed to 0. In all cases, we checked that a transit of another planet of these systems was not expected to occur during the [*Spitzer*]{} observations. - The time of inferior conjunction $T_0$ for all planets of the system. For the potential transiting planets that we considered, $T_0$ corresponds approximatively to the mid-time of the transit searched for by [*Spitzer*]{}. The limb-darkening of the star was modeled by a quadratic law (Claret 2000). For both [*Warm Spitzer*]{} bandpasses (3.6 and 4.5 $\mu$m), values for the two quadratic limb-darkening coefficients $u_1$ and $u_2$ were drawn at each step of the MCMC from normal distributions whose expectations and standard deviations were drawn from the tables of Claret & Bloemen (2011) for the corresponding bandpasses and for the stellar atmospheric parameters given in Tables A.1-4. Five chains of 100000 steps were performed for each analysis, their convergences being checked using the statistical test of Gelman and Rubin (1992). They followed a preliminary chain of 100000 steps, which was performed to estimate the need to rescale the photometric errors. For each light curve, the standard deviation of the residuals was then compared to the mean photometric errors, and the resulting ratios $\beta_w$ were stored. $\beta_w$ represents the under- or overestimation of the white noise of each measurement. On its side, the red noise present in the light curve (i.e. the inability of our model to represent perfectly the data) was taken into account as described in G10, i.e. a scaling factor $\beta_r$ was determined from the standard deviations of the binned and unbinned residuals for different binning intervals ranging from 5 to 120 minutes, the largest values being kept as $\beta_r$. At the end, the error bars were multiplied by the correction factor $CF = \beta_r \times \beta_w$. The derived values for $\beta_r $ and $\beta_w$ are given for each light curves in Tables A.5-11. Results ======= For each planet searched for transit, Table 1 presents the derived posterior full transit probability $P(f,D)$, i.e. the probability that the planet undergoes full transits given the [*Spitzer*]{} data. Bayes theorem shows that $$P(f,D) = \frac{P(f) P(D,f)}{P(f) P(D,f) + P(g) P(D,g) + P(n) P(D,n)}$$ $P(f)$, $P(g)$, and $P(n)$ are the prior (geometric) probabilities of full, grazing and no transit, respectively ($P(g)$ is close to zero, and $P(n) \sim 1 - P(t)$). $P(D,f)$, $P(D,g)$, and $P(D,n)$ are the probabilities (likelihoods) to have the observed data given the three mutually exclusive hypotheses. All the terms of the right-hand side of Eq. 2 are probed by the MCMC analysis, resulting in accurate estimates of $P(f,D)$. Figs. 1-4, 6-14, and 16-20 show the resulting detrended [*Spitzer*]{} light curves and the derived posterior PDFs for the inferior conjunction. Below, we provide relevant details for our 19 targets. BD-061339 --------- BD-061339 (a.k.a. GJ221) is a V=9.7 late-K dwarf around which two planets were detected by HARPS (Lo Curto et al. 2013), a super-Earth on a $\sim$3.9d orbit and a planet of $\sim 50 M_\oplus$ minimal mass on a $\sim$126d orbit. Our model selection process for the RVs (HARPS + PFS) favored a model with a slope in addition to these two planets (Table A.1) . The derived value for this slope is $-0.73 \pm 0.15$ [ ]{}per year, which could correspond to a giant planet in outer orbit or to the imprint of a stellar magnetic cycle. As can be seen in Fig. 1, the high-precision of the [*Spitzer*]{} photometry (42 ppm per 30 min time bin) discards any transit for BD-061339b during our observations, even for an unrealistic pure iron composition. However, the edge of the right wing of the PDF for $T_0$ was not explored by our observations, so a late transit is still possible. The resulting posterior full transit probability is 0.53% (Table 1), high enough to justify a future exploration of the late part of the transit window. \[fig:1\] ![[*Warm Spitzer*]{} photometry for BD-061339 after correction for the instrumental effects, normalization, and binning per 30 min. Models for a central transit of BD-061339b are shown, assuming a pure iron (red solid line) and pure water ice (blue dashed line) composition. The highest posterior probability model, which assumes no transit, is also shown (purple line). The posterior PDF for the transit timing derived from the RV analysis is shown above the figure.](aa29270_fig1a.eps "fig:"){width="9cm"} ![[*Warm Spitzer*]{} photometry for BD-061339 after correction for the instrumental effects, normalization, and binning per 30 min. Models for a central transit of BD-061339b are shown, assuming a pure iron (red solid line) and pure water ice (blue dashed line) composition. The highest posterior probability model, which assumes no transit, is also shown (purple line). The posterior PDF for the transit timing derived from the RV analysis is shown above the figure.](aa29270_fig1b.eps "fig:"){width="9cm"} HD1461 ------ HD1461 is a V=6.5 solar-type star known to host a super-Earth on a 5.77d orbit (Rivera et al. 2010). Recently, a second super-Earth on a 13.5d orbit was announced by Díaz et al. (2015). In our analysis of the HARPS + Keck RVs for this star, the convergence of the MCMC appeared to be significantly improved by removing the 20 first Keck measurements for which the exposure time and the resulting precision were significantly lower than for the rest of the data. Our analysis favored a three planet model (5.77, 13.5, and 377 days period) in addition to a second-order time polynomial, suggesting the presence of a fourth low-frequency signal, and an activity model consisting of a sum of second-order polynomials in the cross-correlation function (CCF, Baranne et al. 1996, Queloz et al. 2001) parameters (Table A.1): width, contrast, and bisector. We inferred that the 377d period, close to the duration of a year, is caused by a systematic effect (the stitching) recently revealed to affect HARPS data (Dumusque et al. 2015). The data used by Díaz et al. (2015) were corrected for this effect. We made several tests that showed that the inclusion or not of this 377d Doppler signal in the RV model did not affect the results for the closest-in planet, including its transit ephemeris, so we kept it in our final analysis. A search for a transit of HD1461b with [*Spitzer*]{} was presented by Kammer et al. in 2014 (program 80220). We first reduced their data and used them with the HARPS+Keck RVs as input for a global MCMC analysis. The resulting posterior full transit probability was 0.5%, suggesting that a small but significant fraction of the transit window was not covered by these [*Spitzer*]{} observations. As can be seen in Fig. 2, a transit that had ended just before the [*Spitzer*]{} observations remained possible. This possibility is amplified by the ramp effect that affected the first hour of the [*Spitzer*]{} data, resulting in an increase of brightness that could be degenerated with a transit egress. We thus complemented these [*Spitzer*]{} archive data with new [*Spitzer*]{} observations covering the first part of the transit window. We performed a global analysis of all [*Spitzer*]{} + RVs that made a transiting configuration for HD1461b very unlikely (see Fig. 2), the resulting posterior full transit probability now being of only 0.14% (Table 1). \[fig:2\] {width="9cm"} {width="9cm"} {width="9cm"} HD10180 ------- HD10180 is a V=7.3 solar-type star hosting a particularly interesting system of six planets of relatively low-masses (Lovis et al. 2011). At first, we did not consider searching for the transit of the 1.18d-period Earth-mass planet candidate HD10180b presented by Lovis et al. (2011) because its Doppler detection was not firmly secured when we planned our [*Spitzer*]{} Cycle 6 observations of the star. Furthermore, we estimated that its transit would, in any case, be too shallow (a few dozens of ppm, at most 100 ppm) to be firmly detected with [*Spitzer*]{}. We thus focused on the planet HD10180c ($M\sin{i} = 13 M_\oplus$, $P$ = 5.76 d), for which the Doppler signal was clearly detected in the HARPS data, and for which the expected transit depth ($> 150$ ppm) was large enough to ensure a sure detection with [*Spitzer*]{}. enabled us to discard a transit of the planet (Fig. 3), the resulting posterior full transit probability being of 0.14% (Table 1). Still, we noticed a shallow structure in the detrended photometry that occurs at a time consistent with a transit of HD10180b (Fig. 4), the best-fit transit depth of $\sim$90 ppm translating into a radius of 1.2 $R_\oplus$ consistent with a planetary mass of $1 M_\oplus$. Based on this result, we decided to observe in our Cycle 9 program 90072 two more transit windows of HD10180b, this time at 4.5 $\mu$m. No transit structure was clearly detected in the resulting light curves (Fig. 4, bottom), and a short MCMC analysis of the data led us to conclude that these new data do not increase the significance of the 2010 tentative detection. We thus conclude that the low-amplitude structure in our Cycle 6 [*Spitzer*]{} light curve is probably just correlated noise of instrumental origin that is not perfectly represented by our baseline model and that can easily be modeled by a shallow enough transit profile. For [*Spitzer*]{}, the average amplitude of these correlated noise structures is of a few dozen of ppm (see Sec. 6), making the firm detection of a unique transit shallower than $\sim$100ppm impossible - similarly to the one expected for HD10180b (see discussion in Sec. 6 and our estimated detection thresholds in Table 1). This so-called red noise limit can be surpassed, but only by gathering more observations of the transit window, as we did here (see also the case of HD40307b in Gillon et al. 2010). \[fig:3\] {width="9cm"} {width="9cm"} \[fig:4\] ![$Top$: Same as Fig. 3, except that a possible transit of HD10180b was included in the global model. Highest posterior probability transit model for HD10180b is shown in purple. The posterior PDF for the transit timing of HD10180b derived from an RV analysis assuming 7 planets is shown above the figure. $Bottom$: 2013 [*Warm Spitzer*]{} photometry for HD10180 binned per 30 min, corrected for the instrumental effects, normalized, and folded on the transit ephemeris for HD10180b based on our tentative detection of a transit and on the results of an RV analysis assuming 7 planets (dT = time from most likely mid-transit time). ](aa29270_fig4a.eps "fig:"){width="9cm"} ![$Top$: Same as Fig. 3, except that a possible transit of HD10180b was included in the global model. Highest posterior probability transit model for HD10180b is shown in purple. The posterior PDF for the transit timing of HD10180b derived from an RV analysis assuming 7 planets is shown above the figure. $Bottom$: 2013 [*Warm Spitzer*]{} photometry for HD10180 binned per 30 min, corrected for the instrumental effects, normalized, and folded on the transit ephemeris for HD10180b based on our tentative detection of a transit and on the results of an RV analysis assuming 7 planets (dT = time from most likely mid-transit time). ](aa29270_fig4b.eps "fig:"){width="9cm"} ![$Top$: Same as Fig. 3, except that a possible transit of HD10180b was included in the global model. Highest posterior probability transit model for HD10180b is shown in purple. The posterior PDF for the transit timing of HD10180b derived from an RV analysis assuming 7 planets is shown above the figure. $Bottom$: 2013 [*Warm Spitzer*]{} photometry for HD10180 binned per 30 min, corrected for the instrumental effects, normalized, and folded on the transit ephemeris for HD10180b based on our tentative detection of a transit and on the results of an RV analysis assuming 7 planets (dT = time from most likely mid-transit time). ](aa29270_fig4c.eps "fig:"){width="9cm"} HD13808 ------- In 2012, we analyzed the HARPS dataset for the $V$=8.4 early K-dwarf HD13808, confirming the existence of the two planets around it with periods of 14.2 d and 53.7 d announced by Mayor et al. (2011, hereafter M11). Furthermore, our analysis revealed the existence of (1) a low-frequency signal that was well-modeled with a quadratic trend, whose origins is the magnetic cycle of the star (Queloz et al. in prep.), and (2) a low-amplitude Doppler signal with a period of 1.091 d which corresponds to a planet of $M_p\sin{i} = 1.5 \pm 0.3$ $M_{\oplus}$ with an interestingly high transit probability of $\sim$20%. The false-alarm probability (FAP) derived by [*Systemic*]{} for this short-period planet candidate was close to 1%. Based on this small FAP and the scientific importance of this putative planet if transiting, we decided to monitor two of its transit windows with [*Spitzer*]{}. Unfortunately, the resulting light curves did not show any convincing transit-like structure (Fig. 5). Our [*Spitzer*]{} data were not acquired during a transit window of the planet at 14.2 d, thus keeping its transiting nature unconstrained. In 2014, we analyzed the updated HARPS dataset. This analysis confirmed the existence of the low-frequency signal of magnetic cycle origins, still well-modeled by a quadratic trend, but it led to a much lower significance (FAP $\sim$ 10%) for the 1.091 d signal. A stronger signal at 18.98 d and at its alias period of 1.056 d emerged with FAP close to 1%. Still, including a polynomial function of the CCF parameters (bisector, contrast, width) in our MCMC modeling, these signals vanished from the residuals, indicating that their origin is stellar (activity), not planetary. The period 18.98 d could then be interpreted as the rotation period of the star, resulting in an equatorial rotation speed $\sim$2.2 km.s$^{-1}$ which is consistent with the measured $v\sin{i}$ of 2 km.s$^{-1}$ (Glebocki & Gnacinski 2005). This stellar signal, combined with aliasing effects, would thus be responsible for creating spurious signals with periods slightly larger than 1 d. \[fig:5\] {width="8cm"} HD20003 ------- HD20003 is a late-G dwarf of magnitude $V$=8.4 for which M11 announced the detection by HARPS of two planets with periods of 11d and 33d. Our analysis of the updated HARPS dataset for HD20003 not only confirmed these planets, but also revealed two longer period signals: one at $\sim$180d that was determined as originating from the stitching effect, and another, at $\sim$10yr, that originates from the magnetic cycle of the star (Udry et al. in prep.). Only HD20003b ($P$=11d) has a significantly eccentric orbit. We used [*Spitzer*]{} to probe its transiting nature. We did not detect a transit (Fig. 6). But, as can be seen in Fig. 6, we did not probe the latest part of its transit window, resulting in a small but still significant posterior full transit probability of 0.54% (Table 1). \[fig:6\] {width="9cm"} {width="9cm"} HD20781 ------- Based on 96 HARPS measurements, M11 announced the discovery of two Neptunes with orbital periods of 29.1d and 85.1d around this nearby $V$=8.4 K0-dwarf. The analysis of our much extended HARPS dataset (212 points) confirmed the existence of these two planets, while revealing the existence of two super-Earths in shorter orbits: a $M_p \sin{i} = 6.3 M_\oplus$ planet at 13.9d period, and a $M_p \sin{i} = 2.1 M_\oplus$ planet at 5.3d period (Udry et al. in prep.). The low minimal mass and relatively high transit probability ($>$7%) of this latter planet made it an especially interesting target for a transit search, so we included it in the target list of our [*Spitzer*]{} program and observed one of its transit window during Cycle 9. Because of the extreme faintness of the Doppler signal ($K \sim 90$ cm.s$^{-1}$), the orbital eccentricity of the planet is poorly constrained from the RVs alone, resulting in a particularly large transit window. As such, we assumed the orbit to be circular to minimize the required [*Spitzer*]{} time, judging a circular orbit as a reasonable assumption, taking into consideration the strong tidal forces exerted by the star at such close distance and the compactness of the planetary system that would make any significantly eccentric orbit unstable. Our [*Spitzer*]{} photometry did not reveal any transit-like structure (Fig. 7), the resulting posterior full transit probability being 0.15% (Table 1). Still, our precision is not high enough to securely detect the transit of this very low-mass planet if its amplitude is below $\sim$150ppm, corresponding to Mercury-like iron-rich compositions. Re-observing the planet’s transit window at higher precision with, for example, CHEOPS will be mandatory to fully exclude a transiting configuration for the planet. \[fig:7\] {width="9cm"} {width="9cm"} HD31527 ------- M11 announced the existence of three Neptune-mass planets around this nearby $V=7.5$ solar-type star, with orbital periods of 16.5d, 51.3d, and 275d. Our analysis of the extended HARPS dataset (242 vs 167 measurements) confirmed the existence of these planets and improved their orbital parameters, while not revealing any other planet (Udry et al. in prep.). We used $\sim$36 hr of continuous [*Spitzer*]{} observation to search for a transit of the innermost planet, HD31527b (geometric transit probability = 4.4%). The resulting light curve did not reveal any transit (Fig. 8), the resulting posterior full transit probability being of 0.45% (Table 1). Compared to the orbital solution that we used to schedule our [*Spitzer*]{} observations, the updated solution presented here (Table 2) results in [*Spitzer*]{} observations that are not well centered on the peak of the posterior PDF for the transit timing. The right wing of this PDF is thus unexplored. Its future exploration with, for example, CHEOPS would be desirable to fully exclude a transiting nature for the planet. \[fig:8\] {width="9cm"} {width="9cm"} HD39194 ------- This star is a $V=8.1$ early K-dwarf around which HARPS detected three super-Earths with orbital periods of 5.6d, 14.0d, and 33.9d (M11, Queloz et al. in prep.). Our analysis of the updated HARPS dataset (261 RVs vs 133 in the discovery paper) fully confirmed the existence of these planets, and revealed a low-amplitude trend. With a geometric transit probability of $\sim$6.5%, theinnermost of these planets was an interesting target for our program, and we monitored one of its transit windows with [*Spitzer*]{} in Dec 2012. As shown in Fig. 9, the resulting light curve was flat, our deduced posterior transit probability being of 0.46% (Table 1). \[fig:10\] {width="9cm"} {width="9cm"} HD45184 ------- M11 reported the detection of a Neptune-mass planet on a 5.9 d period around this bright ($V=6.4$) solar-type star. Based on more than double the HARPS measurements (174 vs 82), our analysis confirms the existence of this planet, while revealing the presence of a second planet of similar mass, $\sim$9.5 $M_\oplus$, on a outer orbit ($P$=13.1 d), and a trend in the RVs that we could relate to the magnetic cycle of the star (Udry et al. in prep.). We monitored the star for more than 11hrs with [*Spitzer*]{} to search for the transit of HD45184b. The resulting light curve is flat (Fig. 10), while its precision would have been high enough to detect the searched transit for any plausible composition. We did not explore the latest part of the transit window, leaving a posterior probability of 1.3% that the planet undergoes full transit (Table 1), for a prior geometric probability of 7.7%. \[fig:11\] {width="9cm"} {width="9cm"} HD47186 ------- HD47186b is a short-period ($P$=4.08d) Neptune-mass planet discovered by HARPS in 2009 (Bouchy et al. 2009) around a $V=7.6$ solar-type star, which is also orbited by a giant planet at a much longer period. Our analysis of the extended HARPS dataset confirmed the existence of these two planets, and enabled us to derive an excellent precision on the time of inferior conjunction of the inner planet ($1\sigma$-error of 30 min for our [*Spitzer*]{} observations). We searched for its transit within our [*Spitzer*]{} cycle 6 program. The resulting flat light curve (Fig. 11) enables us to fully reject the transiting nature of the planet (Table 1). \[fig:12\] {width="9cm"} {width="9cm"} HD51608 ------- We used [*Spitzer*]{} to search for the transit of the $P=14.07$ d Neptune-mass planet HD51608b detected by M11 around this bright ($V=6.3$) late G-type dwarf. Our analysis of the updated HARPS dataset (210 RVs) confirmed the existence of the two planets ($P=14.07$d and $95$d) announced in M11, and favored a low-amplitude trend of probable magnetic cycle origin (Udry et al. in prep.). Our [*Spitzer*]{} photometry (Fig. 12) enabled us to discard a transit of HD51608 for any plausible planetary composition, our posterior full transit probability being 0.11% (Table 1) for a prior geometric transit probability of $\sim$4%. \[fig:13\] {width="9cm"} {width="9cm"} HD93385 ------- The detection of two low-mass planets was reported by M11 for this $V=7.5$ solar-type star, with derived minimal masses of 8.4 and 10.1 $M_\oplus$ and orbital periods of 13.2 d and 46 d, respectively. Our analysis of the updated HARPS dataset (231 RVs) revealed the existence of a third lower-mass ($Mp\sin{i} = 4.0 \pm 0.5 M_\oplus$) planet on a 7.3 d orbit (Queloz et al. in prep., see Table A.3). With [*Spitzer*]{}, we monitored a transit window ($\sim$29hr) of this new planet, HD93385d. Our resulting photometry (Fig. 13) did not reveal any clear transit-like structure, and the posterior full transit probability that we derived from its analysis is 0.23% (Table 1), for a prior geometric transit probability of 7.9%. By injecting transit models in this light curve and analyzing the results with our MCMC code, we concluded that its precision of $\sim$ 40 ppm per half-hour is good enough to discard any central transit of a planet with a density equal or smaller than Earth’s, but cannot firmly discard the transit of denser planet. We thus recommend the re-observation of the transit window at higher photometrical precision with, for example, CHEOPS. \[fig:14\] {width="9cm"} {width="9cm"} HD96700 ------- HD96700b is a $\sim 9 M_\oplus$ planet on a 8.1 d orbit discovered by M11 around a $V=6.5$ solar-type star. Our analysis of the updated HARPS dataset (244 RVs) confirmed its existence and the outer Neptune-mass planet found by M11 on a $\sim$100 d period. It also revealed a low-amplitude trend in the RVs of probable magnetic cycle origin (Queloz et al. in prep.). We searched for the transit of HD96700b with [*Spitzer*]{} in 2013, without success (Fig. 14). The precision of the [*Spitzer*]{} light curve is high enough to discard a transit of HD96700b in it for any possible planetary composition. The resulting posterior transit probability of the planet is 0.09%, for a prior probability of 6.8% (Table 1). \[fig:15\] {width="9cm"} {width="9cm"} HD115617 -------- HD115617 (aka 61 Vir) is a $V$=4.7 solar-type star (G5V) at only 8.5 pc from Earth. A close-in super-Earth ($M_p \sin i = 5 M_\oplus$, $P$ = 4.215 d) and two outer Neptunes ($M_p \sin i = 18$ & $23 M_\oplus$, $P = 38$ & $123$ d) were discovered around it by Vogt et al. (2010) using RVs obtained with Keck/HIRES and the Anglo-Australian Telescope (AAT). We performed a global analysis of the Keck, and AAT RVs that confirmed the existence of the three planets, without revealing any additional object orbiting the star. We used [*Spitzer*]{} to observe a transit window of HD115617 in March 2010. With $K=2.96$, HD115617 is an extremely bright star for [*Spitzer*]{}. At 4.5 $\mu$m, it is nevertheless faint enough to be unsaturated for the shortest available integration time (0.01s). However, the [*Spitzer*]{} Science Center (SSC) informed us that the requested long observation of HD15617 could not be performed at 4.5 $\mu$m for technical reasons. At 3.6 $\mu$m, the star has a flux density of $19000 \pm 4000$ mJy while the saturation limit is 20000 mJy for an integration time of 0.01s. SSC informed us the observations could still be performed without any risk of saturation if the star was placed at the corner of four pixels, leading to an effective flux density that was low enough to avoid saturation. We decided to test this strategy. Unfortunately, the light curve corrected for known systematic effects is corrupted by a clear variability at the level of a few tens of percent (Fig. 15, top). A Lomb-Scargle periodogram (Press et al. 1992) of the corrected photometry reveals a power excess at $\sim$0.11 days (Fig 15, bottom). We could not identify the origin of this variability. Nevertheless, an astrophysical origin is very unlikely, since HD115617 is known to be an old inactive star (Vogt et al. 2010). This variability probably originates from the near-saturation of the detector. Indeed, the center of the stellar image was not located at the corner of four pixels as intended. The instrumental effect at work has a typical timescale similar to the signal we are trying to detect, and its amplitude is much larger than the one of the searched transit. Even worse, it seems highly variable in nature and we could not find any analytical function of external parameters (PSF peak value, center position and width, background, etc.) able to represent it satisfactorily. Without a thorough understanding of the effect, we concluded that searching for a transit of a few hundreds of ppm with such data was illusory. The transiting nature of HD115617b thus remains unconstrained by our project. \[fig:17\] ![$Top$: 61[*Vir*]{} light curve obtained by [*Spitzer*]{} to search for the transit of its planet b, divided for the best-fit phase-pixel model (here a third order $x$ and $y$-position polynomial), unbinned (green dots) and binned to intervals of 0.005d (7.2 min). The start of the second AOR is represented by the red vertical line. $Bottom$: Lomb-Scargle periodogram showing a clear power excess at $\sim$0.11 days and its first harmonic.](aa29270_fig15a.eps "fig:"){width="9cm"} ![$Top$: 61[*Vir*]{} light curve obtained by [*Spitzer*]{} to search for the transit of its planet b, divided for the best-fit phase-pixel model (here a third order $x$ and $y$-position polynomial), unbinned (green dots) and binned to intervals of 0.005d (7.2 min). The start of the second AOR is represented by the red vertical line. $Bottom$: Lomb-Scargle periodogram showing a clear power excess at $\sim$0.11 days and its first harmonic.](aa29270_fig15b.eps "fig:"){width="9cm"} HD125612 -------- HD1215612 is a $V=8.3$ solar-type star around which three planets have been detected so far by RVs, first a $P \sim$ 510 d gas giant in 2007 (Fischer et al.) with Keck+HIRES data, then a close-in ($P$=4.15 d) Neptune-mass planet and a $\sim$3000 d-period gas giant with HARPS (Lo Curto et al. 2010). In April 2010, we used [*Spitzer*]{} to search for the transit of the close-in planet HD125612c. Our initial light curve showed a structure consistent with a transit, so we triggered new observations of the transit window with [*Spitzer*]{} in September 2010 that did not confirm the transit signal. Reanalyzing the April data, we noticed that the transit signal disappeared when we included terms in the PSF widths, revealing that the signal originated from the [*Spitzer*]{} PSF breathing effect (Lanotte et al. 2014). Our global analysis of the two [*Spitzer*]{} light curves allowed us to discard a transiting configuration for the planet for any possible planetary composition (Fig. 16), the resulting posterior probability for a full transit being of only 0.24%, for a prior probability of 9.7% (Table 1). We note that the standard deviations of our [*Spitzer*]{} light curves binned per 30 min intervals are the largest for HD125612 (92 and 102 ppm, see Fig. 16). This is also the case for the RV jitter measured from the HARPS RVs (3.2 [ ]{}), suggesting that HD125612 is significantly more active star than our other targets. \[fig:18\] {width="9cm"} {width="9cm"} {width="9cm"} {width="9cm"} HD134060 -------- HD134060 is a G0-type dwarf with a $V$-magnitude of 6.3 at 24pc from Earth, for which M11 and Udry et al. (in prep.) reported the discovery by HARPS of a Neptune (minimal mass = 11 $M_\oplus$) on a 3.3d orbit, HD134060b, in addition to a giant planet on a much wider orbit. Interestingly, the HARPS RVs showed that HD134060b has a significant eccentricity of $0.40 \pm 0.04$, despite its very short orbit, reminiscent of the still poorly understood eccentricity of the hot Neptune prototype GJ436b (Lanotte et al. 2014). Our analysis of the updated HARPS dataset confirmed this significant eccentricity. The orientation of the elliptic orbit of the planet made its occultation by its host star much more likely than its transit (23 vs 8%, see Table A.4), and its timing much better constrained. We thus decided to use a few hours of [*Spitzer*]{} first to try to detect the occultation of the planet. We choose to perform the measurement at 3.6$\mu$m, based on the clear detection of the occultation of GJ436b at 3.6 $\mu$m and its non-detection at 4.5 $\mu$m (Stevenson et al. 2010, Lanotte et al. 2014). Assuming black-body spectra for the star and the planet, we estimated the expected occultation depth to range from a few tens to more than 200ppm, depending on the unknown planet’s size, albedo, and heat-distribution efficiency. The resulting light curve (Fig. 17, bottom panel) did not reveal a single eclipse, but its precision was not high enough to discard an occultation of the planet for a large range of the plausible planetary parameters cited above. We then attempted a transit search, this time at 4.5 $\mu$m. The resulting light curve was flat (Fig. 17, top panel) and enabled us to fully discard a full transit of the planet (Table 1). \[fig:19\] {width="9cm"} {width="9cm"} {width="9cm"} {width="9cm"} HD181433 -------- Three planets were detected by HARPS around this $V=8.4$ K-dwarf (Bouchy et al. 2008). Our analysis of the updated HARPS dataset confirmed the existence of these three planets, and resulted in a prior transit probability of 5% for the inner most of them, the super-Earth HD181433b ($M_p \sin i \sim 7.5 M_\oplus$, $P$=9.375d). Our [*Spitzer*]{} observations of one of its transit windows led to a transit light curve that did not reveal any transit signature (Fig. 18), while being precise enough to discard a transit for any possible composition of the planet. Our global MCMC analysis of the RVs and [*Spitzer*]{} photometry led to a posterior full transit probability of 0.52% (Table 1), which is probably large enough to justify a future exploration of the first part of the transit window that was left unexplored by our [*Spitzer*]{} observations (see Fig. 18). \[fig:20\] {width="9cm"} {width="9cm"} HD215497 -------- Two planets were detected by HARPS around this nearby ($\sim$44 pc) K-dwarf (Lo Curto et al. 2010), one of them being a giant planet on a significantly eccentric 568d-orbit, and the other being a close-in ($P=$3.93d) super-Earth ($M_p \sin i = 6 M_\oplus$). Our analysis of the updated HARPS dataset, containing only three additional RVs, confirmed the existence of both planets. Based on its interestingly high transit probability of $\sim$ 12% the inner super-Earth HD215497b, we targeted it in our [*Spitzer*]{} transit search. The resulting light curve did not reveal any transit signature and was precise enough to discard a transit for most plausible compositions. Yet, a high impact parameter transit of an iron-dominated planet is not discarded by the data (see Fig. 19), resulting in a small but significant posterior full transit probability of $\sim$ 0.31% (Table 1). \[fig:21\] {width="9cm"} {width="9cm"} HD219828 -------- Discovered by Melo et al. in 2007, HD219828b is a hot Neptune ($M_p \sin i = 20 M_\oplus$) with an interestingly high geometric transit probability of $\sim$14%, thanks to its close-in orbit ($a$=0.05 au, $P$=3.83d), combined with the relatively large size of its evolved G0-type host star ($R_\ast = 1.6 R_\odot$). On the other hand, this large stellar size translated in expected transit depths as small as 100 ppm (pure iron composition), so we monitored two transit windows to reach a photometric precision that was high enough to firmly constrain the (non-)transiting nature of the planet. The resulting light curves, both obtained at 4.5 $\mu$m, are shown in Fig. 20. They do not reveal any transit signature. Our global analysis of the RVs + photometry led to a complete rejection of a full transit configuration (Table 1-. Our analysis of the much extended HARPS dataset compared to the HD219828b discovery paper (91 vs 22 RVs) confirmed the existence of a second, more massive planet ($M_p \sin i \sim 15 M_{Jup}$) planet on a very eccentric ($e=0.81$) long-period ($P$=13.1 years) orbit, as recently announced by Santos et al. (2016). \[fig:22\] {width="9cm"} {width="9cm"} {width="9cm"} {width="9cm"} Discussion and conclusion ========================= The constraint brought by the [*Spitzer*]{} photometry on the transiting nature of a given planet can be directly estimated in Table 1 by comparing the derived prior and posterior transit probabilities. For 16 out of the 19 RV planets targeted here, our [*Spitzer*]{} observations explored the transit window with a coverage and precision high enough to make a transiting configuration very unlikely, the posterior transit probabilities for these 16 targets all being less than 0.55% (Table 1). For HD45184b, this posterior transit probability is still of 1.3%, because our [*Spitzer*]{} observations did not explore the second part of its transit window (see Sect. 5.9 and Fig. 10). The transiting nature of HD13808b and HD115617b is left unexplored by our observations. For the radius of each planet, our MCMC analysis assumed a uniform prior PDF ranging from a pure-iron composition radius to 11 $R_{\oplus}$. The fact that a transiting configuration was disfavored by our MCMC analysis for all targeted planets does not preclude the possibility that one or several transits were in the data, but were just too shallow to be noticed by the Markov Chains. To estimate the actual detection threshold of our observations, we performed the following procedure for each of our targets. We created 50 fake transit light curves based on the multiplication of the actual light curve by a transit model of the targeted RV planets, each transit assuming a circular orbit for the planet, a mid-transit time drawn from the prior transit timing PDF, which was derived from the RV analysis, and an impact parameter drawn from a uniform PDF that ranged from 0 to 0.9. For each transit model, the depth was automatically tuned to have a difference in BIC of +9.2 between models neglecting and taking into account the transit. This difference in BIC corresponds to a Bayes factor of $e^{9.2/2}=100$, indicating a decisive selection of the transit model (Jeffreys 1961), i.e. a firm detection (at better than 3-3.5 sigma) of the transit. We then averaged the transit depths (and planet’s radii) derived for the 50 light curves and adopted the resulting value as the detection threshold for the considered dataset. These detection thresholds are given in Table 1, expressed as transit depth (in ppm) and planetary radius. For each planet, they are compared to the planet’s radius assuming pure-iron and Earth-like compositions. The detection threshold radius is smaller than the ones assuming pure-iron and Earth-like compositions for, respectively, 5 and 14 of the 17 planets for which the transiting nature was constrained by our observations. For 12 and 3 planets, we cannot thus fully reject the absence of a transit in our data, provided very metal-rich and Earth-like compositions, respectively. However, no transiting configuration is even midly favored for any planet by our MCMC analysis (from the comparison of the last columns of Table 1), so, considering the excellent photometric precision of our [*Spitzer*]{} data, the hypothesis of a missed transit is clearly unlikely. Using the posterior transit probabilities shown in Table 1, the formula $1-\prod_{i=1:19} P_i(tr,D)$ indeed results in a posterior probability that none of the probed 19 planets transits of 83%, vs 22% for the corresponding prior probability ($1-\prod_{i=1:19} P_i(tr)$). Our multi-Cycle [*Spitzer*]{} transit search explored the transiting nature of 25 RV planets. It detected one or several transits for the planets HD75732e (aka 55Cnce) (Demory et al. 2011, Gillon et al. 2012) and HD219234b (Motalebi et al 2015), confirmed the transiting nature of HD97658b (Van Grootel et al. 2015), discarded or disfavored the transiting nature of 20 planets (including one presented in Ségransan et al., in prep.), and left the one of two planets unconstrained. By discovering the transits of two planets of a few Earth-masses that are suitable for detailed atmospheric characterization, it brought a significant contribution to the study of super-Earths. Statistically speaking, its final result is normal: considering all the planets listed in Table 1 except HD97658b, which we decided to observe only because we knew that it was probably transiting (Dragomir et al. 2013), the sum of the geometric transit probabilities amounts to 196%, i.e. the project was expected to observe the transits of $\sim 2$ planets. The photometric performances demonstrated by [*Spitzer*]{} in this program are illustrated in Fig. 21. This figure compares as a function of the targets’ $K$-magnitude the standard deviations measured in the detrended light curves for a sampling of 30 min to the corresponding formal errors computed following the instructions of the [*Spitzer*]{} Observation Manual (SOM[^7]). At 4.5$\mu$m, the measured standard deviations are well modeled by the linear relationship $\sigma_{30min} = 32.5 + 11.97\times(K_{mag} - 5)$ ppm, while the mean quadratic difference between the measured standard deviations and the formal errors is $35\pm5$ ppm. This quadratic difference is $64\pm9$ ppm at 3.6 $\mu$m, and $45\pm4$ ppm when neglecting HD125612, which seems to be a more active star than our other targets (Sec. 5.15). These quadratic differences can be attributed to the low-frequency noise of instrumental and astrophysical origins that cannot be represented by our instrumental model. Figure 21 shows that this red noise dominates the photometric precision of [*Spitzer*]{}, especially for the brighter targets. Its values are low enough – a few dozens of ppm – to qualify the photometric precision of [*Spitzer*]{} as excellent, and to make it an optimal facility for the search for very low-amplitude transits on bright nearby stars. The [*Spitzer*]{} mission should come to an end in early 2019. Fortunately, the CHEOPS space mission (Broeg et al. 2015) will arrive just in time to take over the search for the transits of super-Earths discovered by RVs around nearby stars. CHEOPS will not benefit from a targets’ visibility as favorable as [*Spitzer*]{}, but its full dedication to transit observations will more than compensate for its geocentric orbit. \[fig:21\] ![Standard deviations of the detrended [*Spitzer*]{} photometry binned per 30 min (open circles) and the corresponding formal errors (triangles) as a function of the $K$-magnitude of the targets. Blue = 3.6 $\mu$m, red = 4.5 $\mu$m. The dashed red line shows the best-fit linear relationship between the standard deviations measured at 4.5 $\mu$m and the $K$-magnitudes, its equation being $\sigma_{30min} = 32.5 + 11.97\times(K_{mag} - 5)$ ppm. ](aa29270_fig21.eps "fig:"){width="9cm"} \[tab:rvsmallplanets\] [cccccccccccc]{} Planet & $M_s$ & $d$ & $K$ & $M_p \sin{i} $ & Detection & Pure-Fe & Earth-like & Prior & Posterior\ & & & & & threshold & radius & radius & $P(tr)$ & $P(tr,D)$\ & \[$M_\odot$\] & \[pc\] & & \[$M_\oplus$\] & ppm/\[$R_\oplus$\] & \[$R_\oplus$\] & \[$R_\oplus$\] & \[%\] & \[%\]\ HD40307b$^{1}$ & 0.78 & 12.8 & 4.79 & 4.3 & 150/0.91 & 1.13 & 1.44 & 6.6 & 0.19\ GJ3634b$^{2}$ & 0.45 & 19.8 & 7.47 & 7.0 & 500/1.05 & 1.31 & 1.65 & 7.0 & 0.50\ HD75732e$^{3, 4}$ & 0.91 & 12.3 & 4.02 & 7.8 & NA$^{a}$ & 1.34 & 1.70 & 28.9 & 100\ HD97658b$^{5}$ & 0.77 & 21.1 & 5.73 & 7.6 & NA$^{a}$ & 1.33 & 1.69 & 4.3 & 100\ HD219134b$^{6}$ & 0.78 & 6.5 & 5.57 & 4.3 & NA$^{a}$ & 1.15 & 1.44 & 9.5 & 100\ BD-061339b$^{7}$ & 0.63 & 20.3 & 6.31 & 6.9 & 235/1.15 & 1.30 & 1.64 & 7.7 & 0.53\ HD1461b$^{7}$ & 1.04 & 23.2 & 4.90 & 6.7 & 170/1.56 & 1.29 & 1.63 & 8.1 & 0.14\ HD10180c$^{7}$ & 1.06 & 39.0 & 5.87 & 13.1 & 180/1.73 & 1.53 & 1.95 & 8.4 & 0.14\ HD13808b$^{7}$ & 0.77 & 28.6 & 6.78 & 11.8 & NA$^{a}$ & 1.50 & 1.90 & 3.5 & 3.5\ HD20003b$^{7}$ & 0.91 & 43.8 & 6.65 & 11.8 & 280/1.79 & 1.49 & 1.90 & 3.4 & 0.54\ HD20781b$^{7}$ & 0.83 & 35.4 & 6.55 & 2.1 & 220/1.39 & 0.93 & 1.17 & 7.1 & 0.15\ HD31527b$^{7}$ & 0.96 & 38.6 & 6.05 & 10.7 & 190/1.64 & 1.46 & 1.85 & 4.4 & 0.46\ HD39194b$^{7}$ & 0.72 & 25.9 & 6.09 & 4.1 & 210/1.22 & 1.12 & 1.42 & 6.4 & 0.46\ HD45184b$^{7}$ & 1.00 & 21.9 & 4.87 & 12.1 & 150/1.39 & 1.50 & 1.91 & 7.7 & 1.29\ HD47186b$^{7}$ & 1.03 & 39.6 & 6.01 & 23.2 & 175/1.59 & 1.77 & 2.26 & 10.5 & 0.00\ HD51608b$^{7}$ & 0.86 & 34.8 & 6.33 & 13.1 & 195/1.39 & 1.54 & 1.95 & 4.0 & 0.11\ HD93385d$^{7}$ & 1.04 & 42.2 & 6.07 & 4.0 & 200/1.78 & 1.11 & 1.41 & 7.9 & 0.23\ HD96700b$^{7}$ & 0.96 & 25.7 & 5.00 & 9.1 & 155/1.58 & 1.40 & 1.77 & 6.8 & 0.09\ HD115617b$^{7}$& 0.94 & 8.6 & 2.96 & 6.2 & NA$^{a}$ & 1.26 & 1.60 & 9.4 & 9.4\ HD125612c$^{7}$ & 1.09 & 54.2 & 6.84 & 19.3 & 290/1.92 & 1.69 & 2.16 & 9.7 & 0.24\ HD134060b$^{7}$ & 1.07 & 24.2 & 4.84 & 10.0 & 165/1.61 & 1.43 & 1.82 & 8.4 & 0.00\ HD181433b$^{7}$ & 0.86 & 26.8 & 6.09 & 7.5 & 190/1.26 & 1.33 & 1.68 & 4.9 & 0.52\ HD215497b$^{7}$ & 0.87 & 43.6 & 6.78 & 6.1 & 250/1.66 & 1.26 & 1.59 & 11.8 & 0.31\ HD219828b$^{7}$ & 1.18 & 72.3 & 6.53 & 20.2 & 155/2.18 & 1.71 & 2.19 & 14.2 & 0.00\ This work is based in part on observations made with the [*Spitzer Space Telescope*]{}, which is operated by the Jet Propulsion Laboratory, California Institute of Technology under a contract with NASA. Support for this work was provided by NASA. M. Gillon is Research Associate at the Belgian Scientific Research Fund (F.R.S-FNRS). This publication makes use of data products from the Two Micron All Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by the National Aeronautics and Space Administration and the National Science Foundation. Arriagada, P., Anglada-Escudé, G., Butler, R. P., et al. 2013, ApJ, 771, 42 Baranne, A., Queloz, D., Mayor, M. et al. 1996, A&AS, 119, 373 Bonfils, X., Gillon, M., Forveille, T., et al. 2011, A&A, 528, A111 Bouchy F., Mayor M., Lovis C., et al., 2009, A&A, 496, 527 Broeg, C., Fortier, A., Ehrenreich, D., et al. 2013, Hot Planets and Cool Stars, Garching, Germany, Edited by Roberto Saglia; EPJ Web of Conferences, Volume 47 Butler, R. P., Vogt, S. S., Marcy, G. W., et al. 2004, ApJ, 617, 580 Casagrande, L., Schoenrich, R., Asplund, M., et al. 2011, A&A, 530, 138 Charbonneau, D., Berta, Z. K., Irwin, J., et al. 2009, Nature, 462, 891 Claret, A. 2000, A&A, 363, 1081 Claret, A., & Bloemen, S. 2011, A&A, 529, A75 Demory, B.-O., Gillon, M., Deming, D., et al. 2011, A&A, 533, A114 Demory, B.-O., Gillon, M., Seager, S., et al. 2012, ApJL, 751, 28 Díaz, R. F., Ségransan, D., Udry, S., et al. 2016, A&A, 585, A134 Dragomir, D., Matthews, J. M., Eastman, J. D., et al. 2013, ApJL, 772, 2 Dumusque, X., Pepe, F., Lovis, C., & Latham, D. W., 2015, ApJ, 808, 171 Ehrenreich, D., Bourrier, V., Wheatley, P. J., et al. 2015, Nature, 522, 459 Fazio, G. G., Hora, J. L., Allen, L. E., et al., 2004, ApJS, 154, 10 Fischer, D., Marcy, G., Butler, P., et al. 2007, ApJ, 669, 1336 Flower, P. J. 1996, ApJ, 469, 355 Gelman, A., Rubin, D. 1992, Statist. Sci., 7, 457 Gillon M., Pont F., Demory B.-O., et al, 2007, A&A, 472, L13 Gillon, M., Deming, D., Demory, B.-O., et al. 2010, A&A, 518, A25 Gillon, M., Demory, B.-O., Benneke, B., et al. 2012a, A&A, 539, A28 Gillon, M., Triaud, A. H. M. J., Fortney, J. J., et al. 2012b, A&A, 542, A4 Gillon, M., Demory, B.-O., Madhusudhan, N., et al. 2014, A&A, 563, A21 Glebocki, R. & Gnacinski, P. 2005, Catalogue of Stellar Rotational Velocities, ESA, SP-560, 571 Gray, R. O., Corbally, C. J., Garrison, R. F., et al. 2003, AJ, 126, 2048 Gray, R. O., Corbally, C. J., Garrison, R. F., et al. 2006, AJ, 132, 161 Hog, E., Fabricius, C., Makarov, V. V., et al. 2000, A&A, 355, 27 Houk, N. & Cowley, A. P. 1975, Michigan Spectral Survey, Ann Arbor, Dep. Astron., Univ. Michigan, 1 Howard, A. W., Marcy, G. W., Bryson, S. T., et al. 2012, ApJS, 201, 15 Ingalls, J. G., Krick J. E., Carey, S. J., et al. 2012, Proc. SPIE 8442 Jeffreys, H. 1961, [*The Theory of Probability*]{}, 3rd ed. Oxford University Press. Kammer, J. A., Knutson H. A., Howard, A. W., et al. 2014, ApJ, 781, 103 Kharchenko, N. V. 2001, Kinematika i Fizika Nebesnykh Tel, 17, 409 Knutson, H. A., Charbonneau, D., Allen, L. A. et al., 2008, ApJ, 673, 526 Knutson, H. A., Benneke, B., Deming, D. & Homeier, D. 2014a, Nature, 505, 66 Knutson, H. A., Dragomir, D., Kreidberg, L., et al. 2014b, ApJ, 794, 155 Kreidberg, L., Bean, J., Desert, J.-M., et al., 2014, Nature, 505, 69 Lanotte, A. A., Gillon, M., Demory, B.-O., et al. 2014, A&A, 572, 73 Lo Curto G., Mayor, M., Benz, W., et al. 2010, A&A, 512, A48 Lo Curto, G., Mayor, M., Benz, W., et al. 2013, A&A, 551, 59 Lanotte, A., Gillon, M., Demory, B.-O., et al. 2014, A&A, 572, 73 Lovis, C., Ségransan, D., Mayor, M., et al. 2011, A&A, 528, A112 Mandel, J., & Agol, E. 2002, ApJ, 580, 171 Mayor, M., Pepe, F., Queloz, D., et al. 2003, The Messenger, 114, 20 Mayor, M., Marnier, M., Lovis, C., et al. 2011, A&A (submitted), arXiv1109.2497 Melo, C., Santos, N. C., Gieren, W., et al. 2007, A&A, 467, 721 Meschiari, S., Wolf, A. S., Rivera, E. et al., 2009, PASP, 121, 1016 Motalebi, F., Udry, S., Gillon, M., et al. 2015, A&A, 584, A72 Noyes, R. W., Hartmann, L. W., Baliunas, S. L., et al. 1984, ApJ, 279, 763 Press, W. H., Teukolsky, S. A., Vetterling, W. T., Flannery, B. P. 1992, Numerical Recipes in Fortan 77: The Art of Scientific Computing, Cambridge Universty Press Queloz, D., Henry, G. W., Sivan, J. P., et al. 2001, A&A, 379, 279 Rafikov R. R., 2006, ApJ, 648, 666 Rivera, E. J, Butler, R. P., Vogt, S. S., et al. 2010, ApJ, 708, 1492 Rowe, J. F., Bryson, S. T., Marcy, G. W., et al. 2014, ApJ, 784, 45 Santos, N. C., Bouchy, F., Mayor, M., et al. 2004, A&A, 426, L19 Santos, N. C., Sousa, S. G., Mortier, A., et al. 2013, A&A, 556, A150 Santos, N. C., Santerne, A., Faria, J. P., et al. 2016, A&A, 592, A13 Schwarz, G. E. 1978, Annals of Statistics, 6, 461 Seager, S., Kuchner, M., Hier-Majumder, C. A., Militzer, B. 2007, ApJ, 669, 1279 Seager, S. 2010, Exoplanet Atmospheres, Princeton University Press Skrutskie, M. F., Cutri, R. M., Stiening, R., et al. 2006, AJ, 131, 1163 Stevens, D. J. & Gaudi, B. S. 2013, PASP, 125, 933 Stevenson, K. B., Harrington, J., Nymeyer, S., et al. 2010, Nature, 464, 1161 Stevenson, K. B., Harrington, J., Fortney, J., et al. 2012, ApJ, 754, 136 Stetson, P. B. 1987, PASP, 99, 111 Valencia D., Guillot, T., Parmentier, V., Freedman, R. S. 2013, ApJ, 775, 10 Van Grootel, V., Gillon, M., Valencia, D., et al. 2014, ApJ, 786, 2 Van Leeuwen, F. 2007, A&A, 474, 653 Vogt, S. S., Wittenmeyer, R. A., Butler, R. P., et al. 2010, ApJ, 708, 1366 Wisniewski, J. P., Ge, J., Crepp, J. R., et al. 2012, AJ, 143, 107 Winn, J. N., Matthews, J. M., Dawson, R. I., et al. 2011, ApJ, 733, L18 Wright, J. T. 2005, PASP, 117, 657 Tables ====== The following tables describe the targets of this work (Tables A.1 to A.4) and the [*Spitzer*]{} observations (Tables A.5 to A.11). \[tab:targets\] [cccccc]{} Star & BD-061339 & HD1461 & HD10180 & HD13808 & HD20003\ $d$ \[parsec\] & $20.3 \pm 0.7^{(1)}$ & $23.2 \pm 0.3^{(1)}$ & $39.0 \pm 0.6^{(1)}$ & $28.6 \pm 0.5^{(1)}$ & $43.8 \pm 1.2^{(1)}$\ $V$ \[mag\] & $9.67\pm 0.04^{(2)}$ & $6.46 \pm 0.01^{(2)}$ & $7.32 \pm 0.01^{(12)}$ & $8.38 \pm 0.01^{(2)}$ & $8.37 \pm 0.01^{(2)}$\ $K$ \[mag\] & $6.31 \pm 0.02^{(3)}$ & $4.90 \pm 0.02^{(3)}$ &$5.87 \pm 0.02^{(3)}$ & $6.25 \pm 0.02^{(3)}$ & $6.65 \pm 0.02^{(3)}$\ Spectral type & K7V/M0V$^{(4)}$ & G3V$^{(7)}$ & G1V$^{(13)}$ & K2V$^{(13)}$ & G8V$^{(16)}$\ $T_{eff}$ \[K\]$^a$ & $4040 \pm 50^{(5)}$ & $5765 \pm 50^{(8)}$ & $5910 \pm 50^{(14)}$ & $5035 \pm 50^{(8)}$ & $5495 \pm 50^{(8)}$\ \[Fe/H\] \[dex\] & $-0.07 \pm 0.10^{(5)}$ & $0.19 \pm 0.01^{(8)}$ & $0.08 \pm 0.01^{(14)}$ & $-0.21 \pm 0.02^{(8)}$ & $0.04 \pm 0.02^{(8)}$\ $M_\ast$ \[$M_\odot$\] & $0.63 \pm 0.03^{(5)}$ & $1.04 \pm 0.07^{(8)}$ & $1.06 \pm 0.05^{(14)}$ & $0.77 \pm 0.06^{(8)}$ & $0.91 \pm 0.07^{(8)}$\ Bolometric Correction & $-1.07 \pm 0.03^{(6)}$ & $-0.082 \pm 0.009^{(6)}$ & $-0.058 \pm 0.009^{(6)}$ & $-0.290 \pm 0.021^{(6)}$ & $-0.140 \pm 0.013^{(6)}$\ $R_\ast$ \[$R_\odot$\]$^b$ & $0.69 \pm 0.02$ & $1.10 \pm 0.02$ & $1.18 \pm 0.02$ & $0.81 \pm 0.02$ & $0.98 \pm 0.02$\ $\log{g}$ \[cgs\] & $4.56 \pm 0.04$ & $4.37 \pm 0.04$ & $4.32 \pm 0.03$ & $4.51 \pm 0.05$ & $4.41 \pm 0.04$\ RV & & & & &\ Data & HARPS: 102$^{(4)}$+10 & HARPS: 167$^{(9)}$+82$^{(10)}$+5 & HARPS: 190$^{(14)}$+63 & HARPS: 133$^{(9)}$+89 & HARPS: 104$^{(9)}$+77\ & PFS: 15$^{(5)}$ & Keck: 144$^{(11)}$ & & &\ Model & 2 Keplerians & 3 Keplerians & 6 Keplerians & 2 Keplerians & 4 Keplerians\ & + linear trend & + quadratic trend & & + quadratic trend &\ & & +CCF function & & +CCF function &\ Jitter noise \[[ ]{}\] & HARPS: 2.7 & HARPS: 1.5 & 1.5 & 1.7 & 1.4\ & PFS: 2.7 & Keck: 2.4 & 1.6$^e$ & & &\ Planet & BD-061339b$^{(4)}$ & HD1461b$^{(11)}$ & HD10180c$^{(14)}$ & HD13808b$^{(9, 15)}$ & HD 20003b$^{(9, 17)}$\ $M_p \sin{i}$ \[$M_\oplus$\] & $6.93 \pm 0.96 $ & $6.73 \pm 0.47$ & $13.11 \pm 0.62$ & $11.83 \pm 0.88$ &$11.79 \pm 0.61$\ Min. $R_p$ \[$R_{\oplus}$\]$^c$ & $1.30 \pm 0.05$ & $1.29 \pm 0.03$ & $1.53 \pm 0.02$ & $1.50 \pm 0.03$ & $1.49 \pm 0.02$\ Min. $(R_p/R_\ast)^2$ \[ppm\]$^c$ & $298 \pm 29$ & $115 \pm 6$ & $142 \pm 6$ & $ 286 \pm 18$ & $191 \pm 9$\ $T_{eq}$ \[K\]$^d$ & $796 \pm 17$ & $1154 \pm 20$ & $1223 \pm 18$ & $674 \pm 14$ & $836 \pm 12$\ $T_{0}$-2450000 \[BJD$_{TDB}$\] & $6627.48_{-0.16}^{+0.18}$ & $6549.30 \pm 0.12$ & $5212.837_{-0.074}^{+0.059}$ & $6537.49 \pm 0.26$ & $6538.34 \pm 0.36$\ $P$ \[d\] & $3.87310 \pm 0.00037$ & $5.77198 \pm 0.00030$ & $5.75931 \pm 0.00021$ & $14.1853 \pm 0.0019$ & $11.8489 \pm 0.0015$\ $W_{b=0}$ \[min\] & $139 \pm 15$ & $218 \pm 11$ &$238 \pm 10$ & $246 \pm 13$ & $388 \pm 23$\ $K$ \[[ ]{}\] & $3.90 \pm 0.52$ & $2.34 \pm 0.13$ & $4.51 \pm 0.15$ & $3.73 \pm 0.20$ & $3.84 \pm 0.20$\ $a$ \[AU\] & $0.04136 \pm 0.00066$ & $0.0638 \pm 0.0015$ & $0.0641 \pm 0.0010$ & $0.1050 \pm 0.0028$ & $0.09817 \pm 0.00072$\ $e$ & $0.11_{-0.08}^{+0.11}$ & $0.037_{-0.026}^{+0.041}$ &$0.045_{-0.030}^{+0.037}$ & $0.042_{-0.029}^{+0.043}$ & $0.377 \pm 0.047$\ $\omega$ \[deg\] & $192_{-93}^{+69}$ & $134_{-120}^{+110}$ & $320_{-41}^{+51}$ & $272_{-75}^{+85}$ & $267.1_{-8.3}^{+7.6}$\ Prior $P_{transit}$ \[%\] & $7.72_{-0.71}^{+0.77}$ & $8.06 \pm 0.38$ & $8.38_{-0.33}^{+0.30}$ & $3.52 \pm 0.18$ & $3.43 \pm 0.15$\ Prior $P_{occultation}$ \[%\] & $7.9_{-0.7}^{+1.1}$ & $8.00 \pm 0.38$ & $8.78_{-0.32}^{+0.38}$ & $3.68_{-0.18}^{+0.21}$ & $7.50_{-0.56}^{+0.62}$\ \[tab:targets\] [cccccc]{} Star & HD20781 & HD31527 & HD39194 & HD45184\ $d$ \[parsec\] & $35.4 \pm 1.3^{(1)}$ & $38.6 \pm 0.9^{(1)}$ & $21.9 \pm 0.2^{(1)}$\ $V$ \[mag\] & $8.44 \pm 0.01^{(2)}$ & $7.48 \pm 0.01^{(2)}$ & $8.08 \pm 0.01^{(2)}$ & $6.37 \pm 0.01^{(4)}$\ $K$ \[mag\] & $6.55 \pm 0.02^{(3)}$ & $6.05 \pm 0.02^{(3)}$ & $6.09 \pm 0.02^{(3)}$ & $4.87 \pm 0.02^{(3)}$\ Spectral type & K0V$^{(4)}$ & G0V$^{(9)}$ & K0V$^{(4)}$ & G2V$^{(4)}$\ $T_{eff}$ \[K\]$^a$ & $5255 \pm 50^{(5)}$ & $5900 \pm 50^{(5)}$ & $5205 \pm 50^{(5)}$ & $5870 \pm 50^{(5)}$\ \[Fe/H\] \[dex\] & $-0.11 \pm 0.02^{(5)}$ & -$0.17\pm 0.01^{(5)}$ & $-0.61 \pm 0.02^{(5)}$ & $0.04 \pm 0.01^{(5)}$\ $M_\ast$ \[$M_\odot$\] & $0.83 \pm 0.06^{(5)}$ & $0.96 \pm 0.07^{(5)}$ & $0.72 \pm 0.05^{(2)}$ & $1.00 \pm 0.07^{(5)}$\ Bolometric Correction & $-0.208 \pm 0.016^{(6)}$ & $-0.059 \pm 0.009^{(6)}$ & $-0.225 \pm 0.017^{(6)}$ & $-0.058 \pm 0.008^{(6)}$\ $R_\ast$ \[$R_\odot$\]$^b$ & $0.86 \pm 0.02$ & $1.09 \pm 0.02$ & $0.77 \pm 0.02$ & $1.04 \pm 0.02$\ $\log{g}$ \[cgs\] & $4.49 \pm 0.04$ & $4.35 \pm 0.04$ & $4.52 \pm 0.04$ & $4.40 \pm 0.04$\ RV & & & & &\ Data & HARPS: 96$^{(7)}$+117 & HARPS: 167$^{(7)}$+75 & HARPS: 133$^{(7)}$+128 & HARPS: 82$^{(7)}$+92\ Model & 4 Keplerians & 3 Keplerians & 3 Keplerians & 2 Keplerians\ & & & + quadratic trend & + quartic trend\ & & & + CCF function & + CCF function\ Jitter noise \[[ ]{}\] & 1.2 & 1.2 & 1.1 & 1.95\ Planet & HD20781b$^{(8)}$ & HD31527b$^{(7, 8)}$ & HD39194b$^{(7, 9)}$ & HD 45184b$^{(7, 8)}$\ $M_p \sin{i}$ \[$M_\oplus$\] & $2.12 \pm 0.35 $ & $10.68 \pm 0.71$ & $4.08 \pm 0.32$ &$12.08 \pm 0.86$\ Min. $R_p$ \[$R_{\oplus}$\]$^c$ & $0.93 \pm 0.04$ & $1.46 \pm 0.03$ & $1.12 \pm 0.03$ & $1.50 \pm 0.03$\ Min. $(R_p/R_\ast)^2$ \[ppm\]$^c$ & $99 \pm 11$ & $150 \pm 8$ & $ 179 \pm 13$ & $175 \pm 10$\ $T_{eq}$ \[K\]$^d$ & $993 \pm 20$ & $839 \pm 15$ & $935 \pm 19$ & $1143 \pm 20$\ $T_{0}$-2450000 \[BJD$_{TDB}$\] & $6613.92 \pm 0.23$ & $6792.65 \pm 0.36$ & $6285.93 \pm 0.15$ & $6317.67 \pm 0.13$\ $P$ \[d\] & $5.3144 \pm 0.0011$ & $16.5547 \pm 0.0033$ & $5.63675 \pm 0.00044$ & $5.88607 \pm 0.00032$\ $W_{b=0}$ \[min\] & $180 \pm 6$ & $293 \pm 19$ & $174 \pm 10$ & $209 \pm 13$\ $K$ \[[ ]{}\] & $0.88 \pm 0.14$ & $2.78 \pm 0.13$ & $1.83 \pm 0.12$ & $4.33 \pm 0.23$\ $a$ \[AU\] & $0.0560 \pm 0.0014$ & $0.1253 \pm 0.0031$ & $0.0555\pm 0.0013$ & $0.0638 \pm 0.0015$\ $e$ & 0 (fixed) & $0.117 \pm 0.052$ &$0.033_{-0.023}^{+0.035}$ & $0.122_{-0.057}^{+0.053}$\ $\omega$ \[deg\] & - & $42_{-26}^{+23}$ & $224_{-90}^{+97}$ & $178_{-27}^{+30}$\ Prior $P_{transit}$ \[%\] & $7.14 \pm 0.25 $ & $4.40\pm 0.29$ & $6.38 \pm 0.36$ & $7.72_{-0.44}^{+0.51}$\ Prior $P_{occultation}$ \[%\] & $7.14 \pm 0.25$ & $3.81 \pm 0.22$ & $6.57_{-0.33}^{+0.42}$ & $7.67_{-0.43}^{+0.48}$\ \[tab:targets\] [cccccc]{} Star & HD47186 & HD51608 & HD93385 & HD96700 & HD115617\ $d$ \[parsec\] & $39.6 \pm 1.0^{(1)}$ & $34.8 \pm 0.7^{(1)}$ & $42.2 \pm 1.3^{(1)}$ & $25.7 \pm 0.4^{(1)}$ & $8.56 \pm 0.02^{(1)}$\ $V$ \[mag\] & $7.63 \pm 0.01^{(2)}$ & $8.17 \pm 0.01^{(2)}$ & $7.49 \pm 0.01^{(2)}$ & $6.50 \pm 0.01^{(2)}$ & $4.73 \pm 0.01^{(4)}$\ $K$ \[mag\] & $6.01\pm 0.03^{(3)}$ & $6.33 \pm 0.02^{(3)}$ &$6.07 \pm 0.02^{(3)}$ & $5.00 \pm 0.02^{(3)}$ & $2.96 \pm 0.24^{(3)}$\ Spectral type & G5V$^{(4)}$ & G7V$^{(4)}$ & G2/G3V$^{(4)}$ & G1/G2V$^{(4)}$ & G5V$^{(4)}$\ $T_{eff}$ \[K\]$^a$ & $5675 \pm 50^{(5)}$ & $5360 \pm 50^{(5)}$ & $5975 \pm 50^{(5)}$ & $5845 \pm 50^{(5)}$ & $5575 \pm 50^{(5)}$\ \[Fe/H\] \[dex\] & $0.23 \pm 0.02^{(5)}$ & -$0.07 \pm 0.01^{(5)}$ & $0.02 \pm 0.01^{(5)}$ & $-0.18 \pm 0.01^{(5)}$ & $0.01 \pm 0.05^{(5)}$\ $M_\ast$ \[$M_\odot$\] & $1.03 \pm 0.07^{(5)}$ & $0.86 \pm 0.06^{(5)}$ & $1.04 \pm 0.07^{(5)}$ & $0.96 \pm 0.07^{(5)}$ & $0.94 \pm 0.08^{(5)}$\ Bolometric Correction & $-0.100 \pm 0.010^{(6)}$ & $-0.176 \pm 0.016^{(6)}$ & $-0.049 \pm 0.007^{(6)}$ & $-0.069 \pm 0.008^{(6)}$ & $-0.120 \pm 0.012^{(6)}$\ $R_\ast$ \[$R_\odot$\]$^a$ & $1.10 \pm 0.02$ & $0.91 \pm 0.02$ & $1.15 \pm 0.02$ & $1.16 \pm 0.02$ & $0.98 \pm 0.02$\ $\log{g}$ \[cgs\] & $4.37 \pm 0.04$ & $4.45 \pm 0.04$ & $4.33 \pm 0.04$ & $4.29 \pm 0.04$ & $4.43 \pm 0.05$\ RV & & & & &\ Data & HARPS: 66$^{(7)}$+67 & HARPS: 118$^{(8)}$+92 & HARPS: 127$^{(8)}$+106 & HARPS: 146$^{(8)}$+98 & AAT: 126$^{(10)}$\ & & & & & + Keck: 80$^{(10)}$\ Model & 2 Keplerians & 2 Keplerians & 3 Keplerians & 2 Keplerians & 3 Keplerians\ & & + quadratic trend & & + quadratic trend &\ & & + CCF function & & + CCF function &\ Jitter noise \[[ ]{}\] & 0.9 & 1.2 & 1.3 & 1.6 & Keck: 2.3\ & & & & & AAT: 2.2\ Planet & HD47186b$^{(7)}$ & HD51608b$^{(8, 9)}$ & HD93385d$^{(10)}$ & HD96700b$^{(7, 10)}$ & HD 115617b$^{(11)}$\ $M_p \sin{i}$ \[$M_\oplus$\] & $23.2 \pm 1.1 $ & $13.12 \pm 0.77$ & $3.97 \pm 0.48$ & $9.05 \pm 0.64$ &$5.86 \pm 0.66$\ Min. $R_p$ \[$R_{\oplus}$\]$^c$ & $1.77 \pm 0.02$ & $1.54 \pm 0.02$ & $1.11 \pm 0.03$ & $1.40 \pm 0.02$ & $1.24 \pm 0.04$\ Min. $(R_p/R_\ast)^2$ \[ppm\]$^c$ & $217 \pm 9$ & $239 \pm 13$ & $79 \pm 7$ & $ 121 \pm 6$ & $135 \pm 9$\ $T_{eq}$ \[K\]$^d$ & $1277 \pm 22$ & $749 \pm 14$ & $1129 \pm 19$ & $1087 \pm 19$ & $1190 \pm 24$\ $T_{0}$-2450000 \[BJD$_{TDB}$\] & $5179.972 \pm 0.021$ & $6379.95_{-0.16}^{+0.18}$ & $6364.17 \pm 0.27$ & $6521.26 \pm 0.16$ & $5280.71 \pm 0.17$\ $P$ \[d\] & $4.084575 \pm 0.000043$ & $14.0702 \pm 0.0015$ & $7.3422 \pm 0.0014$ & $8.12541 \pm 0.00068$ & $4.21504 \pm 0.00061$\ $W_{b=0}$ \[min\] & $190 \pm 6$ & $255 \pm 11$ &$227_{-24}^{+20}$ & $273_{-13}^{+16}$ & $174 \pm 9$\ $K$ \[[ ]{}\] & $9.1 \pm 1.3$ & $3.85 \pm 0.14$ & $4.51 \pm 0.15$ & $2.97 \pm 0.15$ & $2.44 \pm 0.24$\ $a$ \[AU\] & $0.0505 \pm 0.0012$ & $0.1084 \pm 0.0025$ & $0.0749 \pm 0.0017$ & $0.0780 \pm 0.0019$ & $0.0500 \pm 0.0014$\ $e$ & $0.039\pm 0.014$ & $0.033_{-0.023}^{+0.033}$ &$0.13_{-0.09}^{+0.10}$ & $0.049_{-0.034}^{+0.049}$ & $0.078_{-0.055}^{+0.083}$\ $\omega$ \[deg\] & $57 \pm 23$ & $130_{-66}^{+74}$ & $87 \pm 60$ & $293 \pm 74$ & $58_{-160}^{+73}$\ Prior $P_{transit}$ \[%\] & $10.45 \pm 0.34$ & $3.97_{-0.16}^{+0.19}$ & $7.9_{-0.7}^{+1.1}$ & $6.76_{-0.36}^{+0.33}$ & $9.45_{-0.63}^{+0.99}$\ Prior $P_{occultation}$ \[%\] & $9.84 \pm 0.31$ & $3.84 \pm 0.17$ & $6.68_{-0.61}^{+0.54}$ & $7.12_{-0.34}^{+0.45}$ & $8.89_{-0.72}^{+0.63}$\ \[tab:targets\] [cccccc]{} Star & HD125612 & HD134060 & HD181433 & HD215497 & HD219828\ $d$ \[parsec\] & $54.2 \pm 3.0^{(1)}$ & $24.2 \pm 0.3^{(1)}$ & $26.8 \pm 0.8^{(1)}$ & $43.6 \pm 2.0^{(1)}$ & $72.3 \pm 3.9^{(1)}$\ $V$ \[mag\] & $8.32 \pm 0.02^{(2)}$ & $6.29 \pm 0.01^{(2)}$ & $8.38 \pm 0.01^{(4)}$ & $8.95 \pm 0.02^{(2)}$ & $8.01 \pm 0.01^{(2)}$\ $K$ \[mag\] & $6.84\pm 0.03^{(3)}$ & $4.84 \pm 0.03^{(3)}$ &$6.09 \pm 0.02^{(3)}$ & $6.78 \pm 0.02^{(3)}$ & $6.53\pm 0.02^{(3)}$\ Spectral type & G3V$^{(4)}$ & G3IV$^{(4)}$ & K5V$^{(4)}$ & K3V$^{(4)}$ & G0IV$^{(4)}$\ $T_{eff}$ \[K\]$^a$ & $5915 \pm 50^{(5)}$ & $5965 \pm 50^{(5)}$ & $4880 \pm 50^{(5)}$ & $5000 \pm 100^{(5)}$ & $5890 \pm 50^{(5)}$\ \[Fe/H\] \[dex\] & $0.24 \pm 0.01^{(5)}$ & $0.14 \pm 0.01^{(5)}$ & $0.36 \pm 0.18^{(5)}$ & $0.25 \pm 0.05^{(5)}$ & $0.19 \pm 0.03^{(5)}$\ $M_\ast$ \[$M_\odot$\] & $1.09 \pm 0.07^{(5)}$ & $1.07 \pm 0.07^{(5)}$ & $0.86 \pm 0.17^{(5)}$ & $0.87 \pm 0.11^{(5)}$ & $1.18 \pm 0.08^{(5)}$\ Bolometric Correction & $-0.057\pm 0.008^{(6)}$ & $-0.050 \pm 0.007^{(6)}$ & $-0.360 \pm 0.024^{(6)}$ & $-0.304 \pm 0.048^{(6)}$ & $-0.061 \pm 0.008^{(6)}$\ $R_\ast$ \[$R_\odot$\]$^a$ & $1.03 \pm 0.02$ & $1.15 \pm 0.02$ & $0.84 \pm 0.02$ & $0.97 \pm 0.04$ & $1.60 \pm 0.03$\ $\log{g}$ \[cgs\] & $4.43 \pm 0.03$ & $4.35 \pm 0.04$ & $4.51 \pm 0.11$ & $4.40 \pm 0.08$ & $4.10 \pm 0.04$\ RV & & & & &\ Data & HARPS: 58$^{(7)}$+10 & HARPS: 100$^{(9)}$+50 & HARPS: 107$^{(11)}$+78 & HARPS: 105$^{(7)}$ + 3 & HARPS: 22$^{(21)}$+69\ Model & Keck: 19$^{(8)}$ & 2 Keplerians & 3 Keplerians & 2 Keplerians & 2 Keplerians\ & 3 Keplerians & + CCF & logR’(HK) & & + CCF function &\ & & function & & &\ Jitter noise \[[ ]{}\] & HARPS: 3.2 & 1.3 & 1.0 & 1.3 & 1.5\ & Keck: 4.7 & & & &\ Planet & HD125612c$^{(7)}$ & HD134060b$^{(9, 10)}$ & HD181433b$^{(11)}$ & HD215497b$^{(7)}$ & HD219828b$^{(12)}$\ $M_p \sin{i}$ \[$M_\oplus$\] & $19.3 \pm 2.1 $ & $9.97 \pm 0.60$ & $7.5 \pm 1.1$ & $6.11 \pm 0.78$ &$20.2 \pm 1.2$\ Min. $R_p$ \[$R_{\oplus}$\]$^c$ & $1.69 \pm 0.04$ & $1.43 \pm 0.03$ & $1.33 \pm 0.06$ & $1.26 \pm 0.04$ & $1.71 \pm 0.02$\ Min. $(R_p/R_\ast)^2$ \[ppm\]$^c$ & $226 \pm 14$ & $130 \pm 8$ & $210 \pm 18$ & $ 141 \pm 15$ & $96 \pm 5$\ $T_{eq}$ \[K\]$^d$ & $1269 \pm 21$ & $1469 \pm 24$ & $753_{-23}^{+29}$ & $1101 \pm 39$ & $1596 \pm 27$\ $T_{0}$-2450000 \[BJD$_{TDB}$\] & $5296.79 \pm 0.14$ & $6416.96_{-0.09}^{+0.10}$ & $6265.33 \pm 0.17$ & $5393.96 \pm 0.14$ & $5410.658 \pm 0.050$\ $P$ \[d\] & $4.15514 \pm 0.00044$ & $3.269555_{-0.000080}^{+0.000092}$ & $9.37518 \pm 0.00056$ & $3.93394 \pm 0.00065$ & $3.834863 \pm 0.000094$\ $W_{b=0}$ \[min\] & $172_{-16}^{+12}$ & $309\pm 16$ &$227 \pm 19$ & $152 \pm 19$ & $271 \pm 13$\ $K$ \[[ ]{}\] & $7.33 \pm 0.73$ & $4.69 \pm 0.19$ & $2.72 \pm 0.13$ & $2.81 \pm 0.27$ & $7.43 \pm 0.27$\ $a$ \[AU\] & $0.0520 \pm 0.0011$ & $0.0441 \pm 0.0010$ & $0.0822_{-0.0058}^{+0.0050}$ & $0.0465 \pm 0.0020$ & $0.0507 \pm 0.0012$\ $e$ & $0.093_{-0.064}^{+0.090}$ & $0.480 \pm 0.034$ & $0.380 \pm 0.041$ & $0.215 \pm 0.096$ & $0.063 \pm 0.036$\ $\omega$ \[deg\] & $120_{-71}^{+66}$ & $258.5 \pm 5.2$ & $198.2 \pm 7.1 $ & $122 \pm 30 $ & $228 \pm 39$\ Prior $P_{transit}$ \[%\] & $9.7_{-0.7}^{+1.1}$ & $8.39 \pm 0.32$ & $4.93_{-0.39}^{+0.49}$ & $11.8 _{-1.4}^{+1.8}$ & $14.18 \pm 0.64$\ Prior $P_{occultation}$ \[%\] & $8.86_{-0.69}^{+0.59}$ & $23.2_{-1.5}^{+1.7}$ & $6.25_{-0.51}^{+0.64}$ & $8.57_{-0.87}^{+0.97}$ & $15.32_{-0.67}^{+0.76}$\ \[tab:targets\] [cccccc]{} Star & BD-061339 & HD1461 & HD10180& &\ Program ID & 90072 & (1): 80220$^{(1)}$ & (1): 60027 & &\ & & (2): 90072 & (2) & (3): 90072 & &\ Observation date & 2013-11-30 & (1): 2011-08-31 & (1): 2010-01-16 & &\ & & (2): 2013-09-13 & (2) : 2013-08-10 & &\ & & & (3): 2013-09-09 & &\ Channel \[$\mu$m\] & 4.5 & 4.5 & (1): 3.6 & &\ & & & (2) & (3): 4.5 & &\ AOR(s)$^a$ & (S1): 48815616 & (1): 42790656 & (1): 38139392 & &\ & (1): 48815360 & (S2): 48816128 & (S2): 48596224 & &\ & (2): 48815104 & (2): 48815872 & (2): 48595968 & &\ & & & (S3): 48595712 & &\ & & & (3): 48595456 & &\ pipeline version & S19.1.0 & S19.1.0 & (1): S18.18.0 & &\ & & & (2) & (3): S19.1.0 & &\ Exposure time \[s\] & 0.4 & 0.1 & 0.1 & &\ $N_{BCD}$$^b$ & (1):1038 & (1): 5548 & (1): 4400 & &\ & (2): 970 & (2): 3600 & (2) & (3): 3175 & &\ Duration \[hr\] & (1): 7.9 & (1): 12.9 & (1): 10.4 & &\ & (2): 7.4 & (2): 8.4 & (2) & (3): 7.4 & &\ Photometric aperture \[pixels\] & (1) & (2): 3.5 & (1) & (2): 3.5 & (1): 3 & &\ & & & (2): 2.5 & &\ & & & (3): 3 & &\ Baseline model$^c$ & (1): $p(w_x + [xy]^2) $ +BM & (1): $p(w_x^2 + w_y3 + [xy]^3 + l^2) $ +BM & (1): $p(t + w_x + w_y^2 + [xy]^2 + l)$ + BM &\ & (2): $p(w_x + w_y + [xy]^2)$ + BM & (2): $p(w_x + [xy]^2) $ + BM & (2): $p(w_x + [xy]^2 + l^2)$ + BM &\ & & & (3): $p(w_x + w_y + [xy]^2 + l)$ + BM &\ Error correction factors$^d$ & (1): $\beta_w$ = 0.94, $\beta_r$=1.21 & (1): $\beta_w$ = 0.97, $\beta_r$=1.91 & (1): $\beta_w$ = 0.95, $\beta_r$=1.62 &\ & (2): $\beta_w$ = 0.91, $\beta_r$=1.24 & (2): $\beta_w$ = 0.94, $\beta_r$=1.34 & (2): $\beta_w$ = 0.97, $\beta_r$=1.26 &\ & & & (3): $\beta_w$ = 0.99, $\beta_r$=1.42 &\ \[tab:targets\] [cccccc]{} Star & HD13808 & HD20003 & HD20781& &\ Program ID & 90072 & 90072 & 90072\ Observation date & (1) 2013-08-15 & 2013-09-01 & 2013-11-16\ & (2) 2013-08-27 & &\ Channel \[$\mu$m\] & 4.5 & 4.5 & 4.5\ AOR(s)$^a$ & (S1): 48814848 & (S): 48408576 & (S): 48817664\ & (1): 48814592 & (1): 48408320 & (1): 48817408\ & (S2): 48818176 & (2): 48408064 & (2): 488817152\ & (2): 48817920 & (3): 48407808 & (3): 48816896\ pipeline version & S19.1.0 & S19.1.0 & (1): S19.1.0\ Exposure time \[s\] & 0.4 & 0.4 & 0.4\ $N_{BCD}$$^b$ & (1) & (2): 970 & (1) & (2): 1498 & (1) & (2): 1367\ & & (3): 580 & (3): 580\ Duration \[hr\] & (1) & (2): 7.4 & (1) & (2): 11.5 & (1) & (2): 10.4\ & & (3): 4.4 & (3): 4.4\ Photometric aperture \[pixels\] & (1): 2.75 & (1) & (2) & (3): 3.5 & (1) & (2) & (3): 2.75\ & (2): 3.5 & &\ Baseline model$^c$ & (1): $p(w_x +w_y + [xy]^2) $ +BM & (1): $p(w_x^2 + w_y^2 + [xy]^2) $ +BM & (1): $p(w_x + w_y + [xy]^2 )$ + BM\ & (2): $p(w_x^2 + w_y^2 + [xy])$ + BM & (2): $p(w_x + w_y + [xy]) $ + BM & (2): $p(w_x^2 + w_y + [xy]^2 )$ + BM\ & & (3): $p(w_x + w_y + [xy])$ + BM & (3): $p(w_x + [xy]^2 )$ + BM\ Error correction factors$^d$ & (1): $\beta_w$ = 0.93, $\beta_r$=1.00 & (1): $\beta_w$ = 0.93, $\beta_r$=1.30 & (1): $\beta_w$ = 0.97, $\beta_r$=1.06\ & (2): $\beta_w$ = 0.96, $\beta_r$=1.55 & (2): $\beta_w$ = 0.95, $\beta_r$=1.12 & (2): $\beta_w$ = 0.91, $\beta_r$=1.14\ & & (3): $\beta_w$ = 0.98, $\beta_r$=2.03 & (3): $\beta_w$ = 0.97 $\beta_r$=1.31\ \[tab:targets\] [cccccc]{} Star & HD31527 & HD39194& &\ Program ID & 90072 & 90072\ Observation date & 2014-05-13 & 2012-12-24\ Channel \[$\mu$m\] & 4.5 & 4.5\ AOR(s)$^a$ & (S): 50091776 & (S): 46914816\ & (1): 50091520 & (1): 46914560\ & (2): 50091264 &\ & (3): 50091008 &\ pipeline version & S19.1.0 & (1): S19.1.0\ Exposure time \[s\] & 0.4 & 0.1\ $N_{BCD}$$^b$ & (1) & (2) & (3): 1558 & (1): 4890\ Duration \[hr\] & (1) & (2) & (3): 11.9 & (1): 11.4\ Photometric aperture \[pixels\] & (1) & (2) & (3) : 3.25 & (1): 3.0\ Baseline model$^c$ & (1): $p(w_x^2 + [xy]^2) $ +BM & (1): $p(w_x + w_y + [xy]^2 + l )$ + BM\ & (2): $p(w_x + w_y + [xy]^2)$ + BM &\ & (3): $p(w_x + w_y^3 + [xy]^2)$ + BM &\ Error correction factors$^d$ & (1): $\beta_w$ = 0.93, $\beta_r$=1.19 & (1): $\beta_w$ = 0.97, $\beta_r$=1.15\ & (2): $\beta_w$ = 0.95, $\beta_r$=1.20 &\ & (3): $\beta_w$ = 0.93, $\beta_r$=1.38 &\ \[tab:targets\] [cccccc]{} Star & HD45184 & HD47186 & HD51608\ Program ID & 90072 & 60027 & 90072\ Observation date & 2013-01-24 & 2009-12-14 & 2013-03-27\ Channel \[$\mu$m\] & 4.5 & 3.6 & 4.5\ AOR(s)$^a$ & (S): 46917376 & 38065664 & (S): 48411392\ & (1): 46917120 & & (1): 48411136\ & & & (2): 48410880\ pipeline version & S19.1.0 & S19.1.0 & (1): S19.1.0\ Exposure time \[s\] & 0.1 & 0.1 & 0.4\ $N_{BCD}$$^b$ & (1): 4880 & 2115 & (1): 1498\ & & & (2): 1300\ Duration \[hr\] & (1): 11.4 & 5 & (1): 11.4\ & & & (2): 9.9\ Photometric aperture \[pixels\] & (1): 3.25 & 2.75 & (1) & (2): 2.25\ Baseline model$^c$ & (1): $p(w_x^2 + w_y + [xy]^2 + l) $ +BM & $p(w_x^2 + [xy]^2 + l) $ +BM & (1): $p(w_x + w_y + [xy]^2 )$ + BM\ & & & (2): $p(w_x + w_y + [xy]^2 )$ + BM\ Error correction factors$^d$ & (1): $\beta_w$ = 0.94, $\beta_r$=1.54 & (1): $\beta_w$ = 1.00, $\beta_r$=1.61 & (1): $\beta_w$ = 0.95, $\beta_r$=1.14\ & & & (2): $\beta_w$ = 0.94, $\beta_r$=1.12\ \[tab:targets\] [cccccc]{} Star & HD93385 & HD96700 & HD115617\ Program ID & 90072 & 90072 & 60027\ Observation date & 2013-03-12 & 2013-08-16 & 2010-03-25\ Channel \[$\mu$m\] & 4.5 & 4.5 & 3.6\ AOR(s)$^a$ & (S): 48407552 & (S): 46916864 & (1): 39138816\ & (1): 48407296 & (1): 46916608 & (2): 39139072\ & (2): 48407040 & (2): 46916352 &\ & (3): 48406784 & &\ pipeline version & S19.1.0 & S19.1.0 & S19.1.0\ Exposure time \[s\] & 0.1 & 0.1 & 0.01\ $N_{BCD}$$^b$ & (1) & (2): 4245 & (1): 4880 & (1): 6600\ & (3): 2325 & (2): 1240 & (2): 6640\ Duration \[hr\] & (1) & (2): 11.9 & (1): 11.4 & (1): 6.2\ & (3): 5.4 & (2): 2.9 & (2): 6.3\ Photometric aperture \[pixels\] & (1) & (2) & (3) : 2.5 & (1) & (2): 2.75 & (1) & (2): 2.25\ Baseline model$^c$ & (1): $p(w_x + w_y^2 + [xy] + l) $ +BM & (1): $p(w_x + w_y + [xy]^2) $ +BM & (1): -\ & (2): $p(w_x + w_y + [xy]) $ +BM & (2): $p(w_x + [xy]^2) $ +BM & (2): -\ & (3): $p(w_x + [xy]^2) $ +BM & &\ Error correction factors$^d$ & (1): $\beta_w$ = 0.96, $\beta_r$=1.00 & (1): $\beta_w$ = 0.99, $\beta_r$=1.36 & (1): -\ & (2): $\beta_w$ = 1.00, $\beta_r$=1.10 & (1): $\beta_w$ = 0.97, $\beta_r$=1.65 & (2): -\ & (3): $\beta_w$ = 1.00, $\beta_r$=1.85 & &\ \[tab:targets\] [cccccc]{} Star & HD125612 & HD134060 & HD181433\ Program ID & 60027 & (1): 60027 & 90072\ & & (2): 90072 &\ Observation date & (1): 2010-04-09 & (1): 2010-04-13 & 2012-12-03\ & (2): 2010-09-10 & (2): 2013-05-04 &\ Channel \[$\mu$m\] & 3.6 & (1): 3.6 & 4.5\ & & (2): 4.5 &\ AOR(s)$^a$ & (1): 38110464 & (1): 38110720 & (S): 46915584\ & (2): 40313600 & (S2): 46916096 & (1): 46915328\ & & (2): 46915840 & (2): 46915072\ pipeline version & S18.18.0 & (1): S18.18.0 & S19.1.0\ & & (2): S19.1.0 &\ Exposure time \[s\] & 0.4 & (1): 0.02 & 0.4\ & & (2): 0.1 &\ $N_{BCD}$$^b$ & (1): 1790 & (1): 6200 & (1): 1563\ & (2): 1310 & (2): 4480 & (2): 1498\ Duration \[hr\] & (1): 13.8 & (1): 6 & (1): 11.9\ & (2): 10.2 & (2): 11.5 &(2): 11.4\ Photometric aperture \[pixels\] & (1): 1.75 & (1): 1.75 & (1): 3.25\ & (2): 1.9 & (2): 2.25 & (2): 3.25\ Baseline model$^c$ & (1): $p(w_x^2 + w_y^2 + [xy]^3 + l) $ +BM & (1): $p(w_x + [xy]^2 + l) $ +BM & (1): $p(w_x + w_y + [xy]^2 )$ + BM\ & (2): $p(t + w_x + w_y^2 + [xy] + l^2) $ +BM & (2): $p(w_x^2 + w_y + [xy]^4) $ +BM & (2): $p(w_x + w_y + [xy]^2 )$ + BM\ Error correction factors$^d$ & (1): $\beta_w$ = 0.81, $\beta_r$=2.02 & (1): $\beta_w$ = 1.00, $\beta_r$=1.62 & (1): $\beta_w$ = 0.95, $\beta_r$=1.27\ & (2): $\beta_w$ = 0.72, $\beta_r$=2.34 & (1): $\beta_w$ = 0.95, $\beta_r$=1.30 & (2): $\beta_w$ = 0.93, $\beta_r$=1.11\ \[tab:targets\] [ccc]{} Star & HD215497 & HD219828\ Program ID & 90072 & 60027 & 90072\ Observation date & 2010-07-16 & (1): 2010-08-01\ & & (2): 2013-03-06\ Channel \[$\mu$m\] & 4.5 & 4.5\ AOR(s)$^{a}$ & 38701568 & (1): 38702336\ & & (2S): 46914304\ & & (2A): 46914048\ & & (2B): 46913792\ pipeline version & S18.18.0 & S19.1.0\ Exposure time \[s\] & 0.4 & 0.4\ $N_{BCD}$$^{b}$ & 1750 & (1): 1440\ & & (2A): 1564\ & & (2B): 448\ Duration \[hr\] & 13.4 & (1): 11\ & & (2A): 11.9\ & & (2B): 3.4\ Photometric aperture \[pixels\] & 3 & (1): 2.5\ & & (2A): 3\ & & (2B): 3\ Baseline model$^{c}$ & $p(w_x + w_y + [xy] + l) $ +BM & (1): $p(w_x + w_y + [xy]^2 + l^2) $ +BM\ & & (2A): $p(w_x + [xy]^2 + l) $ +BM\ & & (2B): $p(w_x + [xy]^2 + l^2) $ +BM\ Error correction factors$^{d}$ & $\beta_w$ = 0.95, $\beta_r$=1.20 & (1): $\beta_w$ = 0.93, $\beta_r$=1.53\ & & (2A): $\beta_w$ = 0.96, $\beta_r$=1.01\ & & (2B): $\beta_w$ = 0.94, $\beta_r$=1.14\ [^1]: The photometric and radial velocity time series used in this work are only available in electronic form at the CDS via anonymous ftp to cdsarc.u-strasbg.fr (130.79.128.5) or via http://cdsweb.u-strasbg.fr/cgi-bin/qcat?J/A+A/ [^2]: Our [*Spitzer*]{} program independently revealed the transiting nature of 55Cnce (Demory et al. 2011). and HD97658b (Dragomir et al. 2013) [^3]: Most HARPS measurements used in this work are available on the ESO/HARPS online archive at http://archive.eso.org/wdb/wdb/eso/repro/form [^4]: http://ssc.spitzer.caltech.edu/warmmission/propkit/pet/starpet [^5]: http://ssc.spitzer.caltech.edu/warmmission/propkit/som [^6]: IRAF is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation. [^7]: http://ssc.spitzer.caltech.edu/warmmission/propkit/som/

ArXiv

--- abstract: 'Intercalation of lithium atoms between layers of 2D materials can alter their atomic and electronic structure. We investigate effects of Li intercalation in twisted bilayers of the transition metal dichalcogenide MoS$_2$ through first-principles calculations, tight-binding parameterization based on the Wannier transformation, and analysis of moiré band structures through an effective continuum model. The energetic stability of different intercalation sites for Li between layers of MoS$_2$ are classified according to the local coordination type and the number of vertically aligned Mo atoms, suggesting that the Li atoms will cluster in certain regions of the moiré superlattice. The proximity of a Li atom has a dramatic influence on the interlayer interaction between sulfur atoms, deepening the moiré potential well and leading to better isolation of the flat bands in the energy spectrum. These results point to the usefulness for the use of chemical intercalation as a powerful means for controlling moiré flat-band physics in 2D semiconductors.' author: - Zheyu Lu - Stephen Carr - 'Daniel T. Larson' - Efthimios Kaxiras bibliography: - 'ref.bib' title: 'Lithium intercalation in MoS$_2$ bilayers and implications for moiré flat bands' --- Introduction ============ Strongly correlated insulating behavior and unconventional superconductivity have recently been observed in magic-angle twisted bilayer graphene [@cao2018correlated; @cao2018unconventional]. The small twist angle between the two graphene layers creates a moiré pattern with a characteristic length scale much greater than the lattice constant of the individual layers. The interlayer hybridization of the two layers’ Dirac-cones results in the emergence of flat bands in the low-energy band structure [@rafi2011]. This reduction of the electrons’ kinetic energy favors the usually weak electron-electron interactions and phonon-electron coupling in graphene, leading to twist-induced correlated behavior. Flat bands and correlated physics have been predicted and observed in other moiré superlattices, such as trilayer graphene on top of hexagonal boron nitride [@chen2019evidence; @chen2019signatures] and twisted bilayer-bilayer graphene [@Shen2019; @cao2019electric; @liu2019spin; @Chebrolu2019; @Koshino2019; @Lee2019]. In addition, novel absorption peaks, interpreted as intralayer and interlayer moiré excitons, have been observed in different twisted bilayer transition metal dichalcogenide (TMDC) moiré superlattices due to enhanced electron-hole interactions [@jin2019observation; @seyler2019signatures; @tran2019evidence; @alexeev2019resonantly]. Moiré flat bands were predicted to form at the band edges of twisted bilayer TMDC systems [@wu2018; @wu2019; @Naik2018; @carr2018duality], and they have recently been observed [@Wang2019]. These discoveries have demonstrated how the twist angle can be a powerful tool for engineering new and interesting properties in two-dimensional van der Waals heterostructures. Experimental control of the twist angle can be combined with other tunable perturbations common in nanomaterial experiments. For example, vertical pressure [@yankowitz2019tuning; @Carr2018pressure; @Chittari2018] and external strain [@Bi2019] have been investigated as additional tools to realize flat bands in a wider range of geometries and twist angles. At the same time, intercalation of Li atoms has been used to electrochemically dope the layers in van der Waals heterostructures [@bediako_heterointerface_2018; @larson2018lithium], and has been predicted to enhance the interlayer coupling in the AA-stacked regions of twisted bilayer graphene [@larson2020effects]. In the present work we explore the effects of lithium intercalation in various untwisted, local stacking arrangements of two layers of MoS$_2$, demonstrating significant amplification of the interlayer interactions caused by nearby lithium intercalants. From the calculated changes in local electronic structure, we present a continuum model for the band structure of twisted bilayer MoS$_2$ at small twist angles. The intercalants enhance the moiré potential, leading to better isolation of the flat bands in the energy spectrum. The paper is organized as follows: In Sect. \[methods\] we describe the density functional theory (DFT) and Wannier transformation formalism used to extract *ab initio* parameters used in our modeling. The crystal structure of twisted bilayer MoS$_2$ is described in Sect. \[energetics\], along with our results on the energetics of Li intercalation between the layers. In Sect. \[interlayer\] we study the effect of Li intercalants on the interlayer coupling using both tight-binding and continuum models. We present the moiré band structure, including the effects of Li atoms, in Sect. \[flat-bands\]. Our conclusions are presented in Sect. \[conclusion\]. Computational Methods {#methods} ===================== DFT calculations were performed using the Vienna *Ab initio* Simulation Package (`VASP`) [@kresse1996efficient; @kresse1996efficiency]. The interaction between ionic cores and valence electrons is described by pseudopotentials of the projector augmented wave type. We used the SCAN meta-GGA exchange correlation functional [@sun2015strongly], along with the rVV10 van der Waals functional [@peng2016versatile]. We employed a slab geometry to model double layers with a 22 Å vacuum region between periodic images to minimize the interaction between slabs. The crystal structure was relaxed until Hellmann-Feynman forces were smaller in magnitude than 0.01 eV/Å for each atom. The plane-wave energy cutoff was 350 eV with a reciprocal space grid of size 17$\times$17$\times$1 for the primitive unit cell, and grids of size 9$\times$9$\times$1, 6$\times$6$\times$1, and 4$\times$4$\times$1 for the 2$\times$2, 3$\times$3, and 4$\times$4 supercells, respectively. We calculated only unrotated geometries, and assessed the implications for moiré systems by sampling over different atomic registries between the layers. To extract tight-binding parameters we transform the plane-wave DFT basis into a basis of maximally localized Wannier functions (MLWF) [@Marzari2012] as implemented in the Wannier90 code [@mostofi2008wannier90; @Pizzi_2020]. For the Wannier transformation of bilayer MoS$_2$ with Li intercalants we use the seven highest valence bands and four lowest conduction bands which are composed of Mo $d$-orbitals and S $p$-orbitals [@fang2015]. We do not need to include the Li $s$-orbitals because electrons from Li are primarily transferred to the surrounding layers and raise the Fermi level into the MoS$_2$ conduction bands, but do not form new $s$-bands near $E_F$. The initial projections are chosen to be the atomic $p$- and $d$-orbitals and the final converged Wannier functions remain very similar to the localized atomic orbitals. The effective Hamiltonian in the Wannier basis is interpreted as the full-range *ab initio* tight-binding Hamiltonian (FTBH) [@fang2015]. In practice, MoS$_2$ is an $n$-type semiconductor [@Chhowalla2016] likely due to sulfur vacancies that form during fabrication [@Yang2019]. Lithium intercalation adds electrons (negative charge carriers) and enhances the $n$-type doping. Thus, for prediction of transport properties we focus on the conduction band-edge, which is at the $K$ and $K'$ points of the monolayer Brillouin Zone. This band-edge has very weak spin-splitting in MoS$_2$ ($\sim$3 meV), and so we perform calculations in the absence of spin-orbit coupling. Although this choice reduces the accuracy of our electronic structure calculations, particularly around the valence $K$-point band edge, it greatly simplifies the tight-binding parametrization and the form of the twisted continuum model. Energetics of Lithium intercalation {#energetics} =================================== Each layer of MoS$_2$ is formed by a triangular lattice of Mo atoms sandwiched between two triangular lattices of sulfur atoms. In the naturally occurring 2H bulk phase the S atoms of each layer surround the Mo atoms with trigonal prismatic coordination, and each consecutive layer is rotated by 180$^\circ$ from the one below. An “aligned" bilayer can also be fabricated, where the consecutive layers have the same orientation. Here we focus on the results for a 2H bilayer; the results for the aligned case are similar. We have not studied the 1T structure, where the Mo atoms are octahedrally coordinated within each of the layers. When one layer of a MoS$_2$ bilayer has a relative twist with respect to the other layer, a large scale moiré pattern forms with periodicity represented by the corresponding moiré supercell, as shown in Fig. \[fig:1\]a. Within the supercell the local stacking arrangement will vary, and along the diagonal of the moiré supercell there are three special local stacking patterns with three-fold rotational symmetry. In graphene these three regions are referred to as AA, AB, and BA stacking, where the “A" and “B" labels refer to the two sublattices of the honeycomb lattice. For a 2H bilayer there are more stacking possibilities, which we label by the pairs of atoms that are vertically aligned, as shown in Fig. \[fig:1\]b. For a general TMDC with chemical formula MX$_2$ the 2H structure allows for XMMX, MM, and XX stacking, while the aligned structure allows for MMXX and MX stacking. In XMMX stacking a chalcogen atom of the bottom layer is directly beneath a metal atom of the top layer, and vice versa. For MM and XX the metal atoms or chalcogen atoms are vertically aligned, respectively. ![ **(a)** Top view of a twisted homobilayer of MoS$_2$ in the 2H phase. Different local atomic configurations occur in the small twist-angle moiré supercell. The highlighted regions correspond to local atomic configurations with three-fold rotational symmetry, labeled as XMMX, MM, XX, and XMMX respectively along the diagonal direction. **(b)** Side and top views of the high-symmetry local configurations. XMMX stacking refers to the configuration with S of the top layer aligned with Mo of the bottom layer and Mo of the top layer aligned with S of the bottom layer; MM (XX) stacking region refers to the configuration where Mo (S) of the two layers are aligned.[]{data-label="fig:1"}](Fig1.jpg){width="0.5\linewidth"} Given these different local stacking arrangements in a moiré supercell, it is important to determine the preferred locations for Li intercalants. For a single MoS$_2$ layer, Li adsorption occurs in the hollows of the lattice formed by S atoms, either directly above a Mo atom or an empty site in the Mo lattice. When intercalating between two MoS$_2$ layers a Li atom in a sulfur hollow of one layer can experience several different arrangements of sulfur atoms from the other layer, resulting in octahedral, trigonal prismatic, or tetrahedral coordination (Fig. \[fig:2\]a). In addition to the sulfur coordination of the Li atom, the number of vertically aligned Mo atoms is also important. Based on this argument, we label the possible intercalation sites by their sulfur coordination and number of vertically aligned Mo atoms. For the 2H bilayer there are 5 possibilities: octahedral-0, octahedral-2, trigonal-1, tetrahedral-0, and tetrahedral-1. The aligned case also has 5 intercalation sites: octahedral-1, trigonal-0, trigonal-2, tetrahedral-0, and tetrahedral-1. The intercalation energy of lithium atoms between layers of MoS$_2$ is defined as follows [@shirodkar2016]: $$E_{\rm{I}} = \frac{1}{n}(E_{\rm{MoS_2}} + n E_{\rm{Li}} - E_{\mathrm{MoS}_2 + n \mathrm{Li}}).$$ Here $E_{\rm MoS_2}$ is the energy of the empty, relaxed bilayer of MoS$_2$, $E_{\rm Li}$ is the energy of a single Li atom in bulk bcc lithium, and $E_{\mathrm{MoS}_2 + n \mathrm{Li}}$ is the energy for the bilayer containing $n$ Li atoms. $E_\mathrm{I}$ gives the decrease in the total energy of the system for each Li ion intercalated. We calculate $E_\mathrm{I}$ for a single Li atom in a primitive cell of the bilayer, with one MoS$_2$ layer shifted (but not rotated) relative to the other layer, in order to produce the desired sulfur coordination and alignment with Mo. In order to understand the effect of Li-Li interactions, we repeated the calculations for a single Li atom in larger (still unrotated) bilayer supercells, which is equivalent to decreasing the Li concentration, $d$. For $N\times N$ supercells containing a single Li intercalant the concentration of Li is: $$d = \frac{1}{Na\times Na\times \sin{\frac{\pi}{3}}} = \frac{2\sqrt{3}}{3N^2a^2},$$ where $a=3.19$ Å is the primitive cell lattice constant. Fig. \[fig:2\](b) shows the intercalation energy for a single Li atom in each of the intercalation sites as a function of the size of the supercell. Larger supercells correspond to lower Li concentration. For fixed Li concentration and the same number of vertically aligned Mo atoms, it is not surprising that octahedral and trigonal prismatic coordination, both with six nearest-neighbor ligands, are almost degenerate and more stable than tetrahedral coordination, which has only four nearest-neighbors. The intercalation energy, and hence stability, also increases with the number of vertically aligned Mo atoms, due to a stabilizing charge transfer between the Li and Mo atoms. From Fig. \[fig:2\](b) we see that for the 2H phase the octahedral-2 location, which corresponds to MM stacking, is the most energetically favorable, followed by the trigonal-1 location (XX stacking) and then the octahedral-0 location (XMMX stacking). For each case, $E_\mathrm{I}$ rapidly reaches a plateau with increasing supercell size (decreasing Li concentration), except for tetrahedral coordination in which case the structure is not stable, possessing negative-frequency phonon modes. The energy cost to increase Li concentration from 0.25 to 1.0 Li per bilayer unit-cell is $\sim$200 meV, and is caused by the repulsive nature of neighboring Li$^{+}$ ions. Because the intercalation energy does not change significantly for 2$\times$2 and larger supercells, in subsequent calculations we will use 2$\times$2 supercells where the Li ions are separated by $\sim$6.4 Å. Note that for Li intercalants between twisted layers of graphene the intercalation energy is sensitive to Li-Li separations up to $\sim$15 Å [@larson2020effects]. But even if similar long-range interactions are present for Li atoms between MoS$_2$ layers, the primary contributions to the intercalation energy are captured already by the results of the 2$\times$2 supercells. Furthermore, as will be shown below, the effect of Li atoms on the interlayer couplings are only relevant for S-S pairs within $\sim$3 Å of each Li atom, so sulfur pairs in the 2$\times$2 supercell will be influenced by at most 1 Li atom. ![ **(a)** Atomic geometries of lithium intercalation sites with different local coordination environments: octahedral, trigonal prismatic, and tetrahedral coordination with six, six, and four S atoms symmetrically arranged around the lithium atom. **(b)** Li intercalation energy for both the 2H and aligned phases as a function of supercell size (horizontal axis); the corresponding Li concentration, $d$, is shown on the top. Each curve corresponds to a specific local coordination type and number of vertically aligned Mo atoms. Blue and orange denote octahedral and trigonal prismatic coordination respectively. Solid, dashed, and dotted lines denote 2, 1, and 0 vertically aligned Mo atoms. The intercalation energy does not change significantly for concentrations below 1 lithium per 2$\times$2 supercell, $\sim$ 0.03 Å$^{-2}$. Increasing the concentration from 0.25 Li/Mo ($2 \times 2$ supercell) to 1 Li/Mo ($1 \times 1$ supercell) has an energy cost of $\sim$200 meV.[]{data-label="fig:2"}](Fig2_annotated.jpg){width="0.5\linewidth"} The Li intercalants also modify the distances between the MoS$_2$ layers. Calculations with 1 Li atom in a $2 \times 2$ supercell indicate that Li always increases the interlayer distance, but the amount depends on its local coordination, as shown in Table \[tab:interlayer\_d\]. Unsurprisingly, the tetrahedral coordination shows the largest change in the interlayer distance because there is a sulfur atom directly above the intercalant. Even in $4 \times 4$ supercells the MoS$_2$ layers remain nearly flat with height variations in the Mo atom of only $0.01$ Å, but due to the lower effective Li concentration the layer separations are closer to the unintercalated values. To obtain accurate height profiles in a moiré system requires a twisted supercell calculation [@larson2020effects]. [ P[4cm]{} P[2cm]{} P[2cm]{} P[2cm]{} ]{} &\ Geometry & 0 Li & 1 Li\ 2H octahedral-2 & 6.26 & 6.48\ 2H octahedral-0 & 6.26 & 6.55\ 2H trigonal-1 & 6.85 & 6.90\ 2H tetrahedral-1 & 6.26 & 6.86\ 2H tetrahedral-0 & 6.26 & 7.02\ Effect of lithium on interlayer couplings {#interlayer} ========================================= In bilayer MoS$_2$ the interlayer coupling is dominated by interactions between the closest sulfur $p$-orbitals. If there is no lithium present, such couplings have been shown to be well described by the Slater-Koster two-center approximation [@slater1954; @fang2015]. However, the introduction of Li atoms at varying distances and orientations will require additional parameters to describe accurately the interaction between S atoms in the two MoS$_2$ layers. We specify the location of a Li atom using polar coordinates $(r_\mathrm{Li},\theta)$, where $r_\mathrm{Li}$ is the distance of the Li atom from the center of the sulfur-sulfur bond of interest, and $\theta$ is angle in the $xy$-plane measured relative to the projection of the S-S bond, as shown in Fig. \[fig:3\]. ![ **(a)** Diagram of the atomic sites and relevant variables for the interlayer coupling function under a three-center approximation. The red dashed box on the right provides a top-down view. The red arrow indicates the S-S bond, and the angle $\theta$ is measured from the projection of that vector into the $xy$-plane. **(b)** Interlayer $p_z$-$p_z$ S orbital interaction with respect to $\rm r_{Li}$, with different colors corresponding to different values of $\theta$. The black dotted horizontal line corresponds to the interaction strength without lithium. The red dotted vertical line corresponds to the range of influence of the Li intercalant.[]{data-label="fig:3"}](Fig3.jpg){width="0.5\linewidth"} From the MLWF basis we can extract the sulfur-sulfur matrix elements as a function of Li position, $(r_\mathrm{Li},\theta)$. In Fig. \[fig:3\]b we plot the $p_z$-$p_z$ matrix element as a function of $r_\mathrm{Li}$, with different colors representing different values of $\theta$. It is clear that the presence of Li can have a dramatic effect on the matrix element when Li is nearby, changing both the magnitude and even the sign. When Li is further than $\sim 3$Å from the center of the S-S bond, the effect is negligible. Furthermore, the orientation of the Li atom has a non-negligible but second-order effect. The other combinations of $p$-$p$ matrix elements show similar behavior. We have analyzed how the matrix elements transform under various symmetry operations to constrain the functional form of the interlayer coupling. We considered the following operations as shown in Fig. \[fig:tmp\]: 1. reflection in the $xz$-plane, 2. rotation by $\pi$ about the $y$-axis, 3. inversion (combination of 1 and 2). For example, consider the relation between $\langle1_x|H_{\alpha}|2_y\rangle$ and $\langle1_x|H_{\gamma}|2_y\rangle$, where lower-case Greek letters refer to a given Li location ($r_\mathrm{Li},\theta$). After applying an $xz$-plane reflection, the $\alpha$ configuration will transform into the $\gamma$ configuration and the S atoms are mapped to themselves: 1$\rightarrow$1, 2$\rightarrow$2, and the $p$-orbitals transform as: $|p_x\rangle\rightarrow|p_x\rangle$, $|p_y\rangle\rightarrow-|p_y\rangle$, and $|p_z\rangle\rightarrow|p_z\rangle$. Thus $\langle 1_x | H_{\alpha} | 2_y\rangle = -\langle 1_x | H_{\gamma} | 2_y\rangle$, or, $t_{xy}(\theta)=-t_{xy}(-\theta)$. Following the same procedure, we can derive the transformations of the couplings under all the three symmetry operations. The results are summarized in Table \[tab:1\]. ![ Top-down view of four symmetry-related configurations under three symmetry operations. **(a-d)** correspond to $\alpha$, $\beta$, $\gamma$, and $\delta$ configuration characterized by ($r_{\rm Li}$, $\theta$), ($r_{\rm Li}$, $\pi-\theta$), ($r_{\rm Li}$, $-\theta$), and ($r_{\rm Li}$, $\theta-\pi$) respectively. The left (right) yellow sphere is the sulfur atom from the bottom (top) layer labelled by 1 (2) and the green sphere denotes the lithium ion. $i$ and $j$ denote the atomic $p$ orbitals $p_x$, $p_y$, and $p_z$. The $\alpha$ and $\beta$ configurations are related with a rotation by $\pi$ around the $y$-axis, $\alpha$ and $\gamma$ are related by reflection in the $xz$-plane, and $\alpha$ and $\delta$ are related by inversion. The $\langle1_i|H_k|2_j\rangle$ ($k$ = {$\alpha$, $\beta$, $\gamma$, $\delta$}) denotes the general interlayer sulfur-sulfur $p$-$p$ orbital couplings.[]{data-label="fig:tmp"}](Fig.jpg){width="0.5\linewidth"} [ P[2cm]{} P[2cm]{} P[2cm]{} P[2cm]{} ]{} & $xz$-plane reflection & $y$-axis $\pi$ rotation & inversion\ $t_{p_x, p_x}$ & $+$ & $+$ & $+$\ $t_{p_x, p_y}$ & $-$ & N/A & N/A\ $t_{p_x, p_z}$ & $+$ & N/A & N/A\ $t_{p_y, p_x}$ & $-$ & N/A & N/A\ $t_{p_y, p_y}$ & $+$ & $+$ & $+$\ $t_{p_y, p_z}$ & $-$ & N/A & N/A\ $t_{p_z, p_x}$ & $+$ & N/A & N/A\ $t_{p_z, p_y}$ & $-$ & N/A & N/A\ $t_{p_z, p_z}$ & $+$ & $+$ & $+$\ Based on the Slater-Koster two-center approximation for the $p$-$p$ orbital interaction, we can use two functions ($V_{pp,\pi}$ and $V_{pp,\sigma}$) to describe the nine couplings. They satisfy the following relation [@fang2015]: $$\begin{aligned} V_{pp,\pi}(\textbf{r}) &= \frac{1}{2}\sum_{i}{t_{p_i^{'},p_i}(\textbf{r})}-\frac{1}{2}\sum_{i,j}{t_{p_i^{'},p_j}(\textbf{r})\frac{r_i r_j}{r^2}}\\ V_{pp,\sigma}(\textbf{r}) &= \sum_{i,j}{t_{p_i^{'},p_j}(\textbf{r})\frac{r_i r_j}{r^2}}, \end{aligned}$$ but in the presence of Li also become functions of $(r_\mathrm{Li}, \theta)$. Applying the results of the symmetry analysis, the two Slater-Koster functions have the following symmetries for a fixed displacement between the sulfur atoms (fixed $\textbf{r}$): $$\begin{array}{rcccl} V_{pp, \sigma}(r_{\rm Li}, \theta) &=& V_{pp, \sigma}(r_{\rm Li}, - \theta) &=& V_{pp, \sigma}(r_{\rm Li}, \pi- \theta ) \\ V_{pp, \pi}(r_{\rm Li}, \theta) &=& V_{pp, \pi}(r_{\rm Li}, - \theta) &=& V_{pp, \pi}(r_{\rm Li}, \pi+ \theta ). \end{array}$$ Thus $V_{pp,\sigma}$ and $V_{pp,\pi}$ are completely determined for $\theta$ in the first quadrant. However, even with these constraints the form of the interlayer coupling is still highly complex. Nonetheless, the MLWF procedure captures the microscopic details of how Li affects the interlayer coupling. Effect of Lithium on the moiré potential {#potential} ======================================== The tight-binding approach demonstrates the dramatic influence that Li intercalants can have on interlayer couplings, but a full model accurately incorporating all the additional degrees of freedom introduced by a Li atom would be extremely complicated. To understand the effects that Li can have on the interlayer interactions in a twisted cell, we turn to a simpler continuum model for twisted bilayer TMDCs [@wu2018; @wu2019]. In contrast to the empirical form of interlayer interactions that have been used to study bilayer graphene [@rafi2011], here we use *ab initio* calculations to accurately obtain the interlayer moiré potential by careful study of the DFT band structure [@Jung2014]. In our case, we will focus on the twisted 2H-bilayer MoS$_2$ Because Li intercalation raises the Fermi level into the conduction bands of MoS$_2$, we will construct a continuum model for the electronic states near the conduction band edge which can be described using the effective mass approximation. The two lowest parabolic conduction bands at the $K$ point are a pair of identical bands from each TMDC layer, with an energy splitting caused by interlayer hybridization. Note that for the 2H orientiaton, the $K$ point of the bottom layer corresponds to the $K'$ point of the top layer. However, since we can safely ignore spin at the conduction $K$ points, $K$ and $K'$ give identical band edges. For a twisted bilayer, a simple but robust model is comprised of two monolayer bands with effective mass $m^*$ and an interlayer coupling $V({\bm{r}})$, taken to be a smooth function of the position in the moiré supercell: $$H_{{\bm{k}}} = \begin{pmatrix} - \frac{\hbar {\bm{k}}^2}{2 m^*} & V({\bm{r}}) \\ V^\dagger({\bm{r}}) & - \frac{\hbar {\bm{k}}^2}{2 m^*} \end{pmatrix}$$ A more careful consideration of the problem also takes into account a stacking-dependent onsite energy for both monolayer bands [@wu2019]. More specifically, this captures not just the band splittings but also the bands’ average energy. In contrast to the explicit interlayer coupling $V({\bm{r}})$, this onsite term represents changes to the in-plane electronic structure due to the presence of the neighboring layer, and so also depends on the stacking. Here we ignore this contribution, since our focus is on how lithium intercalation modifies interlayer hybridization to facilitate the appearance of flat-bands, but such a term can play an important role in the complete model. The interlayer interaction $V({\bm{r}})$ acts like a potential energy for the electronic states in the moiré supercell, thus we refer to it as the interlayer moiré potential. From the variation in this potential throughout the moiré supercell we can obtain the band structure for the twisted bilayer. The potential $V({\bm{r}})$ across the entire moiré supercell can be estimated from its value at the three high-symmetry stacking arrangements. The magnitude of $V({\bm{r}})$ is simply one half of the energy splitting of the lowest two states at the $K$ conduction band edge (Fig. \[fig:4\]), while its phase is determined by details of the atomic geometry. Because the conduction band edge has primarily $d$ orbital character, $V({\bm{r}})$ is largest in the non-intercalated case when the metal atoms are aligned (MM configuration). Taking the MM stacking to correspond to ${\bm{r}} = 0$ and expanding the interlayer potential in the lowest harmonics of the reciprocal lattice vectors for the moiré supercell, ${\bm{G^{sc}}}_i$, gives [@rafi2011; @wu2019] $$V({\bm{r}}) = \frac{V_0}{3} \left( 1 + e^{-i {\bm{G^{sc}_1}} \cdot {\bm{r}}} + e^{-i {\bm{G^{sc}_2}} \cdot {\bm{r}}} \right).$$ Compared to the results of Wu et al [@wu2019], who studied TMDC homobilayers in an aligned phase, our system is in the non-aligned 2H phase. Although this provides a different crystal symmetry the expansion is still a good match in the unintercalated case, as the splitting of the conduction band-edge at $K$ mostly depends on the distance between the metal atoms and not the detailed symmetry of the crystal structure. This enters the ${\bm{k}}$-dependent hamiltonian $H_{{\bm{k}}}$ as three interlayer momentum scattering terms of equal strength [@rafi2011]. The prefactor above is chosen as $V_0/3$ such that $V(0) = V_0$. Note that at the other two high-symmetry stackings, ${\bm{r}} = n({\bm{a^{sc}}}_1 + {\bm{a^{sc}}}_2)/3$ for $n=1,2$, the function $V$ is identically zero. Lying along the diagonal direction of the moiré supercell, these are the same three local stacking regions which provide stable intercalation sites for lithium. The most stable intercalation site for each of these local configurations is shown in Fig. \[fig:4\]a-c, with the band structure for a uniform cell with that stacking configuration immediately below in Fig. \[fig:4\]d-f. Combining the results for the band splitting in each different stacking configuration, we can assemble the moiré potentials shown in Fig. \[fig:4\]g-i. The blue line in all three panels shows the moiré potential along the diagonal direction of the moiré supercell in 2H-bilayer MoS$_2$ without lithium intercalation. It has a potential well around the MM stacking configuration which can localize the electronic states and presumably lead to flat bands and correlated physics. However, the potential well is only $\sim$10 meV deep. Introducing Li between the layers, which will first condense around the MM stacking regions because they offer the most energetically favorable intercalation sites, leads to a deeper moiré potential near the MM stacking by a factor of 2. This makes flat bands and correlated interactions more likely. Further increase of the intercalation density results in lithium accumulation in XX stacking regions, deepening the moiré potential but also changing its shape. Moving the origin (${\bm{r}} = 0$) of the moiré potential amounts to a gauge choice in the interlayer coupling that does not affect the moiré band structure in the continuum model [@rafi2011]. However, the lowest harmonic approximation of $V({\bm{r}})$ now fails as the values at MM and XMMX are not identical. Since we only aim to estimate the effect of lithium intercalation on the moiré band structure we ignore this complication and treat the potential well as for the MM case. To calculate the moiré band structure more accurately, it is necessary to include higher harmonics in the interlayer potential. Obtaining these harmonics from DFT calculations requires band structures for additional Li intercalated geometries, which can be challenging to optimize for low-symmetry stacking configurations. {width="\linewidth"} Emerging flat-bands {#flat-bands} =================== The strong effect of lithium intercalation on the interlayer moiré potential is useful in enhancing the flat bands of a twisted TMDC bilayer. In Fig. \[fig:5\] the band structure obtained from the continuum model for the $K$ point conduction bands of twisted 2H-bilayer MoS$_2$ are shown for two different twist angles ($2^\circ$, $1^\circ$) and the three cases of Li intercalation (zero, Li only at MM, and Li at both MM and XX). To match the DFT band structure calculations the value of $V_0$ for three lithium intercalation cases is taken to be 8, 18, and 33 meV, respectively. The effective mass for the conduction band edge at $K$ is obtained by fitting to the monolayer band structure, yielding $m^*/\hbar = 75 $ meV$^{-1}$ Å$^{-2}$. For a $2^\circ$ twist angle, the unintercalated bilayer does not have a splitting between the lowest two superlattice bands, and is thus very far from flat band behavior. With Li intercalation the strength of the moiré potential can be increased by a factor of 4, allowing flat bands to emerge even at this large twist angle. The $1^\circ$ system has flat bands visible even at zero intercalation, but the intercalation greatly increases the gap between the superlattice bands and flattens the bands further. As Li concentration increases, not only does the lowest conduction band become flatter, but the bottom of the moiré potential well shifts from the regions of MM stacking to regions with XX stacking. As the moiré flat-bands host large “spots” of electronic density confined by this potential well, a shifting of the electron density from MM stacking to XX stacking would be visible, for instance, in scanning tunneling microscopy. {width="\linewidth"} Conclusion ========== We have performed first-principles DFT calculations and *ab initio* tight-binding and continuum modeling to explore the structural and electronic properties of lithium intercalation in twisted bilayer MoS$_2$ systems. We found that lithium intercalants have the lowest energy in MM and XX stacking regions in the moiré supercell. The presence of Li dramatically enhances the interlayer interaction, increasing the depth of the effective moiré potential well from $\sim$8 meV to $\sim$20 meV for Li in MM regions, and further to $\sim$33 meV for Li in XX stacking regions. Using a continuum model with an interlayer interaction based on these DFT results, we show that lithium intercalation can better flatten and isolate the conduction bands near the Fermi level. Furthermore, such moiré flat bands can be realized at larger twist angles with the aid of intercalation. Our results demonstrate that intercalation can be a powerful tool for controlling moiré flat bands in twistronic devices. We acknowledge Yiqi Xie for helpful discussions. The computations in this paper were run on the FASRC Odyssey cluster supported by the FAS Division of Science Research Computing Group at Harvard University. S.C. was supported by ARO MURI Award No. W911NF-14-0247 and by the STC Center for Integrated Quantum Materials, NSF Grant No. DMR-1231319.

ArXiv

--- abstract: 'We study the problem of Diophantine approximation on lines in $\mathbb{C}^2$ with numerators and denominators restricted to Gaussian primes. To this end, we develop analogs of well-known results on small fractional parts of $p\gamma$, $p$ running over the primes and $\gamma$ being a fixed irrational, for Gaussian primes.' address: 'Stephan Baier, Jawaharlal Nehru University, Munirka, School of Physical Sciences, Delhi 11067, India' author: - Stephan Baier title: 'Diophantine approximation on lines in $\mathbb{C}^2$ with Gaussian prime constraints - enhanced version' --- Introduction ============ Let $\psi : \mathbb{N} \rightarrow \mathbb{R}^+$ be a non-increasing function. A real number $x$ is called $\psi$-approximable if there exist infinitely many rational numbers $m/n$ with $m\in\mathbb{N}$ and $n\in \mathbb{Z}\setminus\{0\}$ such that $$\left| x- \frac{m}{n} \right| < \frac{\psi(m)}{|n|}.$$ Khinchin’s famous theorem on Diophantine approximation says that if the series $\sum_{m} \psi(m)$ diverges, then almost every real number, in the sense of Lebesgue measure, is $\psi$-approximable, and if the series converges, then almost every real number is not $\psi$-approximable. It is an interesting problem to extend problems of this kind to manifolds in place of the set of real numbers. A nice survey of results on Diophantine approximation on affine subspaces was given by A. Ghosh [@Gho]. It is also natural to restrict the numerators and denominators in these problems to certain sets such as primes. Indeed, a version of Khinchin’s theorem with prime numerators and denominators was proved by G. Harman [@Harman] and later extended to higher dimensions by H. Jones [@Jones]. For plane curves of the form $y=x^{\tau}$, Harman and Jones proved the folling Khinchin-type result with prime-restrictions in their joint work [@HJ]. \[HaJo\] Let $\varepsilon>0$ and $\tau> 1$. Then for almost all positive $\alpha$ there are infinitely many $p$, $q$, $r$, all prime, such that $$0 < p\alpha - r \le p^{-1/6+\varepsilon} \quad \mbox{and} \quad 0< p\alpha^{\tau} - q \le p^{-1/6+\varepsilon}.$$ In [@BG1], A. Ghosh and the author of the present paper considered the same problem for lines in the plane, passing through the orign, establishing the following. \[real\] Let $\varepsilon>0$ and let $c>1$ be an irrational number. Then for almost all positive $\alpha$, with respect to the Lebesgue measure, there are infinitely many triples $(p,q,r)$ with $p$ and $r$ prime and $q$ an integer such that $$0 < p\alpha - r \le p^{-1/5+\varepsilon} \quad \mbox{and} \quad 0< pc\alpha - q \le p^{-1/5+\varepsilon}.$$ Here the condition on $q$ is relaxed, i.e., we allow $q$ to be an integer, but we get a better exponent of $1/5$ in place of $1/6$. Our approach builds on that of Harman and Jones [@HJ], but we use exponential sums with prime variables instead of zero density estimates at a particular point to get the argument work for lines. In [@BG2], we extended this result to lines in higher dimensional spaces. It is very interesting to consider the same circle of problems in the setting of number fields. Indeed, a number field version of Khinchin’s theorem was proved by D. Cantor [@Can]. A new proof for Khinchin’s theorem in the classical case as well as the case of imaginary quadratic number fields was given by D. Sullivan [@Sul] using geodesic flows. However, it seems that number field versions of results of this type with prime restricitions and Diophantine approximation on general manifolds in the number field setting didn’t receive much attention so far. In this paper, we make a step into this direction by considering Diophantine approximation on lines in $\mathbb{C}^2$, where we restrict numerators and denominators to Gaussian primes. What we prove is an analog of Theorem \[real\] in the setting of the number field $\mathbb{Q}(i)$. It is likely that the method could be extended to imaginary quadratic number fields in general, but we confine our investigation to the simplest case, which requires a large amount of extra work and new arguments already. Our result is the following. \[complex\] Let $\varepsilon>0$ and let $c\in \mathbb{C}\setminus\mathbb{Q}(i)$. Then for almost all $\alpha\in \mathbb{C}$, with respect to the Lebesgue measure, there are infinitely many triples $(p,q,r)$ with $p$ and $r$ Gaussian primes and $q$ a Gaussian integer such that $$\label{simultan} |p\alpha - r| \le |p|^{-1/12+\varepsilon} \quad \mbox{and} \quad |pc\alpha - q| \le |p|^{-1/12+\varepsilon}.$$ The structure of our proof resembles that of Theorem \[real\], but the technical details are more involved. In particular, we need to develop results on Diophantine approximation of numbers in $\mathbb{C}\setminus \mathbb{Q}(i)$ by fractions of Gaussian integers with Gaussian prime denominators in sectors of the complex plane, which is the content of section 3. As a by-product, we prove the following result on Diophantine approximation with Gaussian primes. \[coro\] Let $c$ be a complex number such that $c\not\in \mathbb{Q}(i)$, $\varepsilon>0$ be an arbitrary constant and $-\pi\le \omega_1< \omega_2\le \pi$. Then there exist infinitely many Gaussian primes $p$ such that $$\max\{||\Re(pc)||,||\Im(pc)||\}\le |p|^{-1/12+\varepsilon},$$ where $||z||$ denotes the distance of the real number $z$ to the nearest rational integer. We note that $\max\{||\Re(pc)||,||\Im(pc)||\}$ measures the distance of $pc$ to the nearest Gaussian integer in the sense of the supremum norm. Thus, Theorem \[coro\] states that this distance is infinitely often very small as $p$ runs over the Gaussian primes and $c\in \mathbb{C}\setminus\mathbb{Q}(i)$ is fixed. Similar results have been proved for primes in $\mathbb{Q}$ (see [@Mat], for example). A complex analog of Theorem \[HaJo\] for $\tau\in \mathbb{N}\setminus \{1\}$, in particular the case $\tau=2$ of the complex parabola, would certainly be a very interesting problem to consider as well.\ [**Conventions.**]{} (I) Throughout the sequel, we shall assume that $0<|c|\le 1$ in Theorem \[complex\]. The case $|c|>1$ can be treated similarly, by minor modifications of the method.\ (II) Throughout this paper, we follow the usual convention that $\varepsilon$ is a small enough positive real number.\ [**Acknowledgement.**]{} The author would like to thank Prof. Anish Ghosh for useful discussions about this topic at and after a pleasant stay at the Tata Institute in Mumbai in August 2016. He would further like to thank the anonymous referee for useful comments on a first version of this paper which greatly helped to make this paper self-contained. A Metrical approach =================== Our basic approach for a proof of Theorem \[complex\] is an extension of that in [@BG1 section 2] (see also [@BG2 section 2]) and has its origin in [@HJ]. We first establish the metrical lemma below. Our proof follows closesly the arguments in [@HJ Proof of Lemma 1]. Throughout the sequel, we denote by $\mu(\mathcal{C})$ the Lebesgue measure of a measurable set $\mathcal{C}\subseteq \mathbb{C}$ and we write $$D(a,b):=\left\{z\in \mathbb{C}\ :\ a< |z|\le b\right\}$$ and $$D(a,b,\gamma_1,\gamma_2):=\left\{Re^{i\theta}\ :\ a< R\le b,\ \gamma_1< \theta\le \gamma_2\right\}.$$ \[metric\] Let $\mathcal{S}$ be a subset of the positive integers. Assume that $A$ and $B$ are reals such that $0<A<B$ and let $\mathcal{M}:=D(A,B)$. Let $F_N(\alpha)$ be a non-negative real-valued function of $N$, an element of $\mathcal{S}$, and $\alpha$, a complex number. Let further $G_N$ and $V_N$ be real-valued functions of $N\in \mathcal{S}$ such that the following conditions , , , below hold. $$\label{butzel} G_N \rightarrow \infty \quad \mbox{as } N\in \mathcal{S} \mbox{ and } N\rightarrow\infty.$$ $$\label{VNbound} V_N=o\left(G_N\right) \quad \mbox{as } N\in \mathcal{S} \mbox{ and } N\rightarrow\infty.$$ $$\label{FNint} \begin{cases} \mbox{For all } a,b,\gamma_1,\gamma_2 \mbox{ with } A\le a<b\le B \mbox{ and } -\pi\le \gamma_1<\gamma_2\le \pi \mbox{ we have}\\ \limsup\limits\limits_{\substack{N\in \mathcal{S}\\ N\rightarrow \infty}} \int\limits_{\gamma_1}^{\gamma_2} \int\limits_a^b \frac{F_N\left(Re^{i\theta}\right)}{G_N} \ dR\ d\theta \ge (\gamma_2-\gamma_1)\left(b^2-a^2\right). \end{cases}$$ $$\label{constantK} \begin{cases} \mbox{There is a positive constant } K \mbox{ such that, for any measurable set } \mathcal{C}\subseteq\mathcal{M} \\ \mbox{ and any } N\in \mathcal{S},\ \int\limits_{\mathcal{C}} F_N\left(Re^{i\theta}\right) \ dR\ d\theta\le KG_N\mu(\mathcal{C})+V_N. \end{cases}$$ Then for almost all $\alpha\in \mathcal{M}$, we have $$\limsup\limits_{\substack{N\in \mathcal{S}\\ N\rightarrow\infty}} \frac{F_N(\alpha)}{G_N}\ge 1.$$ We write $$H_N(\alpha):=\frac{F_N(\alpha)}{G_N}$$ and suppose that $$\limsup\limits_{\substack{N\in \mathcal{S}\\ N\rightarrow\infty}} H_N(\alpha)<1$$ on a subset of $\mathcal{M}$ with positive measure. Then there must be a set $\mathcal{A}\subset \mathcal{M}$ with positive measure and a constant $c<1$ with $$\label{limsupbound} \limsup\limits_{\substack{N\in \mathcal{S}\\ N\rightarrow\infty}} H_N(\alpha)\le c \quad \mbox{for all } \alpha\in \mathcal{A}.$$ By the Lebesgue density theorem, for each $\varepsilon>0$ there are $a,b,\gamma_1,\gamma_2$ with $A\le a<b\le B$, $-\pi\le \gamma_1<\gamma_2\le \pi$ and $(\gamma_2-\gamma_1)(b^2-a^2)<1$ such that, if we put $\mathcal{B}:=\mathcal{A}\cap \mathcal{Z}$ with $\mathcal{Z}:= D(a,b,\gamma_1,\gamma_2)$, then $$\mu(\mathcal{B})>(1-\varepsilon)\mu(\mathcal{Z})=(1-\varepsilon)(\gamma_2-\gamma_1)(b^2-a^2)$$ and hence $$\mu(\mathcal{Z}\setminus \mathcal{B})< \varepsilon(\gamma_2-\gamma_1)(b^2-a^2)<\varepsilon.$$ Now, using , $$\begin{split} \int\limits_{\mathcal{Z}} H_N(\alpha) \ dR\ d\theta =& \int\limits_{\mathcal{B}} H_N(\alpha) \ dR\ d\theta + \int\limits_{\mathcal{Z}\setminus \mathcal{B}} H_N(\alpha) \ dR\ d\theta\\ \le & \int\limits_{\mathcal{B}} H_N(\alpha) \ dR\ d\theta +K\varepsilon+\frac{V_N}{G_N}, \end{split}$$ where $\arg(\alpha)=\theta$ and $|\alpha|=R$. So if $$\varepsilon:=\frac{(1-c)(\gamma_2-\gamma_1)\left(b^2-a^2\right)}{2K},$$ then, in view of and , it follows that $$\limsup\limits_{\substack{N\in \mathcal{S}\\ N\rightarrow\infty}} \int\limits_{\mathcal{Z}} H_N(\alpha) \ dR\ d\theta \le c\mu(\mathcal{B})+K\varepsilon=c(\gamma_2-\gamma_1)\left(b^2-a^2\right)+K\varepsilon<(\gamma_2-\gamma_1)\left(b^2-a^2\right).$$ This contradicts and so completes the proof. Now let $F_N(\alpha)$ be the number of solutions to with $|p|\le N$ and for $0<A<B$ let $$G_N(A,B):= C\cdot \frac{A}{B} \cdot \frac{N^{5/3+4\varepsilon}}{\log^2 N},$$ where $C>0$ is a suitable constant only depending on $c$. In sections 4 and 5, we will prove the following. \[Theo\] There exists $C=C(c)>0$ and an infinite set $\mathcal{S}$ of natural numbers $N$ such that the following hold.\ [(i)]{} Let $0 < A < B$ be given. Then for all $a,b,\gamma_1,\gamma_2$ with $A\le a<b\le B$ and $-\pi\le \gamma_1<\gamma_2\le \pi$ we have $$\int\limits_{\gamma_1}^{\gamma_2} \int\limits_a^b F_N\left(Re^{i\theta}\right) \ dR\ d\theta \ge (\gamma_2-\gamma_1)\left(b^2-a^2\right)G_N(A,B)$$ if $N\in \mathcal{S}$ and $N$ large enough.\ [(ii)]{} Let $0<A<B$ be given. Then there exists a constant $K=K(A,B)$ such that, for every $\alpha\in \mathbb{C}$ with $A\le |\alpha|\le B$, we have $$F_N(\alpha)\le KG_N(A,B)+J_N(\alpha)$$ with $$\int\limits_{-\pi}^{\pi} \int\limits_A^B \left|J_N\left(Re^{i\theta}\right)\right| dR d\theta=o\left(G_N(A,B)\right)$$ if $N\in \mathcal{S}$ and $N\rightarrow \infty$. Together with Lemma \[metric\], this implies Theorem \[complex\]. Diophantine approximation with Gaussian primes in sectors ========================================================= It will be of key importance to establish results related to the distribution of Gaussian primes in sectors satisfying certain Diophantine properties. This is the content of this section and of independent interest. Results ------- By $\pi(P_1,P_2,\omega_1,\omega_2)$, we denote the number of primes in $D(P_1,P_2,\omega_1,\omega_2)$. The prime number theorem for Gaussian primes in sectors due to Kubylius [@Kub] implies the following. \[PNT\] If $0\le P_1<P_2$ and $\omega_1<\omega_2\le \omega_1+2\pi$, then $$\pi(P_1,P_2,\omega_1,\omega_2) = \frac{2}{\pi}\cdot \frac{(\omega_2-\omega_1)\left(P_2^2-P_1^2\right)+o\left(P_2^2\right)}{\log P_2^2}$$ as $P_2\rightarrow \infty$. Further, for $\delta>0$ and $c\in \mathbb{C}$, we denote by $\pi_c(P_1,P_2,\omega_1,\omega_2;\delta)$ the number of Gaussian primes $p$ contained in $D(P_1,P_2,\omega_1,\omega_2)$ such that $$\min\limits_{q\in \mathbb{Z}[i]} |pc-q|\le \delta,$$ and by $\pi^{\ast}_c(P_1,P_2,\omega_1,\omega_2;\delta)$ we denote the number of Gaussian primes $p$ contained in $D(P_1,P_2,\omega_1,\omega_2)$ such that $$\max(||\Re(pc)||,||\Im(pc)||)=\min\limits_{q\in \mathbb{Z}[i]} \max\left(|\Re(pc-q)|,|\Im(pc-q)|\right) \le \delta.$$ We shall establish the following theorem in the next subsections. \[signi1\] Let $c\in \mathbb{C}\setminus \mathbb{Q}(i)$. Then there exists an increasing sequence of natural numbers $(M_k)_{k\in \mathbb{N}}$ such that the following holds. If $0\le P_1<P_2\le M_k$, $\omega_1<\omega_2\le \omega_1+2\pi$ and $M_k^{\varepsilon-1/12}<\delta_k\le 1/2$, then $$\pi^{\ast}_c(P_1,P_2,\omega_1,\omega_2;\delta_k) = 4\delta_k^2\pi^{\ast}(P_1,P_2,\omega_1,\omega_2)+o\left(\frac{\delta_k^2 M_k^2}{\log M_k}\right)$$ as $k\rightarrow \infty$. Since $$\pi_c(P_1,P_2,\omega_1,\omega_2;\delta)\ge \pi^{\ast}_c(P_1,P_2,\omega_1,\omega_2;\delta/\sqrt{2}),$$ we immediately deduce the following from Theorem \[PNT\] and \[signi1\]. \[signi\] Under the conditions of Theorem \[signi1\], we have $$\pi_c(P_1,P_2,\omega_1,\omega_2;\delta_k) \ge \frac{\delta_k^2(\omega_2-\omega_1)\left(P_2^2-P_1^2\right)+o\left(\delta_k^2 M_k^2\right)}{\log M_k}$$ as $k\rightarrow \infty$. We shall employ Theorem \[PNT\] and Corollary \[signi\] in section 4, in which Theorem \[Theo\](i) is established. Partial results from the present section 3 are also used in section 5, in which Theorem \[Theo\](ii) is established. Historical notes on Diophantine approximation with primes --------------------------------------------------------- Let $\gamma\in \mathbb{R}$ be an irrational number. Then the continued fraction expansion of $\gamma$ yields infinitely many natural numbers $q$ such that $$\left|\gamma-\frac{a}{q}\right| \le q^{-2},$$ where $(a,q)=1$. In other words, for infinitely many $q\in \mathbb{N}$, we have $$||q \gamma||\le q^{-1},$$ where $||x||$ is the distance of $x$ to the nearest integer. The problem of approximating irrational numbers by rational numbers with prime denominator is considerable more difficult und has a long history. The question is for which $\gamma>0$ one can prove the infinitude of primes $p$ such that $$\label{approxi} ||p \gamma||\le p^{-\gamma+\varepsilon}.$$ The first results in this direction were due to Vinogradov [@Vin] who showed that $\gamma=1/5$ is admissable. Vaughan [@Vau] improved this exponent to $\gamma=1/4$ using his famous identity for the von Mangoldt function. It should be noted that using Vaughan’s method, an [*asymptotic*]{} result of the following form can be established. \[Vaughan\] Let $\gamma\in \mathbb{R}$ be irrational and $\varepsilon>0$ be an arbitrary constant. Then there exists an infinite increasing sequence of natural numbers $(N_k)_{k\in \mathbb{N}}$ such that $$\label{asy} \sum\limits_{\substack{p\le N_k\\ ||p\gamma||\le \delta_k}} 1 \sim 2\delta_k \sum\limits_{p\le N_k} 1 \quad \mbox{as } k\rightarrow \infty$$ if $$\label{range} N_k^{-1/4+\varepsilon}\le \delta_k\le 1/2,$$ where $p$ runs over the rational primes. The next important step was Harman’s work [@Har] in which he used his sieve method to show that holds for infinitely many primes $p$ if $\gamma=3/10$. Harman’s method doesn’t imply the asymptotic for $\delta_k=N_k^{-3/10+\varepsilon}$ since it uses a lower bound sieve. However, Harman’s sieve can be employed to recover Vaughan’s result and hence for the same $\delta_k$-range as in . We further mention the work of Heath-Brown and Jia [@HeJ] who used bounds for Kloosterman sums to obtain a further improvement of the exponent to $\gamma=16/49=1/3-0.0068...$. Finally, the exponent $\gamma=1/3$ was achieved in a landmark paper by Matomäki [@Mat] who incorporated the Kuznetsov formula into the method to bound sums of Kloosterman sums. This exponent $\gamma=1/3$ is considered to be the limit of currently available techniques. In the following, we consider an analog problem for Gaussian primes and establish a result corresponding to Theorem \[Vaughan\] in this context, thereby proving Theorem \[signi1\]. This also implies the infinitude of Gaussian primes in sectors satisfying an inequality corresponding to . To this end, we shall apply a version of Harman’s sieve for $\mathbb{Z}[i]$. Our method will require additional counting arguments, as compared to the classical method. The final proof is carried out in subsection \[con\]. Setup ----- Throughout the following, $c$ is a fixed complex number such that $c\not \in \mathbb{Q}(i)$, and we assume that $$\label{assu} 1\le x_1<x_2, \quad -\pi\le \omega_1<\omega_2\le \pi \quad \mbox{and} \quad 0<\delta\le \frac{1}{2}.$$ We compare the quantities $$S_c(x_1,x_2,\omega_1,\omega_2;\delta):=\sum\limits_{\substack{x_1<\mathcal{N}(p)\le x_2\\ ||pc||\le \delta\\ \omega_1<\arg p \le \omega_2}} 1$$ and $$S(x_1,x_2,\omega_1,\omega_2):=\sum\limits_{\substack{x_1<\mathcal{N}(p)\le x_2\\ \omega_1<\arg p\le \omega_2}} 1,$$ where the sums run over Gaussian primes $p$, $\mathcal{N}(n)=|n|^2$ denotes the norm of $n\in \mathbb{Z}[i]$, and we define $$||z||:=\max\left\{||\Re(z)||,||\Im(z)||\right\},$$ where $\Re(z)$ is the real part and $\Im(z)$ is the imaginary part of $z\in \mathbb{C}$. Hence, $||z||$ measures the distance of $z$ to the nearest Gaussian integer with respect to the supremum norm. We note that by Theorem \[PNT\], $$\label{primenumber} S(x_1,x_2,\omega_1,\omega_2) = \frac{2}{\pi} \cdot \frac{(\omega_2-\omega_1)(x_2-x_1)+o(x_2)}{\log x_2} \quad \mbox{as } x_2\rightarrow \infty.$$ Our goal is to construct an infinite increasing sequence $\left(N_k\right)_{k\in \mathbb{N}}$ of natural numbers such that $S_c\left(x,N_k,\omega_1,\omega_2;\delta_k\right)$ is, in a sense, well approximated by the expected quantity $$4\delta_k^2 S(x,N_k,\omega_1,\omega_2)$$ if $x<N_k$ and $N_k^{-\gamma+\varepsilon}\le \delta_k\le 1/2$, where $\gamma$ is a suitable positive number. We shall see that $\gamma=1/24$ is admissable, which corresponds to the exponent $1/12$ in Theorem \[signi1\]. (Recall that $\mathcal{N}(p)=|p|^2$.) Application of Harman’s sieve for $\mathbb{Z}[i]$ ------------------------------------------------- In the following, let $A$ be a finite set of non-zero Gaussian integers, $G$ be a subset of the set $\mathbb{G}$ of Gaussian primes and $z$ be a positive parameter. By $\mathcal{S}(A,G,z)$ we denote the number of elements of $A$ which are coprime to the product of all Gaussian primes in $G$ with norm $\le z$, i.e. $$\mathcal{S}(A,G,z)=\sharp\{n\in A\ :\ p\not| \; n \mbox{ for all } p\in G \mbox{ such that } \mathcal{N}(p)\le z\}.$$ The following is a version of Harman’s sieve for $\mathbb{Z}[i]$. \[Harsie\] Let $A,B$ be finite sets of non-zero Gaussian integers with norm $\le x$. Suppose for any sequences $(a_n)_{n\in \mathbb{Z}[i]}$ and $(b_n)_{n\in \mathbb{Z}[i]}$ of complex numbers satisfying $|a_n|,|b_n|\le 1$ the following hold: $$\label{t1} \sum\limits_{\substack{\mathcal{N}(m)\le M\\ mn\in A}} a_m = \lambda \sum\limits_{\substack{\mathcal{N}(m)\le M\\ mn\in B}} a_m + O\left(Y\right),$$ $$\label{t2} \sum\limits_{\substack{x^{\alpha}<\mathcal{N}(m)\le x^{\alpha+\beta}\\ mn\in A}} a_mb_n = \lambda \sum\limits_{\substack{x^{\alpha}<\mathcal{N}(m)\le x^{\alpha+\beta}\\ mn\in B}} a_mb_n + O\left(Y\right)$$ for some $\lambda,Y>0$, $\alpha>0$, $0<\beta\le 1/2$ and $M>x^{\alpha}$. Then we have $$\label{super} \mathcal{S}(A,\mathbb{G},x^{\beta})=\lambda \mathcal{S}(B,\mathbb{G},x^{\beta})+O\left(Y\log^3 x\right).$$ The proof is parallel to that of Harman’s sieve for the classical case, [@HPD Theorem 3.1.] (Fundamental Theorem) with $R=1$ and $c_r=1$, making repeated use of the Buchstab identity in the setting of $\mathbb{Z}[i]$ (see [@HPD Chapter 11]). Therefore, we omit the details. In the usual terminology, the sums in are referred to as type I bilinear sums, and the sums in as type II bilinear sums. We assume and apply Theorem \[Harsie\] with $x=x_2$ and $\beta=1/2$ to the situation when $$\begin{aligned} A & := &\left\{n\in \mathbb{Z}[i]\ :\ x_1<\mathcal{N}(n)\le x_2,\ \omega_1<\arg n \le \omega_2,\ ||nc || \le \delta\right\},\\ B & := &\left\{n\in \mathbb{Z}[i]\ :\ x_1<\mathcal{N}(n)\le x_2,\ \omega_1<\arg n \le \omega_2\right\} \quad \mbox{and} \quad \beta=\frac{1}{2}.\end{aligned}$$ The parameters $\alpha$ and $M$ will later be chosen suitably. We note that $$\label{Sxd} S_c(x_1,x_2,\omega_1,\omega_2;\delta)=\sum\limits_{\substack{x_1<\mathcal{N}(p)\le x_2\\ ||nc ||\le \delta\\ \omega_1<\arg p\le \omega_2}} 1 = \mathcal{S}(A,\mathbb{G},x_2^{1/2})+O(x_2^{1/2})$$ and $$\label{Sx} S(x_1,x_2,\omega_1,\omega_2)=\sum\limits_{\substack{x_1<\mathcal{N}(p)\le x_2\\ \omega_1<\arg p\le \omega_2}} 1 = \mathcal{S}(B,\mathbb{G},x_2^{1/2})+O(x_2^{1/2}).$$ Detecting small $||nc||$ ------------------------ We observe that $$||nc||\le \delta \Longleftrightarrow \left([\delta-\Re(nc)]-[-\delta-\Re(nc)]\right)\left([\delta-\Im(nc)]- [-\delta-\Im(nc)]\right)=1.$$ Hence, the type I sum in question can be written in the form $$\begin{split} \sum\limits_{\substack{\mathcal{N}(m)\le M\\ mn\in A}} a_m = & \sum\limits_{\mathcal{N}(m)\le M} a_m \cdot \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} \left([\delta-\Re(mnc)]-[-\delta-\Re(mnc)]\right)\times\\ & \left([\delta-\Im(mnc)]-[-\delta-\Im(mnc)]\right). \end{split}$$ Further, using $[x]=x-\psi(x)-1/2$, the inner sum over $n$ can be expressed in the form $$\begin{split} & \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} \left([\delta-\Re(mnc)]-[-\delta-\Re(mnc)]\right)\times\\ & \left([\delta-\Im(nnc)]-[-\delta-\Im(mnc)]\right)\\ = & 4\delta^2 \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} 1-\\ & 2\delta \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} \left(\psi\left(\delta-\Im(mnc)\right)-\psi\left(-\delta-\Im(mnc)\right)\right) - \\ & 2\delta \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} \left(\psi\left(\delta-\Re(mnc)\right)-\psi\left(-\delta-\Re(mnc)\right)\right)+\\ & \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} \left(\psi\left(\delta-\Re(mnc)\right)-\psi\left(-\delta-\Re(mnc)\right)\right)\times\\ & \left(\psi\left(\delta-\Im(mnc)\right)-\psi\left(-\delta-\Im(mnc)\right)\right)\\ = & 4\delta^2 \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} 1 - 2\delta S_1- 2\delta S_2+ S_3, \end{split}$$ say. Next, we approximate the function $\psi(x)$ by a trigonomtrical polynomial using the following lemma due to Vaaler (see [@GKo], Theorem A6). \[Vaaler\] For $0<|t|<1$ let $$W(t)=\pi t(1-|t|) \cot \pi t +|t|.$$ Fix a natural number $J$. For $x\in \mathbb{R}$ define $$\psi^{\ast}(x):=-\sum\limits_{1\le |j|\le J} (2\pi i j)^{-1}W\left(\frac{j}{J+1}\right)e(jx)$$ and $$\sigma(x):=\frac{1}{2J+2} \sum\limits_{|j|\le J} \left(1-\frac{|j|}{J+1}\right)e(jx).$$ Then $\sigma(x)$ is non-negative, and we have $$|\psi^{\ast}(x)-\psi(x)|\le \sigma(x)$$ for all real numbers $x$. Throughout the sequel, $J$ denotes a natural number such that $J\ge \delta^{-1}$ which will be fixed in subsection \[finest\]. From Lemma \[Vaaler\], we deduce that $$\begin{split} S_1\ll & \frac{x_2/\mathcal{N}(m)}{J}+ \sum\limits_{1\le |j|\le J} \frac{1}{|j|} \cdot \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} \left(e\left(j(\delta-\Im(mnc))\right)-e\left(j(-\delta-\Im(mnc)\right) \right) \Big| \\ \ll & \frac{x_2/\mathcal{N}(m)}{J}+ \sum\limits_{1\le |j|\le J} \frac{1}{|j|}\cdot \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} \left(e\left(j\delta\right)-e\left(-j\delta\right)\right) \cdot e\left(-j\Im(mnc)\right)\Big|\\ \ll & \frac{x_2/\mathcal{N}(m)}{J}+ \sum\limits_{1\le |j|\le J} \min\{\delta, |j|^{-1}\} \cdot \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} e\left(j\Im(mnc)\right)\Big|. \end{split}$$ In a similar way, we obtain $$\begin{split} S_2\ll \frac{x_2/\mathcal{N}(m)}{J}+ \sum\limits_{1\le |j|\le J} \min\{\delta, |j|^{-1}\} \cdot \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} e\left(j\Re(mnc)\right)\Big| \end{split}$$ and $$\begin{split} S_3\ll & \frac{x_2/\mathcal{N}(m)}{J^2} + \frac{1}{J}\cdot \sum\limits_{1\le |j_1|\le J} \min\{\delta,|j_1|^{-1}\} \cdot \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} e\left(j_1\Im(mnc)\right)\Big| + \\ & \frac{1}{J}\cdot \sum\limits_{1\le |j_2|\le J} \min\{\delta, |j_2|^{-1}\} \cdot \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} e\left(j_2\Re(mnc)\right)\Big|+\\ & \sum\limits_{\substack{1\le |j_1|\le J\\ 1\le |j_2|\le J}} \min\{\delta, |j_1|^{-1}\} \cdot \min\{\delta, |j_2|^{-1}\} \cdot \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} e\left(j_1\Im(mnc)+j_2\Re(mnc)\right)\Big|. \end{split}$$ Summing over $m$ and using $|a_m|\le 1$ and $J\ge \delta^{-1}$, we get $$\label{typeI} \begin{split} \sum\limits_{\substack{\mathcal{N}(m)\le M\\ mn\in A}} a_m = & 4\delta^2 \sum\limits_{\mathcal{N}(m)\le M} a_m \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} 1\\ & + O\left(\delta x_2^{1+\varepsilon}J^{-1}+\delta (E_1+E_2)+E_3\right)\\ = & 4\delta^2 \sum\limits_{\substack{\mathcal{N}(m)\le M\\ mn\in B}} a_m+ O\left(\delta x_2^{1+\varepsilon}J^{-1}+\delta (E_1+E_2)+E_3\right), \end{split}$$ where $$E_1=\sum\limits_{1\le |j|\le J} \min\{\delta,|j|^{-1}\} \cdot \sum\limits_{\mathcal{N}(m)\le M} \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} e\left(j\Im(mnc)\right)\Big|,$$ $$E_2=\sum\limits_{1\le |j|\le J} \min\{\delta,|j|^{-1}\} \cdot \sum\limits_{\mathcal{N}(m)\le M} \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} e\left(j\Re(mnc)\right)\Big|$$ and $$\begin{split} E_3= & \sum\limits_{1\le |j_1|\le J} \sum\limits_{1\le |j_2|\le J} \min\{\delta,|j_1|^{-1}\} \cdot \min\{\delta,|j_2|^{-1}\} \times\\ & \sum\limits_{\mathcal{N}(m)\le M} \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}}e\left(j_1\Im(mnc)+j_2\Re(mnc)\right)\Big|. \end{split}$$ In a similar way, using $|a_m|,|b_n|\le 1$ and $J\ge \delta^{-1}$, we derive the asymptotic estimate $$\label{typeII} \sum\limits_{\substack{x_2^{\alpha}<\mathcal{N}(m)\le x_2^{\alpha+\beta}\\ mn\in A}} a_mb_n = 4\delta^2 \sum\limits_{\substack{\mathcal{N}(m)\le M\\ mn\in B}} a_mb_n+ O\left(\delta x_2^{1+\varepsilon}J^{-1}+\delta(F_1+F_2)+F_3\right),$$ where $$F_1= \sum\limits_{1\le |j|\le J} \min\{\delta,|j|^{-1}\} \cdot \Big|\sum\limits_{\substack{x_2^{\alpha}<\mathcal{N}(m)\le x_2^{\alpha+\beta}\\ mn\in A}} a_mb_ne\left(j\Im(mnc)\right)\Big|,$$ $$F_2= \sum\limits_{1\le |j|\le J} \min\{\delta,|j|^{-1}\} \cdot \Big|\sum\limits_{\substack{x_2^{\alpha}<\mathcal{N}(m)\le x_2^{\alpha+\beta}\\ mn\in A}} a_mb_ne\left(j\Re(mnc)\right)\Big|$$ and $$\begin{split} F_3= & \sum\limits_{1\le |j_1|\le J} \sum\limits_{1\le |j_2|\le J} \min\{\delta,|j_1|^{-1}\} \cdot \min\{\delta,|j_2|^{-1}\}\times\\ & \Big|\sum\limits_{\substack{x_2^{\alpha}<\mathcal{N}(m)\le x_2^{\alpha+\beta}\\ mn\in A}} a_mb_ne\left(j_1\Im(mnc)+j_2\Re(mnc)\right)\Big|. \end{split}$$ Transformations of the sums $E_i$ and $F_i$ {#trans} ------------------------------------------- We note that $$\label{E1E2} E_1=E_2.$$ We further have, by breaking the $|j|$-range into $O(\log 2J)$ dyadic intervals, $$\label{E1H} E_1\ll (\log 2J) \cdot \sup\limits_{1\le H\le J} \min\{\delta,H^{-1}\} \cdot E_1(H),$$ where $$E_1(H)=\sum\limits_{1\le |j|\le H} \sum\limits_{\mathcal{N}(m)\le M} \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} e\left(j\Im(mnc)\right)\Big|.$$ Similarly, $$\label{E3rel} E_3\ll (\log 2J)^2\cdot \sup\limits_{\substack{1\le |H_1|\le J\\ 1\le |H_2|\le J}} \min\{\delta,H_1^{-1}\} \cdot \min\{\delta,H_2^{-1}\} \cdot E_3(H_1,H_2),$$ where $$E_3(H_1,H_2)=\mathop{\sum\limits_{|j_1|\le H_1} \sum\limits_{|j_2|\le H_2}}_{(j_1,j_2)\not= (0,0)} \sum\limits_{\mathcal{N}(m)\le M} \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}}e\left(j_1\Im(mnc)+j_2\Re(mnc)\right)\Big|.$$ We note that $$\label{Edef} E_3(H_1,H_2)=\sum\limits_{\substack{j\not=0\\ |\Re(j)|\le H_1\\ |\Im(j)|\le H_2}} \sum\limits_{\mathcal{N}(m)\le M} \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} e\left(\Im(jmnc)\right)\Big|$$ and hence, $$\label{E1E3} E_1(H)= E_3(H,1/2).$$ Thus, it suffices to estimate $E_3(H_1,H_2)$ for $H_1\ge 1$ and $H_2\ge 1/2$ to bound $E_1$, $E_2$ and $E_3$. Similarly, $$\label{F1F2} F_1=F_2$$ and $$\label{F1H} F_1\ll (\log 2J) \cdot \sup\limits_{1\le H\le J} \min\{\delta,H^{-1}\} \cdot F_1(H),$$ where $$F_1(H)=\sum\limits_{1\le |j|\le H} \Big|\sum\limits_{\substack{x_2^{\alpha}<\mathcal{N}(m)\le x_2^{\alpha+\beta}\\ mn\in A}} a_mb_ne\left(j\Im(mnc)\right)\Big|,$$ and $$\label{F3rel} F_3= (\log 2J)^2\cdot \sup\limits_{\substack{1\le |H_1|\le J\\ 1\le |H_2|\le J}} \min\{\delta,H_1^{-1}\} \cdot \min\{\delta,H_2^{-1}\} \cdot F_3(H_1,H_2),$$ where $$\label{F3} \begin{split} F_3(H_1,H_2)= & \mathop{\sum\limits_{|j_1|\le H_1} \sum\limits_{|j_2|\le H_2}}_{(j_1,j_2) \not=(0,0)} \Big|\sum\limits_{\substack{x_2^{\alpha}<\mathcal{N}(m)\le x_2^{\alpha+\beta}\\ mn\in A}} a_mb_ne\left(j_1\Im(mnc)+j_2\Re(mnc)\right)\Big|\\ = & \sum\limits_{\substack{j\not=0\\ |\Re(j)|\le H_1\\ |\Im(j)|\le H_2}} \Big|\sum\limits_{\substack{x_2^{\alpha}<\mathcal{N}(m)\le x_2^{\alpha+\beta}\\ mn\in A}} a_mb_ne\left(\Im(jmnc)\right)\Big| \end{split}$$ and hence, $$\label{F1F3} F_1(H)=F_3(H,1/2).$$ Thus, it suffices to estimate $F_3(H_1,H_2)$ for $H_1\ge 1$ and $H_2\ge 1/2$ to bound $F_1$, $F_2$ and $F_3$. So we have reduced the problem to bounding the type I sums $E_3(H_1,H_2)$ and the type II sums $F_3(H_1,H_2)$. Treatment of type II sums {#treat} ------------------------- To treat the type II sums, we first reduce them to type I sums. We begin by splitting $F_3(H_1,H_2)$ into subsums of the form $$\label{subsums} F_3(H_1,H_2,K,K'):=\sum\limits_{\substack{j\not=0\\ |\Re(j)|\le H_1\\ |\Im(j)|\le H_2}} \Big|\sum\limits_{\substack{K<\mathcal{N}(m)\le K'\\ mn\in A}} a_mb_ne\left(\Im(jmnc)\right)\Big|,$$ where $K<K'\le 2K$. Next, we apply the Cauchy-Schwarz inequality, getting $$F_3(H_1,H_2,K,K')^2\ll H_1H_2K\cdot \sum\limits_{\substack{j\not=0\\ |\Re(j)|\le H_1\\ |\Im(j)|\le H_2}} \sum\limits_{\substack{K<\mathcal{N}(m)\le K'}} \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m) \\ \omega_1<\arg(mn)\le \omega_2}} b_ne\left(\Im(jmnc)\right)\Big|^2,$$ where we use the bound $|a_m|\le 1$. Expanding the square and re-arranging summation, we get $$\label{F} \begin{split} & F_3(H_1,H_2,K,K')^2\ll H_1H_2K\cdot \sum\limits_{\substack{j\not=0\\ |\Re(j)|\le H_1\\ |\Im(j)|\le H_2}}\sum\limits_{x_1/K'<\mathcal{N}(n_1),\mathcal{N}(n_2) \le x_2/K} b_{n_1}\overline{b_{n_2}}\times\\ & \sum\limits_{\substack{\max\{K,x_1/\mathcal{N}(n_1),x_1/\mathcal{N}(n_2)\}<\mathcal{N}(m)\le \min\{K',x_2/\mathcal{N}(n_1),x_2/\mathcal{N}(n_2)\}\\ \omega_1<\arg(mn_{1,2})\le \omega_2}} e\left(\Im(jm(n_1-n_2)c)\right)\\ \ll & H_1^2H_2^2Kx+ H_1H_2K\cdot \sum\limits_{\substack{j\not=0\\ |\Re(j)|\le H_1\\ |\Im(j)|\le H_2}} \sum\limits_{\substack{x_1/K'<\mathcal{N}(n_1),\mathcal{N}(n_2) \le x_2/K\\ n_1\not=n_2}} \\ & \Big| \sum\limits_{\substack{\max\{K,x_1/\mathcal{N}(n_1),x_1/\mathcal{N}(n_2)\}<\mathcal{N}(m)\le \min\{K',x_2/\mathcal{N}(n_1),x_2/\mathcal{N}(n_2)\}\\ \omega_1<\arg(mn_{1,2})\le \omega_2}} e\left(\Im(jm(n_1-n_2)c)\right)\Big|\\ \ll & H_1^2H_2^2Kx+ H_1H_2K\cdot \sum\limits_{0<\mathcal{N}(n)\le 4(H_1^2+H_2^2)x_2/K} \sum\limits_{j|n} \sum\limits_{\substack{x_1/K'<\mathcal{N}(n_1),\mathcal{N}(n_2) \le x_2/K\\ n/j=n_1-n_2}} \\ & \Big| \sum\limits_{\substack{\max\{K,x_1/\mathcal{N}(n_1),x_1/\mathcal{N}(n_2)\}<\mathcal{N}(m)\le \min\{K',x_2/\mathcal{N}(n_1),x_2/\mathcal{N}(n_2)\}\\ \omega_1<\arg(mn_{1,2})\le \omega_2}} e\left(\Im(mnc)\right)\Big|.\\ \end{split}$$ Here the second line arrives by isolating the diagonal contribution of $n_1=n_2$ and using the bound $|b_n|\le 1$, and the third line arrives by writing $n=j(n_1-n_2)$. We note that the summation condition $\omega_1<\arg(mn_{1,2})\le \omega_2$ is equivalent to $f_1<\arg m\le f_2$ for suitable $f_1$ and $f_2$ depending on $n_1,n_2,\omega_1,\omega_2$. Estimating sums of linear exponential sums ------------------------------------------ Our next task is to bound linear exponential sums of the form $$\label{1} \begin{split} & \sum\limits_{\substack{\tilde{y}<\mathcal{N}(m)\le y\\ f_1<\arg m\le f_2}} e\left(\Im(m\kappa)\right) = \sum\limits_{\substack{(m_1,m_2)\in \mathbb{Z}^2\\ \tilde{y}<m_1^2+m_2^2\le y\\ f_1<\arg(m_1+im_2)\le f_2}} e\left(\Re(\kappa)m_2+\Im(\kappa)m_1\right)\\ = & \sum\limits_{\substack{(m_1,m_2)\in \mathbb{Z}^2\\ m_1^2+m_2^2\le y\\ f_1<\arg(m_1+im_2)\le f_2}} e\left(\Re(\kappa)m_2+\Im(\kappa)m_1\right) - \sum\limits_{\substack{(m_1,m_2)\in \mathbb{Z}^2\\ m_1^2+m_2^2\le \tilde{y}\\ f_1<\arg(m_1+im_2)\le f_2}} e\left(\Re(\kappa)m_2+\Im(\kappa)m_1\right), \end{split}$$ where $\kappa$ is a complex number and $0\le \tilde{y}<y$. Here we use the following simple slicing argument. We have $$\label{m_1m_2} \begin{split} & \sum\limits_{\substack{(m_1,m_2)\in \mathbb{Z}^2\\ m_1^2+m_2^2\le y\\ f_1<\arg(m_1+im_2)\le f_2}} e\left(\Re(\kappa)m_2+\Im(\kappa)m_1\right)\\ = & \sum\limits_{-\sqrt{y}\le m_1\le \sqrt{y}} \sum\limits_{\substack{-\sqrt{y-m_1^2}\le m_2 \le \sqrt{y-m_1^2}\\ f_1<\arg(m_1+im_2)\le f_2}} e\left(\Re(\kappa)m_2+\Im(\kappa)m_1\right) \\ \ll & \sum\limits_{-\sqrt{y}\le m_1\le \sqrt{y}} \Big|\sum\limits_{\substack{-\sqrt{y-m_1^2}\le m_2 \le \sqrt{y-m_1^2}\\ f_1<\arg(m_1+im_2)\le f_2}} e\left(\Re(\kappa)m_2\right) \Big|\\ \ll & y^{1/2} \cdot \min\left\{|\Re(\kappa)||^{-1},\sqrt{y}\right\}, \end{split}$$ where we observe that the summation condition $$-\sqrt{y-m_1^2}\le m_2 \le \sqrt{y-m_1^2},\quad f_1<\arg(m_1+im_2)\le f_2$$ on $m_2$ is equivalent to $m_2\in I$ for some interval $I\subseteq [-\sqrt{y},\sqrt{y}]$ depending on $m_1,f_1,f_2$ and use the classical bound $$\sum\limits_{a<m\le b} e(mz) \ll \min\left\{b-a+1,||z||^{-1}\right\}$$ for linear exponential sums. Similarly, by interchanging the rules of $m_1$ and $m_2$, we get $$\begin{split} \sum\limits_{\substack{(m_1,m_2)\in \mathbb{Z}^2\\ m_1^2+m_2^2\le y\\ f_1<\arg(m_1+im_2)\le f_2}} e\left(\Re(\kappa)m_2+\Im(\kappa)m_1\right) \ll y^{1/2} \cdot \min\left\{||\Im(\kappa)||^{-1},\sqrt{y}\right\}. \end{split}$$ Taking the geometric mean of these two estimates gives $$\sum\limits_{\substack{(m_1,m_2)\in \mathbb{Z}^2\\ m_1^2+m_2^2\le y\\ f_1<\arg(m_1+im_2)\le f_2}} e\left(\Re(\kappa)m_2+\Im(\kappa)m_1\right) \ll y^{1/2} \cdot \min\left\{||\Im(\kappa)||^{-1},\sqrt{y}\right\}^{1/2} \cdot \min\left\{||\Re(\kappa)||^{-1},\sqrt{y}\right\}^{1/2}.$$ Using , we deduce that $$\label{lin} \sum\limits_{\substack{\tilde{y}<\mathcal{N}(m)\le y\\ f_1<\arg m\le f_2}} e\left(\Im(m\kappa)\right) \ll y^{1/2} \cdot \min\left\{||\Im(\kappa)||^{-1},\sqrt{y}\right\}^{1/2} \cdot \min\left\{||\Re(\kappa)||^{-1},\sqrt{y}\right\}^{1/2}.$$ To bound the sums appearing in subsections \[trans\] and \[treat\], we need to bound sums of linear sums of roughly the shape $$\sum\limits_{\mathcal{N}(n)\sim Z} \left|\sum\limits_{\substack{\mathcal{N}(m)\sim Y\\ f_1(n)<\arg m\le f_2(n)}} e\left(\Im(mnc)\right)\right|.$$ Considering , we are left with bounding expressions of the form $$\label{Gc} G_{c}(y,z):=\sum\limits_{0<\mathcal{N}(n)\le z} \min\left\{||\Im(n c)||^{-1},\sqrt{y}\right\}^{1/2} \cdot \min\left\{||\Re(n c)||^{-1},\sqrt{y}\right\}^{1/2},$$ where $y,z\ge 1$. To this end, we break the above into partial sums $$\begin{split} & G_{c}(y,z,\Delta_1,\Delta_1',\Delta_2,\Delta_2')\\ := & \sum\limits_{\substack{0<\mathcal{N}(n)\le z\\ \Delta_1< ||\Im(n c)|| \le \Delta_1'\\ \Delta_2< ||\Re(n c)||\le \Delta_2'}} \min\left\{||\Im(n c)||^{-1},\sqrt{y}\right\}^{1/2} \cdot \min\left\{||\Re(n c)||^{-1},\sqrt{y}\right\}^{1/2} \end{split}$$ with $0\le \Delta_1< \Delta_1'\le 1/2$ and $0\le \Delta_2< \Delta_2'\le 1/2$ and bound them by $$\label{G} G_{c}(y,z,\Delta_1,\Delta_1',\Delta_2,\Delta_2')\ll \min\left\{\Delta_1^{-1},\sqrt{y}\right\}^{1/2} \cdot \min\left\{\Delta_2^{-1},\sqrt{y}\right\}^{1/2} \cdot \Sigma_{c}(z,\Delta_1',\Delta_2'),$$ where $$\label{Sigma} \Sigma_{c}(z,\Delta_1',\Delta_2') = \sum\limits_{\substack{0<\mathcal{N}(n)\le z\\ ||\Im(n c)|| \le \Delta_1'\\ ||\Re(n c)|| \le \Delta_2'}} 1.$$ In the next subsection, we shall prove that for infinitely many Gaussian integers $q$, a bound of the form $$\label{plug} \Sigma_{c}(z,\Delta_1',\Delta_2')\ll \left(1+\frac{z}{|q|^2}\right) \cdot \left(1+\Delta_1'|q|\right)\left(1+\Delta_2'|q|\right)$$ holds. We shall also see that for these $q$, we have $$\label{also} \Sigma_{c}(z,\Delta_1',\Delta_2')=0 \quad \mbox{if } \max\{\Delta_1',\Delta_2'\}< 1/(\sqrt{8}|q|) \mbox{ and } z\le |q|^2/8.$$ Plugging into gives $$\label{Gtheta} \begin{split} G_{c}(y,z,\Delta_1,\Delta_1',\Delta_2,\Delta_2')\ll & \left(1+\frac{z}{|q|^2}\right)\cdot \min\left\{\Delta_1^{-1},\sqrt{y}\right\}^{1/2} \cdot \min\left\{\Delta_2^{-1},\sqrt{y}\right\}^{1/2} \times\\ & \left(1+\Delta_1'|q|\right)\cdot \left(1+\Delta_2'|q|\right). \end{split}$$ Next, we write $$\begin{split} G_{c}(y,z)=& G_{c}(y,z,0,2^{-L-1},0,2^{-L-1})+\sum\limits_{i=1}^L \sum\limits_{j=1}^L G_{c}(y,z,2^{-i-1},2^{-i},2^{-j-1},2^{-j})+\\ & + \sum\limits_{j=1}^L G_{c}(y,z,0,2^{-L-1},2^{-j-1},2^{-j})+\sum\limits_{i=1}^L G_{c}(y,z,2^{-i-1},2^{-i},0,2^{-L-1}), \end{split}$$ where $L$ satisfies $1/(2\sqrt{y})\le 2^{-L-1}< 1/\sqrt{y}$. Using , we deduce that $$\label{firstcase} \begin{split} G_{c}(y,z)\ll& \left(1+\frac{z}{|q|^2}\right)\left(y^{1/2}+|q|^2\right) \cdot (\log 2y)^2. \end{split}$$ If $z\le |q|^2/8$, then using , we have $$\begin{split} G_{c}(y,z)=& \sum\limits_{i=1}^L \sum\limits_{j=1}^L G_{c}(y,z,2^{-i-1},2^{-i},2^{-j-1},2^{-j})+\\ & + \sum\limits_{j=1}^L G_{c}(y,z,0,2^{-L-1},2^{-j-1},2^{-j})+\sum\limits_{i=1}^L G_{c}(y,z,2^{-i-1},2^{-i},0,2^{-L-1}), \end{split}$$ where $1/(2\sqrt{8}|q|)\le 2^{-L-1}< 1/(\sqrt{8}|q|)$. In this case, using , we deduce that $$\label{secondcase} \begin{split} G_{c}(y,z)\ll \left(|q|y^{1/4}+|q|^{2}\right)\cdot \log^2(2|q|). \end{split}$$ Counting -------- In this subsection, we prove and . To bound the quantity $G_{c}(y,z)$, we need information about the spacing of the points $nc$ modulo 1, where $n\in \mathbb{Z}[i]$. We begin by using the Hurwitz continued fraction development of $c$ in $\mathbb{Z}[i]$ (see [@Hur]) to approximate $c$ in the form $$c=\frac{a}{q}+\gamma,$$ where $a,q\in \mathbb{Z}[i]$, $(a,q)=1$ and $$|\gamma| \le |q|^{-2}.$$ As in the classical case, this continued fraction development yields a sequence of infinitely many $q\in \mathbb{Z}[i]$ satisfying the above. Now it follows that $$\label{dist} \begin{split} & \left|\left|n_1c-n_2c\right|\right|=\left|\left| \frac{(n_1-n_2)a}{q} + (n_1-n_2)\gamma\right|\right| \ge \left|\left| \frac{(n_1-n_2)a}{q} \right|\right| - |n_1-n_2|\cdot |\gamma|\\ \ge & \frac{1}{\sqrt{2}|q|}- \frac{|n_1-n_2|}{|q|^2} \end{split}$$ if $n_1,n_2\in \mathbb{Z}[i]$ such that $n_1\not\equiv n_2\bmod{q}$. We cover the set $$\mathcal{Z}:=\{n\in \mathbb{Z}[i]\ :\ 0<\mathcal{N}(n)\le z\}$$ by $O\left(1+z/|q|^2\right)$ disjoint rectangles $$\mathcal{R}=\{s\in \mathbb{C}\ :\ a_1< \Re(s)\le b_1,\ a_2< \Im(s)\le b_2\},$$ where $|b_i-a_i|\le |q|/4$, so that $$\mathcal{Z}\subset \bigcup\limits_{\mathcal{R}} \mathcal{R}.$$ Note that if $n_1,n_2\in \mathbb{Z}[i]\cap \mathcal{R}$, then $|n_1-n_2|\le |q|/(2\sqrt{2})$ and hence, by , if $n_1,n_2\in \mathbb{Z}[i]\cap \mathcal{R}$ and $n_1\not=n_2$, then $$\label{dist2} \begin{split} \left|\left|n_1c-n_2c\right|\right|\ge \frac{1}{2\sqrt{2}|q|}. \end{split}$$ Now, $$\Sigma_{c}(z,\Delta_1',\Delta_2') \le \sum\limits_{\mathcal{R}} \Sigma_{c}(\mathcal{R},\Delta_1',\Delta_2'),$$ where $$\label{of} \begin{split} & \Sigma_{c}(\mathcal{R},\Delta_1',\Delta_2') := \sum\limits_{\substack{n\in \mathbb{Z}[i]\cap \mathcal{R}\\ ||\Im(n c)||\le \Delta_1'\\ ||\Re(n c)|| \le \Delta_2'}} 1 \\ = & \sum\limits_{\substack{n\in \mathbb{Z}[i]\cap\mathcal{R}\\ \{\Im(n c)\}\le \Delta_1'\\ \{\Re(n c)\} \le \Delta_2'}} 1 + \sum\limits_{\substack{n\in \mathbb{Z}[i]\cap\mathcal{R}\\ \{\Im(n c)\}\ge 1-\Delta_1'\\ \{\Re(n c)\} \le \Delta_2'}} 1 + \sum\limits_{\substack{n\in \mathbb{Z}[i]\cap\mathcal{R}\\ \{\Im(n c)\}\le \Delta_1'\\ \{\Re(n c)\} \ge 1-\Delta_2'}} 1 + \sum\limits_{\substack{n\in \mathbb{Z}[i]\cap\mathcal{R}\\ \{\Im(n c)\}\ge 1-\Delta_1'\\ \{\Re(n c)\} \ge 1-\Delta_2'}} 1. \end{split}$$ If $\{\Im(n_i c)\}\le \Delta_1'\le 1/2$ and $\{\Re(n_i c)\} \le \Delta_2' \le 1/2$ for $i=1,2$, then $$\left|(\{\Re(n_1c)\}+i\{\Im(n_1c)\})-(\{\Re(n_2c)\}+i\{\Im(n_2c)\})\right|\ge \left|\left|n_1c-n_2c\right|\right|,$$ and hence, by , if $n_1,n_2\in \mathbb{Z}[i]\cap \mathcal{R}$ and $n_1\not=n_2$, then $$\left|(\{\Re(n_1c)\}+i\{\Im(n_1c)\})-(\{\Re(n_2c)\}+i\{\Im(n_2c)\})\right|\ge \frac{1}{2\sqrt{2}|q|}.$$ It follows that $$\sum\limits_{\substack{n\in \mathbb{Z}[i]\cap\mathcal{R}\\ \{\Im(n c)\}\le \Delta_1'\\ \{\Re(n c)\} \le \Delta_2'}} 1\ll V_{1/(2\sqrt{2}|q|)}\left(\Delta_1',\Delta_2'\right),$$ where $V_D\left(\Delta_1'\Delta_2'\right)$ is the maximal number of points of distance $\ge D$ that can be put into a rectangle with dimensions $\Delta_1'$ and $\Delta_2'$. The remaining three sums in the last line of can be estimated similarly. It follows that $$\Sigma_{c}(\mathcal{R},\Delta_1',\Delta_2')\ll V_{1/(2\sqrt{2}|q|)}\left(\Delta_1',\Delta_2'\right).$$ Clearly, $$V_D\left(\Delta_1',\Delta_2'\right)\ll \left(1+\frac{\Delta_1'}{D}\right)\left(1+\frac{\Delta_2'}{D}\right).$$ Putting everything together, we obtain . Further, holds because $0<\mathcal{N}(n)\le |q|^2/8$ implies $$\left|\left|nc\right|\right|=\left|\left| \frac{na}{q}+ n\gamma\right|\right| \ge \left|\left| \frac{na}{q}\right|\right| - |n|\cdot |\gamma|\ge \frac{1}{\sqrt{2}|q|}- \frac{\sqrt{|q|^2/8}}{|q|^2}=\frac{1}{\sqrt{8}|q|}.$$ Final estimations of the sums $E_i$ and $F_i$ {#finest} --------------------------------------------- In this section, we set $x:=x_2$ for simplicity. We recall the conditions $H_1\ge 1$ and $H_2\ge 1/2$. Combining , , and , we get $$\begin{split} & F_3(H_1,H_2,K,K')^2 \\ \ll & (H_1H_2x)^{\varepsilon}\cdot \left(H_1^2H_2^2xK+ H_1H_2xK^{1/2}\cdot \left(1+\frac{\left(H_1^2+H_2^2\right)x/K}{|q|^2}\right)\left(K^{1/2}+|q|^2\right)\right), \end{split}$$ where we use the facts that the number $\tau(n)$ of divisors $j$ of $n$ is $O\left(\mathcal{N}(n)^{\varepsilon}\right)$ and that the number of solutions $(n_1,n_2)$ with $x_1/(2K')<\mathcal{N}(n_1),\mathcal{N}(n_2)\le x_2/K$ of the equation $n/j=n_1-n_2$ is $O(x/K)$. Multiplying out and taking square root yields $$\label{F3part} \begin{split} F_3(H_1,H_2,K,K') \ll & (H_1H_2x)^{\varepsilon}\cdot \Big(H_1H_2(xK)^{1/2}+ (H_1H_2)^{1/2} \times\\ & \left((H_1+H_2)x|q|^{-1}+(H_1+H_2)x K^{-1/4}+|q| x^{1/2}K^{1/4}\right)\Big). \end{split}$$ Recall the definition of $F_3(H_1,H_2)$ in . From , we conclude that $$\label{F3est} \begin{split} F_3(H_1,H_2) \ll & (H_1H_2x)^{\varepsilon}\cdot \Big(H_1H_2x^{(1+\alpha+\beta)/2}+ (H_1H_2)^{1/2} \times\\ & \left((H_1+H_2)x|q|^{-1}+(H_1+H_2)x^{1-\alpha/4}+|q|x^{1/2+(\alpha+\beta)/4}\right)\Big) \end{split}$$ by splitting the summation range of $\mathcal{N}(m)$ into $O(\log 2x)$ dyadic intervals $(K,K']$. We also split $E_3(H_1,H_2)$, defined in , into $O(\log 2M)$ parts $$E_3(H_1,H_2,K,K') := \sum\limits_{\substack{j\not=0\\ |\Re(j)|\le H_1\\ |\Im(j)|\le H_2}} \sum\limits_{K<\mathcal{N}(m)\le K'} \Big| \sum\limits_{\substack{x_1/\mathcal{N}(m)<\mathcal{N}(n)\le x_2/\mathcal{N}(m)\\ \omega_1<\arg(mn)\le \omega_2}} e\left(\Im(jmnc)\right)\Big|$$ with $1/2\le K<K'\le 2K$, which, using , and , we estimate by $$\begin{split} & E_3(H_1,H_2,K,K')\\ \ll & \left(x/K\right)^{1/2}\cdot \sum\limits_{\substack{j\not=0\\ |\Re(j)|\le H_1\\ |\Im(j)|\le H_2}} \sum\limits_{K<\mathcal{N}(m)\le K'} \min\left\{||\Re(jmc)||^{-1},\sqrt{x/K}\right\}^{1/2}\cdot \min\left\{||\Im(jmc)||^{-1},\sqrt{x/K}\right\}^{1/2}\\ \ll & x^{\varepsilon} \cdot \left(x/K\right)^{1/2}\cdot \sum\limits_{0<\mathcal{N}(l)\le (H_1^2+H_2^2)K'} \min\left\{||\Re(lc)||^{-1},\sqrt{x/K}\right\}^{1/2}\cdot \min\left\{||\Im(lc)||^{-1},\sqrt{x/K}\right\}^{1/2}\\ \ll & x^{\varepsilon}\cdot \left(x/K\right)^{1/2} \cdot \left(1+\frac{(H_1^2+H_2^2)K}{|q|^2}\right)\left(\left(x/K\right)^{1/2}+|q|^2\right)\\ \ll & x^{\varepsilon} \cdot \left(xK^{-1}+(H_1^2+H_2^2)x|q|^{-2}+(H_1^2+H_2^2)x^{1/2}K^{1/2}+|q|^2x^{1/2}K^{-1/2}\right). \end{split}$$ If $(H_1^2+H_2^2)K'\le |q|^2/8$, then using instead of , we obtain $$\begin{split} E_3(H_1,H_2,K,K') \ll (x|q|)^{\varepsilon}\left(|q|x^{3/4}K^{-3/4}+|q|^2x^{1/2}K^{-1/2}\right). \end{split}$$ We deduce that for all $K\ge 1/2$, $$\begin{split} & E_3(H_1,H_2,K,K')\ll (x|q|)^{\varepsilon} \times\\ & \left((H_1^2+H_2^2)x|q|^{-2}+(H_1^2+H_2^2)x^{1/2}K^{1/2}+|q|x^{3/4}K^{-3/4}+ |q|^2x^{1/2}K^{-1/2}\right) \end{split}$$ which implies $$\label{E3end} \begin{split} & E_3(H_1,H_2)\ll (x|q|)^{\varepsilon}\times\\ & \left((H_1^2+H_2^2)x|q|^{-2}+(H_1^2+H_2^2)x^{1/2}M^{1/2}+|q|x^{3/4}+|q|^2x^{1/2}\right). \end{split}$$ Now, from and , we obtain $$\label{E3ende} E_3\ll (Jx|q|)^{\varepsilon} \cdot \left(\delta^2 J^2x|q|^{-2}+\delta^2 J^2x^{1/2}M^{1/2}+\delta^2|q|x^{3/4}+\delta^2|q|^2x^{1/2}\right),$$ where we use the inequality $$\min\{\delta,H_1^{-1}\}\cdot \min\{\delta,H_2^{-1}\} \le \delta^2,$$ and from and , we obtain $$\label{F3ende} F_3 \ll (Jx)^{\varepsilon}\left(x^{(1+\alpha+\beta)/2}+ \delta Jx|q|^{-1}+\delta J x^{1-\alpha/4}+\delta |q| x^{1/2+(\alpha+\beta)/4}\right),$$ where we use the inequalities $$\min\{\delta,H_1^{-1}\}\cdot \min\{\delta,H_2^{-1}\} \le (H_1H_2)^{-1}$$ and $$\min\{\delta,H_1^{-1}\}\cdot \min\{\delta,H_2^{-1}\} \le \delta(H_1H_2)^{-1/2}$$ (the first for the diagonal, the second for the non-diagonal contribution). Further, from , , and , we infer $$\label{E1E2ende} E_1, E_2\ll (Jx|q|)^{\varepsilon} \cdot \left(\delta J^2x|q|^{-2}+\delta J^2x^{1/2}M^{1/2}+\delta|q|x^{3/4}+\delta |q|^2x^{1/2}\right),$$ where we use the inequality $$\min\{\delta,H^{-1}\}\le \delta,$$ and from , , and , we infer $$\label{F1F2ende} \begin{split} F_1, F_2\ll (Jx|q|)^{\varepsilon}\cdot \left(x^{(1+\alpha+\beta)/2}+ \delta^{1/2} Jx|q|^{-1}+\delta^{1/2} J x^{1-\alpha/4}+\delta^{1/2} |q| x^{1/2+(\alpha+\beta)/4}\right), \end{split}$$ where we use the inequalities $$\min\{\delta,H^{-1}\}\le H^{-1}\quad \mbox{and} \quad \min\{\delta,H^{-1}\}\le \delta^{1/2}H^{-1/2}.$$ Combing , and , we obtain $$\label{typeIest} \begin{split} & \sum\limits_{\substack{\mathcal{N}(m)\le M\\ mn\in A}} a_m = 4\delta^2 \sum\limits_{\substack{\mathcal{N}(m)\le M\\ mn\in B}} a_m+\\ & O\left((Jx|q|)^{\varepsilon}\cdot\left(\delta x J^{-1}+\delta^2 J^2x|q|^{-2}+\delta^2 J^2x^{1/2}M^{1/2}+ \delta^2|q|x^{3/4}+\delta^2|q|^2x^{1/2}\right)\right), \end{split}$$ and combining , and , we obtain $$\label{typeIIest} \begin{split} & \sum\limits_{\substack{x_2^{\alpha}<\mathcal{N}(m)\le x_2^{\alpha+\beta}\\ mn\in A}} a_mb_n = 4\delta^2 \sum\limits_{\substack{\mathcal{N}(m)\le M\\ mn\in B}} a_mb_n+ \\ & O\left((Jx)^{\varepsilon}\cdot \left(\delta xJ^{-1}+x^{(1+\alpha+\beta)/2}+ \delta Jx|q|^{-1}+\delta J x^{1-\alpha/4}+\delta |q| x^{1/2+(\alpha+\beta)/4}\right)\right). \end{split}$$ Now we choose $J:=[\delta^{-1}x^{3\varepsilon}]$, $x:=|q|^{12}$ (and hence $|q|:=x^{1/12}$), $M= x^{2/3}$, $\alpha:=1/3$ and $\beta:=1/2$ so that $$\label{typeIesti} \sum\limits_{\substack{\mathcal{N}(m)\le M\\ mn\in A}} a_m = 4\delta^2 \sum\limits_{\substack{\mathcal{N}(m)\le M\\ mn\in B}} a_m+ O\left(\delta^2x^{1-\varepsilon}+x^{5/6+8\varepsilon}\right)$$ and $$\label{typeIIesti} \begin{split} \sum\limits_{\substack{x_2^{\alpha}<\mathcal{N}(m)\le x_2^{\alpha+\beta}\\ mn\in A}} a_mb_n = 4\delta^2 \sum\limits_{\substack{\mathcal{N}(m)\le M\\ mn\in B}} a_mb_n+ O\left(\delta^2x^{1-\varepsilon}+x^{11/12+8\varepsilon}\right). \end{split}$$ Conclusion {#con} ---------- Having proved and , we deduce that and hold with $Y=\delta^2x_2^{1-\varepsilon}$ if $\delta\ge x_2^{-1/24+5\varepsilon}$. Now using Theorem \[Harsie\], , and , it follows that $$\sum\limits_{\substack{x_1<\mathcal{N}(p)\le x_2\\ ||pc ||< \delta\\ \omega_1<\arg p \le \omega_2}} 1 = 4\delta^2 \sum\limits_{\substack{x_1<\mathcal{N}(p)\le x_2}} 1 + O\left(x_2^{1/2}\right),$$ provided that $x_2=|q|^{12}$, where $a/q$ is a Hurwitz continued fraction approximant of $c$ and $\delta\ge x_2^{-1/24+\varepsilon}$ for any fixed $\varepsilon>0$. So by taking $N_k=|q_k|^{12}$, where $q_k$ is the $k$-th Hurwitz continued fraction denominator for $c$, we have the following result. \[approx\] Let $c$ be a complex number such that $c\not\in \mathbb{Q}(i)$, $\varepsilon>0$ be an arbitrary constant and $-\pi\le \omega_1< \omega_2\le \pi$. Then there exists an infinite increasing sequence of natural numbers $(N_k)_{k\in \mathbb{N}}$ such that $$\sum\limits_{\substack{x<\mathcal{N}(p)\le N_k\\ ||pc ||\le\delta_k\\ \omega_1<\arg p\le \omega_2}} 1 = 4\delta_k^2 \sum\limits_{\substack{x<\mathcal{N}(p)\le N_k\\ \omega_1<\arg p\le \omega_2}} 1 + O\left(N_k^{1/2}\right) \quad \mbox{as } k\rightarrow \infty$$ if $x<N_k$ and $N_k^{-1/24+\varepsilon}\le \delta_k\le 1/2$. This is implies Theorems \[coro\] and \[signi1\]. Notes ----- \(I) The bound for linear exponential sums over $\mathbb{Z}[i]$ was obtained in a very simple way by reduction to one-dimensional linear exponential sums. Certainly, refinements are possible under certain conditions, and this may be useful for other applications. However, it seems that improvements of and the subsequent bounds for averages of linear exponential sums don’t help in this context because the terms that dominate here cannot be removed, in particular, the term $\frac{z}{|q|^2}\cdot y^{1/2}$ in . So improvements of will most likely not lead to progress with regard to the problem considered here.\ (II) It should be possible to improve the exponent 1/12 in Theorems \[complex\] and \[coro\] using lower bound sieves. To improve this exponent in the asymptotic relation in Theorem \[signi1\] as well, different techniques (like bounds for Koosterman-type sums) will be required. This may be an interesting line of future research.\ (III) Another interesting line could be to investigate Diophantine approximation problems of this type for general number fields. Proof of Theorem \[Theo\](i) ============================ In this section, we prove Theorem \[Theo\](i). Following the treatment in section 3, an admissible choice for the $M_k$’s in Theorem \[signi1\] and Corollary \[signi\] are the sixth powers of absolute values of the Hurwitz continued fraction approximants of $c$. Here we note that $M_k:=N_k^{1/2}$. Throughout the remainder of this paper, we assume that the $M_k$’s are of this form, set $$\mathcal{S}:=\{M_1,M_2,...\}$$ and suppose that $N\in \mathcal{S}$. Further, we write $$B_{\Delta}\left(a\right):= \left\{x\in \mathbb{C}\ :\ |x-a|\le \Delta\right\}$$ and keep the notation $\mathbb{G}$ for the set of Gaussian primes. Let $$\mathcal{A}_p=\bigcup_{\substack{r\in \mathbb{G}\\ q\in \mathbb{Z}[i]}} B_{|\eta/p|}\left( \frac{r}{p}\right) \cap B_{|\eta/(cp)|}\left(\frac{1}{c}\cdot \frac{q}{p}\right) \cap D(a,b,\gamma_1,\gamma_2),$$ where $\eta:=|p|^{\varepsilon-1/12}$. Then $$\label{integral} \int\limits_{\gamma_1}^{\gamma_2}\int\limits_{a}^{b} F_N\left(re^{i\theta}\right)\ dr\ d\theta = \sum\limits_{\substack{p \in \mathbb{G}\\ |p|\le N}} \mu(\mathcal{A}_p).$$ Set $$\label{Mdefi} M:=(a+b)/(2a)$$ and $$L:=4\left[\frac{2\pi}{\gamma_2-\gamma_1}\right].$$ Our strategy is to split the summation over $p$ on the right-hand side of into summations over sets of the form $D(P,MP,\omega,\omega+2\pi/L)$ with $MP\le N$ and derive lower bounds. Clearly, $$\label{AB} \sum\limits_{p \in \mathbb{G} \cap D(P,MP,\omega,\omega+2\pi/L) } \mu(\mathcal{A}_p)\ge \sum\limits_{p \in \mathbb{G} \cap D(P,MP,\omega,\omega+2\pi/L) } \mu(\mathcal{B}_p)$$ with $$\mathcal{B}_p=\bigcup_{\substack{r\in \mathbb{G}\\ q\in \mathbb{Z}[i]}} B_{|\eta'/p|}\left( \frac{r}{p}\right) \cap B_{|\eta'/(cp)|}\left(\frac{1}{c}\cdot \frac{q}{p}\right) \cap D(a,b,\gamma_1,\gamma_2),$$ where $$\label{etadef} \eta':=(MP)^{\varepsilon-1/12}.$$ We note that if $|p|\le MP$ and $$\frac{1}{c}\cdot \frac{q}{p} \in B_{|\eta'/p|}\left(\frac{r}{p}\right),$$ which latter is equivalent to $$q \in B_{|\eta'c|}(rc),$$ then $$\mu\left(B_{|\eta'/p|}\left( \frac{r}{p}\right) \cap B_{|\eta'/(cp)|}\left(\frac{1}{c}\cdot \frac{q}{p}\right)\right)\ge \nu,$$ where $$\label{nudef} \nu:=\left(\frac{\pi}{3}-\frac{\sqrt{3}}{2}\right)\cdot \left|\frac{\eta'}{MP}\right|^2=\left(\frac{\pi}{3}-\frac{\sqrt{3}}{2}\right)\cdot (MP)^{2\varepsilon-13/6}.$$ Here we use our condition that $0<|c|\le 1$. Also, for all $p\in D(P,MP,\omega,\omega+2\pi/L)$, $$r\in D(MPa,Pb,\gamma_1+\omega+2\pi/L,\gamma_2+\omega) \Longrightarrow \frac{r}{p}\in D(a,b,\gamma_1,\gamma_2).$$ We thus have $$\label{key} \sum\limits_{p\in D(P,MP,\omega,\omega+2\pi/L)} \mu(\mathcal{B}_p) \ge \nu N(P,\omega),$$ where $N(P,\omega)$ counts the number of $(p,q,r)\in \mathbb{G} \times \mathbb{Z}[i]\times \mathbb{G}$ satisfying $$\label{qprconds} p\in D(P,MP,\omega,\omega+2\pi/L), \quad q\in B_{\delta}(rc), \quad r\in D(MPa,Pb,\gamma_1+\omega+2\pi/L,\gamma_2+\omega),$$ where $$\label{deltadef} \delta:=|\eta' c|=\frac{|c|}{(MP)^{1/12-\varepsilon}}.$$ We note that $$\label{note1} \frac{1}{4}\cdot \left(b^2-a^2\right)\cdot P^2 \le (Pb)^2-(MPa)^2=\frac{3}{4}\cdot \left(b^2-a^2\right) \cdot P^2$$ and $$\label{note2} \frac{\gamma_2-\gamma_1}{2}\le (\gamma_2+\omega)-(\gamma_1+\omega+2\pi/L)\le \gamma_2-\gamma_1.$$ Using Theorem \[PNT\], the number $\pi(P,MP,\omega,\omega+2\pi/L)$ of Gaussian primes\ $p\in D(P,MP,\omega,\omega+2\pi/L)$ is bounded from below by $$\label{pcount} \pi(P,MP,\omega,\omega+2\pi/L) \ge \frac{4}{L}\cdot \frac{(M^2-1)P^2+o(L(MP)^2)}{2\log N}.$$ The number of $(q,r)\in \mathbb{Z}[i]\times \mathbb{G}$ satisfying $$q\in B_{\delta}(rc), \quad r\in D(MPa,Pb,\gamma_1+\omega+2\pi/L,\gamma_2+\omega)$$ equals $\pi_c(MPa,Pb,\gamma_1+\omega+2\pi/L,\gamma_2+\omega;\delta)$ and is, by Corollary \[signi\], bounded from below by $$\label{qrcount} \begin{split} & \pi_c(MPa,Pb,\gamma_1+\omega+2\pi/L,\gamma_2+\omega;\delta)\\ \ge & \frac{\delta^2\left((\gamma_2+\omega)-(\gamma_1+\omega+2\pi/L)\right)((MPa)^2-(Pb)^2)+o(\delta^2 N^2)}{\log N}. \end{split}$$ Combining , , , , , , and , we obtain $$\label{near} \begin{split} & \sum\limits_{p \in \mathbb{G} \cap D(P,MP,\omega,\omega+2\pi/L) } \mu(\mathcal{A}_p)\\ \ge & C(MP)^{4\varepsilon-7/3}\cdot \frac{1}{L}\cdot \frac{(M^2-1)P^2+o(L(MP)^2)}{\log N}\cdot \frac{(\gamma_2-\gamma_1)(b^2-a^2)P^2+o(N^2)}{\log N} \end{split}$$ for some constant $C=C(c)>0$ if $N\in \mathcal{S}$ and $N\rightarrow \infty$. By splitting the interval $(1,N]$ into intervals of the form $(P,MP]=(N/M^k,N/M^{k-1}]$ and summing up, it follows from that $$\begin{split} & \sum\limits_{p \in \mathbb{G} \cap D(1,N,\omega,\omega+2\pi/L)} \mu(\mathcal{A}_p)\\ \ge & \frac{C(M^2-1)(\gamma_2-\gamma_1)(b^2-a^2)N^{5/3+4\varepsilon}}{M^{7/3-4\varepsilon}L\log^2 N} \cdot \sum\limits_{k=1}^{\infty} M^{-(5/3+4\varepsilon)k} \cdot (1+o(1))\\ \ge & \frac{C(\gamma_2-\gamma_1)(b^2-a^2)N^{5/3+4\varepsilon}}{M^{2/3}L\log^2 N} \cdot (1+o(1))\\ \ge & C\cdot \frac{A}{B} \cdot \frac{(\gamma_2-\gamma_1)(b^2-a^2)N^{5/3+4\varepsilon}}{L\log^2 N} \end{split}$$ if $N\in \mathcal{S}$ is large enough, where for the last line, we have used and $A\le a<b\le B$. Splitting the interval $(0,2\pi]$ into intervals of the form $(\omega,\omega+2\pi/L]=(2\pi (k-1)/L,2\pi k/L]$ with $k=1,...,L$, it further follows that $$\sum\limits_{\substack{p \in \mathbb{G}\\ |p|\le N}} \mu(\mathcal{A}_p) = \sum\limits_{k=1}^L \sum\limits_{p \in \mathbb{G} \cap D(1,N,2\pi (k-1)/L,2\pi k/L)} \mu(\mathcal{A}_p) \ge C\cdot \frac{A}{B} \cdot \frac{(\gamma_2-\gamma_1)(b^2-a^2)N^{5/3+4\varepsilon}}{\log^2 N}$$ if $N$ is large enough. Combining this with completes the proof of Theorem \[Theo\](i). Proof of Theorem \[Theo\](ii) ============================= In this section, we prove Theorem \[Theo\](ii), which is the last task to establish our main result, Theorem \[complex\]. Sieve theoretical approach -------------------------- We extend the treatment in [@BG1 section 5] (see also [@BG2 section 4]), which has its origin in [@HJ], to the situation in $\mathbb{Z}[i]$. We point out that there is a mistake in [@BG1 section 5]: The set $\mathcal{A}$ should consist of products of the form $n[n\alpha]$, not of the form $n[n\alpha][nc\alpha]$, and we bound the number of $n$’s such that $n[n\alpha]$ is the product of two primes, not three primes. This mistake, however, doesn’t affect the method and the final result. As in section 3, we define $$||z||:=\max\left\{||\Re(z)||,||\Im(z)||\right\},$$ where $\Re(z)$ is the real part and $\Im(z)$ is the imaginary part of $z\in \mathbb{C}$, and for $x\in \mathbb{R}$, $||x||$ denotes the distance of $x$ to the nearest integer. We further define $$f(z):=\tilde{f}(\Re(z))+\tilde{f}(\Im(z))i$$ if $z\in \mathbb{C}$ and $||z||<1/2$, where $\tilde{f}(x)$ is the integer nearest to $x\not\in \mathbb{Z}+1/2$. First, we split the interval $(0,N]$ into dyadic intervals $(P/2,P]$ with $P:=N/2^k$, $k=0,1,2,...$. Then we write $$\mathcal{A}_P(\alpha)=\left\{n\cdot f(n\alpha) \ :\ n\in \mathbb{Z}[i], P/2< |n|\le P, \ \max\{||n\alpha||,||nc\alpha||\}\le \mu\right\},$$ where $$\label{muedef} \mu:=\left(\frac{P}{2}\right)^{\varepsilon-1/12}$$ if $P^{\varepsilon-1/12}<1/2$, i.e. $$\label{Pcondit} P>2^{1+1/(1/12-\varepsilon)}.$$ It follows that $$\label{upperFN} F_N(\alpha)\le \sum\limits_{\substack{0\le k\le 1+\log_2\left(N/2^{1/(1/12-\varepsilon)}\right)}} \sharp \left(\mathbb{G}_2\cap \mathcal{A}_{N/2^k}(\alpha)\right)+O(1),$$ where $\mathbb{G}_2$ is the set of products of two Gaussian primes. We bound $\sharp\left(\mathbb{G}_2\cap \mathcal{A}_{N/2^k}(\alpha)\right)$ from above using a simple two-dimensional upper bound sieve in the setting of Gaussian integers, which is obtained by a standard application of the Selberg sieve in the setting of Gaussian integers.\ \[uppersieve\] Let $\mathcal{N}$ be a subset of the Gaussian integers and $f_1,f_2:\mathcal{N} \rightarrow \mathbb{Z}[i]$ two functions. For $P\ge 2$ and $d_1,d_2\in \mathbb{Z}[i]\setminus\{0\}$ let $T_P(d_1,d_2)$ be the number of $n\in \mathcal{N}$ such that $$\label{satisfy} f_1(n)\equiv 0 \bmod{d_1}, \quad f_2(n)\equiv 0 \bmod{d_2}, \quad P/2< |n|\le P$$ and $G_P(d_1,d_2)$ the number of $n\in \mathcal{N}$ satisfying such that $f_1(n)$ and $f_2(n)$ are both Gaussian primes. Then for any $X>0$ and $\varepsilon>0$, $$G_P(d_1,d_2)\le \frac{C(\varepsilon)XP^2}{(\log P)^2}+ O\left(\sum\limits_{\substack{d_1,d_2\in \mathbb{Z}[i]\setminus \{0\}\\ 1\le |d_1|,|d_2|\le P^{\varepsilon}}} |d_1d_2|^{\varepsilon}\left|T_P(d_1,d_2) - \frac{XP^2}{|d_1|^2|d_2|^2}\right|\right)$$ as $P\rightarrow \infty$, where $C(\varepsilon)$ is a constant depending only on $\varepsilon$. Here we consider the case when $$\label{max} \mathcal{N}:=\{n\in \mathbb{Z}[i] : \max(||n\alpha||,||nc\alpha||)\le \mu\},$$ $f_1(n)=n$ and $f_2(n)=f(n\alpha)$. We write $T_P(\alpha;d_1,d_2):=T_P(d_1,d_2)$. Clearly, $T_P(\alpha;d_1,d_2)$ equals the number of $n\in \mathbb{Z}[i]$ with $P/2<|n|\le P$ such that $$\label{threeconds} \frac{P}{2|d_1|}<|n|\le \frac{P}{|d_1|}, \quad \left|\left|\frac{nd_1\alpha}{d_2}\right|\right|\le \frac{\mu}{|d_2|}, \quad ||nd_1c\alpha||\le \mu.$$ Heuristically, $T_P(\alpha;d_1,d_2)$ should behave like $12\pi P^2\mu^4/(|d_1|^2|d_2|^2)$. Therefore, we write $$\label{write} T_P(\alpha;d_1,d_2)=\frac{12\pi P^2\mu^4}{|d_1|^2|d_2|^2}+ E_P(\alpha;d_1,d_2).$$ Then, applying Lemma \[uppersieve\] gives $$\sharp\left(\mathbb{G}_2\cap \mathcal{A}_P\right)\le \frac{C(\varepsilon)P^2\mu^4}{\log^2 P} + O\left(\tilde{J}_P(\alpha)\right),$$ where $$\begin{split} \tilde{J}_P(\alpha):= \sum\limits_{1\le |d_1d_2|\le P^{\varepsilon}} |d_1d_2|^{\varepsilon}|E_P(\alpha;d_1,d_2)|. \end{split}$$ Hence, by , to establish the claim in Theorem \[Theo\](ii), it suffices to show that $$\label{average} \begin{split} & \sum\limits_{\substack{0\le k\le 1+\log_2(N/2^{1/(1/12-\varepsilon)})}}\sum\limits_{1\le |d_1d_2|\le N^{\varepsilon}} |d_1d_2|^{\varepsilon}\int\limits_{-\pi}^{\pi} \int\limits_A^B \left|E_{N/2^k}(Re^{i\theta};d_1,d_2)\right|\ dR \ d\theta\\ = & o\left(\frac{N^2\mu^4}{\log^2 N}\right) \end{split}$$ as $N\rightarrow \infty$ and $N\in \mathcal{S}$. Fourier analysis ---------------- Throughout the sequel, we assume that is satisfied. We use Fourier analysis to express $E_P(\alpha;d_1,d_2)$ in terms of trigonometrical polynomials. We have $$\begin{split} T_P(\alpha;d_1,d_2)= \sum\limits_{\substack{n\in \mathbb{Z}[i]\\ P/|2d_1|<|n|\le P/|d_1|}} & \left(\left[\Re\left(\frac{nd_1\alpha}{d_2}\right)+\frac{\mu}{|d_2|}\right]-\left[\Re\left(\frac{nd_1\alpha}{d_2}\right)-\frac{\mu}{|d_2|} \right] \right)\times\\ & \left(\left[\Im\left(\frac{nd_1\alpha}{d_2}\right)+\frac{\mu}{|d_2|}\right]-\left[\Im\left(\frac{nd_1\alpha}{d_2}\right)-\frac{\mu}{|d_2|}\right] \right)\times\\ & \left(\left[\Re\left(nd_1c\alpha\right)+\mu\right]-\left[\Re\left(nd_1c\alpha\right)-\mu\right]\right)\times\\ & \left(\left[\Im\left(nd_1c\alpha\right)+\mu\right]-\left[\Im\left(nd_1c\alpha\right)-\mu\right]\right), \end{split}$$ where $[x]$ is the integral part of $x\in \mathbb{R}$. Writing $\psi(x)=x-[x]-1/2$, it follows that $$\begin{split} & T_P(\alpha;d_1,d_2)= \\ & \sum\limits_{\substack{n\in \mathbb{Z}[i]\\ P/|2d_1|<|n|\le P/|d_1|}} \left(\psi\left(\Re\left(\frac{nd_1\alpha}{d_2}\right)-\frac{\mu}{|d_2|}\right)- \psi\left(\Re\left(\frac{nd_1\alpha}{d_2}\right)+\frac{\mu}{|d_2|}\right) +\frac{2\mu}{|d_2|}\right)\times\\ & \left(\psi\left(\Im\left(\frac{nd_1\alpha}{d_2}\right)-\frac{\mu}{|d_2|}\right)- \psi\left(\Im\left(\frac{nd_1\alpha}{d_2}\right)+\frac{\mu}{|d_2|}\right)+\frac{2\mu}{|d_2|}\right)\times\\ & \left(\psi\left(\Re\left(nd_1c\alpha\right)-\mu\right)-\psi\left(\Re\left(nd_1c\alpha\right)+\mu\right)+2\mu\right)\times\\ & \left(\psi\left(\Im\left(nd_1c\alpha\right)-\mu\right)-\psi\left(\Im\left(nd_1c\alpha\right)+\mu\right)+2\mu\right). \end{split}$$ Let $$\label{J1J2def} J_1:= \left[\frac{N^{\varepsilon}|d_2|}{\mu}\right] \quad \mbox{and} \quad J_2:=\left[\frac{N^{\varepsilon}}{\mu}\right].$$ Then from Lemma \[Vaaler\], using $$\frac{1}{j}\cdot (e(jx)-e(-jx))\ll x$$ for any $j\in \mathbb{N}$ and $x\in \mathbb{R}$, we deduce that $$\label{deduce} \begin{split} T_P(\alpha;d_1,d_2)= & \frac{16\mu^4}{|d_2|^2}\cdot \sum\limits_{\substack{n\in \mathbb{Z}[i]\\ P/|2d_1|<|n|\le P/|d_1|}} 1 + O\left(1+\frac{P^2}{J_1^2J_2^2|d_1|^2}+F_P(\alpha;d_1,d_2)\right)\\ = & \frac{12\pi P^2\mu^4}{|d_1|^2 |d_2|^2} + O\left(\frac{P\mu^2}{|d_1| |d_2|}+\frac{P^2}{J_1^2J_2^2|d_1|^2}+F_P(\alpha;d_1,d_2)\right), \end{split}$$ where $$\begin{split} F_P(\alpha;d_1,d_2)= & \frac{\mu^4}{|d_2|^2}\cdot \sum\limits_{\substack{(m_1,m_2,m_3,m_4)\in \mathbb{Z}^4\setminus\{(0,0,0,0)\} \\ |m_1|\le J_1,\ |m_2|\le J_1\\ |m_3|\le J_2,\ |m_4|\le J_2}} \\ & \Big| \sum\limits_{\substack{n\in \mathbb{Z}[i]\\ P/|2d_1|<|n|\le P/|d_1|}} e\Big(m_1\cdot \Re\Big(\frac{nd_1\alpha}{d_2}\Big)+ m_2\cdot \Im\Big(\frac{nd_1\alpha}{d_2}\Big)\\ & +m_3\cdot \Re(nd_1c\alpha)+ m_4\cdot \Im(nd_1c\alpha)\Big) \Big|\\ = & \frac{\mu^4}{|d_2|^2}\cdot \sum\limits_{\substack{(m_1,m_2,m_3,m_4)\in \mathbb{Z}^4\setminus\{(0,0,0,0)\}\\ |m_1|\le J_1,\ |m_2|\le J_1\\ |m_3|\le J_2,\ |m_4|\le J_2}} \\ & \Big| \sum\limits_{\substack{n\in \mathbb{Z}[i]\\ P/|2d_1|<|n|\le P/|d_1|}} e\Big(\Im\Big(nd_1\alpha \cdot \Big(\frac{m_1+im_2}{d_2}+(m_3+im_4)c\Big)\Big)\Big) \Big|. \end{split}$$ For the last line of , we have used the elementary bound for the error term in the Gauss circle problem, namely $$\sum\limits_{\substack{n\in \mathbb{Z}[i]\\ |n|\le x}} 1 = \pi x^2+O(x).$$ In the following, we will prove that $$\label{end} \int\limits_{-\pi}^{\pi} \int\limits_A^B \left|F_{P}(Re^{i\theta};d_1,d_2)\right|\ dR \ d\theta \ll \frac{N^{2-4\varepsilon}\mu^4}{|d_1|^2|d_2|^2}.$$ In view of , , and , this suffices to prove and therefore establishes the claim of Theorem \[Theo\](ii). We first bound $F_P(\alpha;d_1,d_2)$ for individual $\alpha$. Using with $f_1=-\pi$ and $f_2=\pi$, we have $$\label{lino} \sum\limits_{\substack{n\in \mathbb{Z}[i]\\ \tilde{x}<|n|\le x}} e\left(\Im(n\kappa)\right) \ll x \cdot \min\left\{||\Im(\kappa)||^{-1},x\right\}^{1/2} \cdot \min\left\{||\Re(\kappa)||^{-1},x\right\}^{1/2}.$$ It follows that $$\label{follows} \begin{split} & F_P(\alpha;d_1,d_2)\\ \ll & \frac{P\mu^4}{|d_1|\cdot |d_2|^2}\cdot \sum\limits_{\substack{(m_1,m_2,m_3,m_4)\in \mathbb{Z}^4\setminus\{(0,0,0,0)\}\\ |m_1|\le J_1,\ |m_2|\le J_1\\ |m_3|\le J_2,\ |m_4|\le J_2}} \\ & \min\left\{\left|\left|\Im\left(d_1\Big(\frac{m_1+im_2}{d_2}+(m_3+im_4)c\Big)\alpha\right)\right|\right|^{-1}, \frac{P}{|d_1|}\right\}^{1/2} \times\\ & \min\left\{\left|\left|\Re\left(d_1\Big(\frac{m_1+im_2}{d_2}+(m_3+im_4)c\Big)\alpha\right)\right|\right|^{-1}, \frac{P}{|d_1|}\right\}^{1/2}\\ \le & \frac{P\mu^4}{|d_1|\cdot |d_2|^2}\cdot \sum\limits_{\substack{(n_1,n_2)\in \mathbb{Z}[i]^2\setminus\{(0,0)\}\\ |n_1|\le 2J_1\\ |n_2|\le 2J_2}} \min\left\{\left|\left|\Im\left(d_1\Big(\frac{n_1}{d_2}+n_2c\Big)\alpha\right)\right|\right|^{-1}, \frac{P}{|d_1|}\right\}^{1/2} \times\\ & \min\left\{\left|\left|\Re\left(d_1\Big(\frac{n_1}{d_2}+n_2c\Big)\alpha\right)\right|\right|^{-1}, \frac{P}{|d_1|}\right\}^{1/2}.\\ \end{split}$$ Average estimation for $F_P(\alpha;d_1,d_2)$ -------------------------------------------- To bound the double integral on the left-hand side of , we now use the following lemma. \[av\] Let $z\in \mathbb{C}$ and $Y>0$. Then $$\begin{split} & \int\limits_{-\pi}^{\pi} \int\limits_{A}^{B} \min\left\{\left|\left|\Im\left(zRe^{\theta i}\right)\right|\right|^{-1}, Y\right\}^{1/2} \cdot \min\left\{\left|\left|\Re\left(zRe^{\theta i}\right)\right|\right|^{-1}, Y\right\}^{1/2} \ dR \ d\theta\\ \ll_{A,B} & \max\left\{1,|z|^{-1}\right\}\log (2+Y). \end{split}$$ By change of variables, we have $$\begin{split} & \int\limits_{-\pi}^{\pi} \int\limits_{A}^{B} \min\left\{\left|\left|\Im\left(zRe^{\theta i}\right)\right|\right|^{-1}, Y\right\}^{1/2} \cdot \min\left\{\left|\left|\Re\left(zRe^{\theta i}\right)\right|\right|^{-1}, Y\right\}^{1/2} \ dR \ d\theta\\ = & \frac{1}{|z|^2}\cdot \int\limits_{-\pi}^{\pi} \int\limits_{|z|A}^{|z|B} \min\left\{\left|\left|\Im\left(r e^{\theta i}\right)\right|\right|^{-1}, Y\right\}^{1/2} \cdot \min\left\{\left|\left|\Re\left(re^{\theta i}\right)\right|\right|^{-1}, Y\right\}^{1/2} \ dr \ d\theta. \end{split}$$ Changing from polar to affine coordinates, and using Cauchy-Schwarz, we get $$\begin{split} & \frac{1}{|z|^2} \cdot \int\limits_{-\pi}^{\pi} \int\limits_{|z|A}^{|z|B} \min\left\{\left|\left|\Im\left(re^{\theta i}\right)\right|\right|^{-1}, Y\right\}^{1/2} \cdot \min\left\{\left|\left|\Re\left(re^{\theta i}\right)\right|\right|^{-1}, Y\right\}^{1/2} \ dr \ d\theta\\ \le & \frac{1}{|z|^2}\cdot \int\limits_{|z|A/2}^{|z|B} \int\limits_{|z|A/2}^{|z|B}\min\left\{\left|\left|\Im\left(x+yi\right)\right|\right|^{-1}, Y\right\}^{1/2} \cdot \min\left\{\left|\left|\Re\left(x+yi\right)\right|\right|^{-1}, Y\right\}^{1/2} \ dy \ dx\\ = & \frac{1}{|z|^2}\cdot \int\limits_{|z|A/2}^{|z|B} \int\limits_{|z|A/2}^{|z|B}\min\left\{\left|\left|\Im\left(x+yi\right)\right|\right|^{-1}, Y\right\}^{1/2} \cdot \min\left\{\left|\left|\Re\left(x+yi\right)\right|\right|^{-1}, Y\right\}^{1/2} \ dy \ dx\\ = & \frac{1}{|z|^2}\cdot \left( \int\limits_{|z|A/2}^{|z|B} \min\left\{\left|\left| x\right|\right|^{-1}, Y\right\}^{1/2} \ dx\right)^2\\ \ll & \frac{1}{|z|}\cdot \left(B-\frac{A}{2}\right)\cdot \int\limits_{|z|A/2}^{|z|B} \min\left\{\left|\left| x\right|\right|^{-1}, Y\right\} \ dx\\ = & \left(B-\frac{A}{2}\right)\cdot \int\limits_{A/2}^{B} \min\left\{\left|\left| |z|x\right|\right|^{-1}, Y\right\} \ dx. \end{split}$$ From Lemma 5.1 in [@BG1], it follows that $$\int\limits_{A/2}^{B} \min\left\{\left|\left|\ |z|x\ \right|\right|^{-1}, Y\right\} \ dx \ll_{A,B} \max\left\{1, \left|z\right|^{-1}\right\}\log (2+Y).$$ Putting everything together proves the claim. From and Lemma \[av\], we deduce that $$\label{EP} \begin{split} & \int\limits_{-\pi}^{\pi} \int\limits_A^B \left|F_{P}(Re^{i\theta};d_1,d_2)\right|\ dR \ d\theta\\ \ll_{A,B} & \frac{P\mu^4\log P}{|d_1|\cdot |d_2|^2}\cdot \sum\limits_{\substack{(n_1,n_2)\in \mathbb{Z}[i]^2\setminus\{(0,0)\}\\ |n_1|\le 2J_1\\ |n_2|\le 2J_2}} \max\left\{1, \left|d_1\left(\frac{n_1}{d_2}+n_2c\right)\right|^{-1}\right\}. \end{split}$$ Final estimation ---------------- Clearly, if $0<|d_1|,|d_2|\le P^{\varepsilon}$ and $J_1\ge |d_2|/\mu$, then $$\label{EP1} \begin{split} & \sum\limits_{\substack{(n_1,n_2)\in \mathbb{Z}[i]^2\setminus\{(0,0)\}\\ |n_1|\le 2J_1\\ |n_2|\le 2J_2}} \max\left\{1, \left|d_1\left(\frac{n_1}{d_2}+n_2c\right)\right|^{-1}\right\}\\ \ll & \frac{1}{|d_1|}\cdot \sum\limits_{\substack{n_2\in \mathbb{Z}[i]\setminus\{0\}\\ |n_2|\le 2J_2}} \left(\min\limits_{n_1\in \mathbb{Z}[i]} \left|\frac{n_1}{d_2}+n_2c\right|\right)^{-1} + J_2^2\cdot \sum\limits_{\substack{n\in \mathbb{Z}[i]\setminus\{0\}\\ |n|\le 3J_1}} \max\left\{1, \left|\frac{d_2}{d_1n}\right|\right\}\\ \ll & \frac{1}{|d_1|}\cdot \sum\limits_{\substack{n_2\in \mathbb{Z}[i]\setminus\{0\}\\ |n_2|\le 2J_2}} \left(\min\limits_{n_1\in \mathbb{Z}[i]} \left|\frac{n_1}{d_2}+n_2c\right|\right)^{-1} + J_1^2J_2^2. \end{split}$$ Since $N$ is the sixth power of absolute value of a denominator of the Hurwitz continued fraction approximation of $c$, we have $$c=\frac{a}{q}+O\left(\frac{1}{|q|^2}\right)$$ for some $a,q\in \mathbb{Z}[i]$ with $|q|^6=N$. Hence, $$\left|\frac{n_1}{d_2}+n_2c\right| \ge \frac{1}{|d_2q|}+O\left(\frac{J_2}{|q|^2}\right).$$ Since $$\frac{1}{|d_2q|}\ge P^{-\varepsilon}N^{-1/6} \ge N^{-1/6-\varepsilon},$$ it follows that $$\left|\frac{n_1}{d_2}+n_2c\right| \gg N^{-1/6-\varepsilon}$$ if $n_1\in \mathbb{Z}[i]$, $n_2\in \mathbb{Z}[i]\setminus\{0\}$, $|n_2|\le 2J_2$ and $N$ is large enough, where we recall that $J_2=[N^{\varepsilon}/\mu]\le N^{1/12+\varepsilon}$. Hence, from , we deduce that $$\label{EP2} \sum\limits_{\substack{(n_1,n_2)\in \mathbb{Z}[i]^2\setminus\{(0,0)\}\\ |n_1|\le 2J_1\\ |n_2|\le 2J_2}} \max\left\{1, \left|d_1\left(\frac{n_1}{d_2}+n_2c\right)\right|^{-1}\right\} \ll N^{1/6+\varepsilon}J_2^2 + J_1^2J_2^2.$$ Combining and , we obtain $$\int\limits_{-\pi}^{\pi} \int\limits_A^B \left|F_{P}(Re^{i\theta};d_1,d_2)\right|\ dR \ d\theta\\ \ll_{A,B} \frac{P\mu^4\log P}{|d_1|\cdot |d_2|^2}\cdot \left(N^{1/6+\varepsilon}J_2^2 + J_1^2J_2^2\right),$$ from which follows using and . This completes the proof of Theorem \[Theo\](ii). [3]{} S. Baier, A. Ghosh, [*Diophantine approximation on lines with prime constraints*]{}, Q. J. Math. 66 (2015) 1-12. S. Baier, A. Ghosh, [*Restricted simultaneous Diophantine approximation*]{}, Mathematika 63 (2017) 34–52. D.G. Cantor, [*On the elementary theory of diophantine approximation over the ring of adeles. I.*]{}, Illinois J. Math. 9 (1965) 677–700. S. Graham, G. Kolesnik, [*Van der Corput’s method of exponential sums*]{}, London Mathematical Society Lecture Note Series, 126. Cambridge University Press, Cambridge, 1991. A. Ghosh, [*Diophantine approximation on subspaces of $\mathbb{R}^n$ and dynamics on homogeneous spaces*]{}, arXiv:1606.02399. G. Harman, [*On the distribution of $\alpha p$ modulo one*]{}, J. London Math. Soc. 27 (1983) 9-18. G. Harman, [*Metric diophantine approximation with two restricted variables III. Two prime numbers*]{}, J. Number Theory 29 (1988) 364–375. G. Harman, [*Prime-detecting sieves*]{}, London Mathematical Society Monographs Series, 33, Princeton University Press, Princeton, NJ, 2007. G. Harman, H. Jones, [*Metrical theorems on restricted Diophantine approximations to points on a curve*]{}, J. Number Theory 97 (2002) 45-57. D.R. Heath-Brown, C. Jia, [*The distribution of $\alpha p$ modulo one*]{}, Proc. London Math. Soc. 84 (2002) 79-104. A. Hurwitz, [*Über die Entwicklung complexer Grössen in Kettenbrüche*]{}, Acta Math. 11 (1887) 187-200. H. Jones, [*Khintchins theorem in k dimensions with prime numerator and denominator*]{}, Acta Arith. 99 (2001) 205–225. I. Kubilyus, [*The distribution of Gaussian primes in sectors and contours*]{}, Leningrad. Gos. Univ. U. Zap. Ser. Mat. Nauk 137 (1950) 40-52. K. Matomäki, [*The distribution of $\alpha p$ modulo one*]{}, Math. Proc. Cambridge Philos. Soc. 147 (2009) 267-283. D. Sullivan, [*Disjoint spheres, approximation by imaginary quadratic numbers, and the logarithm law for geodesics*]{}, Acta Math. 149 (1982) 215-237. R.C. Vaughan, [*On the distribution of $\alpha p$ modulo 1*]{}, Mathematika 24 (1977) 135-141. I.M. Vinogradov, [*An elementary proof of a theorem from the theory of prime numbers*]{}, Izvestiya Akad. Nauk SSSR. Ser. Mat. 17 (1953) 3-12.

ArXiv

--- abstract: 'This work is concerned with applying iterative image reconstruction, based on constrained total-variation minimization, to low-intensity X-ray CT systems that have a high sampling rate. Such systems pose a challenge for iterative image reconstruction, because a very fine image grid is needed to realize the resolution inherent in such scanners. These image arrays lead to under-determined imaging models whose inversion is unstable and can result in undesirable artifacts and noise patterns. There are many possibilities to stabilize the imaging model, and this work proposes a method which may have an advantage in terms of algorithm efficiency. The proposed method introduces additional constraints in the optimization problem; these constraints set to zero high spatial frequency components which are beyond the sensing capability of the detector. The method is demonstrated with an actual CT data set and compared with another method based on projection up-sampling.' author: - 'Emil Y. Sidky, Rick Chartrand,  Yuval Duchin, Christer Ullberg,  and Xiaochuan Pan, [^1][^2][^3][^4]' bibliography: - 'lowdose.bib' title: 'High resolution image reconstruction with constrained, total-variation minimization' --- Introduction ============ proceedings focuses on a fundamental issue of resolution in iterative image reconstruction (IIR). IIR is being considered for application in computed tomography (CT), see for example [@ASIR:10], because it is possible to account for noise in the data model and accordingly allow for high quality images with a reduced exposure [@McCollough:09]. IIR methods being considered for CT, all involve implicit solution, see section 15.3 of [@Barrett:FIS], for the image as opposed to explicit solution, derived from exact or approximate inverses to the continuous cone-beam transform. For explicit reconstruction algorithms, which are generally some variant of filtered back-projection (FBP), the reconstructed volume can be obtained one point at a time. Images can be shown on grids of any size and with arbitrarily small grid spacing, for example blowing up a region-of-interest (ROI). Note, that does not mean the resolution is arbitrarily high, because the system resolution is still limited by the discrete data sampling. Reconstruction by implicit solution allows more flexible and realistic data models for the tomographic system, but at a price. As pointed out often, IIR algorithms are generally more computationally intensive. Another important issue is that a complete expansion set for the imaged volume is necessary in order to obtain the reconstruction, and the complete image must be solved for all at once; the image cannot be gotten voxel-by-voxel. More concretely, let us consider a linear imaging model using a voxel expansion of the volume and ideal conditions of perfect data consistency: $$\label{discreteModel} \mathbf{g} = \mathcal{X} \mathbf{f},$$ where $\mathbf{g}$ represents the projection data as a 1D vector; $\mathbf{f}$ is a vector of voxel values; $\mathcal{X}$ is the system matrix, which yields line integrals through the volume and for the present work is based on the ray intersection length as specified in Siddon’s method. This model, which forms the basis of many IIR algorithms, is difficult to solve explicitly for realistic size models of a CT system. The matrix $\mathcal{X}$ is ill-conditioned and can be extremely large, up to a size of $10^9 \times 10^9$. Equation (\[discreteModel\]) is usually solved implicitly by one of many algorithms, conjugate gradients, algebraic reconstruction techniques (ART), etc. Two practical issues arise in the implicit solution of Eq. (\[discreteModel\]): (I) the voxel representation of the image must include the whole region where the measured transmission rays intersect with the support of the subject [@ziegler2008iterative], and (II) to attain the intrinsic resolution of the data the voxels must be small compared with the detector bin width. The combination of these two factors causes a computational burden, because straight-forward voxelization of the reconstruction volume will lead to a very large array. But also, mathematically, the second point that voxels should be small compared to the detector bins leads to an under-determined linear system for Eq. (\[discreteModel\]), because there will be more unknown voxel values than known ray-transmission measurements. In this work, we examine a couple of possible solutions for addressing the second issue – dealing with the large null-space of Eq. (\[discreteModel\]) in a controlled way. In Sec. \[sec:theory\] we present the image reconstruction algorithms and in Sec. \[sec:results\] we demonstrate them on an actual low-intensity CT scan. Constrained, total-variation minimization theory and algorithms {#sec:theory} =============================================================== In recent years, we have been investigating the solution of Eq. (\[discreteModel\]), by solving the constrained TV-minimization problem: $$\label{TVmin} \mathbf{f}^* = \text{argmin} \| \mathbf{f} \|_\text{TV} \text{ such that } |\mathcal{X} \mathbf{f} - \mathbf{g}|^2 \le \epsilon^2 \; \; \; \mathbf{f} \ge 0,$$ where $\| \mathbf{f} \|_\text{TV}$ is the sum over the gradient magnitude image; and $\epsilon$ is a data-error tolerance parameter necessary because the projection data is likely not consistent with the image model. To solve this optimization problem, we have been developing a heuristic algorithm [@SidkyTV:06; @sidky2008image; @sidky2009enhanced; @PanIP:09; @SidkyPC:10; @sidky2010lowintensity] that alternates between projection-onto-convex-sets (POCS), to enforce the constraints, and steepest descent (SD) to reduce image TV. The key point of this algorithm is that the SD step-size is controlled adaptively, so as to not undo progress toward the feasible image set with POCS. This alternating algorithm is called adaptive SD-POCS (ASD-POCS). For investigations where solving Eq. (\[TVmin\]) accurately is important, optimality conditions should be checked [@sidky2008image; @sidky2010lowintensity]. On the other hand, for practical applications it may not be necessary to have an accurate solver [@sidky2009enhanced], and in such cases ASD-POCS yields a clinically useful image within 10-20 iterations. Image reconstruction algorithms for CT based on Eq. (\[TVmin\]) have been shown to be effective for sparse-view projection data [@SidkyTV:06; @song2007sparseness; @chen2008prior; @sidky2008image; @sidky2009enhanced; @jia2010gpu; @SidkyPC:10; @Bergner:10; @Choi:10; @Bian:10], which has obvious implications for patient dose. More recently, we have been interested in how to apply ASD-POCS to what has traditionally accepted as fully angularly-sampled data. For such data, structures on the scale of a detector bin-width are expected to be resolved. For IIR, this requirement poses the above-mentioned problem that the number of voxels may be much larger than the number of data, and direct application of the IIR algorithm may lead to strange noise patterns which can interfere with imaging tasks [@sidky2010lowintensity]. There are likely many ways to resolve this problem; for example, introducing a non-zero cross-section to the ray model in $\mathcal{X}$ may yield nicer noise patterns while maintaining image resolution. At the same time, more realistic system model ling usually comes at the expense of computational speed. As a result, we have been seeking other alternatives. In Ref. [@sidky2010lowintensity], we propose to stabilize the problem in Eq. (\[TVmin\]) by making the following observation: CT resolution is non-uniform, and generally, the angular-sampling is worse than detector resolution. Assuming that the individual projection sampling satisfies Nyquist frequency, each projection can be up-sampled to increase the number of measurement rays at each projection. We will still have a data set which is under-sampled in the angular direction with respect to the very high resolution imaging grid. But we already know that TV-minimization approaches appear to be effective with this type of under-sampling. We refer to this approach as up-sampling ASD-POCS or, in this text, as simply up-sampling. In this proceedings, we propose another approach which may have even a greater advantage in terms of algorithm efficiency. The method comes from realizing that while we need many voxels per bin-width for flexibility of the image representation, we cannot hope to actually attain true super-resolution where structures smaller than the detector bin are visible. To capture this idea mathematically, we impose constraints on the Fourier transform of the image. Specifically, high frequency components of the image are set to zero: $$\label{freqConstraint} \mathcal{F} \mathbf{f} (| \mathbf{\nu} |>\nu_\text{max})= 0,$$ where $\mathcal{F}$ is the discrete Fourier transform (DFT); $\mathbf{\nu}$ represents spatial frequency; and the frequency magnitude $\nu_\text{max}$ is determined by the bin spacing. If the detector bins have width $w$, we set $\nu_\text{max}=1/(2 \pi w)$. Although this choice makes sense physically, $\nu_\text{max}$ can be taken as a algorithm parameter, but varying $\nu_\text{max}$ is beyond the scope of this study. Adding this constraint to the optimization problem in Eq. (\[TVmin\]) leads to frequency-constrained ASD-POCS. This approach involves processing fewer data in exchange for computing DFTs, which can be done efficiently with the fast Fourier transform. The pseudo-code for frequency-constrained ASD-POCS is the same as that reported in Ref. ([@sidky2010lowintensity]) with the additional steps of enforcing Eq. (\[freqConstraint\]) at each line where the image positivity is enforced. For the present results, the frequency constraint was imposed before positivity. Results {#sec:results} ======= ![FBP slice image of a rabbit scanned by a low-dose XCounter CT system. The raw FBP image was smoothed by a Gaussian filter, reducing the image TV by a factor of 8. The indicated rectangles show the ROIs which are used for the algorithm comparisons. In particular, the small one surrounds the wire object used to obtain a sense of resolution. The display window is \[0,0.04\] mm$^{-1}$. []{data-label="fig:FBP"}](fbpoverview.pdf){width="3.45in"} The up-sampling and frequency constrained ASD-POCS algorithms are demonstrated with an XCounter CT scan of a rabbit with a thin wire taped to the outside of the sample holder. The data are low-intensity and contain 1878 projections with a 2266x64 bin detector at a resolution of 0.1 mm. The thin wire provides a good resolution test for the image reconstruction algorithm. For the present purpose, we take the middle row on the detector from this data set and focus on 2D fan-beam CT reconstruction with 1878 projections on a 2266-bin linear detector array. Gaussian-filtered FBP 0.1cm ![(Top) FBP ROI’s for the image presented in Fig. \[fig:FBP\]. (Middle) ASD-POCS image reconstructed from original data for a grid with 0.1 mm pixels. (Bottom) ASD-POCS image reconstructed from original data for a grid with 0.025 mm pixels. The image TV for each case is set to an eighth of that of the raw FBP image. The display window is \[0,0.04\] mm$^{-1}$ on the left ROI and \[0,0.09\] mm$^{-1}$ on the right ROI. []{data-label="fig:problems"}](fbprois.pdf "fig:"){width="3.45in"} 0.5cm ASD-POCS (0.1 mm pixels) 0.1cm ![(Top) FBP ROI’s for the image presented in Fig. \[fig:FBP\]. (Middle) ASD-POCS image reconstructed from original data for a grid with 0.1 mm pixels. (Bottom) ASD-POCS image reconstructed from original data for a grid with 0.025 mm pixels. The image TV for each case is set to an eighth of that of the raw FBP image. The display window is \[0,0.04\] mm$^{-1}$ on the left ROI and \[0,0.09\] mm$^{-1}$ on the right ROI. []{data-label="fig:problems"}](tvmatched.pdf "fig:"){width="3.45in"} 0.5cm ASD-POCS (0.025 mm pixels) 0.1cm ![(Top) FBP ROI’s for the image presented in Fig. \[fig:FBP\]. (Middle) ASD-POCS image reconstructed from original data for a grid with 0.1 mm pixels. (Bottom) ASD-POCS image reconstructed from original data for a grid with 0.025 mm pixels. The image TV for each case is set to an eighth of that of the raw FBP image. The display window is \[0,0.04\] mm$^{-1}$ on the left ROI and \[0,0.09\] mm$^{-1}$ on the right ROI. []{data-label="fig:problems"}](tvnoup.pdf "fig:"){width="3.45in"} An FBP image of this data set is shown in Fig. \[fig:FBP\], where some regularization is performed by Gaussian smoothing with a window 2 pixels wide. The rectangles indicate the regions where comparisons of the different algorithms are shown. Comparisons for each algorithm will be made at a level of image regularization where each image’s TV-norm is an eighth of that of the unregularized FBP image. The images are shown just to give a sense about the behavior of the algorithms; more rigorous evaluation with different levels of regularization will be performed in future work. We illustrate the problems with employing a matched-resolution grid, 1024$\times$1024, and a very high resolution grid, 4096$\times$4096, with the basic ASD-POCS algorithm. The pixel width is 0.1 mm for the former grid and 0.025 mm for the latter. Note that the rabbit support projects onto the middle 1000 bins of the full projection. The FBP image is shown on a 4096$\times$4096 grid, but because FBP is based on an explicit inverse its pixel values are not affected by the grid size. In Fig. \[fig:problems\] the matched grid clearly leads to a loss of resolution relative to FBP, because the image is forced to be uniform over the 0.1mm $\times$ 0.1mm squares. Simply going to a larger array does improve the ASD-POCS image, but the noise pattern shows additional false structure when compared with the FBP image. These structures arise from the fact that the imaging problem is under-sampled by roughly a factor of ten and this under-sampling occurs in both view-angle and detector-bin directions. The random, sparse specks that appear could be confused with micro-calcifications in the context of breast imaging. ASD-POCS (projection up-sampling) 0.1cm ![(Top) ASD-POCS image obtained from the up-sampled data set. (Bottom) ASD-POCS image reconstructed from original data set with additional constraints on the spatial frequency. For both images the pixel width is 0.025 mm. The image TV for each case is set to an eighth of that of the raw FBP image. The display window is the same as that of Fig. \[fig:problems\] []{data-label="fig:newASDPOCS"}](tvup.pdf "fig:"){width="3.45in"} 0.5cm ASD-POCS (spatial frequency constraints) 0.1cm ![(Top) ASD-POCS image obtained from the up-sampled data set. (Bottom) ASD-POCS image reconstructed from original data set with additional constraints on the spatial frequency. For both images the pixel width is 0.025 mm. The image TV for each case is set to an eighth of that of the raw FBP image. The display window is the same as that of Fig. \[fig:problems\] []{data-label="fig:newASDPOCS"}](tvfreq.pdf "fig:"){width="3.45in"} Finally, we show images for ASD-POCS using up-sampling and frequency constraints, and the resulting region-of-interest images are shown in Fig. \[fig:newASDPOCS\]. For the up-sampling method the data are up-sampled at each projection so that the data set size nominally becomes 1878 views by 9064 virtual bins prior to reconstruction by ASD-POCS. The frequency constraint ASD-POCS takes the original 1878$\times$2266 data set as input. As can be seen in the figure, both approaches remove the disturbing noise pattern seen at the bottom of Fig. \[fig:problems\]. Each of these images show some potential advantage over FBP in that the wire appears to be better focused in the up-sampling approach, and the image noise level is reduced approximately 20% for both up-sampling and frequency constrained ASD-POCS. It is possible that the frequency constraint method could lead to better resolution by allowing $\nu_\text{max}$ to be increased. Such a study will be investigated in future work. Conclusion ========== We have developed modifications to ASD-POCS that allow for high-resolution recovery of structures occurring on the scale of the detector bin width. In particular, in this work we show that including constraints on the spatial frequencies of the reconstructed image can improve ASD-POCS images by eliminating false structures, which arise from the large null-space inherent in the imaging model when very high-resolution grids are used to represent the image. Further research, will compare the up-sampling and frequency constraint approaches with many data sets and different levels of image regularity, and we will also look into implementing more realistic ray-models that account for source spot-size and the extent of the detector bins. This realistic modeling in conjunction with frequency extrapolation [@chartrand2010] may allow even further improvement of CT image quality. [^1]: Manuscript received November 21, 2010. This work was supported in part by NIH R01 grants CA120540 and EB000225. The contents of this article are solely the responsibility of the authors and do not necessarily represent the official views of the National Institutes of Health. This work was also supported in part by the U.S. Department of Energy through the LANL/LDRD Program. [^2]: E. Y. Sidky, Y. Duchin, and X. Pan are with the University of Chicago Dept. of Radiology, Chicago, IL 60637 USA. [^3]: R. Chartrand is with Los Alamos National Laboratory, Los Alamos, NM 87545 USA. [^4]: C. Ullberg is with XCounter AB, Danderyd, SE-182 33 Sweden.

ArXiv

--- author: - | Adi Shraibman\ The School of Computer Science\ The Academic College of Tel Aviv-Yaffo\ Email: adish@mta.ac.il title: 'The Corruption Bound, Log Rank, and Communication Complexity' --- \[theorem\][Lemma]{} \[theorem\][Claim]{} \[theorem\][Corollary]{} \[theorem\][Conjecture]{} \[theorem\][Question]{} \[theorem\][Proposition]{} \[theorem\][Axiom]{} \[theorem\][Remark]{} \[theorem\][Example]{} \[theorem\][Exercise]{} \[theorem\][Definition]{} \[theorem\][Observation]{} Introduction ============ The communication complexity literature is mainly concerned with proving lower bounds, and indeed communication complexity lower bounds are used in various areas of theoretical computer science. Nonetheless, upper bounds are also very interesting. Among the most intriguing upper bounds are those expressed in terms of known lower bounds. Let $A$ be a sign matrix, and denote by $D(A)$ the deterministic communication complexity of $A$. Previous upper bounds of this nature on $D(A)$ include: - $D(A) \le (N^0(A)+1)(N^1(A)+1)$. - $D(A) \le \log {\mathrm{rank}}(A) (N^0(A)+1)$. - $D(A) \le \log {\mathrm{rank}}(-A) (N^1(A)+1)$. - $D(A) \le \log {\mathrm{rank}}(A) \log {\mathrm{rank}}_+(A)$. Here, $N^1(A)$ and $N^0(A)$ are the nondeterministic and co-nondeterministic communication complexity of $A$, respectively. Also, ${\mathrm{rank}}(A)$ is the rank of $A$, and ${\mathrm{rank}}_+(A)$ is the positive rank of $A$. All these complexity measures are known lower bounds on the deterministic communication complexity. See [@KN97] or [@Lov90] for a comprehensive survey of these complexity measures and bounds. Recently, Gavinsky and Lovett [@LG13] augmented this repertoire with additional bounds. They proved that the deterministic communication complexity of a sign matrix $A$ is at most $O(CC(A) \log^2 {\mathrm{rank}}(A))$, where $CC(A)$ is either randomized communication complexity, information complexity, or zero-communication complexity. Thus when the rank of the matrix is low, an efficient nondeterministic protocol or a randomized protocol, implies an efficient deterministic protocol. The heart of the proofs in [@LG13] is a clever and simple lemma stating that when the fraction of either $1$’s or $-1$’s in a matrix is small compared to the rank of the matrix then it contains a large monochromatic rectangle. We use this lemma to prove the following bounds: $$D(A) \le O({\mathrm{corr}^{(1)}_{1/4}}(A)\log^2 {\mathrm{rank}}(A)),$$ and $$D(A) \le O({\mathrm{corr}^{(1)}_{1/4}}(-A)\log^2 {\mathrm{rank}}(A)),$$ where ${\mathrm{corr}^{(1)}_{1/4}}(A)$ is a one-sided version of the corruption/rectangle bound with error $1/4$. That is, the corruption bound is the maximum of ${\mathrm{corr}^{(1)}_{1/4}}(A)$ and ${\mathrm{corr}^{(1)}_{1/4}}(-A)$. The corruption bound is smaller than the randomized complexity, information complexity, and zero-communication complexity (see e.g. [@KLLRX12]). Therefore our bounds are smaller than the bounds proved in [@LG13]. The one-sided version of the corruption bound, ${\mathrm{corr}^{(1)}_{1/4}}(A)$, is smaller than the nondeterministic communication complexity. Similarly, ${\mathrm{corr}^{(1)}_{1/4}}(-A)$ is smaller than the co-nondeterministic communication complexity. Thus, our bounds unify the two groups of previous bounds mentioned above, and fall very naturally in this framework. In fact, since $MA$-complexity is bounded by root of the corruption bound [@Kla03], we get as a consequence that $$D(A) \le O(MA(A)^2\log^2 {\mathrm{rank}}(A)),$$ and $$D(A) \le O(MA(-A)^2\log^2 {\mathrm{rank}}(A)).$$ Here $MA(A)$ is the Merlin Arthur (MA) complexity of $A$, with error $1/4$. In this model the players first make a nondeterministic guess and then perform a randomized protocol. This has the nice interpretation that when the rank is low, there is an efficient deterministic protocol, even compared with protocols combining the power of nondeterminism and randomization. The corruption bound (and variants of it) is a central lower bound technique for the randomized communication complexity. It was used already by Yao in [@Yao83], and later e.g. in [@BFS86; @Raz92d]. It is the only lower bound technique (excluding the relatively new partition bound [@JK10]), that has not been transfered also to quantum communication. Log of the rank, on the other hand, is proved to also bound exact quantum communication complexity [@BCW98], even with entanglement [@BW01]. Therefore, the upper bounds in terms of corruption and log-rank imply that the corruption bound is a lower bound on exact quantum communication complexity, if and only if quantum communication is equivalent to deterministic communication complexity. We give the relevant definitions and background in Section \[definitions\], and prove the upper bounds in Section \[upper-bound\]. In Section \[log-rank\] we give a simple proof to the novel bound on the deterministic communication complexity in terms of the rank, proved by Lovett [@Lov13]. Definitions =========== Let $A$ be an $m \times n$ sign matrix, and let $\mu$ be a probability distribution on $[m]\times[n]$. For a set of entries $I \subseteq [m]\times[n]$, let $\mu(I)$ be the sum $\sum_{(i,j) \in I} \mu(i,j)$. A combinatorial rectangle is a subset $S\times T$ of entries, such that $S \subseteq [m]$ and $T \subseteq [n]$. With a slight abuse of notation, for $v \in \{\pm 1\}$ and a combinatorial rectangle $R$, we denote by $\mu(v)=\mu(\{(i,j)|A_{i,j}=v\})$, and $\mu(v, R)=\mu(\{(i,j)\in R|A_{i,j}=v\})$. We also write $\mu(v | R)$ for the conditional probability equal to $\mu(v,R)/\mu(R)$. Finally, let $$err(R,\mu,v) = \mu(-v | R).$$ Intuitively, $err(R,\mu,v)$ represents the probability of error in the rectangle $R$, assuming a randomized protocol answers $v$ on $R$. Let $A$ be a sign matrix, $0 \le {\epsilon}\le 1$, and $v \in \{-1,1\}$. Define $${\mathrm{size}^{(v)}_{{\epsilon}}}(A, \mu) = \max_R \{\mu(R) : err(R,\mu,v) \le \epsilon \},$$ where the maximum is over all combinatorial rectangles $R$. Usually, when defining the corruption bound, balanced distributions are considered. That is, distributions for which the probability of $-1$ and the probability of $1$ are not too small. Here, we consider strictly balanced distributions that have their support on a combinatorial rectangle. We also add the restriction that the distributions are uniform on the set of entries equal to $-1$, and also on the set of entries equal to $1$. Formally, we say that a distribution $\mu$ on $[m]\times[n]$ is uniformly-balanced, if it satisfies: - The set of entries $(i,j)$ for which $\mu(i,j) > 0$ is a combinatorial rectangle. - $A_{i,j} = A_{x,y}$ implies that $\mu(i,j) = \mu(x,y)$, if both are nonzero. - $\mu(1) = \mu(-1) = \frac{1}{2}$. The corruption bound is: Let $A$ be a sign matrix, $0 \le {\epsilon}\le 1$, and $v \in \{-1,1\}$. Define $${\mathrm{corr}^{(v)}_{{\epsilon}}}(A) = \max_{\mu} \log 1/{\mathrm{size}^{(v)}_{{\epsilon}}}(A, \mu),$$ where $\mu$ runs over all uniformly-balanced distributions. Also define $${\mathrm{corr}_{{\epsilon}}}(A) = \max_{v \in \{-1,1\}} \{{\mathrm{corr}^{(v)}_{{\epsilon}}}(A) \}.$$ When $A$ is monochromatic, define ${\mathrm{corr}_{{\epsilon}}}(A) = 0$. If $A$ is all $1$’s then define ${\mathrm{corr}^{(1)}_{{\epsilon}}}(A) = 0$, and otherwise define ${\mathrm{corr}^{(-1)}_{{\epsilon}}}(A) = 0$. Note that when $A$ is monochromatic then there are no balanced distributions on $A$, thus an explicit definition is required for this special case. Note that the above version of the corruption bound is smaller than the usual definition, since we consider only a subset of the distributions. But, as we are interested in proving an upper bound, this works to our advantage, as long as we can still prove the properties we need. The upper bound {#upper-bound} =============== We prove the following upper bound: \[main\] Let $A$ be a sign matrix. Then $$D(A) \le O({\mathrm{corr}^{(1)}_{1/4}}(A) \log^2 {\mathrm{rank}}(A) ),$$ and $$D(A) \le O({\mathrm{corr}^{(-1)}_{1/4}}(A) \log^2 {\mathrm{rank}}(A) ).$$ Observe that, ${\mathrm{corr}^{(-1)}_{1/4}}(A)= {\mathrm{corr}^{(1)}_{1/4}}(-A)$. The proof of the upper bound is composed of three main ingredients: - An amplification lemma for the error in the corruption bound, [@Kla03]. - A lemma relating the size of monochromatic rectangles to the size of corrupted rectangles with small error (relative to the rank), [@LG13; @Lov13]. - Nisan and Wigderson’s deterministic protocol, [@NW95]. This protocol uses the existence of a large monochromatic rectangle. A sketch of the proof is: If the corruption bound is small, then for every probability distribution there is a large rectangle with a small fraction of either $1$’s or $-1$’s, by definition. Using the amplification lemma, this fraction can be made small even relative to the rank. Then, by the second part, there is a large monochromatic rectangle, allowing the use of the protocol of Nisan and Wigderson. In combining these three parts together, a few modifications and adjustments are required, most of which are contained in Corollary \[cor:corr\_mono\] relating the corruption bound to the size of a monochromatic rectangle. We start with the amplification lemma for the corruption bound. \[lem:amp\] Let $A$ be a sign matrix, $0 \le \epsilon < 1/2$ [^1] , and $\ell \in \mathbb{N}$. Then $${\mathrm{corr}^{(1)}_{\epsilon^{\ell}}}(A) \le O(\ell \cdot {\mathrm{corr}^{(1)}_{{\epsilon}}}(A)).$$ For convenience, we sketch the proof. We also observe that the proof in [@Kla03] gives more than revealed in the statement of the lemma [^2]. Let $k = {\mathrm{corr}^{(1)}_{{\epsilon}}}(A)$, and fix any uniformly-balanced distribution $\mu$ on the entries of $A$. Then, there is a rectangle $R$ of measure at least $2^{-k}$ and with at most an ${\epsilon}$-fraction of $-1$’s, with respect to $\mu$. Now, let $\mu_1$ be the uniformly-balanced distribution supported on $R$. Again, there is a rectangle $R_1$ in $R$ with measure at least $2^{-k}$ and at most an ${\epsilon}$-fraction of $-1$’s, with respect to $\mu_1$. The measure of $R_1$ with respect to $\mu$ is at least $2^{-2k}$, and the fraction of $-1$’s is at most $O({\epsilon}^2)$ (see [@Kla03] for the details). Repeating this process gives the proof. Intuitively, picking the new distribution $\mu_1$ allows to “hide” the small fraction of $-1$’s, and search for a smaller fraction inside this set. It is left to observe that when the starting distribution $\mu$ is uniformly-balanced, the distribution $\mu_1$ picked in the process is also uniformly balanced. Thus, the proof in [@Kla03] is valid for our restricted version of the corruption bound. The following lemma implies that amplification achieves a monochromatic rectangle when the error is proportional to $1/{\mathrm{rank}}(A)$. \[claim:corr\_to\_mono\] Let $A$ be an $m \times n$ sign matrix with ${\mathrm{rank}}(A)=r$. Assume that the fraction of $1$’s or the fraction of $-1$’s in $A$ is at most $\frac{1}{2r}$. Then $A$ contains a monochromatic rectangle $R$ such that $|R| \ge \frac{mn}{8}$. We note that the proof of Claim \[claim:corr\_to\_mono\] is purely algebraic, and thus it holds for the rank over any field. In particular, it holds with rank over $\mathbb{F}_2$ which can be much smaller than the rank over the reals. This is worth noting since the log rank conjecture is false for the rank over $\mathbb{F}_2$, e.g. the Inner Product (Hadamard) matrix. Thus, in any proof of the log rank conjecture, at least one ingredient must use inimitable properties of the rank over $\mathbb{R}$. Lemma \[lem:amp\] and Claim \[claim:corr\_to\_mono\], imply the following corollary: \[cor:corr\_mono\] Let $A$ be a sign matrix, $q_1=\log {\mathrm{rank}}(A) \cdot {\mathrm{corr}^{(1)}_{1/4}}(A)$, $q_2=\log {\mathrm{rank}}(A) \cdot {\mathrm{corr}^{(1)}_{1/4}}(-A)$, and also let $q = \min \{q_1, q_2\}$. Then $A$ contains a monochromatic rectangle of normalized size $2^{-O(q)}$. Let ${\mathrm{rank}}(A)=r$, and assume without loss of generality that $q_1 \le q_2$. If the fraction of either $-1$’s or $1$’s in $A$ is smaller than $1/2r$, we can apply Claim \[claim:corr\_to\_mono\] and conclude. Otherwise, let $\mu$ be the uniformly-balanced distribution supported on all the entries of $A$, and denote by $u$ the uniform distribution on the entries of $A$. Let $\delta = u(-1)$, be the fraction of $-1$’s in $A$. Then for every entry $(i,j)$ it holds that $$\begin{aligned} \label{1} A_{i,j} = -1 &\Rightarrow& \mu(i,j) = \frac{u(i,j)}{2\delta},\\ A_{i,j} = 1 &\Rightarrow& \mu(i,j) = \frac{u(i,j)}{2(1-\delta)}. \nonumber\end{aligned}$$ Let ${\mathrm{corr}^{(1)}_{1/4r}}(A) = k$. Since $\mu$ is uniformly-balanced, there is a rectangle $R$ such that $\mu(R) \ge 2^{-k}$ and $err(R,\mu,1) \le 1/4r$. By Equation (\[1\]) and the fact that $1/2r \le \delta \le 1-1/2r$, we have that $$u(R) \ge \frac{1}{r}2^{-k}=2^{-k-\log r}.$$ In addition, $\mu(-1,R) \le 1/4r(\mu(-1,R)+\mu(1,R))$. By (\[1\]) again, the latter inequality gives $$\begin{aligned} \frac{u(-1,R)}{2\delta} &\le& \frac{1}{4r}(\frac{u(-1,R)}{2\delta}+\frac{u(1,R)}{2(1-\delta)}).\end{aligned}$$ Assume $\delta \le 1/2$, then multiplying both sides by $2\delta$ yields, $$\begin{aligned} u(-1,R) &\le& \frac{1}{4r}(u(-1,R)+\frac{\delta u(1,R)}{(1-\delta)})\\ &\le& \frac{1}{4r}(u(-1,R)+u(1,R))\\ &=& \frac{1}{4r} u(R).\end{aligned}$$ If $\delta > 1/2$ then $$u(-1,R) \le 2\mu(-1,R) \le \mu(R)/2r \le u(R)/2,$$ and we can repeat the argument on $R$. This decreases the size of the rectangle that we find by at most an extra factor of $2^{-k}$. In both cases we therefore find a combinatorial rectangle $R$ of normalized size at least $2^{-2k-\log r}$, and a fraction of $-1$’s at most $1/4r$. Now, by Claim \[claim:corr\_to\_mono\], there exists a monochromatic rectangle $R'$ such that $$|R'| \ge |R|/8 \ge 2^{-2k-\log r-3}.$$ To conclude the proof, apply Lemma \[lem:amp\] with $\ell = \lceil \log_4 4r \rceil$, which gives $${\mathrm{corr}^{(1)}_{1/4r}}(A) \le O({\mathrm{corr}^{(1)}_{1/4}}(A)\log r).$$ It is left to state the result of Nisan and Wigderson [@NW95]. We use a similar formulation to that used in [@LG13; @Lov13]. Though the statement is slightly different than that of [@NW95], the proof remains the same. \[claim:gen\_bound\] Let $A$ be a sign matrix and let ${\mathrm{rank}}(A)=r$. Assume that for every sub-matrix of $A$ there exists a monochromatic rectangle of normalized size at least $2^{-q}$. Then, $$D(A) \le O(\log^2 r+q\log r).$$ Here too we note that the use of the rank in the proof is algebraic. Thus the claim holds with the rank over any field. For the proof of Theorem \[main\] we combine Corollary \[cor:corr\_mono\], Claim \[claim:gen\_bound\] and the fact that the corruption bound is monotone in sub-matrices. Let ${\mathrm{rank}}(A)=r$. We show that $$D(A) \le O({\mathrm{corr}^{(1)}_{1/4}}(A)\log^2 r).$$ For every sub-matrix $B$ of $A$ and every $0 \le {\epsilon}\le 1$, ${\mathrm{corr}^{(1)}_{{\epsilon}}}(B) \le {\mathrm{corr}^{(1)}_{{\epsilon}}}(A)$. Let $q= {\mathrm{corr}^{(1)}_{1/4}}(A)\log r$. Then by Corollary \[cor:corr\_mono\], every sub-matrix of $A$ contains a monochromatic rectangle of normalized size $2^{-O(q)}$. The proof now follows using Claim \[claim:gen\_bound\]. The proof that $D(A) \le O({\mathrm{corr}^{(-1)}_{1/4}}(A)\log^2 r)$ is similar. In fact, it also follows by symmetry, since $D(A)=D(-A)$ and ${\mathrm{corr}^{(-1)}_{1/4}}(A)={\mathrm{corr}^{(1)}_{1/4}}(-A)$. On the log rank conjecture {#log-rank} ========================== As mentioned in the introduction, upper bounds in communication complexity are particularly interesting when stated in terms of known lower bounds, as this gives an alternative characterization of the underlying communication complexity measure. A fundamental question in communication complexity, posed by Lovász and Saks [@LS88; @LS93], is whether deterministic communication complexity is bounded by a polynomial in the log of the rank of the matrix. There is a simple deterministic protocol using ${\mathrm{rank}}(A)$ bits of communication. Despite significant efforts, no better bounds were found until recently. A breakthrough result of Lovett [@Lov13] (following also [@LG13] and [@BLZ12]) is the first improvement on this simple upper bound. We give a simple proof of this result using the framework described in the previous sections. The proof is essentially observing that discrepancy corresponds to the error in corruption, and the error can be amplified as seen in Section \[upper-bound\]. We therefore prove the following: \[cor-disc\] Let $A$ be a sign matrix, and let $disc(A)=1/d$. Then $${\mathrm{corr}_{1/4}}(A) \le O(d\log d).$$ Here $disc(A)$ is the discrepancy of $A$, defined as follows: Let $\mu$ be a distribution on the entries of $A$. The discrepancy with respect to $\mu$ is the maximal discrepancy between the measure of $1$’s and the measure of $-1$’s, over combinatorial rectangles in $A$. The discrepancy of $A$ is the minimal discrepancy over all probability distributions. The discrepancy is often used to lower bound communication complexity in different models, and it is also equivalent (up to a constant) to the reciprocal of margin complexity. See [@LMSS07; @LS08b] for the definitions and proof of the equivalence of these measures. This equivalence was used in [@LS08b] to prove that $1/disc(A) \le O(\sqrt{{\mathrm{rank}}(A)})$. Combined with Lemma \[cor-disc\] it implies that ${\mathrm{corr}_{1/4}}(A) \le O(\sqrt{{\mathrm{rank}}(A)} \log {\mathrm{rank}}(A))$. Using this in Theorem \[main\] gives the upper bound of [@Lov13] on the deterministic communication complexity in terms of the rank (up to log factors). We prove Lemma \[cor-disc\] in the remainder of this section. We use the following lemma, generalizing the amplification lemma of [@Kla03] [^3]. \[lem:amp2\] Let $A$ be a sign matrix. Then, for $1/4 < \epsilon < 1/2$ it holds that $${\mathrm{corr}^{(1)}_{1/4}}(A) \le O(\frac{1}{(1/2-{\epsilon})}\cdot {\mathrm{corr}^{(1)}_{{\epsilon}}}(A)),$$ and, for $0 \le \epsilon \le 1/4$ and $\ell \in \mathbb{N}$ $${\mathrm{corr}^{(1)}_{{\epsilon}^{\ell}}}(A) \le O(\ell\cdot {\mathrm{corr}^{(1)}_{{\epsilon}}}(A)).$$ We use the above lemma to amplify the discrepancy between the $1$’s and $-1$’s in $A$ in order to get the bound on the corruption bound: By definition of the discrepancy, for every probability distribution $\mu$ on the entries of $A$, there is a combinatorial rectangle R, such that $$\label{2} |\sum_{(i,j)\in R} \mu(i,j)A_{i,j} | \ge \frac{1}{d}.$$ Thus, for every probability distribution there is a combinatorial rectangle for which the probability of either $1$’s or $-1$’s in it is at most $1/2-1/2d$. We show that this implies that ${\mathrm{corr}^{(1)}_{1/2-1/6d}}(A) = O(\log d)$. Proving similarly for ${\mathrm{corr}^{(-1)}_{1/2-1/6d}}(A)$ and applying Lemma \[lem:amp2\] gives the bound ${\mathrm{corr}_{1/4}}(A) \le O(d\log d)$. It is left to prove the bound on ${\mathrm{corr}^{(1)}_{1/2-1/6d}}(A)$. Let $\mu$ be a uniformly-balanced distribution. We first observe that we can assume without loss of generality that $$\sum_{(i,j)\in R} \mu(i,j)A_{i,j} \ge \frac{1}{3d}.$$ Otherwise, the sum in Equation (\[2\]) is negative, and since $\mu$ is uniformly-balanced $$\sum_{(i,j)\in \bar{R}} \mu(i,j)A_{i,j} \ge \frac{1}{d},$$ where $\bar{R}$ is the complement of $R$. But $\bar{R}$ can be partitioned into three combinatorial rectangles, and thus there is a rectangle $R'$ such that $$\sum_{(i,j)\in R'} \mu(i,j)A_{i,j} \ge \frac{1}{3d}.$$ Now, $\sum_{(i,j)\in R} \mu(i,j)A_{i,j} = \mu(1,R)-\mu(-1,R)$, and $\mu(R) = \mu(1,R)+\mu(-1,R)$. Therefore, $$\begin{aligned} \mu(-1,R) &=& \frac{1}{2}\mu(R)-\frac{1}{2}\sum_{(i,j)\in R} \mu(i,j)A_{i,j} \\ &\le& \frac{1}{2}\mu(R)-\frac{1}{6d}\\ &=& \frac{1}{2}\mu(R)-\frac{\mu(R)}{6d\mu(R)}\\ &=& (\frac{1}{2}-\frac{1}{6d\mu(R)})\mu(R)\\ &\le& (\frac{1}{2}-\frac{1}{6d})\mu(R).\end{aligned}$$ This concludes the proof, as obviously $\mu(R) \ge \frac{1}{d}$. Proof of Lemma \[lem:amp2\] {#sec:proof} =========================== Let $A$ be a sign matrix. Then, for $1/4 < \epsilon < 1/2$ it holds that $${\mathrm{corr}^{(1)}_{1/4}}(A) \le O(\frac{1}{(1/2-{\epsilon})}\cdot {\mathrm{corr}^{(1)}_{{\epsilon}}}(A)),$$ and, for $0 \le \epsilon \le 1/4$ and $\ell \in \mathbb{N}$ $${\mathrm{corr}^{(1)}_{{\epsilon}^{\ell}}}(A) \le O(\ell\cdot {\mathrm{corr}^{(1)}_{{\epsilon}}}(A)).$$ Note that a unified way to state the lemma is: for every $0 < {\epsilon}_1 < {\epsilon}< 1/2$ it holds that $${\mathrm{corr}^{(1)}_{{\epsilon}_1}}(A) \le O(\log_{{\epsilon}} {\epsilon}_1 \cdot {\mathrm{corr}^{(1)}_{{\epsilon}}}(A)).$$ We start with a single step of refinement. Let ${\mathrm{corr}^{(1)}_{{\epsilon}}}(A) = k$, and pick a uniformly-balanced distribution $\mu$. By definition there is a rectangle $R$ such that $\mu(R) \ge 2^{-k}$ and $err(R,\mu,1) \le {\epsilon}$, we show that there is a rectangle $R_1$ such that $\mu(R_1) \ge 2^{-2k}$ and $err(R_1,\mu,1) \le 2{\epsilon}^2$. Let $\mu_1$ be the uniformly balanced distribution on $R$. Then for every entry $(i,j)$ it holds that $$\begin{aligned} \label{2} A_{i,j} = -1, (i,j) \in R &\Rightarrow& \mu_1(i,j) = \frac{\mu(i,j)}{2\mu(-1,R)},\\ A_{i,j} = 1, (i,j) \in R &\Rightarrow& \mu_1(i,j) = \frac{\mu(i,j)}{2(\mu(R)-\mu(-1,R))}. \nonumber\end{aligned}$$ Since $\mu_1$ is uniformly-balanced, there is a rectangle $R_1$ such that $\mu_1(R_1) \ge 2^{-k}$ and $err(R_1,\mu_1,1) \le {\epsilon}$. By Equation (\[2\]): $$\begin{aligned} \mu(R_1) &=& \sum_{(i,j)\in R_1} \mu(i,j)\\ &=& \sum_{A_{i,j}=-1, (i,j)\in R_1} \mu(i,j) + \sum_{A_{i,j}=1, (i,j)\in R_1} \mu(i,j) \\ &=& \sum_{A_{i,j}=-1, (i,j)\in R_1} 2\mu(-1,R) \mu_1(i,j) + \sum_{A_{i,j}=1, (i,j)\in R_1} 2(\mu(R)-\mu(-1,R)) \mu_1(i,j) \\ &=& 2\mu(-1,R) \sum_{A_{i,j}=-1, (i,j)\in R_1} \mu_1(i,j) + 2(\mu(R)-\mu(-1,R))\sum_{A_{i,j}=1, (i,j)\in R_1} \mu_1(i,j) \\ &=& 2\mu(-1,R) \mu_1(-1,R_1) + 2(\mu(R)-\mu(-1,R))(\mu_1(R_1)-\mu_1(-1,R_1)) \\ &=& 2[\mu(-1|R)\mu(R) \mu_1(-1|R_1)\mu(R_1) + (\mu(R)-\mu(-1|R)\mu(R))(\mu_1(R_1)-\mu_1(-1|R_1)\mu(R_1))]\\ &=& 2\mu(R)\mu(R_1)[\mu(-1|R)\mu_1(-1|R_1) + (1-\mu(-1|R))(1-\mu_1(-1|R_1))].\end{aligned}$$ Recall that $\mu(-1|R) \le {\epsilon}< 1/2$ and $\mu_1(-1|R_1) \le {\epsilon}< 1/2$. The function $f(x,y)=1-x-y+2xy$ satisfies $f(x,y) \ge 1/2$ for $x,y \in [0,1/2]$. Thus $$\begin{aligned} \mu(R_1) &=& 2\mu(R)\mu(R_1)[\mu(-1|R)\mu_1(-1|R_1) + (1-\mu(-1|R))(1-\mu_1(-1|R_1))]\\ &\ge& \mu(R)\mu_1(R_1)\\ &\ge& 2^{-2k}.\end{aligned}$$ It is left to consider the error $err(R_i,\mu,1)$, for $1 \le i \le T$. Again we start with a single step, recall that $err(R_1,\mu,1) = \mu(-1|R_1) = \mu(-1,R_1)/\mu(R_1)$. Using Equation (\[2\]) we get $$\begin{aligned} \mu(-1,R_1) &=& \sum_{A_{i,j}=-1, (i,j)\in R_1} \mu(i,j)\\ &=& \sum_{A_{i,j}=-1, (i,j)\in R_1} 2\mu(-1,R) \mu_1(i,j)\\ &=& 2\mu(-1,R) \sum_{A_{i,j}=-1, (i,j)\in R_1} \mu_1(i,j)\\ &=& 2\mu(-1,R) \mu_1(-1,R_1)\\ &\le& 2{\epsilon}^2\mu(R)\mu(R_1).\end{aligned}$$ Combining the last inequality with the inequality $\mu(R_1) \ge \mu(R)\mu_1(R_1)$ proved earlier, we get $$\begin{aligned} \mu(-1|R_1) &=& \mu(-1,R_1)/\mu(R_1)\\ &\le& \frac{2{\epsilon}^2\mu(R)\mu(R_1)}{2\mu(R)\mu_1(R_1)}\\ &=& 2{\epsilon}^2.\end{aligned}$$ Similarly, after $T$ steps of refinement, finding finer and finer rectangles $R_1, R_2, \ldots, R_T$, we get $\mu(R_T) \ge 2^{-(T+1)k}$, and the error satisfies $\mu(-1|R_T) \le 2^{T}{\epsilon}^{T+1}$. Therefore, for ${\epsilon}\le 1/4$, after $T=2l$ steps we achieve error at most $2^{2l}{\epsilon}^{2l+1} \le {\epsilon}^l$. For $1/4 < {\epsilon}< 1/2$ let $\delta = 1/2 - {\epsilon}$. After $T$ steps the error is at most $(1-2\delta)^T$, and for $T=c/\delta$ steps the error is smaller than $1/4$, for some constant $c$. Acknowledgements {#acknowledgements .unnumbered} ================ I thank Troy Lee and Michal Parnas for their help in writing this manuscript. [10]{} L. Babai, P. Frankl, and J. Simon. Complexity classes in communication complexity theory. In [*Proceedings of the 27th IEEE Symposium on Foundations of Computer Science*]{}. IEEE, 1986. Eli Ben-Sasson, Shachar Lovett, and Noga Ron-Zewi. An additive combinatorics approach relating rank to communication complexity. In [*Foundations of Computer Science (FOCS), 2012 IEEE 53rd Annual Symposium on*]{}, pages 177–186, 2012. H. Buhrman, R. Cleve, and A. Wigderson. Quantum vs. classical communication and computation. In [*Proceedings of the 30th ACM Symposium on the Theory of Computing*]{}, pages 63–68. ACM, 1998. H. Buhrman and [R. de]{} Wolf. Communication complexity lower bounds by polynomials. In [*Proceedings of the 16th IEEE Conference on Computational Complexity*]{}, pages 120–130, 2001. D. Gavinsky and S. Lovett. En route to the log-rank conjecture: New reductions and equivalent formulations. Electronic Colloquium on Computational Complexity (ECCC), vol. 20, p. 80, 2013. R. Jain and H. Klauck. The partition bound for classical communication complexity and query complexity. In [*IEEE Conference on Computational Complexity*]{}, pages 247–258. IEEE Computer Society, 2010. I. Kerenidis, S. Laplante, V. Lerays, J. Roland, and D. Xiao. Lower bounds on information complexity via zero-communication protocols and applications. In [*FOCS*]{}, pages 500–509. IEEE Computer Society, 2012. H. Klauck. Rectangle size bounds and threshold covers in communication complexity. In [*Proceedings of the 18th IEEE Conference on Computational Complexity*]{}. IEEE, 2003. E. Kushilevitz and N. Nisan. . Cambridge University Press, 1997. N. Linial, S. Mendelson, G. Schechtman, and A. Shraibman. Complexity measures of sign matrices. , 27(4):439–463, 2007. N. Linial and A. Shraibman. Learning complexity versus communication complexity. In [*Proceedings of the 23rd IEEE Conference on Computational Complexity*]{}, pages 384–393. IEEE, 2008. L. Lovász. Communication complexity: A survey. In B. Korte, L. Lovász, H. Pr[ö]{}mel, and A. Schrijver, editors, [*Paths, flows, and [VLSI]{}-layout*]{}, pages 235–265. Springer-Verlag, 1990. L. Lovász and M. Saks. M[ö]{}bius functions and communication complexity. In [*Proceedings of the 29th IEEE Symposium on Foundations of Computer Science*]{}, pages 81–90. IEEE, 1988. L. Lovász and M. Saks. Communication complexity and combinatorial lattice theory. , 47:322–349, 1993. S. Lovett. Communication is bounded by root of rank. Technical Report arXiv:1306.1877, arXiv, 2013. N. Nisan and A. Wigderson. A note on rank [vs.]{} communication complexity. , 15(4):557–566, 1995. A. Razborov. On the distributional complexity of disjointness. , 106:385–390, 1992. A. Yao. Lower bounds by probabilistic arguments. In [*Proceedings of the 24th IEEE Symposium on Foundations of Computer Science*]{}, pages 420–428, 1983. [^1]: Here ${\epsilon}$ must be bounded away from $1/2$. A more precise lemma, dealing also with the case of ${\epsilon}$ close to $1/2$, is stated in Section \[log-rank\]. [^2]: We provide a complete and independent proof for a more general lemma in Section \[sec:proof\]. [^3]: Studying the proof of Lemma \[lem:amp\] in [@Kla03], it can be adapted to give a proof for this generalized lemma. We give an independent proof in Section \[sec:proof\].

ArXiv

--- abstract: 'The manual creation of gold standards for named entity recognition and entity linking is time- and resource-intensive. Moreover, recent works show that such gold standards contain a large proportion of mistakes in addition to being difficult to maintain. We hence present [<span style="font-variant:small-caps;">Bengal</span>]{}, a novel automatic generation of such gold standards as a complement to manually created benchmarks. The main advantage of our benchmarks is that they can be readily generated at any time. They are also cost-effective while being guaranteed to be free of annotation errors. We compare the performance of 11 tools on benchmarks in English generated by [<span style="font-variant:small-caps;">Bengal</span>]{} and on 16 benchmarks created manually. We show that our approach can be ported easily across languages by presenting results achieved by 4 tools on both Brazilian Portuguese and Spanish. Overall, our results suggest that our automatic benchmark generation approach can create varied benchmarks that have characteristics similar to those of existing benchmarks. Our approach is open-source. Our expe-rimental results are available at <http://faturl.com/bengalexpinlg> and the code at <https://github.com/dice-group/BENGAL>.' author: - | Axel-Cyrille Ngoma Ngomo$^{1,2}$ Michael Röder$^{1}$ Diego Moussallem$^{1,2}$ Ricardo Usbeck$^{1}$ René Speck$^{1,2}$\ $^{1}$Data Science Group, University of Paderborn, Germany\ $^{2}$AKSW Research Group, University of Leipzig, Germany\ [{first.lastname}@upb.de]{}\ [lastname@informatik.uni-leipzig.de]{} bibliography: - 'inlg2018.bib' title: '[BENGAL]{}: An Automatic Benchmark Generator for Entity Recognition and Linking' --- Introduction {#intro} ============ The creating of gold standard is of central importance for the objective assessment and development of approaches all around computer science. For example, evaluation campaigns such as BioASQ [@bioasq] have led to an improvement of the F-measure achieved by bio-medical question answering systems by more than 5%. While the manual creation of and gold standards (also called benchmarks) has the advantage of yielding resources which reflect human processing, it also exhibits significant disadvantages: *Annotation mistakes*: Human annotators have to read through every sentence in the corpus and often (a) miss annotations or (b) assign wrong resources to entities for reasons as various as fatigue or lack of background knowledge (and this even when supported with annotation tools). For example,   was able to determine that up to 38,453 of the annotations in commonly used benchmarks (see GERBIL [@gerbil] for a list of these benchmarks) were erroneous. A manual evaluation of 25 documents from the ACE2004 benchmark revealed that 195 annotations were missing and 14 of 306 annotations were incorrect. Similar findings were reported for AIDA/CONLL [@conll2003] and OKE2015 [@okechallenge]. *Volume*: Manually created benchmarks are usually small (commonly $<2,500$ documents, see Table \[tab:features\]). Hence, they are of little help when aiming to benchmark the scalability of existing solutions (especially when these solutions use caching). *Lack of updates*: Manual benchmark generation approaches lead to static corpora which tend not to reflect the newest reference knowledge graphs (also called s). For example, several of the benchmarks presented in GERBIL [@gerbil] link to outdated versions of Wikipedia or DBpedia. *Popularity bias*: show that manual benchmarks are often biased towards popular resources. *Lack of availability*: The lack of benchmarks for resource-poor languages inhibits the development of corresponding and solutions. Automatic methods are a *viable and supplementary* approach for the generation of gold standards for and , especially as they address some of the weaknesses of the manual benchmark creation process. *The main contribution of our paper is a novel approach for the automatic generation of benchmarks for and* dubbed [<span style="font-variant:small-caps;">Bengal</span>]{}. Our approach relies on the abundance of structured data in on the Web and is based on techniques which verbalize such data to generate automatically annotated natural language statements. Our automatic benchmark creation method addresses the drawbacks of manual benchmark generation aforementioned as follows: It alleviates the human annotation error problem by relying on data in which explicitly contain the entities to find. [<span style="font-variant:small-caps;">Bengal</span>]{}is able to generate arbitrarily large benchmarks. Hence, it can enhance the measurement of both the accuracy and the scalability of approaches. [<span style="font-variant:small-caps;">Bengal</span>]{}can be updated easily to reflect the newest terminology and reference s. Hence, it can generate corpora that reflect the newest s. [<span style="font-variant:small-caps;">Bengal</span>]{}is not biased towards popular resources as it can choose entities to include in the benchmark generated following a uniform distribution. [<span style="font-variant:small-caps;">Bengal</span>]{}can be ported to any token-based language. This is exemplified by porting [<span style="font-variant:small-caps;">Bengal</span>]{}to Portuguese and Spanish. Related Work {#sec:related} ============ Gold Standards for NER and EL ----------------------------- According to GERBIL [@gerbil], the 2003 CoNLL shared task [@conll2003] is the most used benchmark dataset for recognition and linking. The ACE2004 and MSNBC [@Cucerzan07] news datasets were used by Ratinov et al. [@rat:rot] to evaluate their seminal work on linking to Wikipedia. Another often-used corpus is AQUAINT, e.g., used by Milne and Witten [@milne2008learning]. Detailed dataset statistics on some of these benchmarks can be found in Table \[tab:features\]. A recent uptake of publicly available corpora [@N3; @yovisto] based on RDF has led to the creation of many new datasets. For example, the Spotlight corpus and the KORE 50 dataset were proposed to showcase the usability of RDF-based annotations [@spotlight]. The multilingual N3 collection [@N3] was introduced to widen the scope and diversity of NIF-based corpora. Another recent observation is the shift towards gold standards for micropost documents like tweets. For example, the Microposts2014 corpus [@cano2014] was created to evaluate on smaller pieces of text. Semi-automatic approaches to benchmark creation are commonly crowd-based. They use one or more recognizers to create a first set of annotations and then hand over the tasks of refinement and/or linking to crowd workers to improve the quality. Examples of such approaches include and CALBC [@rebholz2010calbc]. introduced a voting-based algorithm which analyses the hyperlinks presented in the input texts retrieved from different disambiguation systems such as Babelfy [@moro2014multilingual]. Each entity mention in the input text is linked based on the degree of agreement across three systems. [<span style="font-variant:small-caps;">Bengal</span>]{}is the first automatic approach that makes use of structured data and can be replicated on any RDF KB for benchmarks. NLG for the Web of Data ----------------------- A plethora of works have investigated the generation of texts from such as  . However, the generation of NL from RDF has only recently gained momentum. This attention comes from the great number of published works such as [@cimiano2013exploiting; @duma2013generating; @ell2014language; @biran2015discourse] which used as an input data and achieved promising results. Moreover, the works published in the WebNLG [@colin2016webnlg] challenge, which used deep learning techniques such as [@sleimi2016generating; @mrabet2016aligning], also contributed to this interest. has also been showing promising benefits to the generation of benchmarks for evaluating systems, e.g., [@gardent2017creating; @perez2016building; @mohammed2016category; @schwitter2004controlled; @hewlett2005effective; @sun2006domain]. However, RDF has never been used for creating NER and NEL benchmarks. [<span style="font-variant:small-caps;">Bengal</span>]{} addresses this research gap. The [<span style="font-variant:small-caps;">Bengal</span>]{}approach {#sec:bengal} ==================================================================== {width="\textwidth"} [<span style="font-variant:small-caps;">Bengal</span>]{}is based on the observation that more than 150 billion facts pertaining to more than 3 billion entities are available in machine-readable form on the Web (i.e., as RDF triples).[^1] The basic intuition behind our approach is hence as follows: *Given that and are often used in pipelines for the extraction of machine-readable facts from text, we can invert the pipeline and go from facts to text*, thereby using the information in the facts to produce a gold standard that is *guaranteed to contain no errors*. In the following, we begin by giving a brief formal overview of . Thereafter, we present how we use to generate and benchmarks automatically and at scale. Preliminaries and Notation {#subsec:notation} -------------------------- ### RDF The notation presented herein is based on the RDF 1.1 specification. An RDF graph $G$ is a set of facts. Each fact is a triple $t = (s, p, o) \in (R \cup B) \times P \times (R \cup B \cup L)$ where $R$ is the set of all resources (i.e., things of the real world), $P$ is the set of all predicates (binary relations), $B$ is the set of all blank nodes (which basically express existential quantification) and $L$ is the set of all literals (i.e., of datatype values). We call the set $R \cup P \cup L \cup B$ our universe and call its elements entities. A fragment of DBpedia[^2] is shown below. We will use this fragment in our examples. For the sake of space, our examples are in English. However, note that we ported [<span style="font-variant:small-caps;">Bengal</span>]{}to Portuguese and Spanish so as to exemplify that it is not biased towards a particular language. Also, the morphological richness of both led us to choose them as languages. :Albert_Einstein dbo:birthPlace :Ulm . :Albert_Einstein dbo:deathPlace :Princeton . :Albert_Einstein rdf:type dbo:Scientist . :Albert_Einstein dbo:field :Physics . :Ulm dbo:country :Germany. :Albert_Einstein rdfs:label "Albert Einstein"@en. ### Benchmarks We define a benchmark as a set $C$ of annotated documents $D_i$. Each document $D_i$ is a sequence of characters $s_{i1} \ldots s_{in}$. Each subsequence $s_{ij} \ldots s_{ik}$ (with $j < k$) of the document $D_i$ which stands for a resource $r \in R$ is assumed to be marked as such. We model the marking of resources by the function $m: C \times \mathbb{N} \times \mathbb{N} \rightarrow R$ and write $m(D_i, j, k) = r$ to signify that the substring $s_{ij} \ldots s_{ik}$ stands for the resource $r$. In case the substring $s_{ij} \ldots s_{ik}$ does not stand for a resource, we write $m(D_i,j,k) = \epsilon$. Let $D_0$ be the example shown in Listing \[lst:text\]. We would write $m(D_0, 0, 14) = \texttt{:AlbertEinstein}$. Albert Einstein was born in Ulm. Verbalization ------------- The notation and formal framework for verbalization in [<span style="font-variant:small-caps;">Bengal</span>]{}are based on SPARQL2NL [@ngo+13a]. Let $W$ be the set of all words in the dictionary of our target language (e.g., English). We define the realization function $\rho: R \cup P \cup L \rightarrow W^*$ as the function which maps each entity to a word or sequence of words from the dictionary. Formally, the goal of our NLG approach is to devise an extension of $\rho$ to conjunctions of RDF triples. This extension maps all triples $t$ to their realization $\rho(t)$ and defines how these atomic realizations are to be combined. We denote the extension of $\rho$ by the same label $\rho$ for the sake of simplicity. We adopt a rule-based approach to devise the extension of $\rho$, where the rules extending $\rho$ to RDF triples are expressed in a conjunctive manner. This means that for premises $P_1,\ldots,P_n$ and consequences $K_1,\ldots,K_m$ we write $P_1 \wedge \ldots \wedge P_n \Rightarrow K_1 \wedge \ldots \wedge K_m$. The premises and consequences are explicated by using an extension of the Stanford dependencies.[^3] We rely especially on the constructs explained in Table \[tab:dependencies\]. For example, a possessive dependency between two phrase elements $e_1$ and $e_2$ is represented as $\texttt{poss}(e_1,e_2)$. For the sake of simplicity, we sometimes reduce the construct $\texttt{subj(y,x)}\wedge \texttt{dobj(y,z)}$ to the triple `(x,y,z)` $\in W^3$. **Dependency** **Explanation** ---------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ `cc` Stands for the relation between a conjunct and a given conjunction (in most cases [and]{} or [or]{}). For example in the sentence [John eats an apple and a pear]{}, `cc(PEAR,AND)` holds. We mainly use this construct to specify reduction and replacement rules. `conj`$^*$ Used to build the *conjunction* of two phrase elements, e.g. `conj(subj(EAT,JOHN), subj(DRINK,MARY))` stands for [John eats and Mary drinks]{}. `conj` is not to be confused with the logical conjunction $\wedge$, which we use to state that two dependencies hold in the same sentence. For example `subj(EAT,JOHN) \wedge dobj(EAT,FISH)` is to be read as [John eats fish]{}. `dobj` Dependency between a verb and its *direct object*, for example `dobj(EAT,APPLE)` expresses [to eat an/the apple]{}. `nn` The *noun compound modifier* is used to modify a head noun by the means of another noun. For instance `nn(FARMER,JOHN)` stands for [farmer John]{}. `poss` Expresses a possessive dependency between two lexical items, for example `poss(JOHN,DOG)` expresses [John’s dog]{}. `subj` Relation between *subject* and verb, for example `subj(BE,JOHN)` expresses [John is]{}. Approach -------- [<span style="font-variant:small-caps;">Bengal</span>]{}assumes that it is given (1) an RDF graph $G \subseteq (R \cup B) \times P \times (R \cup B \cup L)$, (2) a number of documents to generate, (3) a minimal resp. maximal document size (i.e., number of triples to use during the generation process) $d_{min}$ resp. $d_{max}$, (4) a set of restrictions pertaining to the resources to generate and (5) a strategy for generating single documents. Given the graph $G$, [<span style="font-variant:small-caps;">Bengal</span>]{}begins by selecting a set of *seed resources* from $G$ based on the restrictions set using parameter (4). Thereafter, it uses the strategy defined via parameter (5) to select a subgraph of $G$. This subgraph contains a randomly selected number $d$ of triples with $d_{min} \leq d \leq d_{max}$. The subgraph is then verbalized. The verbalization is annotated automatically and finally returned as a single document. Each single document then may be paraphrased if this option is chosen in the initial phase. This process is repeated as many times as necessary to reach the predefined number of documents. In the following, we present the details of each step underlying our benchmark generation process displayed in . ### Seed Selection Given that we rely on RDF, we model the seed selection by means of a SPARQL SELECT query with one projection variable. Note that we can use the wealth of SPARQL to devise seed selection strategies of arbitrary complexity. However, given that NER and EL frameworks commonly focus on particular classes of resources, we are confronted with the condition that the seeds must be instances of a set of classes, e.g., `:Person`, `:Organization` or `:Place`. The SPARQL query for our example dataset would be as follows: SELECT ?x WHERE { {?x a :Person.} UNION {?x a :Organization.} UNION {?x a :Place.} } ### Subgraph Generation Our approach to generating subgraphs is reminiscent of SPARQL query topologies as available in SPARQL query benchmarks. As these queries (e.g., FEASIBLE[^4] queries) describe real information needs, their topology must stand for the type of information that is necessitated by applications and humans. We thus distinguish between three main types of subgraphs to be generated from RDF data: (1) *star graphs* provide information about a particular entity (e.g, the short biography of a person); (2) *path graphs* describe the relations between two entities (e.g., the relation between a gene and a side-effect); (3) *hybrid graphs* are a mix of both and commonly describe a specialized subject matter involving several actors (e.g., a description of the cast of a movie). *Star Graphs.* For each $s_i \in S$, we gather all triples of the form $t = (s_i, p, o) \in R \times P \times (R \cup L)$.[^5] The triples are then added to a list $L(s_i)$ sorted in descending order according to a hash function $h$. After randomly selecting a document size $d$ between $d_{min}$ and $d_{max}$, we select $d$ random triples from $L(s_i)$. For the dataset shown in Listing \[lst:example\] and $d=2$, we would for example get Listing \[lst:star\]. :AlbertEinstein :birthPlace :Ulm . :AlbertEinstein :deathPlace :Princeton . *Symmetric Star Graphs.* As above with $t \in \{(s_i, p, o) \in G \vee (o, p, s_i) \in G$}. *Path Graphs.* For each $s_i \in S$, we begin by computing list $L(s_i)$ as in the symmetric star graph generation. Then, we pick a random triple $(s_i, p, o)$ or $(o, p, s_i)$ from $L(s_i)$ that is such that $o$ is a resource. We then use $o$ as seed and repeat the operation until we have generated $d$ triples, where $d$ is randomly generated as above. For the example dataset shown in Listing \[lst:example\] and $d=2$, we would for example get Listing \[lst:path\]. :AlbertEinstein :birthPlace :Ulm . :Ulm :country :Germany . *Hybrid Graphs.* This is a 50/50-mix of the star and path graph generation approaches. In each iteration, we choose and apply one of the two strategies above randomly. For example, the hybrid graph generation can generate: :AlbertEinstein :birthPlace :Ulm . :AlbertEinstein :deathPlace :Princeton . :Ulm :country :Germany . *Summary Graph Generation.* This last strategy is a specialization of the star graph generation where the set of triples to a resource is not chosen randomly. Instead, for each class (e.g., `:Person`) of the input KB, we begin by filtering the set of properties and only consider properties that (1) have the said class as domain and (2) achieve a coverage above a user-set threshold (60% in our experiments) (e.g., `:birthPlace`, `:deathPlace`, `:spouse`). We then build a property co-occurence graph for the said class in which the nodes are the properties selected in the preceding step and the co-occurence of two properties $p_1$ and $p_2$ is the instance $r$ of the input class where $\exists o_1, o_2: (r, p_1, o_1) \in K \wedge (r, p_2, o_2) \in K$. The resulting graph is then clustered (e.g., by using the approach presented by ). We finally select the clusters which contain the properties with the highest frequencies in $K$ that allow the selection of at least $d$ triples from $K$. For example, if `:birthPlace` (frequency = 10), `:deathPlace` (frequency = 10) were in the same cluster while `:spouse` (frequency = 8) were in its own cluster, we would choose the pair (`:birthPlace`, `:deathPlace`) and return the corresponding triples for our input resource. Hence, we would return Listing \[lst:star\] for our running example. ### Verbalization module The verbalization (micro-planning) strategy for the first four strategies consists of verbalizing each triple as a single sentence and is derived from SPARQL2NL [@ngo+13a]. To verbalize the subject of the triple $t=(s, p, o)$, we use one of its labels according to (e.g., the `rdfs:label`). If the object $o$ is a resource, we follow the same approach as for the subject. Importantly, the verbalization of a triple $t = (s, p, o)$ depends mostly on the verbalization of the predicate `p` (see Table \[tab:dependencies\] for semantics). If `p` can be realized as a noun phrase, then a possessive clause can be used to express the semantics of $(s, p, o)$. For example, if $p$ can be verbalized as a nominal compound like `birth place`, then the verbalization $\rho(s, p, o)$ of the triple is as follows: `poss`($\rho(p)$,$\rho(s)$) $\wedge$ `subj`(,$\rho(p)$) $\wedge$ `dobj`(,$\rho(o)$). In case `p`’s realization is a verb, then the triple can be verbalized as `subj`($\rho(p)$,$\rho(s)$) $\wedge$ `dobj`($\rho(p)$,$\rho(o)$). In our example, verbalizing (`:AlbertEinstein`, `dbo:birthPlace`, `:Ulm`) would thus lead to `Albert Einstein’s birth place is Ulm.`, as `birth place` is a noun. In the case of summary graphs, we go beyond the verbalization of single sentences and merge sentences that were derived from the same cluster. For example, if $p_1$ and $p_2$ can be verbalized as nouns, then we apply the following rule: $\rho(s, p_1, o_1)\land\rho(s, p_2, o_2) \Rightarrow \text{\texttt{conj}}(\text{\texttt{poss}}(\rho(p_1),\rho(s)) \land \text{\texttt{subj}}(\text{\texttt{BE}}_1,\rho(p_1)) \land \text{\texttt{dobj}}(\text{\texttt{BE}}_1,\rho(o_1)) \nonumber \land \text{\texttt{poss}}(\rho(p_2),\rho(\text{\texttt{pronoun}}(s)))\nonumber \land \text{\texttt{subj}}(\text{\texttt{BE}}_2,\rho(p_2)) \land \text{\texttt{dobj}}(\text{\texttt{BE}}_2,\rho(o_2))\nonumber$. Note that `pronoun(s)` returns the correct pronoun for a resource based on its type and gender. Therewith, we can generate `Albert Einstein’s birth place is Ulm and his death place is Princeton`. ### Paraphrasing With this step, [<span style="font-variant:small-caps;">Bengal</span>]{}avoids the generation of a large number of sentences that share the same terms and the same structure. Additionally, this step makes the use of reverse engineering strategies for the generation more difficult as it increases the diversity of the text in the benchmarks. Our paraphrasing is largely based on  and runs as follows: 1. Change the structure of the sentence: We use the location of verbs in each sentence to randomly change passive into active structures and vice-versa. Sentences which describe type information (e.g., `Einstein is a person`) are not altered. 2. Replace synonyms: We use POS tags to select alternative labels from the knowledge base and a reference dictionary to replace entity labels by a synonym. An example of a paraphrase generated by [<span style="font-variant:small-caps;">Bengal</span>]{}is shown in Listing \[lst:parasummary\]. (*\bfseries Original:*) Edmund Pettus Bridge is a bridge. It crosses Alabama River. Its type is Through arch bridge. It was declared a National Historic Landmark on March 11, 2013. (*\bfseries Paraphrased:*) Edmund Pettus Bridge is a bridge. It crosses Alabama River. Through arch bridge is its type. Pettus was declared a National Historic Landmark on March 11, 2013. Experiments and Results {#sec:eval} ======================= We generated 13 datasets in English (B1-B13), 4 datasets in Brazilian Portuguese and 4 datasets in Spanish to evaluate our approach.[^6] B1 to B10 were generated by running our five sub-graph generation methods with and without paraphrasing. The number of documents was set to 100 while $(d_{min}, d_{max})$ was set to $(1,5)$. B11 shows how [<span style="font-variant:small-caps;">Bengal</span>]{}can be used to evaluate the scalability of approaches.[^7] Here, we used the hybrid generation strategy to generate 10,000 documents. B12 and B13 comprise 10 longer documents each with $d_{min}$ set to 90. For B12, we focused on generating a high number of entities in the documents while B13 contains less entities but the same number of documents. We compared B1-B13 with the 16 manually created gold standards for English found in GERBIL. The comparison was carried out in two ways. First, we assessed the features of the datasets. Then, we compared the micro F-measure of 11 NER and EL frameworks on the manually and automatically generated datasets. We chose to use these 11 frameworks because they are included in GERBIL. This inclusion ensures that their interfaces are compatible and their results comparable. In addition, we assessed the performance of multi-lingual NER and EL systems on the datasets P1-P4 to show that [<span style="font-variant:small-caps;">Bengal</span>]{}can be easily ported to languages other than English. English Dataset features ------------------------ The first aim of our evaluation was to quantify the variability of the datasets B1–B13 generated by [<span style="font-variant:small-caps;">Bengal</span>]{}. To this end, we compared the distribution of the part of speech (POS) tags of the [<span style="font-variant:small-caps;">Bengal</span>]{}datasets with those of the 16 benchmark datasets. An analysis of the Pearson correlation of these distributions revealed that the manually created datasets (D1–D16) have a high correlation (0.88 on average) with a minimum of 0.61 (D10–D16). The correlation of the POS tag distributions between [<span style="font-variant:small-caps;">Bengal</span>]{}datasets and a manually created dataset vary between 0.34 (D7–B11) and 0.89 (D14–B9) with an average of 0.67. This shows that [<span style="font-variant:small-caps;">Bengal</span>]{}datasets can be generated to be similar to manually created datasets (D14–B9) as well as to be very different to them (D7–B11). Hence, [<span style="font-variant:small-caps;">Bengal</span>]{}can be used for testing sentence structures that are not common in the current manually generated benchmarks.[^8] We also studied the distribution of entities and tokens across the datasets in our evaluation. Table \[tab:features\] gives an overview of these distributions, where $E$ is the set of entities in the corpus $C$. The distribution of values for the different features is very diverse across the different manually created datasets. This is mainly due to (1) different ways to annotate entities and (2) the domains of the datasets (news, description of entities, microposts). As shown in Table \[tab:features\], [<span style="font-variant:small-caps;">Bengal</span>]{}can be easily configured to generate a wide variety of datasets with similar quality and number of documents to those of real datasets. This is mainly due to our approach being able to generate benchmarks ranging from (1) benchmarks with sentences containing a large number of entities without any filler terms (high entity density) to (2) benchmarks which contain more information pertaining to entity types and literals (low entity density). \[tab:features\] ID Name Doc. $|C|$ Tokens $|T|$ Entities $|E|$ $|T|/|C|$ $|E|/|C|$ $|E|/|T|$ ----- ------------------------------------------------------------------------ ------------ -------------- ---------------- ----------- ----------- ----------- D1 ACE2004 57 21312 306 373.9 5.4 0.01 D2 AIDA/CoNLL-Complete 1393 245008 34929 175.9 25.1 0.14 D8 IITB 104 66531 18308 639.7 176.0 0.28 D11 Microposts2014-Train 2340 40684 3822 17.4 1.6 0.09 D15 OKE 2015 Task 1 evaluation 101 3064 664 30.3 6.6 0.22 B1 [<span style="font-variant:small-caps;">Bengal</span>]{}Path 100 100 1202 362 12.02 3.6 0.30 B2 [<span style="font-variant:small-caps;">Bengal</span>]{}Path Para 100 100 1250 362 12.5 3.6 0.29 B3 [<span style="font-variant:small-caps;">Bengal</span>]{}Star 100 100 3039 880 30.39 8.8 0.29 B5 [<span style="font-variant:small-caps;">Bengal</span>]{}Sym 100 100 2718 725 27.18 7.25 0.26 B9 [<span style="font-variant:small-caps;">Bengal</span>]{}Summary 100 100 2033 637 20.33 6.37 0.31 B11 [<span style="font-variant:small-caps;">Bengal</span>]{}Hybrid 10000 10000 556483 165254 55.6 16.5 0.30 B12 [<span style="font-variant:small-caps;">Bengal</span>]{}Hybrid Long 10 10 9162 2417 241.7 916.2 0.26 B13 [<span style="font-variant:small-caps;">Bengal</span>]{}Star Long 10 10 7369 316 31.6 736.9 0.04 Annotator performance --------------------- We used GERBIL to evaluate the performance of 11 annotators on the manually created as well as the [<span style="font-variant:small-caps;">Bengal</span>]{}datasets. We evaluated the annotators within an A2KB (annotation to knowledge base) experiment setting: Each document of the corpora was sent to each annotator. The annotator had to find and link all entities to a reference KB (here DBpedia). We measured both the performance of the NER and the EL steps. Table \[tab:f1-scores\] shows the micro F1-score of the different annotators on chosen datasets. The manually created datasets showed diverse results. We analyzed the results further by using the F1-scores of the annotators as features of the datasets. Based on these feature vectors, we calculated the Pearson correlations between the datasets to identify datasets with similar characteristics.[^9] The Pearson correlations of the F-measures achieved by the different annotators on the AIDA/CoNLL datasets (D2–D5) are very high (0.95–1.00) while the correlation between the results on the Spotlight corpus (D7) and N3-Reuters-128 (D13) is around -0.62. The results on D1 and D12–D15 have a correlation to the AIDA/CoNLL results (D2–D5) that is higher than 0.5. In contrast, the correlations of D7 and D8 to the AIDA/CoNLL datasets range from -0.54 to -0.36. These correlations highlight the diversity of the manually created datasets and suggest that creating an approach which emulates all datasets is non-trivial. Like the correlations between the manually created datasets, the correlations between the results achieved on [<span style="font-variant:small-caps;">Bengal</span>]{}datasets and hand-crafted datasets vary. The results on [<span style="font-variant:small-caps;">Bengal</span>]{}correlate most with the results on the OKE 2015 data. The highest correlations were achieved with the OKE 2015 Task 1 dataset and range between 0.89 and 0.92. This suggests that our benchmark can emulate entity-centric benchmarks. The correlation of [<span style="font-variant:small-caps;">Bengal</span>]{}with OKE is however reduced to 0.82 in D13, suggesting that [<span style="font-variant:small-caps;">Bengal</span>]{}can be parametrized so as to diverge from such benchmarks. A similar observation can be made for the correlation D12 and ACE2004, where the correlation increased with the size of the documents in the benchmark. The correlation between the results across [<span style="font-variant:small-caps;">Bengal</span>]{}datasets varies between 0.54 and 1, which further supports that [<span style="font-variant:small-caps;">Bengal</span>]{}can generate a wide range of diverse datasets. Annotator Performance on Spanish and Brazilian Portuguese --------------------------------------------------------- We implemented [<span style="font-variant:small-caps;">Bengal</span>]{}for Brazilian Portuguese by using the RDF verbalizer presented in  and ran four multilingual and (MAG [@Moussallem2017], DBpedia Spotlight, Babelfy, and PBOH [@Ganea:2016:PBM:2872427.2882988]) frameworks thereon. We also evaluated the performance of these annotators on subsets of the HAREM datasets [@freitas2010second][^10]. We then extended this verbalizer to Spanish using the adaption of SimpleNLG to Spanish [@soto2017adapting]. We generated Spanish [<span style="font-variant:small-caps;">Bengal</span>]{}datasets and evaluated the aforementioned and systems on them. [^11] We also included VoxEL [@voxeliswc2018], a recent gold standard for Spanish. While the extension of [<span style="font-variant:small-caps;">Bengal</span>]{}to Portuguese is an important result in itself, our results also provide additional insights in the and performance of existing solutions. Our results suggest that existing solutions are mostly biased towards a high precision but often achieve a lower recall on this language. For example, both Spotlight’s and Babelfy’s recall remain below 0.6 in most cases while their precision goes up to 0.9. This clearly results from the lack of training data for these resource-poor languages. In contrast, the Spanish annotators presented low but consistent results, which confirms the lack of training data of these approaches on Spanish. Discussion and Conclusion {#sec:discussion} ========================= We presented and evaluated [<span style="font-variant:small-caps;">Bengal</span>]{}, an approach for the automatic generation of NER and EL benchmarks. Our results suggest that our approach can generate diverse benchmarks with characteristics similar to those of a large proportion of existing benchmarks in several languages. Overall, our results suggest that [<span style="font-variant:small-caps;">Bengal</span>]{}benchmarks can ease the development of NER and EL tools (especially for resource-poor languages) by providing developers with insights into their performance at virtually no cost. Hence, [<span style="font-variant:small-caps;">Bengal</span>]{}can improve the push towards better NER and EL frameworks. In future work, we plan to extend the ability of [<span style="font-variant:small-caps;">Bengal</span>]{}to generate longer and more complex sentences as well as the capability of generating different surface forms for a given entity by relying on referring expression models such as NeuralREG model [@thiagoacl2018]. We also intend to provide thorough evaluations of annotators across other resource-poor languages and create corresponding datasets to push the development of tools to process these languages. Acknowledgements {#acknowledgements .unnumbered} ================ This work has been supported by the H2020 project HOBBIT (GA no. 688227) as well as the BMVI projects LIMBO (project no. 19F2029C), OPAL (project no. 19F20284) and also supported by the German Federal Ministry of Education and Research (BMBF) within ’KMU-innovativ: Forschung für die zivile Sicherheit’ in particular ’Forschung für die zivile Sicherheit’ and the project SOLIDE (no. 13N14456). The authors gratefully acknowledge financial support from the German Federal Ministry of Education and Research within Eurostars, a joint programme of EUREKA and the European Community under the project E! 9367 DIESEL and E! 9725 QAMEL. [^1]: <http://stats.lod2.eu> [^2]: <http://dbpedia.org> [^3]: For a complete description of the vocabulary, see <http://nlp.stanford.edu/software/dependencies_manual.pdf>. [^4]: <http://aksw.org/Projects/Feasible> [^5]: Note that we do not consider blank nodes as they cannot be verbalized due to the existential quantification they stand for. [^6]: All [<span style="font-variant:small-caps;">Bengal</span>]{}datasets can be found at <https://hobbitdata.informatik.uni-leipzig.de/bengal/> [^7]: The scalability results are available at <https://goo.gl/9mnbwC> and cannot be presented herein due to space limitations. [^8]: Our complete results at <https://goo.gl/TBDxCa>. [^9]: All values are at <http://goo.gl/Mg3rE1>. [^10]: All Portuguese results at <http://faturl.com/bengalpt>. [^11]: All Spanish results at <http://faturl.com/bengales>.

ArXiv

--- abstract: 'The essential feature of weak measurements on quantum systems is the reduction of measurement back-action to negligible levels. To observe the non-classical features of weak measurements, it is therefore more important to avoid additional back-action errors than it is to avoid errors in the actual measurement outcome. In this paper, it is shown how an optical weak measurement of diagonal (PM) polarization can be realized by path interference between the horizontal (H) and vertical (V) polarization components of the input beam. The measurement strength can then be controlled by rotating the H and V polarizations towards each other. This well-controlled operation effectively generates the back-action without additional decoherence, while the visibility of the interference between the two beams only limits the measurement resolution. As the experimental results confirm, we can obtain extremely high weak values, even at rather low visibilities. Our method therefore provides a realization of weak measurements that is extremely robust against experimental imperfections.' address: | $^1$ Graduate school of Advanced Sciences of Matter, Hiroshima University\ 1-3-1 Kagamiyama, Higashi-Hiroshima, Japan, 739-8530,\ $^2$ JST,Crest, Sanbancho 5, Chiyoda-ku, Tokyo 102-0075, Japan author: - 'Masataka Iinuma$^1$, Yutaro Suzuki$^1$, Gen Taguchi$^1$, Yutaka Kadoya$^1$, and Holger F. Hofmann$^{1,2}$' title: 'Weak measurement of photon polarization by back-action induced path interference' --- submitted to [*New J. Phys*]{} Introduction ============ In ideal quantum measurements, there is a trade-off between the information obtained about the measured observable and the back-action suffered by observables that do not share any eigenstates with the measured observable. A fully resolved strong measurement has a maximal back-action since it completely removes any coherences between the eigenstates of the measured observable. On the other hand, a weak measurement with low resolution can have negligible back-action, leaving the coherences of the initial state almost completely intact. As first pointed out by Aharonov et al. [@Aharonov88], it is then possible to obtain measurement results far outside the spectrum of the eigenvalues of the measured observable by post-selecting a specific final measurement outcome. In the limit of negligible back-action, these post-selected results only depend on the initial state, the final state, and the operator of the measured observable. It is therefore possible to define the measurement result as the weak value of the measured observable for the specific combination of initial state and final state defined by state preparation and post-selection. It was soon realized that photon polarization was an ideal system for the experimental realization of weak measurements, since optics provides optimal control of coherence using well-established technologies [@Duck89; @Ritchie91]. At first, the implications and the usefulness of weak values were unclear. However, recent advances in quantum technologies have revived the interest in the unusual properties of weak values, with possible applications in precision measurements [@Hosten08; @Dixon09; @Bru10; @Hof10a], realizations using quantum logic gates [@Pryde05; @Ralph06], resolution of quantum paradoxes [@Resch04; @Wil08; @Lundeen09; @Yokota09; @Goggin09] and more fundamental implications for quantum statistics and quantum physics [@Tamate09; @Hofmann10; @Hosoya10]. Because of the wide range of problems that can be addressed by weak measurements, it seems to be desirable to develop simple and efficient technological implementations that are not too sensitive to experimental errors. In the following, we therefore present an experimental setup for the weak measurement of photon polarization that uses a basic two path interferometer as meter system. In such a setup, the essential problem is the limited visibility of the interferences between the two paths. If the two paths correspond to the eigenstates of the measured observable, the limited visibility causes an additional back-action on coherent superpositions of these eigenstates, limiting the magnitude of the weak values observed in the experiment. We therefore propose an alternative realization of weak measurements where the interference occurs between eigenstates of the back-action observable defined by the post-selection. As we discuss below, it is then possible to control the back-action precisely by implementing it separately in each path. The measurement effect is then obtained because this back-action induces an interference effect that depends on the quantum coherences between the back-action observable. The result is a conventional weak measurement (or variable strength measurement), but now the errors caused by finite visibility of the interference only reduce the measurement resolution, without causing any additional back-action. Our setup is therefore ideally suited for weak measurements in the presence of experimental imperfections. The rest of the paper is organized as follows. In sec. \[sec:setup\] we describe the principle of back-action induced interference and the experimental setup used to realize it. In sec. \[sec:results\], we present the experimental results obtained for the weak values of photon polarization and show that the errors are close to the theoretical limit for the measurement strength used in the experiment. In sec. \[sec:urel\], we present experimental results for the trade-off between resolution and back-action in our setup. The results show that the visibility only limits measurement resolution, without contributing to the back-action. Sec. \[sec:conclusion\] summarizes the results and concludes the paper. Quantum measurement by back-action induced interference {#sec:setup} ======================================================= In our experiment, we realize a measurement of photon polarization with variable measurement strength by making use of the fact that the diagonal polarization is determined by the phase coherence between the horizontal (H) and vertical (V) polarizations. The positive (P) and the negative (M) superpositions can therefore be distinguished by interference between the H and V components of the photon state. Although path interference between these two components cannot occur if the beams corresponding to the H and V polarizations can still be distinguished by their orthogonal polarization states, it is possible to induce a well-controlled amount of interference by erasing the HV information in the beams using a coherent rotation of polarization towards a common diagonal polarization. The increase of interference as the polarizations become less and less distinguishable corresponds to the trade-off between measurement information and back-action in the quantum measurement. Significantly, the final interference that results in a correlation between the output path and the PM polarization of the input state does not change the HV polarization at all, regardless of the visibility of the interference. The flips of HV polarization caused by the measurement back-action are therefore limited to the flips caused by the rotation of the polarization in the two arms of the interferometer. This method thus ensures optimal control of the back-action, permitting arbitrarily low back-action even in the presence of significant experimental errors. Figure \[fig:setup\] shows our setup in more detail. The photon path is split by a polarizing beam splitter (PBS) into a V polarized path a1 and an H polarized path a2. The polarizations are then rotated in opposite directions by half-wave-plates (HWP) mounted in each path. Specifically, the HWP in path a1 is rotated by an angle of $- \theta $ from the horizontal/vertical alignment, whereas the HWP in path a2 is rotated by an angle of $ + \theta $. Finally, the two polarization components interfere at a beam splitter (BS) with 50 % reflectivity for all polarizations, resulting in the output beams b1 and b2. A glass plate is used to compensate path length differences between path a1 and path a2, and a HWP is inserted in b2 to compensate for the phase shift caused by the difference in the number of reflections between the H and V components. The input photons were generated by using a CW titanium-sapphire laser(wavelength 830 nm, output power 300 mW) and passed through a Glan-Thompson prism to select photons with horizontal polarization. Neutral density(ND) filters were used to obtain intensities suitable for single photon counting with typical count rates around 100 kHz. The initial state of photon polarization was prepared by rotating the HWP upstream of the PBS. The number of output photons in path b1 and path b2 were counted using the single photon counting modules ( SPCM-AQR-14 ) D1 and D2, which were optically coupled to the paths b1 and b2 through fiber couplers and optical fibers. To keep track of fluctuations in the input light, the input intensity was monitored by detecting photons reflected by a pellicle beam splitter( reflectivity=8% ) with another single photon counting module D3. Experimentally, the ratio of counts in D3 to total counts in D1 and D2 was found to be 0.020. Post-selection was realized by inserting polarizers in the output beams to select only the desired output polarization in both paths. ![Setup of the polarization measurement using back-action induced interference. Interferences between the V polarized path a1 and the H polarized path a2 is induced using oppositely rotated HWP to reduce the angle between the polarization and therefore the distinguishability of the paths in terms of polarization. The output beams then distinguish positive and negative superpositions of H and V, corresponding to diagonal P and M polarization.[]{data-label="fig:setup"}](figure1.eps){width="100mm"} The weak measurement is realized by the interferometer setup between the PBS and the BS. If the polarizations in path a1 and a2 are orthogonal at the final beam splitter, no interference will be observed in the output probabilities of $b1$ and $b2$. The interference at the final beam splitter will simply restore the original superposition of H and V polarization of the input state. By rotating the polarization in both arms towards the same diagonal polarization P, the distinguishability of the two paths is reduced and the phase coherence between the HV components is converted into interferences between the paths. As a result, the probability of finding the photon in path b1 increases for positive superpositions of H and V (P polarization), and decreases for negative superpositions (M polarization). In the absence of experimental imperfections, the interference at the final beam splitter can be expressed in terms of the polarization vectors in paths a1 and a2 given in the $HV$-basis, $$\begin{aligned} \mid b1 \rangle & = & \frac{1}{\sqrt{2}} C_{\mathrm{H}} \left [ \begin{array}{l} \cos 2 \theta \\ \sin 2 \theta \end{array} \right ] + \frac{1}{\sqrt{2}} C_{\mathrm{V}} \left [ \begin{array}{l} \sin 2 \theta \\ \cos 2 \theta \end{array} \right ] \nonumber %%%%%%%% \\ %%%%%%%% \mid b2^\prime \rangle & = & \frac{1}{\sqrt{2}} C_{\mathrm{H}} \left [ \begin{array}{l} \cos 2 \theta \\ \sin 2 \theta \end{array} \right ] - \frac{1}{\sqrt{2}} C_{\mathrm{V}} \left [ \begin{array}{l} \sin 2 \theta \\ \cos 2 \theta \end{array} \right ] ,\end{aligned}$$ where $C_{\mathrm{H}}$ and $C_{\mathrm{V}}$ are the probability amplitudes of the H and V polarized components of the input state. For $\theta=0$, $\mid b1 \rangle$ reproduces the input polarization, while the polarization in $\mid b2^\prime \rangle$ is changed by a phase flip between the H and the V components. The HWP in b2 compensates this phase flip, resulting in the non-normalized output states $\mid b1 \rangle$ and $\mid b2 \rangle$ in the output beams of the measurement setup, $$\begin{aligned} \mid b1 \rangle = %\hspace{0.5cm} \frac{1}{\sqrt{2}} \left [ \begin{array}{rr} \cos 2 \theta & \: \: \: \: \sin 2 \theta \\ \: \: \: \: \sin 2 \theta & \cos 2 \theta \end{array} \right ] \left [ \begin{array}{c} C_{\mathrm{H}} \\ C_{\mathrm{V}} \end{array} \right ] & = & \hat{M}_{\mathrm{b1}} \, | \psi_{i} \rangle \nonumber %%%%%%% \\ %%%%%%% \mid b2 \rangle = \frac{1}{\sqrt{2}} \left [ \begin{array}{rr} \cos 2 \theta & -\sin 2 \theta \\ -\sin 2 \theta & \cos 2 \theta \end{array} \right ] \left [ \begin{array}{c} C_{\mathrm{H}} \\ C_{\mathrm{V}} \end{array} \right ] &=& \hat{M}_{\mathrm{b2}} \, | \psi_i \rangle, \label{eqn:b12}\end{aligned}$$ where $ \mid \psi_i \rangle $ is the input state defined by $C_{\mathrm{H}} $ and $C_{\mathrm{V}}$ and the measurement operators $\hat{M}_{\mathrm{b1}}$ and $\hat{M}_{\mathrm{b2}}$ represent the effects of the measurement described by their matrix representation in the $HV$-basis. It is easy to see that the eigenstates of the measurement operators are the positive and negative superpositions of $\mid H \rangle$ and $\mid V \rangle$, corresponding to the diagonal polarization states, $| P \rangle$ and $| M \rangle$. In terms of the Stokes parameter $\hat{S}_{\mathrm{PM}} = | P \rangle \langle P | - | M \rangle \langle M | $, the positive operator measure defining the probabilities of finding the photon in b1 or b2 is therefore given by $$\begin{aligned} \hat{M}_{\mathrm{b1}}^{\dagger} \hat{M}_{\mathrm{b1}} &=& \frac{1}{2} \left ( \hat{1} + \epsilon_{\mathrm{PM}} \, \hat{S}_{\mathrm{PM}} \right ) \nonumber \\ \hat{M}_{\mathrm{b2}}^{\dagger} \hat{M}_{\mathrm{b2}} &=& \frac{1}{2} \left ( \hat{1} - \epsilon_{\mathrm{PM}} \, \hat{S}_{\mathrm{PM}} \right ), \label{eqn:eb12}\end{aligned}$$ where $\epsilon_{\mathrm{PM}} = \sin 4 \theta $ determines the measurement resolution. Without post-selection, the difference between the output probabilities in b1 and b2 is directly related to the PM polarization of the input state, $$\begin{aligned} P(\mathrm{b1}) - P(\mathrm{b2}) & = & \langle \psi_{i} | {\hat{M}_{\mathrm{b1}}}^{\dagger} \hat{M}_{\mathrm{b1}} | \psi_{i} \rangle - \langle \psi_{i} | {\hat{M}_{\mathrm{b2}}}^{\dagger} \hat{M}_{\mathrm{b2}} | \psi_{i} \rangle \nonumber \\ & = & \epsilon_{\mathrm{PM}} \, \langle \psi_{i} | \hat{S}_{\mathrm{PM}} | \psi_{i} \rangle. \label{eqn:nops}\end{aligned}$$ Experimentally, it is therefore possible to determine the polarization of the input light by dividing the difference in output probability by a constant value $\epsilon_{\mathrm{PM}}$, where the proper value of $\epsilon_{\mathrm{PM}}$ can be determined from the visibility obtained for maximally polarized inputs. In the case of output post-selection, the difference of the conditional output probabilities can now be interpreted as a conditional measurement of PM polarization. Experimentally, the conditional value $ \langle \hat{S}_{\mathrm{PM}} \rangle_{\mathrm{exp}} (m_f)$ obtained by post-selecting an output polarization state $\mid m_f \rangle$ is determined from the output probabilities by $$\langle \hat{S}_{\mathrm{PM}} \rangle_{\mathrm{exp}} (m_f) = \frac{1}{\epsilon_{\mathrm{PM}}} \left ( P(\mathrm{b1} | \mathrm{m}_{f}) - P(\mathrm{b2} | \mathrm{m}_{f}) \right ). \label{eqn:def-con}$$ In the limit of negligible back-action ($\epsilon_{\mathrm{PM}} \to 0$), this experimental value is equal to the theoretically predicted weak value, $$\langle \hat{S}_{\mathrm{PM}} \rangle_{\mathrm{weak}} = \mathrm{Re} \left [ \frac{\langle \mathrm{m}_{f} | \hat{S}_{\mathrm{PM}} | \psi_{i} \rangle} {\langle \mathrm{m}_{f} | \psi_{i} \rangle} \right ].$$ However, the finite measurement back-action for non-zero measurement resolutions $\epsilon_{\mathrm{PM}}$ modifies this result even in the case of an ideal measurement. Using Eq. (\[eqn:def-con\]) to determine the conditional probabilities, the experimental value expected at finite back-action is $$\begin{aligned} \langle \hat{S}_{\mathrm{PM}} \rangle_{\mathrm{exp}} (m_f) & = & \frac{1}{\epsilon_{\mathrm{PM}}} \frac{| \langle \mathrm{m}_{f} | \hat{M}_{\mathrm{b1}} | \psi_{i} \rangle |^{2} - | \langle \mathrm{m}_{f} | \hat{M}_{\mathrm{b2}} | \psi_{i} \rangle |^{2}} {| \langle \mathrm{m}_{f} | \hat{M}_{\mathrm{b1}} | \psi_{i} \rangle |^{2} + | \langle \mathrm{m}_{f} | \hat{M}_{\mathrm{b2}} | \psi_{i} \rangle |^{2}} \nonumber \\ & = & \frac{| \langle \mathrm{m}_{f} | \psi_{i} \rangle |^{2}}{| \langle \mathrm{m}_{f} | \psi_{i} \rangle |^{2}+\eta_{\mathrm{HV}} \Delta_{\mathrm{flip}}} \langle \hat{S}_{\mathrm{PM}} \rangle_{\mathrm{weak}} \label{eqn:ps}\end{aligned}$$ where $\eta_{\mathrm{HV}}=\sin^2(2\theta)$ is equal to the transition probability between H and V polarization given by the measurement operators $\hat{M}_{\mathrm{b1}}$ and $\hat{M}_{\mathrm{b2}}$, and $\Delta_{\mathrm{flip}}$ is the change in the post-selection probability caused by a polarization flip described by the operator $\hat{S}_{\mathrm{PM}}$, given by $$\Delta_{\mathrm{flip}} = | \langle \mathrm{m}_{f} | \hat{S}_{\mathrm{PM}} | \psi_{i} \rangle |^{2} - | \langle \mathrm{m}_{f} | \psi_{i} \rangle |^{2}.$$ As shown in Eq. (\[eqn:ps\]), the experimental value is approximately equal to the weak value if the back-action induced change in the post-selection probability given by $\eta_{\mathrm{HV}} \Delta_{\mathrm{flip}}$ is sufficiently smaller than the original post-selection probability of $| \langle \mathrm{m}_{f} | \psi_{i} \rangle |^{2}$. However, extremely large weak values can only be obtained when the original post-selection probability goes to zero. To achieve extremely enhanced experimental weak values, it is therefore essential to keep the transition probability $\eta_{\mathrm{HV}}$ as small as possible. In particular, it is necessary to avoid additional errors from dephasing between the P and M polarized components. In our setup, we achieve extremely small values of $\eta_{\mathrm{HV}}$ by limiting the use of path interferences to an interference between a path associated with the initial H polarization and a path associated with the initial V polarization, therefore avoiding the HV transitions that would be caused by finite visibility interferences between the P and M polarized eigenstates of the measurement operators. As a result, our setup enables us to measure extremely high weak values, even at low visibilities of the path interference. Experimental demonstration of the weak measurement {#sec:results} ================================================== For the experimental demonstration of the weak measurement, we chose a variable input state given by $C_{\mathrm{H}}=\sin \phi$ and $C_{\mathrm{V}}=\cos \phi$. Post-selection was implemented by inserting polarization filters selecting only the H polarized output components between the output ports and the detectors. Ideally, we should then be able to observe a theoretical weak value of $\langle \hat{S}_{\mathrm{PM}}\rangle_{\mathrm{weak}}=1/\tan \phi$. However, the measurement back-action modifies the directly determined experimental values to $$\langle \hat{S}_{\mathrm{PM}} \rangle_{\mathrm{exp}} = \frac{\sin\phi \cos\phi}{\sin^2 \phi + \eta_{\mathrm{HV}} (\cos^2\phi-\sin^2 \phi)}, \label{eqn:exp-wv}$$ where $\eta_{\mathrm{HV}}$ is the transition probability between H and V polarizations, including both the uncertainty limited back-action and additional effects of experimental imperfections in the setup. As discussed in the previous section, the back-action effects summarized by $\eta_{\mathrm{HV}}$ limit the magnitude of the experimental weak values that can be observed experimentally. For small $\eta_{\mathrm{HV}}$, the maximal value is $\langle \hat{S}_{\mathrm{PM}} \rangle_{\mathrm{exp}}=1/\sqrt{4 \eta_{\mathrm{HV}}}$, obtained at an input polarization angle of $\phi=\sqrt{\eta_{\mathrm{HV}}}$. ![Experimental results of the weak measurement. The experimental weak values $\langle \hat{S}_{\mathrm{PM}} \rangle_{\mathrm{exp}}$ are shown as a function of initial polarization angle $\phi$. The open circles indicate the experimental data obtained from the conditional probabilities of the post-selected results, the broken line shows the expected effects of back-action given by Eq.(\[eqn:exp-wv\]), and the solid line shows the theoretical weak value.[]{data-label="fig:result1"}](figure2.eps){width="100mm"} Fig. \[fig:result1\] shows the experimental results for the weak values obtained with the measurement resolution obtained by setting the HWPs to $\theta = 0.5^\circ$. Theoretically, this corresponds to a measurement resolution of $\epsilon_{\mathrm{PM}}=0.035$ and a back-action related transition probability of $\eta_{\mathrm{HV}}=0.0003$. The experimental results are in good agreement with the theoretical weak values up to and including the measurement values obtained at $\phi = \pm 4^\circ$. The three measurement values obtained close to $\phi=0$ are consistent with the theoretical prediction for the back-action effects given by Eq.(\[eqn:exp-wv\]) for $\eta_{\mathrm{HV}}=0.0003$. This correspondence suggests that almost all the flips in HV polarization are caused by the rotation of the HWPs in the measurement setup, with only negligible contributions from additional error sources. The extremal weak values observed in the experiment were found at $\pm 20$. Since the input angles for these values are at $\phi=\pm 2^\circ$, this is lower than the maximal value of $\pm 28.6$ theoretically predicted for angles of about $\phi=\pm 1^\circ$. However, even the achievement of a 20 fold enhancement of the weak value requires a transition probability below $0.0006$. If the weak measurement was realized by a separation of the P and M polarizations followed by an interference between the paths to partially erase the measurement information, the visibility of the interference needed to obtain a 20 fold enhancement of the weak value would have to be as high as $\mathrm{V_{PM}}=1-2\eta_{\mathrm{HV}}=0.9988$. It is therefore essential that our setup only uses interferences between the H and V polarized paths, avoiding the errors that would be introduced by limited visibilities in path interferences between P and M polarization. In our setup, the visibility of the path interference between the H and V polarized components was found to be $\mathrm{V_{HV}}=0.71$. The effects of this error reduce the measurement resolution by introducing transitions between the P and M polarizations. As a result, the measurement resolution is $\epsilon_{\mathrm{PM}}=0.025$ instead of the ideal value of 0.035 predicted from $\theta$. However, this reduced resolution has no effects on the observation of weak values at low $\theta$, since weak values are always obtained from averages over a sufficiently high number of low resolution measurements. The experimental results thus confirm the main merit of our method for the realization of weak measurements. Oppositely, the method is not as suitable for strong measurements, where back-action is always maximal and an optimization of measurement resolution is desirable. Since we can vary the measurement strength of our setup continuously between weak and strong measurements, we can illustrate the performance of our setup in these very different operating regimes in terms of the experimental errors observed in measurement resolution and back-action as the measurement strength is varied by rotation the HWPs from $\theta=0^\circ$ to $\theta=22.5^\circ$. Relation between measurement resolution and back-action {#sec:urel} ======================================================= In principle, the measurement resolution $\epsilon_{\mathrm{PM}}$ and the measurement back-action given by $\eta_{\mathrm{HV}}$ should be defined in terms of the experimental input-output relations of the measurement setup. For a specific input state with a PM polarization of $\langle \hat{S}_{\mathrm{PM}}\rangle=2 \mathrm{Re} [C_{\mathrm{H}}^* C_{\mathrm{V}}^*]$, the measurement resolution is given by the ratio between the output probability difference and the expectation value of the Stokes parameter in the input, $$\epsilon_{\mathrm{PM}}=\frac{P(\mathrm{b1})-P(\mathrm{b2})}{2 \mathrm{Re} [C_{\mathrm{H}}^* C_{\mathrm{V}}^*]},$$ where $P(\mathrm{b1})$ and $P(\mathrm{b2})$ are obtained from the total number of counts in $b1$ and $b2$. Likewise, the measurement back-action flips H and V polarization, reducing the input HV polarization of $\langle \hat{S}_{\mathrm{HV}}\rangle=|C_{\mathrm{H}}|^2-|C_{\mathrm{V}}|^2$ by a factor of $1-2 \eta_{\mathrm{HV}}$. If a measurement of HV polarization is performed in the output, the experimental measurement back-action is obtained from $$%2 \eta_{\mathrm{HV}}=1-\frac{P(\mathrm{H})-P(\mathrm{V})}{|C_{\mathrm{H}}|^2-|C_{\mathrm{V}}|^2}, 1 - 2 \eta_{\mathrm{HV}}=\frac{P(\mathrm{H})-P(\mathrm{V})}{|C_{\mathrm{H}}|^2-|C_{\mathrm{V}}|^2},$$ where $P(\mathrm{\mathrm{H}})$ and $P(\mathrm{V})$ are the total H and V polarized counts summed over both $b1$ and $b2$. For consistency, it is convenient to define the measurement back-action as $2 \eta_{\mathrm{HV}}$, since a complete randomization of HV polarization ($P(H)=P(V)$) then corresponds to a back-action of $1$. In the absence of experimental errors, our measurement setup would have a measurement resolution of $ \epsilon_{\mathrm{PM}} = \sin 4 \theta $ and a back-action given by $1-2 \eta_{\mathrm{HV}} = \cos 4 \theta$, depending on the angles $\theta$ of the HWPs. This result achieves the uncertainty limit for resolution and measurement back-action in two level systems [@Englert96], as given by the uncertainty relation $$\epsilon_{\mathrm{PM}}^2 + (1-2 \eta_{\mathrm{HV}})^2 \leq 1. \label{eqn:uncertainty}$$ In the actual experiment, linear decoherence effects reduce the values of $\epsilon_{\mathrm{PM}}$ and $1-2\eta_{\mathrm{HV}}$ from their ideal values to values below the uncertainty limit. If these reductions are expressed in terms of experimental visibilities, $ \epsilon_{\mathrm{PM}} = \mathrm{V_{HV}} \sin 4 \theta $ and $1-2 \eta_{\mathrm{HV}} = \mathrm{V_{PM}} \cos 4 \theta $, the actual relation between back-action and resolution can be described by $$\frac{\epsilon_{\mathrm{PM}}^2}{\mathrm{V_{HV}}^2} + \frac{(1-2\eta_{\mathrm{HV}})^2}{\mathrm{V_{PM}}^2} = 1. \label{eqn:relation}$$ If the values obtained for $\epsilon_{\mathrm{PM}}$ and $2 \eta_{\mathrm{HV}}$ are shown for different measurement strengths $\theta$, they should therefore lie on an ellipse around ($\epsilon_{\mathrm{PM}}=0, \, 2 \eta_{\mathrm{HV}}=1$), where $\mathrm{V_{HV}}$ determines the resolution in the strong measurement limit at $\theta=22.5^\circ$, and $\mathrm{V_{PM}}$ determines the back-action in the weak measurement limit at $\theta=0^\circ$. ![Relation between measurement back-action $\eta_{\mathrm{HV}}$ and measurement resolution $\epsilon_{\mathrm{PM}}$. Open circles indicate experimental values obtained for different measurement strengths $\theta$. The broken line shows the relation expected for a visibility of $0.7$ in our setup. The solid line indicates the uncertainty limit given by Eq.(\[eqn:uncertainty\]).[]{data-label="fig:result2"}](figure3.eps){width="100mm"} Fig. \[fig:result2\] shows the experimental results obtained with an input state at $\phi = 25^\circ$. The results reproduce the relation between resolution and back-action expected for $\mathrm{V_{HV}} = 0.7$ and $\mathrm{V_{PM}} = 1$ as shown by the broken line in the graph, except for some discrepancy in the values obtained in the strong measurement limit. Since the strong measurement limit is very sensitive to the visibilities of our interferometer, it is possible that these discrepancies may have been caused by instabilities in the interferometer. In general, the result is consistent with the values of $\mathrm{V_{HV}}=0.71$ and $ \mathrm{V_{PM}}>0.9988$ estimated from the weak measurement results. Since there is no experimentally resolvable limitation to the reduction of measurement back-action at low $\theta$, the experimentally obtained relation between resolution and back-action confirms that our setup is particularly suited for weak measurements. Fig. \[fig:result2\] thus illustrates the specific feature of our measurement setup in terms of the noise characteristics at different measurement strengths. Conclusions {#sec:conclusion} =========== We realized a weak measurement of diagonal (PM) photon polarization by path interference between the H and V polarized components. In this case, the visibility of the path interference depends on the amount of back-action induced by gradually rotating the orthogonal polarizations of the paths towards each other. It is therefore possible to control the amount of back-action precisely, while errors caused by the limited visibility of the path interference only affect the measurement resolution. This situation is ideal for the realization of weak measurements, since the achievement of extremely high weak values depends critically on the limitation of the total back-action to error rates below the post-selection probability. Our results show that we can achieve 20 fold enhancement of the weak values, even though the visibility of the path interference was only $0.71$. This robustness against experimental errors can be achieved because the measurement resolution is not relevant for the measurement of weak values. The requirements for operating in the weak measurement regime are therefore quite different from the requirements for operating in the strong measurement regime. We have characterized this difference in the experimental requirements by measuring the resolution and back-action of our setup at different measurement strengths. The present setup achieves the uncertainty limit in the weak measurement regime, but not in the strong measurement regime, where its measurement resolution is limited by the visibility of path interference. The characterization of errors for different measurement strengths thus confirms the specific usefulness of our approach for weak measurements. The setup presented here is easy to realize and allows the observation of extreme weak values even in the presence of significant experimental imperfections. It may therefore be useful in the characterization and control of quantum processes by weak measurements. In particular, it may greatly simplify the integration of weak measurements into optical quantum circuits, and the performance of multiple weak measurements in a cascaded system. We hope that these simplifications will help to establish weak measurements as part of the quantum information toolbox, leading to better insights into the fundamental properties of emerging quantum technologies. Acknowledgements {#acknowledgements .unnumbered} ================ Part of this work has been supported by the Grant-in-Aid program of the Japanese Society for the Promotion of Science, JSPS. Reference {#reference .unnumbered} ========= [99]{} Aharonov Y, Albert D Z and Vaidman L 1988 [*Phys. Rev. Lett.*]{} [**60**]{} 1351–1354 Duck I M, Stevenson P M and Sudarshan E C G 1989 [*Phys. Rev. D*]{} [**40**]{} 2112–2117 Ritchie N W, Story J G and Hulet R G 1991 [*Phys. Rev. Lett.*]{} [**66**]{} 1107 Hosten O and Kwiat P 2008 [*Science*]{} [**319**]{} 787 Dixon P B, Starling D J, Jordan A N, and Howell J C 2009 [*Phys. Rev. Lett.*]{} [**102**]{} 173601 Brunner N and Simon C 2010 [*Phys. Rev. Lett.*]{} [**105**]{} 010405 Hofmann F H 2010 [*e-print*]{} arXiv quant-ph/1005.0654V1 Pryde G J, O’Brien J L, White A G, Ralph T C and Wiseman H M 2005 [*Phys. Rev. Lett.*]{} [**94**]{} 220405 Ralph T C, Bartlett S D, O’Brien J L, Pryde G J and Wiseman H M 2006 [*Phys. Rev. A*]{}, [**73**]{} 012113 Resch K J, Lundeen J S and Steinberg A M 2004 [*Phys. Lett. A*]{} [**324**]{} 125 Williams N S and Jordan A N 2008 [*Phys. Rev. Lett.*]{} [**100**]{} 026804 Lundeen J S and Steinberg A M 2009 [*Phys. Rev. Lett.*]{} [**102**]{} 020404 Yokota K, Yamamoto T, Koashi M, and Imoto N 2009 [*New J. Phys.*]{} [**11**]{} 033011 Goggin M E, Almeida M P, Barbieri M, Lanyon B P, O’Brien J L, White A G and Pryde G J 2009 [*e-print*]{} arXiv quant-ph/0907.1679V1 Tamate S, Kobayashi H, Nakanishi T, Sugiyama K, and Kitano M 2009 [*New J. Phys.*]{} [**11**]{} 093025 Hofmann F H 2010 [*Phys. Rev. A*]{} [**81**]{} 012103 Hosoya A and Shikano Y 2010 [*J. Phys. A: Math. Theor.*]{} [**43**]{} 385307 Englert E-G 1996 [*Phys. Rev. Lett.*]{} [**77**]{} 2154–2157

ArXiv

--- abstract: 'Scaling up the convolutional neural network (CNN) size (e.g., width, depth, etc.) is known to effectively improve model accuracy. However, the large model size impedes *training* on resource-constrained edge devices. For instance, federated learning (FL) on edge devices cannot tackle large CNN training demands, even though there is a strong practical need for FL due to its privacy and confidentiality properties. To address the resource-constrained reality, we reformulate FL as a group knowledge transfer (GKT) training algorithm. GKT designs a variant of the alternating minimization approach to train small CNNs on edge nodes and periodically transfer their knowledge by knowledge distillation to a large server-side CNN. GKT consolidates several advantages in a single framework: reduced demand for edge computation, lower communication cost for large CNNs, and asynchronous training, all while maintaining model accuracy comparable to FL. To simplify the edge training, we also develop a distributed training system based on our GKT. We train CNNs designed based on ResNet-56 and ResNet-110 using three distinct datasets (CIFAR-10, CIFAR-100, and CINIC-10) and their non-IID variants. Our results show that GKT can obtain comparable or even slightly higher accuracy. More importantly, GKT makes edge training affordable. Compared to the edge training using FedAvg, GKT demands 9 to 17 times less computational power (FLOPs) on edge devices and requires 54 to 105 times fewer parameters in the edge CNN.' author: - | Chaoyang HeMurali AnnavaramSalman Avestimehr\ University of Southern California\ Los Angeles, CA 90007\ `chaoyang.he@usc.eduannavara@usc.eduavestime@usc.edu`\ bibliography: - 'references.bib' title: | Group Knowledge Transfer:\ Collaborative Training of Large CNNs on the Edge --- Introduction ============ The size of convolutional neural networks (CNN) matters. As seen in both manually designed neural architectures (ResNet [@he2016deep]) and automated architectures discovered by neural architecture search (DARTS [@liu2018darts], MiLeNAS [@he2020milenas], EfficientNets [@tan2019efficientnet]), scaling up CNN size (e.g., width, depth, etc.) is known to be an effective approach for improving model accuracy. Unfortunately, training large CNNs is challenging for resource-constrained edge devices (e.g., smartphones, IoT devices, and edge servers). However, the demand for edge-based training is increasing as evinced by a recent upsurge of research interest in Federated Learning (FL) [@kairouz2019advances]. FL is a distributed learning paradigm that can collaboratively train a common model for many edge devices without centralizing any device’s dataset [@mcmahan2016communication; @he2019central; @wang2020federated]. FL can boost model accuracy through collaborative training on private data in situations when a single organization or user does not have sufficient or relevant data. Consequently, many FL services have been deployed (e.g., Google has improved the accuracy of item ranking and language models on Android smartphones by using FL [@bonawitz2019towards]). FL is also a promising solution when data centralization is undesirable or infeasible due to privacy and regulatory constraints [@kairouz2019advances]. *However*, one significant impediment in edge training is the gap between the computational demand of large CNNs and the meager computational power on edge devices. FL approaches (FedAvg [@mcmahan2016communication], FedProx [@li2018federated], and FedMA [@wang2020federated]) can reduce communication frequency by local SDG and model averaging [@yu2019parallel], but they only evaluate the convergence property on small CNNs, or assume the client has enough computational power and run large CNNs with the GPU accelerator, which is unrealistic in real-world system. To tackle the computational limitation of edges, model parallelism-based split learning (SL) [@gupta2018distributed; @vepakomma2018split] partitions a large model and offloads the larger portion of the neural architecture to the cloud. Nevertheless, the obvious drawback of SL is that it has a severe straggler problem because of its highly frequent synchronization mechanism in which a single mini-batch iteration requires multiple rounds of communication between the server and edges. In this paper, we propose Group Knowledge Transfer (GKT), an efficient collaborative training framework for resource-constrained edge devices. GKT aims to incorporate benefits from both FedAvg [@mcmahan2016communication] and SL [@gupta2018distributed; @vepakomma2018split] by training locally as in FL but also computing efficiently as in SL. GKT can transfer knowledge from many compact CNNs to a large CNN. The essence of GKT is that it reformulates FL as an alternating minimization (AM) approach [@ortega1970iterative; @bertsekas1989parallel; @bolte2014proximal; @attouch2010proximal; @wright2015coordinate; @razaviyayn2013unified], which optimizes two random variables (the edge model and the server model) by alternatively fixing one and optimizing another. Under this reformulation, GKT further develops a new knowledge distillation (KD) paradigm, group knowledge transfer, to boost the performance of the server model. As shown in Fig. \[fig:overview\_left\], the compact CNN contains a lightweight feature extractor and classifier inside and is thus efficiently trainable on edge devices using their private data (*1 - local training*). After local training, all the edge nodes agree to generate *exactly* the same tensor dimensions as an output from the feature extractor. The larger server model is trained by taking features extracted from the edge-side model and also computing a KD-based loss function using the ground truth and soft label (the probabilistic prediction in KD [@hinton2015distilling; @bucilu2006model; @ba2014deep; @romero2014fitnets]) predicted from the edge-side model (*2 - periodic transfer*). To boost the edge model’s performance, the server sends its predicted soft labels to the edge, then the edge also trains its local dataset with a KD-based loss function using server-side soft labels (*3 - transfer back*). Thus, the server’s performance is essentially boosted by knowledge transferred from the edge models and vice-versa. Once the training is complete after many rounds of communication, each edge’s final model is a combination of its local feature extractor and shared server model (*4 - edge-sided model*). The primary trade-off is that GKT shifts the computing burden from edge devices to server-side since the large CNN must be trained on the GPU-equipped server. GKT unifies multiple advantages into a single framework: 1. GKT is computationally and memory-efficient, similar to SL; 2. GKT can train in a local SGD manner like FL to reduce the communication frequency; 3. Exchanging hidden features as in SL, as opposed to exchanging the entire model as in FL, reduces the communication cost. 4. GKT naturally supports asynchronous training, which circumvents the severe synchronization issue in SL. As shown in Fig. \[fig:overview\_left\], the server model can immediately start training when it receives inputs from any client. We develop a distributed training system for GKT and comprehensively evaluate GKT using edge and server CNNs designed based on ResNet  [@he2016deep] (as shown in Fig. \[fig:overview\_right\]). We train on three datasets with varying training difficulties (CIFAR-10 [@krizhevsky2009learning], CIFAR-100 [@krizhevsky2009learning], and CINIC-10 [@darlow2018cinic]) and their non-IID (non identical and independent distribution) variants. As for the model accuracy, our results on both IID and non-IID datasets show that GKT can obtain accuracy comparable to FedAvg [@mcmahan2016communication]. More importantly, GKT makes edge training affordable. Compared to FedAvg, GKT demands 9 to 17 times less computational power (FLOPs) on edge devices and requires 54 to 105 times fewer parameters in the edge CNN. To understand GKT comprehensively, asynchronous training and ablation studies are also performed. Related Works ============= **Federated Learning**. Existing FL methods such as FedAvg [@mcmahan2016communication], FedProx [@li2018federated], and FedMA [@wang2020federated] face significant hurdles in training large CNNs on resource-constrained devices. Recent works [@hsieh2019non] [@reddi2020adaptive] analyze the performance of FedAvg in large CNNs, but they rely on GPU training to complete the evaluations. Others [@bernstein2018signsgd; @wangni2018gradient; @tang2018communication; @alistarh2017qsgd; @lin2017deep; @wang2018atomo] optimize the communication cost without considering edge computational limitations. Model parallelism-based split learning [@gupta2018distributed; @vepakomma2018split] attempts to break the computational constraint, but it requires frequent communication with the server. **Knowledge Distillation (KD)**. We use KD [@hinton2015distilling] in a different way from existing works. Previous works only consider transferring knowledge from a large network to a smaller one [@hinton2015distilling; @bucilu2006model; @ba2014deep; @romero2014fitnets], or they transfer knowledge from a group, but each member in the group shares the same large model architecture or a large portion of the neural architecture with specific tail or head layers [@zhang2018deep; @anil2018large; @song2018collaborative; @jeong2018communication; @chen2019online; @park2019distilling]. Moreover, all teachers and students in distillation share the same dataset [@chen2019online; @tran2020hydra; @zhu2018knowledge; @vongkulbhisal2019unifying], while in our setting each member (client) can only access its own independent dataset. Previous methods use centralized training, but we utilize an alternating training method. **Efficient On-device Deep Learning**. Our work also relates to efficient deep learning on edge devices, such as model compression [@han2015deep; @he2018amc; @yang2018netadapt], manually designed architectures (MobileNets [@howard2017mobilenets], ShuffeNets [@zhang2018shufflenet], SqueezeNets [@iandola2016squeezenet]), or even efficient neural architecture search (EfficientNets [@tan2019efficientnet], FBNet [@wu2019fbnet]). However, all of these techniques are tailored for the inference phase rather than the training phase. Group Knowledge Transfer ======================== ![Reformulation of FL: An Alternating Minimization Perspective[]{data-label="fig:am"}](figures/NN.pdf){width="\textwidth"} Preliminary ----------- We aim to collaboratively train large convolutional neural networks (e.g., ResNet) on many resource-constrained devices that are not equipped with GPU accelerators without centralizing each device’s dataset to the server side. We specifically consider supervised learning with $C$ categories in the entire dataset $\mathcal{D}$. We assume that there are $K$ clients (edge devices) in the network. The $k$th node has its own dataset $\mathcal{D}^{k}:=\left\{\left(\boldsymbol{X}_{i}^{k}, y_{i}\right)\right\}_{i=1}^{N^{(k)}}$, where $\boldsymbol{X}_{i}$ is the $i$th training sample, $y_{i}$ is the corresponding label of $\boldsymbol{X}_{i}$, $y_i \in\{1,2, \ldots, C\}$ (a multi-classification learning task), and $N^{(k)}$ is the sample number in dataset $\mathcal{D}^{k}$. $\mathcal{D} = {\{\mathcal{D}_{1}, \mathcal{D}_{2}, ..., \mathcal{D}_{k}\}}$, $N = \sum_{k=1}^{K} N^{(k)}$. FedAvg [@mcmahan2016communication] formulates the objective function of DNN-based federated learning as: $$\small{\min_{\boldsymbol{W}} F(\boldsymbol{W}) \stackrel{\text { def }}{=} \min_{\boldsymbol{W}} \sum_{k=1}^{K} \frac{N^{(k)}}{N} \cdot f^{(k)}(\boldsymbol{W}), \text{where } \space f^{(k)}(\boldsymbol{W}) = \frac{1}{N^{(k)}} \sum_{i=1}^{N^{(k)}} \ell(\boldsymbol{W}; \boldsymbol{X}_{i}, y_{i}) \label{eq:FL}}$$ where $\boldsymbol{W}$ represents the network weight of a global CNN in each client. $f^{(k)}(\boldsymbol{W})$ is the $k$th client’s local objective function that measures the local empirical risk over the heterogeneous dataset $\mathcal{D}^k$. $\ell$ is the loss function of the global CNN model. The objective function is viewed as a distributed optimization problem and can also be solved using off-the-shelf algorithms (e.g., FedProx [@li2018federated] and FedMA [@wang2020federated]). These methods utilize local SDG and model averaging [@yu2019parallel] for communication-efficient training. However, as shown in Fig. \[fig:am\](a), the main drawback of FL is that existing methods cannot address the resource-constrained nature of edge devices. Model parallelism-based split learning [@gupta2018distributed; @vepakomma2018split], as shown in Fig. \[fig:am\](b), attempts to break the computational constraint by splitting $\boldsymbol{W}$ into two portions and offloading the larger portion into the server-side, but a single mini-batch iteration requires remote forward propagation and backpropagation. For edge computing, such a highly frequent synchronization mechanism may lead to the severe straggler problem that significantly slows down the training process. Reformulation ------------- **Non-convex Optimization**. To solve the resource-constrained problem in existing FL, we reconsider another methodology to solve the FL optimization problem. As illustrated in Fig. \[fig:am\](c), we divide the global CNN $\boldsymbol{W}$ in Eq. into two partitions: a small feature extractor model $\boldsymbol{W}_e$ and a large-scale server-side model $\boldsymbol{W}_s$, and put them on the edge and the server, respectively. We also add a classifier $\boldsymbol{W}_c$ for $\boldsymbol{W}_e$ to create a small but fully trainable model on the edge. Consequently, we reformulate a single global model optimization into an non-convex optimization problem that requires us to solve the edge model and the server model simultaneously. Our reformulation is as follows: [$$\begin{aligned} \underset{\boldsymbol{W}_s} {\operatorname{argmin}} F_s(\boldsymbol{W}_s, \boldsymbol{W}_e^*) &= \underset{\boldsymbol{W}_s} {\operatorname{argmin}} \sum_{k=1}^{K} \sum_{i=1}^{N^{(k)}} \ell_{s}\left( f_s(\boldsymbol{W}_s; \boldsymbol{H}_i^{(k)}), y_i^{(k)} \right) \label{eq:server_am} \\ \textit{subject to:}~\boldsymbol{H}_i^{(k)} &= f_e^{(k)}(\boldsymbol{W}_{e}^{(k)}; \boldsymbol{X}_i^{(k)})\label{eq:hiden_features}\\ \underset{(\boldsymbol{W}_{e}^{(k)}, \boldsymbol{W}_{c}^{(k)})} {\operatorname{argmin}} F_c(\boldsymbol{W}_{e}^{(k)}, \boldsymbol{W}_{c}^{(k)}) & = \underset{(\boldsymbol{W}_{e}^{(k)}, \boldsymbol{W}_{c}^{(k)})} {\operatorname{argmin}} \sum_{i=1}^{N^{(k)}} \ell_{c}\left(f^{(k)}\big((\boldsymbol{W}_{e}^{(k)}, \boldsymbol{W}_{c}^{(k)}); \boldsymbol{X}_i^{(k)}\big), y_i^{(k)} \right) \\ &\small{ = \underset{(\boldsymbol{W}_{e}^{(k)}, \boldsymbol{W}_{c}^{(k)})} {\operatorname{argmin}} \sum_{i=1}^{N^{(k)}} \ell_{c} \big(f_c^{(k)}(\boldsymbol{W}_{c}^{(k)}; \underbrace{f_e^{(k)}(\boldsymbol{W}_{e}^{(k)}; \boldsymbol{X}_i^{(k)}}_{\boldsymbol{H}_i^{(k)}})), y_i^{(k)}\big)} \label{eq:client_am}\end{aligned}$$]{} Where $\ell_s$ and $\ell_c$ are general loss functions for the server model and the edge model, respectively. $f_s$ is the server model, and $f^{(k)}$ is the edge-side model which consists of feature extractor $f_e^{(k)}$ followed by a classifier $f_c^{(k)}$. $\boldsymbol{W}_s$, $\boldsymbol{W}_e^{(k)}$, $\boldsymbol{W}_c^{(k)}$ are the network weights of $f_s$, $f_e^{(k)}$, $f_c^{(k)}$, respectively. $\boldsymbol{H}^{(k)}$ is the feature map (a hidden vector or tensor) output by feature extractor $f_e^{(k)}$ (Eq. ). The $k$th client model $f^{(k)}$ is trained on its local dataset (Eq. ), while the server model $f_s$ is trained using $\boldsymbol{H}^{(k)}$ as input features (Eq. ). The final trained model architecture for client $k$ is stacked by the architecture of the feature extractor $f_e^{(k)}$ and the architecture of the server model $f_s$. **Advantages and Challenges**. The core advantage of the above reformulation is that when we assume the model size of $f^{(k)}$ is multiple orders of magnitude smaller than that of $f_s$, the edge training is affordable. Moreover, as discussed in [@gupta2018distributed; @vepakomma2018split], for large CNN training, the communication bandwidth for transferring $\boldsymbol{H}^{(k)}$ to the server is substantially less than communicating the whole model parameters as is done in traditional federated learning. Conversely, we also observe the difficulty of the reformulated optimization problem. First, each client is expected to adequately solve the inner optimization (Eq. ). Namely, each client must train its feature extractor $f_e^{(k)}$ well to accurately generate $\boldsymbol{H}^{(k)}$ for any given input image using Eq. . However, in the FL setting, the dataset on each edge device is small and thus may be inadequate in training a CNN-based feature extractor solely based on the local dataset. Second, the outer optimization Eq. and inter optimization Eq. are correlated. This correlation further makes the outer optimization difficult to converge if the individual client-side feature extractors are not trained adequately. Group Knowledge Transfer (GKT) ------------------------------ **Scaling Edge Dataset Limitations with Knowledge Transfer**. Given the above challenges, we incorporate knowledge distillation loss into the optimization equations to circumvent the optimization difficulty. The intuition is that knowledge transferred from the the server model can boost the optimization on the edge (Eq. ). As such, we propose to transfer group knowledge from edge-side CNNs to the server CNN. To be more specific, in Eq. and , we design $\ell_s$ and $\ell_c$ as follows. $$\begin{aligned} \ell_s = \ell_{CE} + \sum_{k=1}^{K} \ell_{KD} \left(\boldsymbol{z}_{s}, \boldsymbol{z}_{c}^{(k)} \right) = \ell_{CE} + \sum_{k=1}^{K} D_{K L}\left(\boldsymbol{p}_{k} \| \boldsymbol{p}_{s}\right) \label{eq:loss_server} \\ \ell_c^{(k)} = \ell_{CE} + \ell_{KD} \left(\boldsymbol{z}_{s}, \boldsymbol{z}_{c}^{(k)}\right) = \ell_{CE} + D_{KL} \left(\boldsymbol{p}_{s} \| \boldsymbol{p}_{k}\right) \label{eq:loss_client}\end{aligned}$$ $\ell_{CE}$ is the cross-entropy loss between the predicted values and the ground truth labels. $D_{KL}$ is the Kullback Leibler (KL) Divergence function that serves as a term in the loss function $\ell_s$ and $\ell_c$ to transfer knowledge from a network to another. $\boldsymbol{p}_{k}^i=\frac{\exp \left(z_{c}^{(k,i)}/T\right)}{\sum_{i=1}^{C} \exp \left(z_{c}^{(k,i)}/T\right)}$ and $\boldsymbol{p}_{s}^i=\frac{\exp \left(z_{s}^{i}/T\right)}{\sum_{i=1}^{C} \exp \left(z_{s}^{i}/T\right)}$. They are the probabilistic prediction of the client model $f^{(k)}$ and the server model $f_s$, respectively. They are calculated with the softmax of logits $\boldsymbol{z}$. The logit $\boldsymbol{z}$ is the output of the last fully connected layer in a neural network, and T is the temperature hyperparameter of the softmax function. Intuitively, the KL divergence loss attempts to bring the soft label and the ground truth close to each other. In doing so, the server model absorbs the knowledge gained from each of the edge models. Similarly, the edge models try to bring their predictions closer to the prediction of server model and thereby absorb the server model knowledge to improve their feature extraction capability. **Improved Alternating Minimization**. After plugging Eq. and into our reformulation (Eq. and ), we propose a variant of Alternating Minimization (AM) [@ortega1970iterative; @bertsekas1989parallel; @bolte2014proximal; @attouch2010proximal; @wright2015coordinate; @razaviyayn2013unified] to solve the reformulated optimization problem as follows: [$$\begin{aligned} &\underset{\boldsymbol{W}_s} {\operatorname{argmin}} F_s(\boldsymbol{W}_s, \boldsymbol{W}_e^{(k)*}) = \underset{\boldsymbol{W}_s} {\operatorname{argmin}} \sum_{k=1}^{K} \sum_{i=1}^{N^{(k)}} \ell_{CE}\big( f_s(\boldsymbol{W}_s; \underbrace{f_e^{(k)}(\boldsymbol{W}_{e}^{(k)*}; \boldsymbol{X}_i^{(k)}}_{\boldsymbol{H}_i^{(k)}}), y_i^{(k)} \big) + \sum_{k=1}^{K} \ell_{KD} \big(\boldsymbol{z}_c^{(k)*}, \boldsymbol{z}_s \big) \label{eq:server_kd}\\ &\textit{where }~ \boldsymbol{z}_c^{(k)*} = f_c^{(k)}(\boldsymbol{W}_{c}^{(k)}; \underbrace{f_e^{(k)}(\boldsymbol{W}_{e}^{(k)*}; \boldsymbol{X}_i^{(k)}}_{\boldsymbol{H}_i^{(k)}})), \textit{and } \boldsymbol{z}_s = f_s(\boldsymbol{W}_s; \boldsymbol{H}_i^{(k)}) \\ &\underset{\boldsymbol{W}^{(k)}} {\operatorname{argmin}} F_c(\boldsymbol{W}_s^*, \boldsymbol{W}^{(k)}) = \underset{\boldsymbol{W}^{(k)}} {\operatorname{argmin}} \sum_{i=1}^{N^{(k)}} \ell_{CE} \big(f_c^{(k)}(\boldsymbol{W}_{c}^{(k)}; \underbrace{f_e^{(k)}(\boldsymbol{W}_{e}^{(k)}; \boldsymbol{X}_i^{(k)}}_{\boldsymbol{H}_i^{(k)}})), y_i^{(k)}\big) + \ell_{KD} \big(\boldsymbol{z}_s^*, \boldsymbol{z}_c^{(k)} \big) \label{eq:client_kd}\\ &\textit{where }~ \boldsymbol{z}_c^{(k)} = f_c^{(k)}(\boldsymbol{W}_{c}^{(k)}; \underbrace{f_e^{(k)}(\boldsymbol{W}_{e}^{(k)}; \boldsymbol{X}_i^{(k)}}_{\boldsymbol{H}_i^{(k)}})), \textit{and } \boldsymbol{z}_s^* = f_s(\boldsymbol{W}_s^*; \boldsymbol{H}_i^{(k)})\end{aligned}$$]{} Where the $*$ superscript notation in above equations presents related random variable $\boldsymbol{W}$ is fixed. $\boldsymbol{W}^{(k)}$ is the combination of $\boldsymbol{W}_{e}^{(k)}$ and $\boldsymbol{W}_{c}^{(k)}$. AM is a solver in convex and non-convex optimization theory and practice that optimize two random variables alternatively. In Eq. , we fix $\boldsymbol{W}^{(k)}$ and optimize (train) $\boldsymbol{W}_{s}$ for epochs, and then we switch to to fix $\boldsymbol{W}_{s}$ and optimize $\boldsymbol{W}^{(k)}$ for epochs. This optimization persists many rounds between Eq. and until reaching a convergence state. [2]{} **Server executes:** $X_e^k, Z_k \leftarrow $ **ClientLocalTraining**$(k)$ $Z_s \leftarrow $ empty dictionary $\boldsymbol{w}_s \leftarrow \boldsymbol{w}_s-\eta_s \nabla \ell_s(b;\boldsymbol{w}_s,\boldsymbol{\alpha}_s)$ $Z_s[k][\textit{b\_idx}] \leftarrow $ $f_s(x_e; \boldsymbol{w}_s, \boldsymbol{\alpha}_s) $ // *illustrated as “transfer back” in Fig. \[fig:overview\_left\]* send the server logits $Z_s[k]$ to client $k$ **ClientLocalTraining**($k$): // *Run on client $k$* // *illustrated as “local training ”in Fig. \[fig:overview\_left\]* $\boldsymbol{w}_k \leftarrow \boldsymbol{w}_k-\eta_k \nabla \ell_c^{(k)}(b;\boldsymbol{w}_k,\boldsymbol{\alpha}_k)$ *// extract features and logits* $X_e^k, Z_k \leftarrow $ empty dictionary $x_e^k \leftarrow $ $f_e^k(\boldsymbol{x}_i; \boldsymbol{w}_e^k, \boldsymbol{\alpha}_e^k)$ $z_k \leftarrow $ $f_c(f_e^k(\boldsymbol{x}_i; \boldsymbol{w}_e^k, \boldsymbol{\alpha}_e^k), y_i; \boldsymbol{w}_c, \boldsymbol{\alpha}_c)$ $X_e^k[\textit{b\_idx}] \leftarrow x_e^k $ $Z_k[\textit{b\_idx}] \leftarrow z_k $ return $X_e^k$, $Z_k$ to server \[alg:GKT\] **Key Insight**. The essence of our reformulation is that the alternating minimization (Eq. and Eq. ) uses knowledge distillation across all edges to simplify the optimization, which scales the dataset limitation on each edge in federated learning. In particular, we achieve this objective using a local cross-entropy loss computed based only on the ground truth and the model output, and a second loss that uses the KL divergence across edges and the server, which effectively captures the contribution (knowledge) from multiple client datasets. Moreover, each minimization subproblem can be solved with SGD and its variants (e.g., SGD with momentum [@qian1999momentum], ADAM [@kingma2014adam; @zou2019sufficient]). **Training Algorithm.** To implement GKT as described above, we design an alternating and periodical training algorithm. To elaborate, we illustrate the entire training algorithm in Fig. \[fig:overview\_left\] and summarize it as Algorithm \[alg:GKT\]. During each round of training, the client uses local SGD to train several epochs and then sends the extracted features and its logits to the server. When the server receives extracted features and logits from each client, it trains the much larger server-side CNN. The server then sends back its global logits to each client. This process iterates over multiple rounds, and during each round the knowledge of all clients is transferred to the server model and vice-versa. For the GKT training framework, the remaining step is designing specific neural architectures for the client model and the server model. To evaluate the effectiveness of GKT, we design CNN architectures based on ResNet [@he2016deep], which are shown in Fig. \[fig:overview\_right\]. More details can also be found in Appendix \[ap:architecture\]. Experiments =========== Experimental Setup ------------------ **A Distributed Training Library to Support GKT**. Note that the popular deep learning frameworks PyTorch [@paszke2019pytorch] and TensorFlow [@abadi2016tensorflow] cannot fully support our GKT algorithm (e.g. they only support gradient/model exchanging, while our GKT requires exchanging soft labels and hidden vectors). Thus, to simplify the usage of our GKT training framework, we use `FedML` [@chaoyanghe2020fedml], an open source federated learning library that simplifies the new algorithm development, and deploy it in a distributed computing environment. Our server node has 4 NVIDIA RTX 2080Ti GPUs with sufficient GPU memory for large model training. We use several CPU-based nodes as clients training small CNNs. **Task and Dataset**. Our training task is image classification on CIFAR-10 [@krizhevsky2009learning], CIFAR-100 [@krizhevsky2009learning], and CINIC-10 [@darlow2018cinic]. We also generate their non-IID variants by splitting training samples into $K$ unbalanced partitions. Details of these three datasets are introduced in Appendix \[app:dataset\]. The test images are used for a global test after each round. For different methods, we record the top 1 test accuracy as the metric to compare model performance. Note that we do not use FL benchmark datasets published recently (e.g., LEAF [@caldas2018leaf]) because the benchmark models provided are tiny models (CNN with only a few layers) or the datasets they contains are too easy for CNNs (e.g., MNIST), which are unable to adequately evaluate our algorithm running on large CNN models. **Baselines**. We compare GKT with state-of-the-art FL method FedAvg [@mcmahan2016communication], and a centralized training approach. Split Learning-based method [@gupta2018distributed; @vepakomma2018split] is used to compare the communication cost. Note that we do not compare with FedProx [@li2018federated] because it performs worse than FedAvg in the large CNN setting, as demonstrated in [@wang2020federated]. FedMA [@wang2020federated] can work on VGG [@simonyan2014very], but it does not support batch normalization. Thus, it cannot work on ResNet. **Model Architectures**. Two modern CNN architectures are evaluated: ResNet-56 and ResNet-110 [@he2016deep]. The baseline FedAvg requires all edge nodes to train using these two CNNs. For GKT, the edge and server-sided models are designed based on these two CNNs. On the edge, we design a tiny CNN architecture called ResNet-8, which is a compact CNN containing 8 convolutional layers (described in Fig. \[fig:overview\_right\] and Table \[table:resnet8\] in Appendix). The server-sided model architectures are ResNet-55 and ResNet-109 (Table \[table:resnet55\] and \[table:resnet109\] in Appendix), which have the same input dimension to match the output of the edge-sided feature extractor. For split learning, we use the extractor in ResNet-8 as the edge-sided partition of CNNs, while the server-side partitions of CNN is also ResNet-55 and ResNet-109. Result of Model Accuracy ------------------------ For standard experiments, we run on 16 clients and a GPU server for all datasets and models. Fig. \[fig:training\_curve\] shows the curve of the test accuracy during training on ResNet-56 model with 3 datasets. It includes the result of centralized training, FedAvg, and GKT. We also summarize all numerical results of ResNet-56 and ResNet-110 in Table \[table:comparison\]. In both IID and non-IID setting, GKT obtains comparable or even better accuracy than FedAvg. **Hyperparameters**. There are four important hyper-parameters in our GKT framework: the communication round, as stated in line \#2 of Algorithm \[alg:GKT\], the edge-side epoch number, the server-side epoch number, and the server-side learning rate. After a tuning effort, we find that the edge-side epoch number can be just 1. The server epoch number depends on the data distribution. For IID data, it is 20, and for non-IID, the value depends on the level of data bias. For IID, Adam optimizer [@kingma2014adam] works better than SGD with momentum [@qian1999momentum], while for non-IID, SGD with momentum works better. During training, we reduce the learning rate once the accuracy has plateaued [@li2019exponential; @shi2020learning]. We use the same data augmentation techniques for fair comparison (random crop, random horizontal flip, and normalization). More details of hyper-parameters are described in Appendix \[app:hpo\]. Efficiency Evaluation --------------------- \[fig:compute\_eff\] \[table:com\_eff\] To compare the computational demand on training, we count the number of FLOPs (floating-point operations) performed on edge using prior methods [@AI_and_Compute; @hernandez2020measuring]. We report the result on CIFAR-100 in Fig. \[fig:compute\_eff\]. Compared to the FedAvg baseline, the computational cost on the edge of our GKT (ResNet-8) is 9 times less than that of ResNet-56 and 17 times less than that of ResNet-110 (The memory cost comparison can be roughly compared by the model parameter number: ResNet-8 has 11K parameters, which is 54 times less than that of ResNet-56 and 105 times less than that of ResNet-110. We also test the CPU running time per mini-batch (batch size is 64) forward-backward propagation on Intel i7 CPU (which has a more aggressive performance than current edge devices). The results show that ResNet-8 requires only 3% of ResNet-110’s training time (30 ms v.s. 950 ms). To compare communication costs, we use SL [@gupta2018distributed; @vepakomma2018split] as the baseline, which also exchanges hidden feature maps rather than the entire model. The communication cost is calculated using Eq. and in Appendix \[sec:communication\_cost\_method\] without using data compression techniques. The results are shown in Fig. \[table:com\_eff\] (X-axis units: GBytes). GKT uses fewer feature map exchanges with the server than SL. Ablation Study: Understanding GKT under Different Settings ---------------------------------------------------------- \[table:lossfunction\] **The Effectiveness of Bi-directional Knowledge Transfer**. In this study we evaluate the efficacy of using knowledge distillation loss, $\ell_{KD}$ in Eq. and Eq. . Table \[table:loss functions\] shows the results. We created one scenario labeled *None* in the table where both the client and server only use $\ell_{CE}$ and do not use $\ell_{KD}$. In another scenario, only the clients use $\ell_{CE}$ to update their local models, but the server does not (*only S-&gt;E*). If we do not use knowledge transfer at all, the training will diverge ( as shown for *None*). Moreover, the transfer from the server to the edge always helps (*only S-&gt;E*), while the transfer from the edge to the server is more helpful as the dataset becomes increasingly difficult (CIFAR-100). **Asynchronous Training**. Since the server does not need to wait for updates from all clients to start training, GKT naturally supports asynchronous training. We present the experimental results in Table \[table:Asynchronous Training\]. It proves that asynchronous training does not negatively affect model accuracy. This demonstrates the advantage of our method over SL, in which every edge requires multiple synchronizations for each mini-batch iteration. **GKT with Different Edge Number.** To understand the scalability of GKT, we evaluate its performance with varying edge nodes. The test accuracy results are shown in Table \[table:Edge Number\]. In general, adding more edge nodes does not negatively affect accuracy. **Smaller Architectures.** We test the performance of GKT using even smaller edge models: ResNet-4, ResNet-6 on CIFAR-10. ResNet-4 and ResNet-6 use one and two BasicBlock components (including two convolutional layers), respectively. The result is shown in Table \[table:small CNN\]. While reducing the edge model size to ResNet-8 did not reduce accuracy, when the model size is reduced even more substantially, it does reduce the overall accuracy. Conclusion ========== In this work, to tackle the resource-constrained reality, we reformulate FL as a group knowledge transfer (GKT) training algorithm. GKT can efficiently train small CNNs on edges and periodically transfer their knowledge by knowledge distillation to a server-side CNN with a large capacity. GKT achieves several advantages in a single framework: reduced demand for edge computation, lower communication cost for large CNNs, and asynchronous training, all while maintaining model accuracy comparable to FL. To simplify the edge training, we also develop a distributed training system based on our GKT. We evaluate GKT by training modern CNN architectures (ResNet-56 and ResNet-110) on three distinct datasets (CIFAR-10, CIFAR-100, and CINIC-10) and their non-IID variants. Our results show that GKT can obtain comparable or even slightly higher accuracy. More importantly, GKT makes edge training affordable. Compared to the edge training using FedAvg, GKT costs 9 to 17 times less computational power (FLOPs) and requires 54 to 105 times fewer parameters. Broader Impact {#broader-impact .unnumbered} ============== GKT can efficiently train large deep neural networks (CNNs) in resource-constrained edge devices (such as smartphones, IoT devices, and edge servers). Unlike past FL approaches, GKT demonstrates the feasibility of training a large server-side model by using many small client models. GKT preserves the data privacy requirements of the FL approach but also works within the constraints of an edge computing environment. Smartphone users may benefit from this technique because their private data is protected, and they may also simultaneously obtain a high-quality model service. Organizations such as hospitals, and other non-profit entities with limited training resources, can collaboratively train a large CNN model without revealing their datasets while achieving significant training cost savings. They can also meet the requirements regarding the protection of intellectual property, confidentiality, regulatory restrictions, and legal constraints. As for the potential risks of our method, a client can maliciously send incorrect hidden feature maps and soft labels to the server, which may potentially impact the overall model accuracy. These effects must be detected and addressed to maintain overall system stability. Second, the relative benefits for each client may vary. For instance, in terms of fairness, edge nodes which have smaller datasets may obtain more model accuracy improvement from collaborative training than those which have a larger amount of training data. Our training framework does not consider how to balance this interest of different parties. Datasets ======== A Summary of Dataset Used in Experiments {#app:dataset} ---------------------------------------- **CIFAR-10** [@krizhevsky2009learning] consists of 60000 32$\times$32 colour images in 10 classes, with 6000 images per class. There are 50000 training images and 10000 test images. **CIFAR-100** [@krizhevsky2009learning] has the same amount of samples as CIFAR-10, but it is a more challenging dataset since it has 100 classes containing 600 images each. **CINIC-10** [@darlow2018cinic] has a total of 270,000 images, 4.5 times that of CIFAR-10. It is constructed from two different sources: ImageNet and CIFAR-10. It is not a guaranteed that the constituent elements are drawn from the same distribution. This characteristic fits for federated learning because we can evaluate how well models cope with samples drawn from similar but not identical distributions. CINIC-10 has three sub-datasets: training, validation, and testing. We train on the training dataset and test on the testing, without using the validation dataset for all experiments. Our source code provides link to download these three datasets. For the non-IID dataset, the partition is unbalanced: sampling $\mathbf{p}_{c} \sim \operatorname{Dir}_{J}(0.5)$ and allocating a $\mathbf{p}_{c, k}$ proportion of the training samples of class $c$ to local client $k$. Heterogeneous Distribution (non-IID) in Each Client {#app:non-iid-distribution} --------------------------------------------------- We fix the non-IID distribution to fairly compare different methods. Table \[tab:non-iid-distribution\] is a specific distribution used in experiments. We also conduct experiments in other non-IID distributions and see that our GKT method also beats baselines. To generate the different distribution, we can change the random seed in *main.py* of our source code. Extra Experimental Results and Details ====================================== Computational Efficiency on CIFAR-10 and CINIC-10 ------------------------------------------------- \[fig:com\_efficiency\] The Method of Communication Cost Calculation {#sec:communication_cost_method} -------------------------------------------- For split learning (SL), the method to calculate the communication cost is: $$\begin{aligned} \textit{Communication Cost of SL} = (\textit{the size of the hidden feature map} + \textit{the size of the gradient in the split layer}) \notag \\ \times \textit{(number of samples in dataset)} \times \textit{(number of epochs)} \label{eq:com_sl}\end{aligned}$$ For GKT, the method to calculate the communication cost is: $$\begin{aligned} \textit{Communication Cost of GKT} = (\textit{the size of the hidden feature map} + \notag \\ \textit{the size of soft labels received from the server side}) \times \textit{(number of samples in dataset)} \notag \\ \times \textit{(number of communication rounds)} \label{eq:com_gkt}\end{aligned}$$ Details of Convolutional Neural Architecture on Edge and Server --------------------------------------------------------------- ResNet-8 is a compact CNN. Its head convolutional layer (including batch normalization and ReLU non-linear activation) is used as the feature extractor. The remaining two Bottlenecks (a classical component in ResNet, each containing 3 convolutional layers) and the last fully-connected layer are used as the classifier. \[ap:architecture\] Hyperparameters {#app:hpo} --------------- In table \[table:hpcifar10\], \[table:hpcifar100\], and \[table:hpcinic10\], we summarize the hyperparameter settings for all experiments. If applying our GKT framework to a new CNN architecture with different datasets, we suggest tuning all hyper-parameters based on our hyperparameters.

ArXiv

--- abstract: 'Bose-Einstein condensates can be used to produce work by tuning the strength of the interparticle interactions with the help of Feshbach resonances. In inhomogeneous potentials, these interaction ramps change the volume of the trapped gas allowing one to create a thermodynamic cycle known as the Feshbach engine. However, in order to obtain a large power output, the engine strokes must be performed on a short timescale, which is in contrast with the fact that the efficiency of the engine is reduced by irreversible work if the strokes are done in a non-adiabatic fashion. Here we investigate how such an engine can be run in the Thomas-Fermi regime and present a shortcut to adiabaticity that minimizes the irreversible work and allows for efficient engine operation.' author: - Tim Keller - Thomás Fogarty - Jing Li - Thomas Busch bibliography: - 'tf\_engine\_literature.bib' title: 'A Feshbach engine in the Thomas-Fermi regime' --- Introduction ============ Understanding and exploring concepts in quantum thermodynamics is currently a highly active topic with implications for the future development of quantum technologies [@OurManSteve]. Within this area, the creation and operation of quantum engines which implement the Otto cycle and use cold atoms as their working medium, has received special attention, as they can be treated instructively and lend themselves to experimental realisation [@Campo2015Bang; @Beau2016ScalingUp; @li2018efficient; @Niedenzu2019quantized; @Rossnagel2016IonEngine]. Similar to classical thermodynamical engines, quantum engines will achieve maximum efficiency if they are run without creating irreversible work. While this can be achieved by adiabatic evolution, this mode of operation has the drawback that it requires the external parameter changes to be very slow [@born1928beweis]. As reliable, fast and simple control is needed e.g. for the development of new technologies [@acin2018quantum], more recently the use of shortcuts to adiabaticity (STA) has received increased attention [@guery2019shortcuts]. Shortcut protocols provide a way to mimic adiabatic evolution in a finite time, mostly by requiring different parameter ramps or additional levels of control. While a large number of shortcuts have been found for single particle systems, shortcuts for interacting many-particle settings are still rare [@Diao2018STAFermi; @Fogarty2019twoatoms; @Kahan2019CompositeSTA]. However, first experiments have demonstrated the viability of shortcuts protocols for atomic Bose-Einstein condensates (BECs) in the repulsive mean-field limit [@schaff2011shortcut; @rohringer2015non] and for fermionic systems in the unitary limit [@PDeng2018STAFermiExp]. Utilizing STA protocols to efficiently drive the dynamics through the control of the external potential parameters, such as trapping frequencies, can therefore increase the performance of finite time quantum heat engines [@Campo2015Bang; @Abah_2017]. Furthermore, recent works have extended this idea to engines that are driven by changing internal parameters of the working medium, such as changes to the interparticle interaction strength [@li2018efficient; @Chen2019]. In ultracold atoms these are described by a scattering length, which can be controlled experimentally by varying an external magnetic field about a Feshbach resonance [@feshbach1958unified; @courteille1998observation; @chin2010feshbach]. While applying STA protocols to drive interactions is not a trivial task, Li *et al.* showed that it is possible in the case of a bright solitonic BEC which can be frictionlessly compressed and expanded using designed Feshbach pulses [@li2018efficient]. Even though this can lead to an efficient Otto cycle, the operational range of the engine was very limited due to the possibility of BEC collapse in the presence of driven attractive interactions [@gerton2000BECcollapse; @carr2008multidimensional]. It is therefore interesting to extend the idea of the Feshbach engine to BECs in the stable Thomas-Fermi regime of large particle numbers and strong repulsive interactions [@dalfovo1999theory]. For this we derive in this work a novel interaction ramp that allows for the frictionless compression and expansion of such a Thomas-Fermi BEC in an almost arbitrarily short time, and show that these ramps can act as STAs in a Feshbach engine. The presentation is organized as follows: In Sec. \[sec:sta\] we introduce a scaling ansatz to derive an interaction ramp for a harmonically trapped $d$-dimensional BEC in the Thomas-Fermi limit which ensures that the system follows an adiabatic path at all times. We then verify that the ramp is working as intended, up to some minimum time, by numerically simulating the dynamics using the full Gross-Pitaveskii equation and comparing it to a non-optimized reference ramp. In Sec. \[sec:feshbach\_engine\] we show that the shortcut ramp can indeed be used to increase the power and efficiency of the engine and in Sec. \[sec:modulational\_instability\] we perform a stability analysis to derive the minimum time in which the interaction ramp can be performed before a modulational instability leads to a condensate collapse. Shortcut to adiabaticity {#sec:sta} ======================== In this section we derive an interaction ramp for a BEC in the Thomas-Fermi limit that can act as an STA [@chen2010fast; @guery2019shortcuts] and evaluate its performance in compressing and expanding a BEC compared to a smooth non-optimized reference ramp. Interaction ramp ---------------- ![Interaction ramps obtained from the shortcut to adiabaticity according to Eq.  in three dimensions and for different values of $T_f$. The black dashed line shows the time-rescaled adiabatic reference (TRA). []{data-label="fig:ramps"}](figures/TF_3D_ramps.pdf){width="0.9\linewidth"} For simplicity we start by considering a one-dimensional BEC trapped in an inhomogeneous potential, $V(x)$, and generalise the results to higher dimensions later. The condensate can be described by a single wave function, $\psi(x)$, whose dynamics is governed by the Gross-Pitaevskii equation [@pethick_smith_2008] $$i\hbar \frac { \partial \psi } { \partial t } = \left[- \frac { \hbar ^ { 2 } } { 2 m } \frac{\partial^2}{\partial x^2} + V ( x ) + g(t)| \psi | ^ { 2}\right]\psi \, , \label{eq:gpe_cc}$$ where $m$ is the mass of an individual particle in the condensate, $g(t)$ is the nonlinear interaction strength, which may vary in time, and the wave function is normalized to the number of particles, $\int dx |\psi(x)|^2=N$. Since we are interested in compressing or expanding the width of the wave function without changing its general shape, we choose a scaling ansatz [@theocharis2003modulational] of the form $$\psi(x,t) = \frac{1}{\sqrt{a(t)}}e^{i\varphi(x,t)}\phi(y(x,t),\tau(t)) \, ,$$ where the spatial coordinate is rescaled as $y(x,t) = x/a(t)$ and we have also introduced a rescaled time $\tau(t)$. Inserting this ansatz into the Gross-Pitaevskii equation and choosing the phase as $$\varphi(x,t) = \frac{m}{2\hbar}\frac{\dot{a}(t)}{a(t)}x^2 \, , \label{eq:ChoicePhase}$$ to eliminate the term proportional to $\partial \phi/\partial y$, we get $$\begin{aligned} i\hbar \frac {\partial \phi } {\partial\tau }\frac{\partial\tau}{\partial t} = &\left[- \frac { \hbar^2} {2m} \frac{1}{a^2}\frac{\partial^2}{\partial y^2} +V (x) + \frac{g(t)}{a}| \phi |^2\right]\phi\nonumber\\ &+ \left[\frac{i\hbar}{2}\frac{\dot{a}}{a}+\hbar\dot{\varphi}-\frac{i\hbar^2}{2m}\frac{\partial^2\varphi}{\partial x^2} + \frac{\hbar^2}{2m}\left(\frac{\partial\varphi}{\partial x}\right)^2\right]\phi \, .\end{aligned}$$ The choice of phase made in Eq.  can be interpreted as a gauge transformation $\hat{U}=e^{i\varphi(x,t)}$, which adjusts the momentum of the expanding or shrinking system as $$\hat{p} \rightarrow \hat{U}\hat{p}\hat{U}^\dagger = \hat{p} - m\frac{\dot{a}}{a}\hat{x} \, ,$$ where $\dot{a}x/a$ is the local velocity [@castin1996bose; @del2013shortcuts]. The same transformation is also commonly found in other scaling problems, e.g. as the optimal solution in a variational approach [@kagan1996evolution; @perez1996low; @li2016shortcut]. Assuming the external potential is given by a harmonic trap, $V(x)=\frac{1}{2}m\omega^2x^2$, we obtain $$\begin{aligned} i\hbar \frac{\partial \phi} {\partial\tau }\frac{\partial\tau}{\partial t} = \left[- \frac { \hbar ^ { 2 } } { 2 m } \right.& \frac{1}{a^2}\frac{\partial^2}{\partial y^2}\nonumber\\ + \frac{1}{2}m&\left(\ddot{a}+\omega^2 a\right)ay^2 +\left. \frac{g(t)}{a}| \phi | ^ { 2}\right]\phi \, . \label{eq:scaling_gpe}\end{aligned}$$ Choosing the rescaled time $\tau$ and the term $\ddot{a}+\omega^2 a$ in such a way that it leads to a solvable Gross-Pitaevskii equation then allows one to design control pulses for a frictionless evolution of the BEC, e.g. by varying either the trap frequency $\omega$, the interaction strength $g$ or both [@muga2009frictionless; @stefanatos2012frictionless; @del2013shortcuts; @deffner2014classical]. Many experiments involving repulsively interacting Bose-Einstein condensates are carried out in the so-called Thomas-Fermi (TF) regime, where the potential and the interaction energies are much larger than the kinetic energy, $Ng\gg \sqrt{\hbar^3\omega/m}$ [@pethick_smith_2008]. This allows one to neglect the kinetic energy term in the Gross-Pitaevskii equation and obtain an analytical solution of the form $$\psi(x,t) = \sqrt{\frac{1}{g}\left(\mu-V(x)\right)}e^{-i\mu t/\hbar} \quad\text{ for }\quad \mu>V(x) \label{eq:thomas_fermi_wf}$$ and $\psi(x,t)\equiv 0$ otherwise. The chemical potential $\mu$ is determined via the normalization condition. Considering the TF limit in the scaling GPE Eq.  and choosing scaling functions as [@ozcakmakli2012shortcuts] $$\begin{split} \ddot{a} + \omega^2a &= \omega^2\frac{g(t)}{gi}\frac{1}{a^2} \quad\text{ and }\\ \tau = \int_0^t dt' \frac{g(t')}{g_ia(t')} &= \int_0^t dt' \frac{a(t')}{\omega^2}\left(\ddot{a}(t') + \omega^2a(t')\right)\, , \end{split} \label{eq:scaling_functions}$$ with some initial interaction strength $g_i$, then gives $$i\hbar\frac{\partial\phi}{\partial\tau}= \left[\frac{1}{2}m\omega^2y^2 + g_i | \phi | ^ { 2}\right]\phi \, .$$ This again has the aforementioned Thomas-Fermi solution $$\phi(y,\tau) = e^{-i\mu_i\tau/\hbar}\sqrt{\frac{1}{g_i}\left(\mu_i-\frac{1}{2}m\omega^2y^2\right)}\;,$$ with $\mu_i = \left(\frac{9}{32}m\omega^2N^2g_i^2\right)^{1/3}$. Inserting everything back into the scaling ansatz gives us an analytic expression for the evolution of the wave function $$\begin{split} \psi(x,t) = &\frac{1}{\sqrt{a(t)}}e^{i\frac{m}{2\hbar}\frac{\dot{a}(t)}{a(t)}x^2}e^{-i\frac{\mu_i}{\hbar}\int_0^tdt'\frac{g(t')}{g_ia(t')}}\times\\ &\sqrt{\frac{1}{g_i}\left(\mu_i-\frac{1}{2}m\omega^2\frac{x^2}{a^2(t)}\right)} \, , \end{split} \label{eq:wf_analytical}$$ and choosing a suitable scaling function $a(t)$ then allows us to reverse-engineer an interaction ramp that will take the system along this evolution. For this we rearrange Eq.  for $g(t)$ and get $$g(t) = g_i\frac{a^2(t)}{\omega^2}\left(\ddot{a}(t)+\omega^2a(t)\right)\;.$$ Choosing appropriate boundary conditions for $a(t)$ of the form $$\begin{split} a(0) &= a_i = 1 \, ,\\ a(T_f) &= a_f = \left(g_f/g_i\right)^{1/3} \;,\\ \dot{a}(0) &=\dot{a}(T_f)=\ddot{a}(0)=\ddot{a}(T_f) = 0 \, , \end{split} \label{eq:tf_a_cond}$$ we can drive the system from the Thomas-Fermi ground state at an initial interaction strength $g_i$ to the ground state at a final value $g_f$ in an almost arbitrarily short time $T_f$, while mimicking an adiabatic evolution. It is worth noting that by doing this the system also acquires an additional, but irrelevant phase that depends on $T_f$. The boundary conditions for $a(t)$ can be fulfilled by a fifth-order polynomial $a(t) = a_i + (a_f-a_i)\left[10s^3 -15s^4 + 6s^5\right]$ with $s=t/T_f$, which has the form of a *smoother step*-function [@perlin2002improving]. This shortcut to adiabaticity is easily generalised to a $d$-dimensional BEC in an isotropic harmonic trap in the Thomas-Fermi limit by driving the interaction strength according to $$g(t) = g_i \frac{a^{d+1}(t)}{\omega^2}\left(\ddot{a}(t)+\omega^2 a(t)\right) \label{eq:g_ramp}$$ and requiring $a(t_F)=\left(g_f/g_i\right)^{1/(d+2)}$ as well as replacing $g_ia(t')$ with $g_ia^d(t')$ in the time scaling $\tau$ in Eq. . In the following we will use dimensionless units and scale lengths by $x_0=\sqrt{\hbar/m\omega}$, energies by $\hbar\omega$, time by $\omega$ and interaction strengths by $\hbar\omega x_0^d$. Evaluation ---------- As a reference for benchmarking the shortcut performance we define a time-rescaled adiabatic (TRA) stroke, which can be obtained by letting $T_f\rightarrow\infty$ or equivalently by setting $\ddot{a}(t)\equiv 0$ in the interaction ramp $g(t)$. A comparison between the TRA ramp for compressing a three-dimensional Thomas-Fermi BEC from $g_i=1$ to $g_f=0.8$ with STA ramps for varying $T_f$ is shown in Fig. \[fig:ramps\], and one can immediately notice that faster ramps require larger changes in the interaction strength over the duration of the STA. ![**Upper row:** Irreversible work $W_\mathrm{irr}$ and fidelity $F$ after compressing a 3D BEC consisting of $N=10^4$ atoms from an initial interaction $g_i=1$ to $g_f=0.8$ in a time $T_f$ using the shortcut to adiabaticity (STA) and an adiabatic reference protocol (TRA). **Lower row:** Identical plots but for the reverse expansion stroke and $N=8000$ atoms.[]{data-label="fig:strokes"}](figures/TF_3D_compression_stroke_g1_wirr.pdf "fig:"){width="24.00000%"}![**Upper row:** Irreversible work $W_\mathrm{irr}$ and fidelity $F$ after compressing a 3D BEC consisting of $N=10^4$ atoms from an initial interaction $g_i=1$ to $g_f=0.8$ in a time $T_f$ using the shortcut to adiabaticity (STA) and an adiabatic reference protocol (TRA). **Lower row:** Identical plots but for the reverse expansion stroke and $N=8000$ atoms.[]{data-label="fig:strokes"}](figures/TF_3D_compression_stroke_g1_fid.pdf "fig:"){width="24.00000%"}\ ![**Upper row:** Irreversible work $W_\mathrm{irr}$ and fidelity $F$ after compressing a 3D BEC consisting of $N=10^4$ atoms from an initial interaction $g_i=1$ to $g_f=0.8$ in a time $T_f$ using the shortcut to adiabaticity (STA) and an adiabatic reference protocol (TRA). **Lower row:** Identical plots but for the reverse expansion stroke and $N=8000$ atoms.[]{data-label="fig:strokes"}](figures/TF_3D_expansion_stroke_g1_wirr.pdf "fig:"){width="24.00000%"}![**Upper row:** Irreversible work $W_\mathrm{irr}$ and fidelity $F$ after compressing a 3D BEC consisting of $N=10^4$ atoms from an initial interaction $g_i=1$ to $g_f=0.8$ in a time $T_f$ using the shortcut to adiabaticity (STA) and an adiabatic reference protocol (TRA). **Lower row:** Identical plots but for the reverse expansion stroke and $N=8000$ atoms.[]{data-label="fig:strokes"}](figures/TF_3D_expansion_stroke_g1_fid.pdf "fig:"){width="24.00000%"} We can assess the performance of the shortcut by numerically simulating the Gross-Pitaevskii equation with the calculated interaction ramps and evaluating the irreversible work and fidelity, $$W_\mathrm{irr} = E(T_f) - E_f \quad \text{ and } \quad F = \left|\braket{\psi(T_f)|\psi_\mathrm{target}}\right|^2\;,$$ at the end of the stroke as a measure of how close the system’s state $\ket{\psi(T_f)}$ with energy $E(T_f)$ after the evolution is to the desired target state $\ket{\psi_\mathrm{target}}$ with energy $E_f$. For a three-dimensional BEC consisting of $N=10^4$ atoms and a compression going from $g_i=1$ to $g_f=0.8$ we show these quantities for the STA and TRA strokes in the upper row of Fig. \[fig:strokes\]. For comparison we also show them for the reverse expansion stroke going from $g_i=0.8$ to $g_f=1$ in the lower row of Fig. \[fig:strokes\], but for $N=8000$ instead. We note we use the actual ground states for the full Gross-Pitaevskii equation as the initial and target states for the evolution, which differ slightly from the Thomas-Fermi wave function in Eq. , especially around the condensate edges. Therefore their energy is slightly higher than the Thomas-Fermi value of $$E = \frac{5}{7}N\mu \quad\text{ with }\quad \mu = \left(\frac{15Ng}{16\sqrt{2}\pi}\right)^{2/5} \;. \label{eq:bec_energy}$$ Nevertheless, one can see that in each case the STA outperforms the TRA stroke by several orders of magnitude in the irreversible work for almost any stroke duration $T_f$ above some threshold $T_f^\mathrm{min}$. Similarly, above this threshold the shortcut always achieves a perfect fidelity of $F=1$ with the target state while the TRA falls short. The sharp dips in the irreversible work for some stroke times as well as the near-perfect fidelity for the TRA around $T_f\approx\pi$ are accidental and can be attributed to the underlying dynamics in the harmonic trapping potential [@haque2013slow]. The sudden increase in irreversible work and the accompanying drop of the fidelity to basically zero if the shortcut is performed too fast is due to a modulational instability that exists for the nonlinear Schrödinger equation (see for example [@sulem2007nonlinear]). It is triggered by the shortcut ramp driving the BEC at attractive interactions for extended periods of time, resulting in a collapse of the condensate. This collapse can create trains of bright solitons [@strecker2002formation; @cornish2006formation; @everitt2017observation] which lead to a clear deviation from the adiabatic path. In the particular example considered here and in the next section, the threshold below which the modulational instability appears is roughly $T_f^\mathrm{min} \approx 0.05$ and in Section \[sec:modulational\_instability\] we present a more detailed stability analysis to derive a general criterion for a given dimensionality, chemical potential, change in interaction and stroke time. Feshbach engine {#sec:feshbach_engine} =============== ![The energy of the system in the Thomas-Fermi regime according to Eq. , as a function of the interaction strength $g$ for $N=10^4$ (upper dashed line) and $N=8000$ (lower dashed line). The solid lines indicate the Otto engine cycle consisting of two adiabatic strokes between $g_i=1$ and $g_f=0.8$, performing compression and expansion work, $\langle W_C\rangle$ and $\langle W_E\rangle$ respectively, and two isochoric strokes adding or removing heat $\langle Q_{N_+}\rangle$ or $\langle Q_{N_-}\rangle$ respectively by adding or removing particles to and from the condensate.[]{data-label="fig:TF_3D_g1_cycle"}](figures/TF_3D_g1_cycle.pdf){width="45.00000%"} **(a)(b)** ![(a) Efficiency $\eta$ and (b) power $P$ of the three-dimensional Feshbach engine as a function of the engine cycle duration $\tau = 2T_f$, comparing the performance of the shortcut to adiabaticity (STA) for the adiabatic strokes with an adiabatic reference (TRA). The maximally attainable efficiency for this specific engine cycle is $\eta_{\mathrm{AD}} = 0.0854$. (c) Plot of efficiency against engine power for both the STA and TRA adiabatic strokes and (d) relative increase in engine power by using the STA strokes as a function of cycle duration $\tau$.[]{data-label="fig:performance"}](figures/TF_3D_g1_eff.pdf "fig:"){width="24.00000%"}![(a) Efficiency $\eta$ and (b) power $P$ of the three-dimensional Feshbach engine as a function of the engine cycle duration $\tau = 2T_f$, comparing the performance of the shortcut to adiabaticity (STA) for the adiabatic strokes with an adiabatic reference (TRA). The maximally attainable efficiency for this specific engine cycle is $\eta_{\mathrm{AD}} = 0.0854$. (c) Plot of efficiency against engine power for both the STA and TRA adiabatic strokes and (d) relative increase in engine power by using the STA strokes as a function of cycle duration $\tau$.[]{data-label="fig:performance"}](figures/TF_3D_g1_pow.pdf "fig:"){width="24.00000%"}\ **(c)(d)** ![(a) Efficiency $\eta$ and (b) power $P$ of the three-dimensional Feshbach engine as a function of the engine cycle duration $\tau = 2T_f$, comparing the performance of the shortcut to adiabaticity (STA) for the adiabatic strokes with an adiabatic reference (TRA). The maximally attainable efficiency for this specific engine cycle is $\eta_{\mathrm{AD}} = 0.0854$. (c) Plot of efficiency against engine power for both the STA and TRA adiabatic strokes and (d) relative increase in engine power by using the STA strokes as a function of cycle duration $\tau$.[]{data-label="fig:performance"}](figures/TF_3D_g1_EMP.pdf "fig:"){width="24.00000%"}![(a) Efficiency $\eta$ and (b) power $P$ of the three-dimensional Feshbach engine as a function of the engine cycle duration $\tau = 2T_f$, comparing the performance of the shortcut to adiabaticity (STA) for the adiabatic strokes with an adiabatic reference (TRA). The maximally attainable efficiency for this specific engine cycle is $\eta_{\mathrm{AD}} = 0.0854$. (c) Plot of efficiency against engine power for both the STA and TRA adiabatic strokes and (d) relative increase in engine power by using the STA strokes as a function of cycle duration $\tau$.[]{data-label="fig:performance"}](figures/TF_3D_g1_pow_ratio.pdf "fig:"){width="24.00000%"} A Feshbach engine operates an Otto cycle which consists of two adiabatic interaction ramps that compress and expand a BEC, and which are connected by two isochoric strokes that add or remove particles [@li2018efficient]. The latter can be realized by cooling or heating the thermal cloud of atoms surrounding any BEC and thus prompting atoms to condense into or evaporate from the condensate, respectively. In the following we will evaluate the Otto engine cycle driven in a harmonically trapped, three-dimensional BEC, where the interaction strength goes between $g_i=1$ and $g_f=0.8$ for the adiabatic strokes and the particle number between $N_i=10^4$ and $N_f=8000$ for the isochoric strokes (see Fig. \[fig:TF\_3D\_g1\_cycle\]). We quantify the performance of the engine by calculating its efficiency and power $$\eta = -\frac{\langle W_C \rangle + \langle W_E\rangle}{\langle Q_{N_+}\rangle} \quad \quad P = -\frac{\langle W_C \rangle + \langle W_E\rangle}{\tau} \, ,$$ where the compression and expansion work $\langle W_C\rangle=E(N_i,g_f)-E(N_i,g_i)$ and $ \langle W_E\rangle = E(N_f,g_i)-E(N_f,g_f)$ as well as the heat $\langle Q_{N_+}\rangle = E(N_i,g_i)-E(N_f,g_i)$ are always calculated from the actual values obtained after performing the strokes and not from the adiabatic values indicated in Fig. \[fig:TF\_3D\_g1\_cycle\]. For the cycle duration $\tau$ we assume that the isochoric strokes can be performed a lot faster than the work strokes and that we can therefore approximate it as $\tau\approx 2T_f$ with the work stroke duration $T_f$. An analytical expression for the maximally attainable adiabatic efficiency of such a Thomas-Fermi Feshbach engine can be obtained by using the Thomas-Fermi wave functions to calculate the energy of the BEC at the engine cycle end points and is given as a function of the compression ratio $g_f/g_i$ by $$\eta_\mathrm{AD} = 1 - \left(\frac{g_f}{g_i}\right)^\gamma \, ,$$ with $\gamma = 2/3$ (1D BEC), $\gamma = 1/2$ (2D BEC) and $\gamma = 2/5$ (3D BEC). It is worth noting that for the one-dimensional case this is less efficient than the bright soliton Feshbach engine using the same compression ratio, which has an exponent of $\gamma = 2$ [@li2018efficient]. The efficiency and power of the engine cycle, using the adiabatic strokes shown in Fig. \[fig:strokes\], are plotted in Fig. \[fig:performance\](a) and (b) respectively. One can immediately see that the STA enables the engine to reach its maximum adiabatic efficiency of $\eta_{\mathrm{AD}} = 0.0854$ already for cycle times four to five times shorter than in the TRA case and leads to a considerable increase in engine power for all cycle times considered, as long as the modulational instability is not triggered. Plotting the ratio $P_\mathrm{STA}/P_\mathrm{TRA}$ in Fig. \[fig:performance\](d) shows that this increase can reach up to $60\%$. As expected, the advantage decreases and the STA and TRA perform equally well once the cycle times are increased to more and more adiabatic values. Finally, plotting the engine efficiency versus power in Fig. \[fig:performance\](c) shows that the shortcut enables the engine to run with high efficiency even at high power output, which is an important factor in the operation of any heat engine [@lebon2008understanding]. Modulational Instability {#sec:modulational_instability} ======================== To better understand the limit of the TF Feshbach engine, we perform in the following a stability analysis to determine the threshold times, $T_f^\mathrm{min}$, below which the shortcuts to adiabaticity derived in Sec. \[sec:sta\] fail due to the appearance of a modulational instability. These threshold times act as an intrinsic quantum speed limit [@deffner2017quantum] for the manipulation of the Thomas-Fermi BECs considered here. For simplicity, we will again first perform the analysis for the compression of a one-dimensional BEC and show later that the obtained stability criterion can easily be extended to higher dimensions. This is confirmed by comparison with numerical simulations. The general reason for an instability to occur is that for short manipulation times $T_f$, the system needs to be driven into the regime of attractive interactions, $g<0$, for a certain amount of time (see Fig. \[fig:MI\_ramp\]), and entering this regime too deeply will lead the condensate to collapse [@carr2004spontaneous]. Using the same parameters of $N=10^4$, $g_i=1$ and $g_f=0.8$ as in the previous sections, we find that the modulational instability for compressing a one-dimensional BEC via the shortcut is triggered around $T_f^\mathrm{min}\approx 0.45$. ![Interaction ramp given by the shortcut to adiabaticity for reducing the interactiong strength of a one-dimensional BEC from $g_i=1$ to $g_f=0.8$ in time $T_f$ according to Eq. . The time-rescaled adiabatic reference (TRA) obtained by setting $\ddot{a}\equiv 0$ is used for comparison. For these parameters, the modulational instability seems to be triggered once the ramp’s minimum goes below $-g_i$ (dashed line), which happens roughly around $T_f\approx 0.45$. []{data-label="fig:MI_ramp"}](figures/MI_ramp.pdf){width="40.00000%"} ![Emergence of the modulational instability in the condensate density for compressing a one-dimensional BEC with $N=10^4$ atom in the Thomas-Fermi regime from $g_i=1$ to $g_f=0.8$ by the interaction STA in Eq. in a time $T_f=0.4$.[]{data-label="fig:MI_energy_combined"}](figures/TF_1D_density.pdf){width="47.00000%"} To understand this instability, let us first carefully examine its appearance in the Gross-Pitaevskii dynamics. From the density distributions shown in Fig. \[fig:MI\_energy\_combined\] one can see that the instability manifests itself by the appearance of rapid oscillations that develop in the center of the condensate where the density is maximal. Inserting the analytical solution for the wave function dynamics from Eq. into the GPE gives $$i \frac { \partial \psi } { \partial t } = \left[- \frac { 1 } { 2 } \frac{\partial^2}{\partial x^2} -\frac{1}{2}\frac{\ddot{a}(t)}{a(t)}x^2 + \mu_i\left(a\ddot{a}+a^2\right)\right]\psi \, , \label{eq:GPE_stability}$$ from which, or by directly setting $x=0$ in Eq. , one can see that close to the trap center, where the density is approximately homogeneous and the kinetic and potential energies can be neglected, the wave function evolves according to $$\psi_\mathrm{hom}(t) = \sqrt{\frac{\mu_i}{a(t)g_i}}e^{-i\int_0^td\tau \mu(\tau)} \;,$$ where $\mu(t) = \mu_i(\ddot{a}a + a^2)$. It is worth noting that from Eq.  one can see that the interaction ramp accomplishes the condensate rescaling by creating a time-varying harmonic trapping potential with a time-varying ground state energy. We assume that the modulational instability can be described by a perturbation to the homogeneous solution of the form of $\psi(x,t) = \psi_\mathrm{hom}(t)\left[1+u(t)\cos(kx)\right]$, with complex amplitude $u(t) = u_\mathrm{re}(t) + iu_\mathrm{im}(t)$ and wave number $k$. Inserting this into the Gross-Pitaevskii equation and keeping only terms up to first order in $u$ then leads to $$\frac{\partial}{\partial t}\begin{pmatrix} u_\mathrm{re}\\ u_\mathrm{im} \end{pmatrix} = \begin{pmatrix} \frac{\dot{a}}{2a} & \frac{k^2}{2}\\ -\left[\frac{k^2}{2}+2\mu(t)\right] & \frac{\dot{a}}{2a} \end{pmatrix} \begin{pmatrix} u_\mathrm{re}\\ u_\mathrm{im} \end{pmatrix} + \begin{pmatrix}\frac{\dot{a}}{2a}\frac{1}{\cos(kx)}\\ 0 \end{pmatrix} . \label{eq:ODE_1}$$ Since we are interested in perturbations with a length scale $k\sim\frac{1}{\zeta}$ comparable to the condensate’s healing length $\zeta = 1/\sqrt{2\mu}$ [@dalfovo1999theory], all the terms proportional to $\dot{a}/2a$ are negligibly small and we can consider the ODE $$\ddot{u}_\mathrm{re} +\frac{k^2}{2}\left(\frac{k^2}{2} + 2\mu(t)\right)u_\mathrm{re} = 0 \, , \label{eq:ODE_2}$$ which is similar to Hill’s equation [@magnus2013hill]. Note that the same equation can be obtained for the $2D$ and $3D$ case by replacing $\mu_i$ and $a_f$ in $\mu(t)$ with the appropriate values. Let us note that if $\mu(t)$ was periodic, i.e. if we had choosen a sinusoidal function to fulfil the boundary conditions [@fogarty2019fast], Floquet theory would provide exact conditions for the stability of Eq.  [@magnus2013hill; @teschl2012ordinary]. However, a ramp resulting from such an approach generally has maxima and minima with larger magnitude than the corresponding polynomial ramp and therefore triggers the instability even earlier. Furthermore, the instability would usually occur around $t\approx T_f/4$, when the ramp reaches its minimum, which means that the periodicity of the ramp would not come into play. ![Minimum time for the compression of a one-dimensional BEC in the Thomas-Fermi regime consisting of $N=10^4$ atoms by changing its interaction from $g_i=1$ (blue curve) or $g_i=2$ (red curve) to a final value of $g_f$ via the STA in Eq. . The curves represent the lines $\Delta=10^{14}$ for the stability criterion . The yellow curve shows $T_f^\mathrm{min}$ for a three-dimensional BEC with $N=10^4$, $g_i=1$ and the same $\Delta = 10^{14}$. The circles show numerically obtained values for $T_f^\mathrm{min}$ from simulating the full Gross-Pitaevskii equation in each case and using the sharp increase in $W_\mathrm{irr}$ as an indicator for the instability.[]{data-label="fig:TF_stability_analysis"}](figures/TF_stability_analysis.pdf){width="45.00000%"} Since we are considering a complex amplitude $u(t)$, it is helpful to do a change of coordinates. For this we rewrite Eq. in polar coordinates using $u(t)=r(t)e^{i\varphi(t)}$ and omitting the terms proportional to $\dot{a}/2a$ to get $$\begin{aligned} \dot{r} &= -2\mu(t)\cos(\varphi)\sin(\varphi)r = -\mu(t)\sin(2\varphi)r\;,\\ \dot{\varphi} &= -\left[\frac{k^2}{2}+2\mu(t)\cos^2(\varphi)\right]\;.\end{aligned}$$ In this form it is easy to see that for $\mu(t)<0$ and $\varphi = \arccos\left(\frac{k}{2\sqrt{|\mu(t)|}}\right)$ we have $\dot{\varphi} = 0 $ and the amplitude of the perturbation can increase exponentially according to $$\dot{r} = rk\sqrt{|\mu(t)|-\frac{k^2}{4}}\;,$$ if $k<2\sqrt{|\mu(t)|}$. The increase is maximal for $k=\sqrt{2|\mu(t)|}$, for which we have $$\dot{r} = \tilde{\mu}(t) r \quad \text{ or } \quad r(t) = e^{\int_0^td\tau \tilde{\mu}(\tau)}r(0)\;,$$ and where we have defined $$\tilde{\mu}(t) = \begin{cases} \left|\mu(t)\right| & \text{for } \mu(t) \leq 0 \\ 0 & \text{otherwise}\;. \end{cases}$$ This allows us to define the total relative increase in the perturbation’s amplitude after the interaction ramps as $$\Delta = \exp\left(\int_0^{T_f} d\tau \tilde{\mu}(\tau)\right) \, . \label{eq:stability_criterion}$$ The previous observation that the modulational instability starts forming at the center of the condensate indicates that it is triggered or seeded by noise [@carr2004pulsed; @nguyen2017formation]. In our numerical simulations this is numerical noise, which appears on the level of $10^{-14}$. Since it grows exponentially, a good choice as a criterion for stability is to set $\Delta = 10^{14}$, i.e. the point at which the small perturbation on the BEC’s wavefunction becomes comparable to magnitude of the BEC wavefunction itself. The resulting curves for a compression stroke are shown in Fig. \[fig:TF\_stability\_analysis\]. The minimal compression times $T_f^\mathrm{min}$ from an initial interaction $g_i$ to some final value $g_f<g_i$ agree well with the actual times obtained from numerically simulating the GPE both for one-dimensional and three-dimensional BECs. While we found $\Delta = 10^{14}$ to be the best fit for our numerical data, changing the criterion even by several orders of magnitude only leads to slight deviations in the resulting stability curves. Conclusion ========== By using a scaling ansatz we have exactly solved the dynamics of an atomic Bose-Einstein condensate subject to a specific interaction ramp in the repulsive Thomas-Fermi limit. This interaction ramp provides a shortcut to adiabaticity for driving the condensate from one interaction strength to another and thereby compressing or expanding it in a short amount of time while avoiding unwanted excitations. We have shown how these shortcuts can increase the efficiency of a Feshbach engine by using them as the adiabatic strokes of its Otto cycle. Using numerical simulations of the full condensate dynamics, we have shown that the speed-up of the condensate manipulation is limited by a modulational instability leading to a condensate collapse. The instability is caused by the need to drive the condensate at increasingly attractive interactions for longer periods of time as the ramp becomes shorter and shorter. Finally, we have performed a stability analysis and determined a criterion that provides an accurate limit $T_f^\mathrm{min}$ for a given initial chemical potential and final interaction strength. This work has been supported by the Okinawa Institute of Science and Technology Graduate University and used the computing resources of the Scientific Computing and Data Analysis section. Numerical simulations of the three-dimensional BEC dynamics were performed using the GPUE codebase [@schloss2018gpue].

ArXiv

A\ [rene.gy@numericable.com]{}\  \  \ **Abstract** For any prime $p$ and any natural numbers $\ell, n$ such that $p$ does not divide $n$, it holds that $\sum_{i\ge \ell+1}(-1)^i {\lfloor \frac{n}{p}\rfloor p \choose i}{n-1+i(p-1) \choose n-1+\ell(p-1)} \equiv 0\bmod p$. Our proof involves Stirling numbers. =12.875pt Introduction, notations and preliminaries ========================================= The purpose of the present paper is to establish that for any prime $p$ and any natural numbers $\ell, n$ such that $p$ does not divide $n$, we have $$\sum_{i\ge \ell+1}(-1)^i {\lfloor \frac{n}{p}\rfloor p \choose i}{n-1+i(p-1) \choose n-1+\ell(p-1)} \equiv 0\bmod p.$$ In the following, $[[x^n]]f(x)$ denotes the coefficient of $x^n$ in $f(x)$, where $f$ is a formal power series with the argument $x$ and $Df(x)$ is the derivative of $f(x)$ with respect to $x$. If $x$ is a real number, we denote $\lfloor x \rfloor$ the largest integer smaller or equal to $x$. We also use the Iverson bracket notation: $\big[\mathfrak{P}\big]=1$ when proposition $\mathfrak{P}$ is true, and $\big[\mathfrak{P}\big]=0$ otherwise. \[ce\] Let $f(w)$ be a formal power series, and $\alpha$ a natural number, we have $[[w^n]]\frac{f(w)^\alpha}{\alpha}=[[w^{n-1}]]\frac{f(w)^{\alpha-1} Df(w)}{n}$. This is clear, since $[[w^n]]f(w)=\frac{[[w^{n-1}]]Df(w)}{n}$. We recall some basic properties of the binomial coefficients and Stirling numbers, which can be found for instance in [@Graham94]. The binomial coefficients ${n\choose k}$, are defined by $\sum_{k}{n\choose k}x^k =(1+x)^n$, whatever the sign of integer $n$. They obviously vanish when $k<0$. They are easily obtained by the basic recurrence relation ${{n}\choose {k}}={{n-1}\choose {k}}+{{n-1}\choose{k-1}}$. When $n > 0$, we have ${-n\choose k}= (-1)^k{n+k-1\choose n-1}$.\ The cycle (or first kind) Stirling numbers ${n\brack k}$, $n\ge0$ may be defined by the horizontal generating function $$\begin{aligned} \label{gf} \sum_{k}{n\brack {k}}x^k&=\prod_{j=0}^{n-1}(x+j), \end{aligned}$$ where an empty product is meant to be $1$. Alternatively, they have the exponential generating function $$\begin{aligned} \label{egfs1} \sum_n {n \brack k} \frac{x^n}{n!}&=\frac {(-1)^k \big(\ln(1-x)\big)^k}{k!}.\end{aligned}$$ They obviously vanish when $k<0$ and $k>n$. They are easily obtained by the basic recurrence ${{n}\brack {k}}=(n-1){{n-1}\brack {k}}+{{n-1}\brack {k-1}}$, valid for $n \ge 1$, with ${{0}\brack {k}}=[k=0]$.\ We let ${n\brace {k}}$, $n\ge0$, be the partition (or second kind) Stirling number. They also vanish when $k<0$ and $k>n$. Their basic recurrence is ${{n}\brace {k}}=k{{n-1}\brace {k}}+{{n-1}\brace {k-1}}$ for $n\ge1$, with $ {0\brace {k}}=[k=0]$. They have the following exponential generating function $$\begin{aligned} \label{egfs2} \sum_n {n \brace k} \frac{x^n}{n!}&=\frac {(e^x-1)^k}{k!}\end{aligned}$$ and the following explicit expression $$\begin{aligned} \label{aaa} {n\brace {k}}&= \frac{(-1)^k}{k!}\sum_{j\ge 0}(-1)^j\binom{k}{j}j^n.\end{aligned}$$ \[st2\] For any natural number $n$, $n>0$, we have $$\label{lab} {n \brace {p-1}} \equiv \big[p-1 \text{ divides }n\big] \pmod p \text{.}$$ We recall the Wilson theorem which states that $(p-1)! \equiv -1 \bmod p$ for any prime $p$, and some other well-known congruences, valid for any prime $p$: $$\begin{aligned} \binom{p-1}{j} &\equiv (-1)^j \big[0\le j\le p-1 \big]\pmod p\text{,} \\ \sum_{p-1\ge j\ge 1}j^k &\equiv - \big[p-1 \text{ divides }k\big]\pmod p\text{,} \end{aligned}$$ so that, the claim readily follows from Equation . $p$-congruences for Stirling numbers of first kind. =================================================== Let $p$ be a prime number. We consider $x^{\bar{p}}:=\prod_{j=0}^{p-1}(x+j) = \sum_{k}{p\brack {k}}x^k $ as a polynomial in $ \mathbb Z / p\mathbb Z$. In that ring, we have $ \sum_{k}{p\brack {k}}x^k= x^p-x $, since the polynomials on both sides have have same degree $p$, same coefficient for $x^p$ and same roots: $0,-1,-2, \cdot \cdot \cdot, -(p-1)$. In particular, for $k$ such that $1<k \le p-1$, we have ${p\brack {k}}\equiv 0 \bmod p$.\ Let $n$ be a non negative integer and $r$, resp $q$ be the residue, resp. the quotient of the Euclidean division of $n$ by $p$, such that $n=q p+r$, with $0 \le r \le p-1$. We may explicitly write and regroup the factors of $x^{\bar{n}}$ so that $$x^{\bar{n}}= \prod_{t=0}^{q-1} \left((x+tp)(x+tp+1)\cdot \cdot\cdot (x+tp+(p-1)\right) \prod_{u=0}^{r-1}(x+q+u),$$ with the convention that when $q=0$ or $r=0$ the empty products are meant to be equal to $1$. Then, reducing modulo $ p$, we have $$\begin{aligned} \sum_{k}{n\brack {k}}x^k &\equiv \prod_{t=0}^{q-1} \left(x(x+1)\cdot \cdot\cdot (x+(p-1)\right) \prod_{u=0}^{r-1}(x+q p+u) \pmod p\\ & \equiv \left(x(x+1)\cdot \cdot\cdot (x+(p-1)\right)^{q} \prod_{u=0}^{r-1}(x+u) \pmod p \\ & \equiv \left(x^p-x\right)^{ q} \prod_{u=0}^{r-1}(x+u) \pmod p \\ & \equiv x^{q} \left(x^{p-1}-1\right)^{q} \prod_{u=0}^{r-1}(x+u) \pmod p \\ & \equiv x^{q} \sum_{m=0}^{q} (-1)^{q-m}{q \choose m} x^{m(p-1)} \prod_{u=0}^{r-1}(x+u) \pmod p \\ \sum_{k}{n\brack {k}}x^{k-q} & \equiv \left(\sum_{m=0}^{q} (-1)^{q-m}{q\choose m} x^{m(p-1)}\right)\left(\sum_{\ell=0}^{r}{r\brack {\ell}} x^{\ell}\right) \pmod p \\ & \equiv \sum_{k}\sum_{\underset{m(p-1)+\ell=k-q}{m,l}}(-1)^{q-m}{q \choose m}{r\brack {\ell}} x^{k-q}\pmod p. \\\end{aligned}$$ Then $$\begin{aligned} {n\brack {k}}&\equiv \sum_{\underset{m(p-1)+\ell=k-q}{m,l}}(-1)^{q-m}{q \choose m}{r\brack {\ell}} \pmod p. \end{aligned}$$ Now if $p$ divides $n$, we have $r=0$, and there may exist only one possible solution in non-negative integers $ \ell,m$ to the equation $ m(p-1)+\ell=k-q $ which may have a non-zero contribution to the sum on the right hand side of the above congruence: this is when $p-1$ divides $k-\frac{n}{p}$ and we have $\ell=0$ and $m=\frac{k-q}{p-1}$.\ Otherwise, $1\le r <p$ and since $\ell \le r$ and $r<p$, we have $\ell < p$. But also $\ell >0$, since ${r\brack 0}=0$, since $r>0$. Then, there exists at most one solution in non-negative integers $ \ell,m$ to the equation $ m(p-1)+\ell=k-q $. Indeed, let $\rho$ be the residue of the Euclidean division of $k-q$ by $p-1$. We have $0 \le \rho <p-1$.\ If $\rho=0$, then the unique solution is $\ell=p-1$ and $m= \frac{k-q}{p-1}-1$.\ If $0<\rho \le r$, then the unique solution is $\ell = \rho$ and $m= \frac{k-q-\rho}{p-1}$.\ And finally if $r<\rho <p-1$, there is no solution.\ Putting everything together, we have the following theorem for the $p$-congruence of the Stirling numbers of the first kind \[3.1\]Let $p$ be a prime number and $n, k$ non-negative integers such that $0\le k \le n$ and let $r$, (resp. $q$), be the residue, (resp. the quotient) of the Euclidean division of $n$ by $p$ and let $\rho$ be the residue of the Euclidean division of $k-q$ by $p-1$. We have $$\begin{aligned} \label{3.1} {n \brack k} \equiv (-1)^{q -\frac{k-q-j}{p-1}} {r \brack j} \binom{q}{\frac{k-q-j}{p-1}} \pmod p\end{aligned}$$ with $j=\rho+[\rho=0][r=p-1](p-1)$. [**Remark.**]{} Note that when $\rho=0$, we have ${n \brack k} \equiv 0 \bmod p$ unless when $r=p-1$ in which case, we have ${n \brack k} \equiv (-1)^{q+1 -\frac{k-q}{p-1}}\binom{q}{\frac{k-q}{p-1}-1} \bmod p$.\ \[3.1.1.c\] Let $p$ be a prime number, and $i,m$ and $s$ three natural numbers. We have $$\begin{aligned} \label{3.1.1} {m+s+m(p-1) \brack m+s +i(p-1)} &\equiv (-1)^{m-i}{m+\lfloor\frac{s}{p}\rfloor \choose i +\lfloor\frac{s}{p}\rfloor} \pmod p .\end{aligned}$$ We apply Theorem \[3.1\] with $n= mp+s$ and $k= m+s +i(p-1)$. We have $q= m+\lfloor\frac{s}{p}\rfloor$ and $k-q= s +i(p-1)-\lfloor\frac{s}{p}\rfloor=(i+\lfloor\frac{s}{p}\rfloor)(p-1)+r$. Then, when $s -p\lfloor\frac{s}{p}\rfloor=r <p-1$ we have $r=\rho<p-1$ and then $j=r$ and then Congruence reduces to Congruence . Otherwise $s -p\lfloor\frac{s}{p}\rfloor= r = p-1$ and then $\rho=0$ and then $j=p-1$ and then Congruence also reduces to Congruence . An identity involving both kinds of Stirling numbers ==================================================== Now, we present an identity involving binomial coefficients and the Stirling numbers of both kinds, which we believe is new. \[3.2\]Let $p$ be a positive integer and $n, k$ non-negative integers. We have $$\begin{aligned} \label{3.3} (-1)^{p-1}{n-1 \choose p-1}{n-p+1 \brack k}&= \sum_{i}(-1)^i {k-1+i \choose i}{i \brace p-1} {n \brack i+k}.\end{aligned}$$ [**Remark.**]{} It is interesting to compare this identity to Equation (6.28) in [@Graham94]. This is another three parameters identity but involving only the second kind of Stirling numbers, which is rather easily obtained from their exponential generating function. We reproduce it hereafter: under the condition that $\ell,m,n \ge 0$, we have $$\begin{aligned} \label{kn} {\ell+m \choose \ell}{n \brace \ell+m}= \sum_k {k\brace \ell}{n-k \brace m}{n \choose k}. \end{aligned}$$ If we replace $m$ by $-m$ and $n$ by $-n$ in Equation , taking into account that ${-n \choose k}=(-1)^k{n+k-1\choose k}$ and the known duality ${-a \brace -b}$ = ${b \brack a}$ (see [@Graham94]), we obtain our Equation with $p-1= \ell$. This shows that Equation (6.28) from [@Graham94] holds under the only condition $\ell \ge 0$, the conditions $m,n \ge 0 $ given in [@Graham94] are not needed.\  \ [*Proof of Theorem 3.1.*]{} The following proof, using the [*coefficient extractor*]{} method and the generating functions and is due to Marko Riedel [@Riedel20]. Let $S$ be the right hand side of . We have $$\begin{aligned} S&= \sum_{i=p-1}^{n-k}(-1)^i {k-1+i \choose i}{i \brace p-1} {n \brack i+k}\\ &= \sum_{i=p-1}^{n-k}(-1)^i {k-1+i \choose i}i![[z^i]]\frac{(e^z-1)^{p-1}}{(p-1)!} n![[w^n]]\frac{\left(\log\frac{1}{1-w} \right)^{i+k}}{(i+k)!}\\ &= \frac{n!}{(p-1)!(k-1)!} [[w^n]]\sum_{i=p-1}^{n-k}(-1)^i [[z^i]](e^z-1)^{p-1} \frac{\left(\log\frac{1}{1-w} \right)^{i+k}}{(i+k)}. \end{aligned}$$ Then by Lemma \[ce\], we have $$\begin{aligned} S&= \frac{(n-1)!}{(p-1)!(k-1)!}[[w^{n-1}]]\frac{1}{1-w}\sum_{i=p-1}^{n-k}(-1)^i[[z^i]](e^z-1)^{p-1} \left(\log\frac{1}{1-w} \right)^{i+k-1}\\ &= \frac{(-1)^{k-1}(n-1)!}{(p-1)!(k-1)!} [[w^{n-1}]]\frac{1}{1-w}\sum_{i=p+k-2}^{n-1}(-1)^{i}[[z^i]]z^{k-1}(e^z-1)^{p-1}\left(\log\frac{1}{1-w} \right)^{i}\\ &= \frac{(-1)^{k-1}(n-1)!}{(p-1)!(k-1)!}[[w^{n-1}]] \frac{1}{1-w}\sum_{i\ge p+k-2}\left(-\log\frac{1}{1-w} \right)^{i}[[z^i]]z^{k-1}(e^z-1)^{p-1}. \end{aligned}$$ since $\left(\log\frac{1}{1-w} \right)^{i}= w^i+ \cdot \cdot$.\  \ Now, we have $(e^z-1)^{p-1}=z^{p-1}+ \cdot \cdot$, then the lowest power of $z$ in the power series of $z^{k-1}(e^z-1)^{p-1}$ is $z^{p+k-2}$ and then\ $$\begin{aligned} S&= \frac{(-1)^{k-1}(n-1)!}{(p-1)!(k-1)!}[[w^{n-1}]] \frac{1}{1-w}\sum_{i\ge 0}\left(-\log\frac{1}{1-w} \right)^{i}[[z^i]]z^{k-1}(e^z-1)^{p-1}\\ &= \frac{(-1)^{k-1}(n-1)!}{(p-1)!(k-1)!} [[w^{n-1}]] \frac{1}{1-w}\left(-\log\frac{1}{1-w} \right)^{k-1}(e^{-\log\frac{1}{1-w}}-1)^{p-1}\\ &= \frac{(-1)^{k-1}(n-1)!}{(p-1)!(k-1)!} [[w^{n-1}]] \frac{1}{1-w}\left(-\log\frac{1}{1-w} \right)^{k-1}(-w)^{p-1}\\ &= \frac{(-1)^{p-1}(n-1)!}{(p-1)!(k-1)!} [[w^{n-1}]] \frac{1}{1-w}\left(\log\frac{1}{1-w} \right)^{k-1}w^{p-1}\\ &= \frac{(-1)^{p-1}(n-1)!}{(p-1)!(k-1)!} [[w^{n-p}]] \frac{1}{1-w}\left(\log\frac{1}{1-w} \right)^{k-1}\\ &= \frac{(-1)^{p-1}(n-1)!}{(p-1)!(k-1)!}(n-p+1) [[w^{n-p}]] \frac{1}{1-w}\frac{\left(\log\frac{1}{1-w} \right)^{k-1}}{n-p+1}\\ &= \frac{(-1)^{p-1}(n-1)!}{(p-1)!(k-1)!}(n-p+1) [[w^{n-p+1}]]\frac{\left(\log\frac{1}{1-w} \right)^{k}}{k} \ \ \ \ \text{\ by Lemma \ref{ce}}\\ &= \frac{(-1)^{p-1}(n-1)!}{(p-1)!(n-p)!}(n-p+1)! [[w^{n-p+1}]]\frac{\left(\log\frac{1}{1-w} \right)^{k}}{k!}\\ &= (-1)^{p-1}{n-1 \choose p-1}{n-p+1 \brack k}. \end{aligned}$$ As a corollary, we have another $p$-congruence involving the Stirling numbers of first kind. \[3.4\] Let $p$ be a prime number, and $n,k$, non-negative integers, the following congruence holds $$\begin{aligned} \label{3.4} \sum_{\underset{p-1 \vert i}{i>0}}{k-1+i \choose k-1}{n \brack i+k} &\equiv \big[p \text{ divides n}\big]{n-p+1 \brack k} \pmod p .\end{aligned}$$ Consider Equation when $p$ is prime. When $i=0$, we have ${i \brace p-1} =0$, and when $i>0$, by Lemma \[st2\], we have ${i \brace p-1} \equiv \big[p-1 \text{ divides } i\big] \bmod p$. Moreover $(-1)^{p-1} \equiv 1 \bmod p$, then, modulo $p$, the rhs of Equation is $$\sum_{\underset{p-1 \vert i}{i>0}}{k-1+i \choose k-1}{n \brack i+k}.$$ Now, for the lhs of Equation , when $p$ does not divide $n$, by Kummer theorem, whenever $n-p \equiv 0 \bmod p$ we have ${n-1 \choose p-1} \equiv 0 \pmod p$, since in this case the addition $(n-p)+(p-1)$ in base $p$ has at least one carry (at the lowest digit). And eventually, when $p$ divides $n$, we have ${n-1 \choose p-1} \equiv {-1 \choose p-1}=(-1)^{p-1} \equiv 1 \pmod p$. Back to binomial coefficients ============================= We can now prove our claim from the introduction. \[3.1.2\] Let $p$ be a prime number, and $m,\ell$ and $s$ three natural numbers such that $m>l$. We have $$\sum_{i\ge \ell+1}(-1)^{m-i} {m+\lfloor \frac{s}{p}\rfloor \choose i+\lfloor \frac{s}{p}\rfloor}{m+s-1+i(p-1) \choose m+s-1+\ell(p-1)}$$ $$\begin{aligned} \label{3.1.2}& \equiv \big[p \text{ divides s}\big](-1)^{m-1-\ell} {m-1+\frac{s}{p} \choose \ell+\frac{s}{p}}\pmod p . \end{aligned}$$ In particular, for any natural number $n$ such that $p$ does not divides $n$, we have $$\label{3.1.3} \sum_{i\ge \ell+1}(-1)^i {\lfloor \frac{n}{p}\rfloor p \choose i}{n-1+i(p-1) \choose n-1+\ell(p-1)} \equiv 0\pmod p .$$ If we substitute $m+s+\ell(p-1)$ for $k$ and $mp+s$ for $n$, Congruence becomes $$\begin{aligned} &\sum_{\underset{p-1 \vert i}{i>0}}{ m+s-1+\ell(p-1)+i \choose m+s-1+\ell(p-1)}{mp+s \brack i+ m+s+\ell(p-1)}\\ &\equiv \big[p \text{ divides s}\big]{(m-1)p+s+1 \brack m+s+\ell(p-1)} \pmod p . \end{aligned}$$ which, by the appropriate index change, is $$\begin{aligned} &\sum_{i\ge\ell +1}{ m+s-1+i(p-1)\choose m+s-1+\ell(p-1)}{mp+s \brack m+s+i(p-1)}\\ & \equiv \big[p \text{ divides s}\big]{(m-1)p+s+1 \brack m+s+\ell(p-1)} \pmod p . \end{aligned}$$ Finally, we make use of Corollary \[3.1.1.c\] and we have $$\begin{aligned} &\sum_{i\ge\ell +1}{ m+s-1+i(p-1)\choose m+s-1+\ell(p-1)} (-1)^{m-i}{m+\lfloor\frac{s}{p}\rfloor \choose i +\lfloor\frac{s}{p}\rfloor}\\ & \equiv \big[p \text{ divides s}\big](-1)^{m-1-\ell}{m-1+\lfloor\frac{s+1}{p}\rfloor \choose \ell +\lfloor\frac{s+1}{p}\rfloor} \pmod p \end{aligned}$$ which is the claim, since when $p$ divides $s$, we have $\lfloor\frac{s+1}{p}\rfloor=\frac{s}{p}$. Eventually, Congruence is obtained from by letting $s$ be the residue of the Euclidean division of $n$ by $p$ and substituting $\lfloor \frac{n}{p}\rfloor p $ for $m$. \ [**Acknowledgement**]{}: The author is indebted to Marko Riedel for the proof of Theorem (3.1). [9]{} R.L. Graham, D.E. Knuth and O. Patashnik, [*Concrete Mathematics*]{}, Adison-Wesley Publishing Company, 2nd Edition (1994). M. Riedel, A three-parameters identity involving Stirling numbers of both kinds, URL (version: 2020-03-24): https://math.stackexchange.com/q/3592354.

ArXiv

--- abstract: 'We study the behavior at infinity of ring $Q$-homeomorphisms with respect to $p$-modulus for $p>n$.' address: - | 01024, Ukraine\ Kiev-4,\ 3, Tereschenkivska st. - | 01024, Ukraine\ Kiev-4,\ 3, Tereschenkivska st. author: - Ruslan Salimov - Bogdan Klishchuk date: 'January 1, 2004' title: | On the Asymptotic Behavior\ of Ring $Q$-homeomorphisms\ with Respect to $P$-modulus --- Introduction ============ Let us recall some definitions, see [@Vai]. Let $\Gamma$ be a family of curves $\gamma$ in $\Bbb R^{n}$, $n\geqslant2$. A Borel measurable function $\rho:{\Bbb R^{n}}\to[0,\infty]$ is called [*admissible*]{} for $\Gamma$, (abbr. $\rho\in{\rm adm}\,\Gamma$), if $$\int\limits_{\gamma}\rho(x)\,ds\ \geqslant\ 1$$ for any curve $ \gamma\in\Gamma$. Let $p\in (1,\infty)$. The quantity $$M_p(\Gamma)\ =\ \inf_{\rho\in\mathrm{adm}\,\Gamma}\int\limits_{{\Bbb R^{n}}}\rho^p(x)\,dm(x)\,$$ is called [*$p$–modulus*]{} of the family $\Gamma$. For arbitrary sets $E$, $F$ and $G$ of $\Bbb R^{n}$ we denote by $\Delta(E, F, G)$ a set of all continuous curves $\gamma: [a, b] \rightarrow \Bbb R^{n}$ that connect $E$ and $F$ in $G$, i.e., such that $\gamma(a) \in E$, $\gamma(b) \in F$ and $\gamma(t) \in G$ for $a < t < b$. Let $D$ be a domain in $\Bbb R^{n}$, $n\geqslant2$, $x_0\in D$ and $d_0 = {\rm dist}(x_{0}, \partial D)$. Set $$\mathbb{A}(x_{0}, r_{1}, r_{2}) = \{ x \in \Bbb R^{n}: r_{1} < |x- x_{0}| < r_{2}\} \,,$$ $$S_{i}=S(x_0,r_{i})=\{x\in \Bbb R^{n}: \, |x-x_0| = r_{i}\}\,, \quad i = 1,\,2\,.$$ Let a function $Q: D\rightarrow [0,\infty]$ be Lebesgue measurable. We say that a homeomorphism $f: D \rightarrow \Bbb R^{n}$ is ring $Q$-homeomorphism with respect to $p$-modulus at $x_0 \in D$ if the relation $$M_p(\Delta(fS_{1}, fS_{2}, fD))\ \leqslant\ \int\limits_{{\mathbb{A}}}Q(x)\,\eta^{p}(|x-x_{0}|)\,dm(x)\,$$ holds for any ring $\mathbb{A} = \mathbb{A}(x_{0}, r_{1}, r_{2})$, $0< r_{1} < r_{2} < d_0$, $d_0 = {\rm dist}(x_{0}, \partial D)$, and for any measurable function $\eta: (r_{1}, r_{2}) \rightarrow [0,\infty]$ such that $$\int\limits_{r_{1}}^{r_{2}} \eta(r)\, dr = 1\,.$$ The theory of $Q$-homeomorphisms for $p=n$ was studied in works [@RS]–[@S1], for $1<p<n$ in works [@G1]–[@S6] and for $p>n$ in works [@SK1]–[@SK4], see also [@CR1], [@CR2]. Denote by $\omega_{n-1}$ the area of the unit sphere $\mathbb{S}^{n-1} = \{x\in \Bbb R^{n}: \, |x| = 1\}$ in $\Bbb R^{n}$ and by $q_{x_0}(r)= \frac{1}{\omega_{n-1}\, r^{n-1}}\int\limits_{S(x_0, r)}Q(x)\, d\mathcal{A}$ the integral mean over the sphere $S(x_0,r)=\{x\in \Bbb R^{n}: \, |x-x_0| = r\}$, here $d\mathcal{A}$ is the element of the surface area. Now we formulate a criterion which guarantees for a homeomorphism to be the ring $Q$-homeomorphisms with respect to $p$-modulus for $p>1$ in $\Bbb R^{n}$, $n\geqslant2$. [**Proposition 1.**]{} *Let $D$ be a domain in $\Bbb R^{n}$, $n\geqslant 2$, and let $Q:D\to[0,\infty]$ be a Lebesgue measurable function such that $q_{x_0}(r) \neq \infty$ for a.e. $r\in(0, d_0)$, $d_0 = {\rm dist}(x_{0}, \partial D)$. A homeomorphism $f: D \rightarrow \Bbb R^{n}$ is ring $Q$-homeomorphism with respect to $p$-modulus at a point $x_{0} \in D$ if and only if the quantity* $$M_{p}\left(\Delta(fS_{1}, fS_{2}, f\mathbb{A})\right) \leqslant \frac{\omega_{n-1}}{\left(\int\limits_{r_{1}}^{r_{2}}\frac{dr}{r^\frac{n-1}{p-1}\,q_{x_0}^{\frac{1}{p-1}}(r)}\right)^{p-1}}\,$$ holds for any $0 < r_{1} < r_{2} < d_0$ (see [@S2], Theorem 2.3). Following the paper [@MRV], a pair $\mathcal{E}=(A,C)$ where $A\subset\Bbb R^{n}$ is an open set and $C$ is a nonempty compact set contained in $A$, is called [*condenser*]{}. We say that a condenser $\mathcal{E}=(A,C)$ lies in a domain $D$ if $A\subset D$. Clearly, if $f:D\to\Bbb R^{n}$ is a homeomorphism and $\mathcal{E}=(A,C)$ is a condenser in $D$ then $(fA,fC)$ is also condenser in $fD$. Further, we denote $f\mathcal{E}=(fA,fC)$. Let $\mathcal{E}=(A,C)$ be a condenser. Denote by $\mathcal{C}_0(A)$ a set of continuous functions $u:A\to\mathbb{R}^1$ with compact support. Let $\mathcal{W}_0(\mathcal{E})=\mathcal{W}_0(A,C)$ be a family of nonnegative functions $u:A\to\mathbb{R}^1$ such that 1) $u\in \mathcal{C}_0(A)$, 2) $u(x)\geqslant1$ for $x\in C$ and 3) $u$ belongs to the class ${\rm ACL}$ and $$ u= (\_[i=1]{}\^[n]{}( )\^[2]{})\^. $$ For $p \geqslant1$ the quantity $${\rm cap}_p\,\mathcal{E}={\rm cap}_p\,(A,C)=\inf\limits_{u\in \mathcal{W}_0(\mathcal{E})}\, \int\limits_{A}\,\vert\nabla u\vert^p\,dm(x)\,$$ is called [*$p$-capacity*]{} of the condenser $\mathcal{E}$. It is known that for $p>1$ $$\label{EMC} {\rm cap}_p\,\mathcal{E}=M_p(\Delta(\partial A,\partial C; A\setminus C)),$$ see in ([@Sh],Theorem 1). For $p>n$ the inequality $$\label{eqks2.8} {\rm cap}_{p}\, (A,C) \geqslant n\, \Omega_n^{\frac{p}{n}}\, \left(\frac{p-n}{p-1}\right)^{p-1} \, \left[m^{\frac{p-n}{n(p-1)}}(A) - m^{\frac{p-n}{n(p-1)}}(C)\right]^{1-p}\,$$ holds where $\Omega_n$ is a volume of the unit ball in $\mathbb{R}^{n}$ (see, e.g., the inequality 8.7 in [@Maz]). Main results ============ Now we consider the main result of our paper on the behavior at infinity of ring $Q$-homeomorphisms with respect to $p$-modulus for $p>n$. The case $p=n$ was studied in the work [@SS1]. Let $$L(x_0, f, R) = \sup\limits_{|x - x_{0}|\leqslant R}\, |f(x) - f(x_{0})|\,.$$ Suppose that $f:\Bbb R^{n} \rightarrow \Bbb R^{n}$ is a ring $Q$-homeomorphism with respect to $p$-modulus at a point $x_0$ with $p>n$ where $x_0$ is some point in $\Bbb R^{n}$ and for some numbers $r_0 >0$, $K>0$ the condition $$\label{b2H} q_{x_{0}}(t) \leqslant K\,t^{\alpha}\,$$ holds for a.e. $t\in [r_0, +\infty)$. If $\alpha\in [0, p-n)$ then $$\varliminf\limits_{R \rightarrow \infty}\, \frac{L(x_0, f, R)}{R^{\frac{p-n-\alpha}{p-n}}} \geqslant K^{\frac{1}{n-p}}\, \left(\frac{p-n}{p-n-\alpha}\right)^{\frac{p-1}{p-n}} > 0\,.$$ If $\alpha = p-n$ then $$\varliminf\limits_{R \rightarrow \infty}\, \frac{L(x_0, f, R)}{\left(\ln R\right)^{\frac{p-1}{p-n}}} \geqslant K^{\frac{1}{n-p}}\, \left(\frac{p-n}{p-1}\right)^{\frac{p-1}{p-n}} > 0\,.$$ Consider a condenser $\mathcal{E}=\left(A, C\right)$ in $\Bbb R^{n}$, where $A=\{x\in \Bbb R^{n}: |x-x_0|< R\}$, $C=\{x\in \Bbb R^{n}: |x-x_0|\leqslant r_{0}\}$, $0< R< r_{0} <\infty$. Then $f\mathcal{E} = \left(fA,fC\right)$ is a ringlike condenser in $\Bbb R^{n}$ and by (\[EMC\]) we have equality $${\rm cap}_{p}\, f\mathcal{E}= \mathrm{M}_{p}\left(\Delta(\partial fA, \partial fC; f(A\setminus C))\right).$$ Due to the inequality (\[eqks2.8\]) $${\rm cap}_{p}\, (fA,fC) \geqslant n\, \Omega_n^{\frac{p}{n}}\, \left(\frac{p-n}{p-1}\right)^{p-1} \, \left[m^{\frac{p-n}{n(p-1)}}(fA) - m^{\frac{p-n}{n(p-1)}}(fC)\right]^{1-p}\,$$ we obtain $$\label{d3} {\rm cap}_{p}\, (fA,fC) \geqslant n\, \Omega_{n}^{\frac{p}{n}}\, \left(\frac{p-n}{p-1}\right)^{p-1}\, \left[m(fA) \right]^{\frac{n-p}{n}}\,.$$ On the other hand, by Proposition 1, one gets $$\label{d4} {\rm cap}_{p}\, (fA,fC) \leqslant \frac{\omega_{n-1}}{\left(\int\limits_{r_0}^{R}\frac{dt}{t^\frac{n-1}{p-1}\,q_{x_0}^{\frac{1}{p-1}}(t)} \right)^{p-1}}\,.$$ Combining the inequalities (\[d3\]) and (\[d4\]), we obtain $$n\, \Omega_{n}^{\frac{p}{n}}\, \left(\frac{p-n}{p-1}\right)^{p-1}\, \left[m(fA) \right]^{\frac{n-p}{n}} \leqslant \frac{\omega_{n-1}}{\left(\int\limits_{r_0}^{R}\frac{dt}{t^\frac{n-1}{p-1}\,q_{x_0}^{\frac{1}{p-1}}(t)} \right)^{p-1}}\,.$$ Due to $\omega_{n-1} = n\, \Omega_{n}$, the last inequality can be rewritten as $$\label{a7} %\begin{split} \Omega_{n}^{\frac{p}{n}- 1}\, \left(\frac{p-n}{p-1}\right)^{p-1}\, \left[m(fA) \right]^{\frac{n-p}{n}} \leqslant \left(\int\limits_{r_0}^{R}\frac{dt}{t^\frac{n-1}{p-1}\,q_{x_0}^{\frac{1}{p-1}}(t)} \right)^{1-p}\,. %\end{split}$$ Consider a case when $\alpha \in [0, p-n)$. Then from the condition (\[b2H\]) the estimate $$\Omega_{n}^{\frac{p}{n}- 1}\, \left(\frac{p-n}{p-1}\right)^{p-1}\, \left[m(fA) \right]^{\frac{n-p}{n}} \leqslant K\, \left(\frac{p-n-\alpha}{p-1}\right)^{p-1}\, \left(R^{\frac{p-n-\alpha}{p-1}} - r_0^{\frac{p-n-\alpha}{p-1}}\right)^{1-p}\,$$ holds. Therefore $$\label{a9} %\begin{split} m(fB(x_0, R)) \geqslant \Omega_{n}\, K^{\frac{n}{n-p}}\, \left(\frac{p-n}{p-n-\alpha}\right)^{\frac{n(p-1)}{p-n}}\, \left(R^{\frac{p-n-\alpha}{p-1}} - r_0^{\frac{p-n-\alpha}{p-1}}\right)^{\frac{n(p-1)}{p-n}}\,. %\end{split}$$ Due to $$\label{a10} %\begin{split} m(fB(x_0, R)) \leqslant \Omega_{n}\, L^{n}(x_0, f, R)\,, %\end{split}$$ from the inequality (\[a9\]) we have $$L(x_0, f, R) \geqslant K^{\frac{1}{n-p}}\, \left(\frac{p-n}{p-n-\alpha}\right)^{\frac{p-1}{p-n}}\, \left(R^{\frac{p-n-\alpha}{p-1}} - r_0^{\frac{p-n-\alpha}{p-1}}\right)^{\frac{p-1}{p-n}}\,.$$ Dividing the last inequality by $R^{\frac{p-n-\alpha}{p-n}}$ and taking the lower limit for $R \rightarrow\infty$, we conclude $$\varliminf\limits_{R \rightarrow \infty}\, \frac{L(x_0, f, R)}{R^{\frac{p-n-\alpha}{p-n}}} \geqslant K^{\frac{1}{n-p}}\, \left(\frac{p-n}{p-n-\alpha}\right)^{\frac{p-1}{p-n}}\,.$$ Now we consider a case when $\alpha= p-n$. Then from (\[a7\]) we get $$\Omega_{n}^{\frac{p}{n}- 1}\, \left(\frac{p-n}{p-1}\right)^{p-1}\, \left[m(fA) \right]^{\frac{n-p}{n}} \leqslant K\, \left(\ln\frac{R}{r_0}\right)^{1-p}\,.$$ Therefore $$m(fB(x_0, R)) \geqslant \Omega_{n}\, K^{\frac{n}{n-p}}\, \left(\frac{p-n}{p-1}\right)^{\frac{n(p-1)}{p-n}}\, \left(\ln\frac{R}{r_0}\right)^{\frac{n(p-1)}{p-n}}\,.$$ Due to the estimate (\[a10\]) we obtain $$L(x_0, f, R) \geqslant K^{\frac{1}{n-p}}\, \left(\frac{p-n}{p-1}\right)^{\frac{p-1}{p-n}}\, \left(\ln\frac{R}{r_0}\right)^{\frac{p-1}{p-n}}\,.$$ Finally, dividing the last inequality by $\left(\ln R\right)^{\frac{p-1}{p-n}}$ and taking the lower limit for $R \rightarrow\infty$, we conclude $$\varliminf\limits_{R \rightarrow \infty}\, \frac{L(x_0, f, R)}{\left(\ln R\right)^{\frac{p-1}{p-n}}} \geqslant K^{\frac{1}{n-p}}\, \left(\frac{p-n}{p-1}\right)^{\frac{p-1}{p-n}}\,.$$ This completes the proof of Main Theorem. Let us consider some examples. [*Example 2.1.*]{} Let $f_1:\mathbb{R}^{n} \to \mathbb{R}^{n}$, where $$f_1(x)=\begin{cases} K^{\frac{1}{n-p}} \left(\frac{p-n}{p-n-\alpha}\right)^{\frac{p-1}{p-n}}\, |x|^{\frac{p-n-\alpha}{p-n}}\, \frac{x}{|x|} \, ,& x \neq 0\\ 0,& x=0 \,.\end{cases}\, %\end{equation}$$ It can be easily seen that $\lim\limits_{x \rightarrow \infty}\, \frac{|f(x)|}{|x|^{\frac{p-n-\alpha}{p-n}}} = K^{\frac{1}{n-p}}\, \left(\frac{p-n}{p-n-\alpha}\right)^{\frac{p-1}{p-n}}$. Let us show that the mapping $f_1$ is a ring $Q$-homeomorphism with respect to $p$-modulus with the function $Q(x) = K\,|x|^{\alpha}$ at the point $x_{0}= 0$. Clearly, $q_{x_{0}}(t) = K\,t^{\alpha}$. Consider a ring $\mathbb{A}(0, r_{1}, r_{2})$, $0< r_{1} < r_{2} < \infty$. Note that the mapping $f_{1}$ maps the ring $\mathbb{A}(0, r_{1}, r_{2})$ onto the ring $\widetilde{\mathbb{A}}(0, \widetilde{r}_{1}, \widetilde{r}_{2})$, where $$%\begin{equation} %\label{a7} \widetilde{r}_{i} = K^{\frac{1}{n-p}}\, \left(\frac{p-n}{p-n-\alpha}\right)^{\frac{p-1}{p-n}}\, r_{i}^{\frac{p-n-\alpha}{p-n}}\,, \quad i = 1,\,2. %\end{equation}$$ Denote by $\Gamma$ a set of all curves that join the spheres $S(0,r_{1})$ and $S(0,r_{2})$ in the ring $\mathbb{A}(0, r_{1}, r_{2})$. Then one can calculate $p$-modulus of the family of curves $f_{1}\Gamma$ in implicit form: $$%\begin{equation} %\label{a8} \mathrm{M}_{p}(f_{1}\Gamma) = \omega_{n-1}\, \left(\frac{p-n}{p-1}\right)^{p-1}\, \left(\widetilde{r}_{2}^{\frac{p-n}{p-1}} - \widetilde{r}_{1}^{\frac{p-n}{p-1}}\right)^{1-p}\, %\end{equation}$$ (see, e.g., the relation (2) in [@Ge]). Substituting in the above equality the values $\widetilde{r}_{1}$ and $\widetilde{r}_{2}$, defined above, one gets $$\mathrm{M}_{p}(f_{1}\Gamma) = \omega_{n-1}\, K\, \left(\frac{p-n-\alpha}{p-1}\right)^{p-1}\, \left(r_{2}^{\frac{p-n-\alpha}{p-1}} - r_{1}^{\frac{p-n-\alpha}{p-1}}\right)^{1-p}\,.$$ Note that the last equality can be written by $$%\begin{equation} %\label{a30} \mathrm{M}_{p}(f_{1}\Gamma) = \frac{\omega_{n-1}}{\left(\int\limits_{r_{1}}^{r_{2}}\frac{dt}{t^\frac{n-1}{p-1}\,q_{x_0}^{\frac{1}{p-1}}(t)}\right)^{p-1}}\,, %\end{equation}$$ where $q_{x_0}(t) = K\,t^{\alpha}$. Hence, by Proposition 1, the homeomorphism $f_{1}$ is a ring $Q$-homeomorphism with respect to $p$-modulus for $p>n$ with the function $Q(x) = K\,|x|^{\alpha}$ at the point $x_{0}= 0$. [*Example 2.2.*]{} Let $\alpha= p-n$ and $f_2:\mathbb{R}^{n} \to \mathbb{R}^{n}$, where $$f_2(x)=\begin{cases} K^{\frac{1}{n-p}} \left(\frac{p-n}{p-1}\right)^{\frac{p-1}{p-n}}\, \left(\ln|x|\right)^{\frac{p-1}{p-n}}\, \frac{x}{|x|} \, ,& x \neq 0\\ 0,& x=0 \,.\end{cases}\, %\end{equation}$$ It can be easily seen that $\lim\limits_{x \rightarrow \infty}\, \frac{|f(x)|}{\left(\ln|x|\right)^{\frac{p-1}{p-n}}} = K^{\frac{1}{n-p}}\, \left(\frac{p-n}{p-1}\right)^{\frac{p-1}{p-n}}$. By analogy with Example 2.1, we can show that the mapping $f_2$ is a ring $Q$-homeomorphism with respect to $p$-modulus with the function $Q(x) = K\,|x|^{p-n}$. [*Remark 2.1.*]{} Examples 2.1 and 2.2 show that the estimates in Main Theorem are exact, i.e. are attained on the above mappings. This work was supported by the budget program “Support of the development of priority trends of scientific researches”(KPKVK6541230). [1]{} -2pt J. Väisälä, [*Lectures on n-dimensional quasiconformal mappings.*]{} Lecture Notes in Math., [**229**]{}, Springer-Verlag, Berlin, 1971. V. I. Ryazanov, E. A. Sevost’yanov, [*Equicontinuous classes of ring $Q$-homeomorphisms.*]{} Siberian Mathematical Journal **48**:6 (2007), 1093–1105. O. Martio, V. Ryazanov, U. Srebro, E. Yakubov, [*Q-homeomorphisms.*]{} Complex analysis and dynamical systems Contemp. Math. **364** (2004), Amer. Math. Soc., Providence, RI, 193–203. O. Martio, V. Ryazanov, U. Srebro, E. Yakubov, [*On Q-homeomorphisms.*]{} Ann.Acad. Sci. Fenn. Math. **30**:1 (2005), 49–69. O. Martio, V. Ryazanov, U. Srebro, E. Yakubov, [*Moduli in modern mapping theory.*]{} Springer Math. Monogr., New York, 2009. R. Salimov, [*ACL and differentiability of a generalization of quasiconformal maps.*]{} Izvestiya: Mathematics **72**:5 (2008), 977–984. A. Golberg, [*Differential properties of $(\alpha, Q)$-homeomorphisms.*]{} Further Progress in Analysis, Proc. 6th ISAAC Congr (2009), 218–228. A. Golberg, [*Integrally quasiconformal mappings in space.*]{} Transactions of Institute of Mathematics, the NAS of Ukraine **7**:2 (2010), 53–64. A. Golberg, R. Salimov, [*Logarithmic Holder continuity of ring homeomorphisms with controlled p-module.*]{} Complex Variables and Elliptic Equations **59**:1 (2014), 91–98. A. Golberg, R. Salimov, E. Sevost’yanov, [*Distortion estimates under mappings with controlled p-module.*]{} Ann. Univ. Bucharest, Ser. Math 5 (LXIII) (2014), P. 95–114. R. Salimov, [*On finitely Lipschitz space mappings.*]{} Siberian Elecronic Mathematical Reports **8** (2011), 284–295. R. Salimov, [*Estimation of the measure of the image of the ball.*]{} Siberian Mathematical Journal **53**:4 (2012), 920–930. R. Salimov, [*To a theory of ring $Q$-homeomorphisms with respect to a $p$-modulus.*]{} Ukrainian Mathematical Bulletin **10**:3 (2013), 379–396. R. Salimov, [*Оne property of ring $Q$-homeomorphisms with respect to a $p$-module.*]{} Ukrainian Mathematical Journal **65**:5 (2013), 728–733. R. Salimov, B. Klishchuk, [*The extremal problem for the area of an image of a disc.*]{} // Reports of the NAS of Ukraine **10** (2016), 22–27. B. Klishchuk, R. Salimov, [*Lower bounds for the area of the image of a circle.*]{} Ufa Mathematical Journal **9**:2 (2017), 55–61. R. Salimov, B. Klishchuk, [*Extremal problem for the area of the image of a disk.*]{} Zapiski Nauchnykh Seminarov POMI **456** (2017), 160–171. R. Salimov, B. Klishchuk, [*An extremal problem for the volume functional.*]{} Matematychni Studii **50**:1 (2018), 36–43. B. Klishchuk, R. Salimov, [*Lower bounds for the volume of the image of a ball.*]{} Ukrainian Mathematical Journal **71**:6 (2019), 774–785. M. Cristea, [*Some properties of open, discrete generalized ring mappings.*]{} Complex Variables and Elliptic Equations **61**:5 (2016), 623–643. M. Cristea, [*The limit mapping of generalized ring homeomorphisms.*]{} Complex Variables and Elliptic Equations **61**:5 (2016), 608–622. R. Salimov, E. Smolovaya, [*On the order of growth of ring Q-homeomorphisms at infinity.*]{} Ukrainian Mathematical Journal **62**:6 (2010), 829–836. O. Martio, S. Rickman, and J. Väisälä, [*Definitions for quasiregular mappings.*]{} Ann. Acad. Sci. Fenn. Ser. A1. Math. **448** (1969), 1–40. V. A. Shlyk, [*The equality between $p$-capacity and $p$-modulus.*]{} Siberian Mathematical Journal **34**:6 (1993), 216–221. V. Maz’ya, [*Lectures on isoperimetric and isocapacitary inequalities in the theory of Sobolev spaces.*]{} Contemp. Math. **338** (2003), 307–340. F. W. Gehring, [*Lipschitz mappings and the $p$-capacity of ring in $n$-space.*]{} Advances in the theory of Riemann surfaces (Proc. Conf. Stonybrook, N.Y., 1969), Ann. of Math. Studies **66** (1971), 175–193.

ArXiv

--- abstract: 'While observations of many high-precision radio pulsars of order $\lesssim1~\mu$s across the sky are needed for the detection and characterization of a stochastic background of low-frequency gravitational waves (GWs), sensitivity to single sources of GWs requires even higher timing precision. The Argentine Institute of Radio Astronomy (IAR; Instituto Argentino de Radioastronomía) has begun observations of the brightest-known millisecond pulsar, J0437$-$4715. Even though the two antennas are smaller than other single-dish telescopes previously used for pulsar timing array (PTA) science, the IAR’s capability to monitor this pulsar daily coupled with the pulsar’s brightness allows for high-precision pulse arrival-time measurements. While upgrades of the facility are currently underway, we show that modest improvements beyond current plans will provide IAR with unparalleled sensitivity to this pulsar. The most stringent upper limits on single GW sources come from the North American Nanohertz Observatory for Gravitational Waves (NANOGrav). Observations of [PSR J0437$-$4715]{}will provide a significant sensitivity increase in NANOGrav’s “blind spot” in the sky where fewer pulsars are currently being observed. With state-of-the-art instrumentation installed, we estimate the array’s sensitivity will improve by a factor of $\approx$2-7 over 10 years for 20% of the sky with the inclusion of this pulsar as compared to a static version of the PTA used in NANOGrav’s most recent limits. More modest instrumentation result in factors of $\approx$1.4-4. We identify four other candidate pulsars as suitable for inclusion in PTA efforts. International PTA efforts will also benefit from inclusion of these data given the potential achievable sensitivity.' author: - 'M.T.Lam' - 'J.S.Hazboun' title: 'Precision Timing of [PSR J0437$-$4715]{}with the IAR Observatory and Implications for Low-Frequency Gravitational Wave Source Sensitivity' --- Introduction {#sec:intro} ============ The search for gravitational waves (GWs) using an array of recycled millisecond pulsars (MSPs) is a key science goal for the largest radio telescopes in the world. Pulsars act as high-precision clocks and the passage of a GW along the line of sight is expected to cause slight variations in the arrival times of their observed emission that can be measured using high-precision pulsar timing [@Sazhin1978; @Detweiler1979; @fb1990]. Large telescopes are needed to measure high signal-to-noise ratio (S/N) pulses and estimate their arrival times to high precision [@Ransom+2019]. While current pulsar-timing efforts are dominated by large single-dish (60- to 300-m-class) telescopes, the use of larger, more sensitive telescopes and interferometers is becoming an increasingly important contribution to pulsar timing array (PTA) science [@IPTADR2]. The expected first detection of low-frequency ($\sim$nHz- $\mu$Hz) GWs is from a stochastic background of unresolved supermassive black hole binaries [e.g., @Rosado+2015]. Detection and characterization of this background require observations of many pulsars to build pairwise quadrupolar correlations in the observed arrival-time perturbations. The best-timed MSPs are used in these searches, with timing precision $\lesssim$1 $\mu$s. Individual sources of GWs, such as from a single resolved supermassive black hole binary inspiral or merger, require MSPs with precision an order-of-magnitude or better such that specific waveforms can be recovered [e.g., @blf2011]. Observing many well-timed pulsars both increases the signal-to-noise ratio of a GW detection and future characterization, and improves upon the coverage of events across the sky. The North American Nanohertz Observatory for Gravitational Waves ---------------------------------------------------------------- The North American Nanohertz Observatory for Gravitational Waves (NANOGrav; @McLaughlin2013 [@Ransom+2019]) collaboration is working towards the detection and characterization of low-frequency GWs by means of monitoring an array of nearly 80 precisely-timed MSPs using three telescopes: the 305-m William E. Gordon Telescope of the Arecibo Observatory (Arecibo), the 100-m Robert C. Byrd Green Bank Telescope of the Green Bank Observatory (GBO), and the 27-element 25-m antennas of the National Radio Astronomy Observatory’s (NRAO) Karl G. Jansky Very Large Array (VLA). NANOGrav has already placed upper limits on the amount of GWs in the Universe, which have provided strong astrophysical constraints on merging galaxies and cosmological models [@NG11GWB; @NG11CW; @NG11BWM; @3C66B]. NANOGrav is poised to detect and begin characterization of the stochastic GW background from an ensemble of unresolved supermassive black hole binaries (SMBHBs) within the next 3-5 years (@Rosado+2015 [@Taylor+2016; @NG11GWB]; Arzoumanian et al. in prep). The first detection of a continuous-wave (CW) source of GWs from a single SMBHB is expected by the late 2020s [@Rosado+2015; @Mingarelli+2017; @Kelley+2018; @NG11CW]. In addition, NANOGrav is sensitive to the direct signature of SMBHB mergers, known as a “burst with memory” (BWM; @NG11BWM). The Argentine Institute for Radio Astronomy ------------------------------------------- The Argentine Institute of Radio Astronomy (IAR; Instituto Argentino de Radioastronomía) has begun upgrades of two 30-meter antennas primarily for the purpose of pulsar observations [@IAR]. Located at latitude $\approx -37^\circ$, the observatory covers declinations $\delta < -10^\circ$. The two antennas, A1 and A2, are 30-m equatorial-mount telescopes separated by 120 m. Both are capable of observations around 1.4 GHz, each with current specifications listed in Table \[table:A1A2\]. While 30-meter antennas fall below the typical size of radio telescopes used in precision timing experiments [e.g., @IPTADR2], the IAR Observatory has several advantages for high-precision timing, specifically in the area of GW detection and characterization. With access to the Southern sky below $\delta < -10^\circ$ and many galactic-disk pulsars, and the ability to track objects for nearly four hours, IAR can and has demonstrated the capability of high-cadence timing on a number of those pulsars, specifically of the bright millisecond pulsar [J0437$-$4715]{}[@IAR]. Their preliminary observations suggest that future measurements will be a significant contributor to PTA sensitivity. [l|cc]{} Maximum Tracking Time ($T_{\rm obs}$) & 3 hr 40 min & 3 hr 40 min\ Circular Polarizations ($N_{\rm pol}$) & 1 & 2\ Receiver Temperature ($T_{\rm rcvr}$) & 100 K & 110 K\ Gain ($G$) & 11.9 Jy/K & 13.0 Jy/K\ Frequency Range (MHz) & 1100–1510 & 1200–1600\ Bandwidth ($B$) & 112 MHz & 56 MHz \[table:A1A2\] The Millisecond Pulsar [J0437$-$4715]{} --------------------------------------- Discovered in the Parkes Southern Pulsar Survey [@Johnston+1993], [PSR J0437$-$4715]{}is the brightest-known MSP and has the lowest dispersion measure (DM; the integrated line-of-sight electron density) of any MSP; it is the second-lowest of any pulsar [@PSRCAT]. It has a close distance of $156.8 \pm 0.3$ pc measured from the derivative of the binary orbital period [@Reardon+2016]. As such, high S/N pulses can be precisely timed, making this pulsar an excellent addition to PTAs. Some pulsar properties are listed in Table \[table:J0437\]. In addition to the ability to measure high S/N pulses, given its low DM, we expect that unmodeled chromatic (i.e., radio-frequency-dependent) interstellar propagation effects are reduced for this pulsar [@cs2010], thereby limiting one of the significant sources of timing noise found in many other pulsars. While the pulsar has measurable achromatic noise consistent with rotational spin fluctuations [@sc2010; @NG9EN], its high-precision timing nonetheless allows it to be used in tests for deterministic GW signals. In this paper, we demonstrate the unique capabilities of the IAR to observe [PSR J0437$-$4715]{}and provide unparalleled sensitivity to the pulsar for the purpose of single-source CW detection and future characterization. In §\[sec:sensitivity\], we provide the framework for our time-of-arrival (TOA) sensitivity estimates. We consider the current IAR observations of [PSR J0437$-$4715]{}in §\[sec:current\_obs\] and then extrapolate to future possible operating modes in §\[sec:future\_obs\], comparing the modes to observations of the pulsar by other international observatories as well. In §\[sec:projections\], we demonstrate how the IAR will improve sensitivity to single-source CWs when the data are combined with that of NANOGrav. As the collaboration observing the most number of pulsars [@Ransom+2019] and with the most recent and stringent CW limits [@NG11CW; @3C66B], we will focus our analyses primarily on improving upon a [*projected*]{} NANOGrav data set with observations of [PSR J0437$-$4715]{}, noting that upcoming analyses (e.g., with @NG12 [@NG12WB], which already include more time baseline and more pulsars) imply greater sensitivity in the future. We describe other possible pulsar targets in §\[sec:other\_MSPs\] and conclude in §\[sec:conclusion\]. The Sensitivity toward Continuous Waves {#sec:sensitivity} ======================================= In this section, we describe the mathematical framework for determining the precision of an individual pulsar and then the sensitivity toward CW sources from the sum of correlated GW signals between pulsar pairs. As the components of this framework have been discussed significantly throughout the literature, we will lay out each piece but only briefly, and will point the reader to other references with more detailed discussion. Single-Pulsar Timing Sensitivity -------------------------------- [l|c]{} \[table:J0437\] Spin Period$^{\rm a}$ ($P$) & 5.76 ms\ Dispersion Measure$^{\rm a}$ (DM) & 2.64 pc cm$^{-3}$\ Effective Width$^{\rm b}$ (${W_{\rm eff}}$) & 667.6 $\mu$s\ Flux Density at 400 MHz$^{\rm a}$ ($S_{400}$, ) & 550.0 mJy\ Flux Density at 1000 MHz$^{\rm c}$ ($S_{400}$) & 223 mJy\ Flux Density at 1400 MHz$^{\rm a}$ ($S_{1400}$) & 160.0 mJy\ Spectral Index$^{\rm c}$ ($\alpha$) & $-$0.99\ Jitter rms for 1 hour at 730 MHz ($\sigma_{\rm J,730}$) & $61 \pm 9$ ns\ Jitter rms for 1 hour at 1400 MHz ($\sigma_{\rm J,1400}$) & $48.0 \pm 0.6$ ns\ Jitter rms for 1 hour at 3100 MHz ($\sigma_{\rm J,3100}$) & $41 \pm 2$ ns\ Phase Bins (${N_\phi}$, assumed) & 512\ Median $\sigma_{\rm S/N}$ for Antenna 1 ($\sigma_{\rm S/N, A1}$) & 200 ns\ Median $\sigma_{\rm S/N}$ for Antenna 2 ($\sigma_{\rm S/N, A2}$) & 330 ns\ Median S/N for Antenna 1$^{\rm c}$ ($\tilde{S}_{\rm A1}$) & 148.8\ Median S/N for Antenna 2$^{\rm c}$ ($\tilde{S}_{\rm A2}$) & 90.5\ We use the framework of @optimalfreq to estimate the noise contributions to pulsar TOA uncertainties, or equivalently how precisely we can measure the TOAs. @IAR show timing residuals (the observed TOAs minus a model for the expected arrival times) for [PSR J0437$-$4715]{}and demonstrate excess noise beyond the uncertainties derived from the common practice of matching a pulse template to the observed pulsar profile to estimate their TOAs [@Taylor1992]. ### Short-timescale Noise {#sec:shorttimenoise} @NG9WN describe a model for white noise in timing residuals on short timescales, $\lesssim 1$ hr, i.e., the time of typical observations. In that analysis, they assume perfect polarization calibration and radio-frequency-interference (RFI) removal; the former is currently impossible in the case of A1 with only one measured polarization channel. However, assuming both are true, the three white-noise components described are the template-fitting, jitter, and scintillation noise terms. Template-fitting errors rely on the matched-filtering assumption that the observed pulse profiles are an exact match to a template shape with some additive noise. These are the [*minimum*]{} possible errors for TOA estimates, and are given by \_[S/N]{}(S) = , \[eq:TFerror\] where ${W_{\rm eff}}$ is the effective width of the pulse, ${N_\phi}$ is the number of samples (bins) across the profiles, and $S$ is the signal-to-noise ratio (written this way in equations for clarity), taken to be the peak to off-pulse rms. The effective width is related to the spin period of the pulsar $P$ and the template profile normalized to unit height, $U(\phi)$, by [W\_[eff]{}]{}= . \[eq:Weff\] Pulse profiles are frequency-dependent and so this error will apply to each profile measured at a given frequency. However, given the low-bandwidths currently available at the IAR Observatory, it is sufficient given the S/N-regime of the observations to average the data across not only time but frequency as well to obtain a single TOA per epoch per antenna. Similarly, many other frequency-dependent variations within the band will be small and we will ignore those here. While ${W_{\rm eff}}$ and ${N_\phi}$ are constants, the pulse S/N will vary on short timescales due to diffractive scintillation, with a known probability density function (PDF) written as [@cc1997; @NG9WN] f\_S(S S\_0, [n\_]{}) = e\^[-S[n\_]{}/S\_0]{}(S), \[eq:sn\_pdf\] where $S_0$ is the mean S/N, $\Gamma$ is the Gamma function, $\Theta$ is the Heaviside step function, and ${n_{\mathrm{ISS}}}$ is the number of scintles in the observation of length $T$ and bandwidth $B$, given by [n\_]{}(1+\_t )(1+\_). \[eq:N\_DISS\] The scintillation parameters ${\Delta t_{\mathrm{d}}}$ and ${\Delta \nu_{\mathrm{d}}}$ describe the characteristic timescale and bandwidth of intensity maxima, or scintles, in a dynamic spectrum. The $\eta_t$ and $\eta_\nu$ parameters are the scintle filling factors $\approx 0.2$ [@cs2010; @Levin+2016]. @Gwinn+2006 found two diffractive scintillation scales for [PSR J0437$-$4715]{}made at an observing frequency of 328 MHz. The first has ${\Delta t_{\mathrm{d}}}= 1000$ s and ${\Delta \nu_{\mathrm{d}}}= 16$ MHz while the second has ${\Delta t_{\mathrm{d}}}= 90$ s and ${\Delta \nu_{\mathrm{d}}}= 0.5$ MHz. For a Kolmogorov medium, we can scale these quantities as ${\Delta t_{\mathrm{d}}}\propto \nu^{6/5}$ and ${\Delta \nu_{\mathrm{d}}}\propto \nu^{22/5}$ [@cr1998; @optimalfreq]. Therefore, when scaled to an observing frequency of 1.4 GHz, we have ${\Delta t_{\mathrm{d}}}= 5700$ s and ${\Delta \nu_{\mathrm{d}}}= 9.5$ GHz for the first scale, and ${\Delta t_{\mathrm{d}}}= 510$ s and ${\Delta \nu_{\mathrm{d}}}= 300$ MHz for the second scale. @Keith+2013 measured the scintillation parameters at 1.5 GHz for [PSR J0437$-$4715]{}; when scaled to 1.4 GHz, they found ${\Delta t_{\mathrm{d}}}= 2290$ s and ${\Delta \nu_{\mathrm{d}}}= 740$ MHz, in between the two scales observed by @Gwinn+2006. Using these measurements of the scintillation parameters, we will estimate ${n_{\mathrm{ISS}}}$ and thus the predicted impact on the PDF of $S$. Given the current small bandwidths of the receiver, since $B < {\Delta \nu_{\mathrm{d}}}$ in all cases, the second of the two components in Eq. \[eq:N\_DISS\] will be $\approx 1$ and so we have ${n_{\mathrm{ISS}}}\approx 1 + \eta_t T/{\Delta t_{\mathrm{d}}}$. For $T = 13200$ s, this quantity will range between 1.5 and 5.2. The higher ${n_{\mathrm{ISS}}}$ is, the more that the PDF will tend towards the mean S/N value $S_0$, whereas when ${n_{\mathrm{ISS}}}$ is close to 1, the distribution becomes exponential. Therefore, this value will heavily dictate the timing uncertainties achievable, which we will briefly discuss when applying to the real data. The two other short-timescale contributions to the white noise discussed in @NG9WN are the jitter and scintillation-noise terms; the latter is separate from the S/N change due to scintillation. Both cause stochastic deviations to the observed pulse shapes such that the matched-filtering assumption of template fitting no longer applies. Jitter, due to single-pulse stochasticity, becomes a significant noise contribution for well-timed, high-S/N MSPs [@NG12WN]. @sod+14 measured the timing uncertainty due to jitter for an hour-long observation to be $\sigma_{\rm J} = 48.0 \pm 0.6$ ns, significantly less than the template-fitting errors shown in @IAR. In addition, since the rms jitter scales inversely with the square root of the number of pulses, this contribution to the TOA uncertainty for a 3 hr 40 min observation will be even smaller in IAR data, and therefore can be ignored. Scintillation noise is caused by the finite-scintle effect and results in pulse-shape stochasticity due to an imperfectly known [*pulse broadening function*]{} [@cwd+90; @cs2010]. Its maximum value is approximately the scattering timescale ${\tau_{\mathrm{d}}}$, which is inversely proportional to ${\Delta \nu_{\mathrm{d}}}$ by [@cr1998; @cs2010] [\_]{}= , where $C_1 \approx 1$ is a coefficient that depends on the geometry of the intervening medium. Since scattering tends to be more significant at higher DMs [@Bhat+2004], we already expect ${\tau_{\mathrm{d}}}$ to be small for [PSR J0437$-$4715]{}. For the minimum value of ${\Delta \nu_{\mathrm{d}}}$ above, we find that maximum value of ${\tau_{\mathrm{d}}}\lesssim 1$ ns, and therefore scintillation noise can also be ignored for this pulsar. ### Dispersion Measure Estimation Estimation of DM and the subsequent correction of dispersion is critical for proper precision timing. Since the scattering timescale is small, the dominant component to DM estimation errors is that due to the white noise described above [@optimalfreq]. For measurements taken at two frequencies $\nu_1$ and $\nu_2$, with the frequency ratio defined as $r \equiv \nu_2 / \nu_1 \ge 1$, the timing uncertainty due to DM estimation is [see Appendix A of @DMnu_response] \_[[$\widehat{\mathrm{DM}}$]{}]{} = , \[eq:DMerr\] where the $\sigma_\nu$ values are the timing uncertainties at each frequency. Both @cs2010 and @optimalfreq describe the matrix formalism for calculating this uncertainty when multi-channel measurements are available. Estimating the DM with measurements taken at multiple frequencies results in reduced uncertainties even if the covered range $r$ is the same. If DM variations are not accounted for in a timing model, as in @IAR, then additional uncertainties arise. Many types of effects give rise to both stochastic and systematic variations in DM [@DMt]. For example, the Earth-pulsar line of sight passing close to the Sun will cause an increase in the DM from the increased electron density of the solar wind [@You+07b; @NG11SW; @Tiburzi+2019]; while the motion through this electron density profile gives rise to systematic variations, any fluctuations in the solar wind will cause variations that are random in nature. The turbulent interstellar medium will also cause rise to DM fluctuations that are random though correlated over time [@fc90; @pw91]. The turbulent medium is typically parameterized by a power-law electron density wavenumber spectrum over many orders of magnitude [@Armstrong+1995]. An equivalent formulation that one can derive from this spectrum is the DM structure function [e.g., @DMt], given by D\_[[$\mathrm{DM}$]{}]{}() \^2, where $\tau$ is the time lag and the brackets denote the ensemble average. The structure function of the timing perturbations $\delta t$ is related to the DM structure function by D\_[[[t]{}]{}]{}() = K\^2 \^[-4]{} D\_[$\mathrm{DM}$]{}(), where $K \approx 4.149 \times 10^3~\mu$s GHz$^2$ pc$^{-1}$ cm$^3$ is the dispersion constant in observationally convenient units, i.e., for a DM of 1 pc cm$^{-3}$ and radio emission at 1 GHz, we expect 4149  $\mu$s of delay. This is a significantly larger delay than the achievable timing precision since DM [*estimates*]{} are known much more precisely and therefore the delay can be corrected for to high accuracy. We can directly relate the rms timing fluctuations for a given time lag to the structure function by \_[[t]{}]{}() = \^[1/2]{}. \[eq:sigma\_tau\] For a Kolmogorov turbulent spectrum, we can write the structure function in the power-law form [@DMt] D\_[[t]{}]{}() = ()\^[5/3]{} = 0.0253 \^2 \_[GHz]{}\^[-2]{} ()\^[5/3]{}. For [PSR J0437$-$4715]{}, we will first consider the middle value of ${\Delta t_{\mathrm{d}}}$ provided by @Keith+2013, with ${\Delta t_{\mathrm{d}}}= 2290$ s. On the timescale of the $\tau \approx$100 days as shown in @IAR, the rms timing perturbations for unaccounted for DM variations would be approximately 110 ns at 1.4 GHz and so should be a negligible part of their error budget shown. Nonetheless, if consider this error for longer timespans, as for 1000 and 10000 days, the rms perturbations are 740 ns and 5.0 $\mu$s, respectively. Note that this is not the total rms variations here, only those on the timescale of $\tau$ but find the estimate instructive nonetheless. The full rms can be obtained by relating the structure function to the power spectrum of the timing perturbations [@DMt] and integrating to find the total variance. ### Additional Sources of Uncertainty @IAR discuss the RFI environment around the Observatory, including the amplitude and approximate impact on the pulsar observations. While the effect appears to be minimal, RFI is a growing concern at all observatories worldwide. IAR’s proximity to the urban center of Buenos Aires means that active real-time mitigation will need to be implemented in the future to reduce any growing impact on the error budget. Also importantly for the IAR Observatory, polarization miscalibration can be a significant contribution to the timing uncertainty, including when both polarizations are measured. @IAR discuss the development of current and future polarization calibration routines as a goal for the IAR. In addition to the TOA uncertainty estimates, [PSR J0437$-$4715]{}is known to show red noise in its timing residuals [e.g., @Kerr+2020], consistent with rotational spin noise [@sc2010; @NG9EN]. While red noise will impact the sensitivity towards CWs, the use of matched filtering in detection analyses coupled with the requirement to observe a correlated signal in multiple pulsars means that individual red noise is of less critical significance in CW analyses versus stochastic-background analyses. Pulsar GW Sensitivities ----------------------- GW detection relies on measuring the signature of TOA perturbations amongst pulsars. In the case of a stochastic background, for instance from the unresolvable sum of signals from SMBHBs, one relies on cross-correlations between these perturbations. For individual SMBHBs, and other single-source signals, one uses a deterministic model. In both cases the sky location and sensitivity of individual pulsars affect the ability of the full PTA to detect gravitational waves. We use the framework discussed in @Hazboun+2019 and @optimalobs to estimate the sensitivity to single-sources, which we describe here briefly. While the ability to claim a detection of GWs from a single SMBHB requires more than one pulsar, the sensitivity to a single source depends on the signal strength and noise in each individual pulsar. These signals are then added up across the network to build a robust detection. There a number of statistics defined in the literature to search for single-source GWs, [@Babak+2012; @Ellis+2012a; @Ellis+2012b; @Taylor+2016]. Here we focus on assessing and optimizing the sensitivity of PTAs [@optimalobs; @Hazboun+2019] given different pulsars and changes in the observing strategies. The framework of @Hazboun+2019 uses a match-filter statistic tailored to studying sky-dependent sensitivity. The [S/N]{}, $\rho(\hat k)$, is dependent on the response of an individual pulsar’s sky position and noise characteristics but has been averaged over inclination angle and GW polarization \^2(k)\_[inc]{} = 2 T\_[obs]{}\_[f\_[L]{}]{}\^[f\_[H]{}]{} df \_i . \[e:rho2(k)\] Here $S_h(f)$ is the strain power spectral density (PSD) of a monochromatic SMBHB signal, $T_i/T_{\rm obs}$ is the fraction of total observation time for the PTA covered by a particular pulsar labeled by $i$, and ${S_i(f,\hat k)}$ is the strain-noise PSD for a particular pulsar. The dependence on sky location, $\hat k$, comes from the quadrupolar respone function of pulsars to GWs. This [S/N]{} shows that the signals from individual pulsars add independently, but this really only tells part of the story. While the quantity $\rho(\hat k)$ captures the sensitivity of a PTA to individual sources one would need a significant signal in a few pulsars in order to claim a detection. The PSD ${S_i(f,\hat k)}$ is related to the usual noise spectra, written in @optimalobs as the sum of white noise, red noise, and the stochastic GW background, P\_i(f) = P\_[[$\mathrm{W}$]{},i]{}(f) + P\_[,i]{}(f) + P\_(f), \[eq:noise\_spectra\] where the stochastic background term takes a power law form of $P_{{\mathrm{SB}}}(f) \propto f^{-\beta}$ with $\beta = 13/3$ for an ensemble of SMBHBs [@Jenet+2006]. The PSD ${S_i(f,\hat k)}$ also takes into account the timing model fit of the pulsar, through the inverse noise-weighted transmission function and the response function, which maps the strain from the GWs to the induced residuals in the pulsar’s TOAs. See @Hazboun+2019 for more details. For uniform white noise given by $\sigma_{{\ensuremath{\mathrm{W}}},i}$ between observations with cadence $c_i$, the white-noise term is simply $P_{{\ensuremath{\mathrm{W}}},i}(f) = 2 \sigma_{{\ensuremath{\mathrm{W}}},i}^2/c_i$. This form demonstrates the strength of the IAR Observatory in observing [PSR J0437$-$4715]{}: the cadence can be a factor of $\sim$30 larger than for other telescopes, thus significantly reducing the white noise term. The strain power spectrum for a single CW source with GW frequency $f_0$ and amplitude $h_0$ is [@tr2013; @optimalobs; @Hazboun+2019] S\_h(f) [h\_0\^2]{}. \[e:Sh(f)\_CW\] where $\delta_T(f)$ is the finite-time approximating function of a Dirac delta function, \_T(f) , \_[T ]{} \_T(f) = (f). That is, given the simple CW approximation often used in PTA analyses of an unchanging orbital frequency, over an infinite time, the power spectrum would be a delta function at $f_0$. While more realistic signals, such as those that include a phase shift from the delayed pulsar term and frequency evolution, are usually used in real PTA analyses [e.g., @NG11CW], the sensitivity is accurately estimated with a circular model. Additionally, it should be noted that this framework includes the signal from the time-delayed pulsar term. An analysis using the pulsar term requires very accurate pulsar distances in order to gain an appreciable signal; perfect knowledge of the distance boosting the squared [S/N]{} by a factor of $2$. While standard analyses [@NG11CW] use the pulsar term, it is unclear how much is gained by its inclusion. This has no effect on the ratios of detection thresholds examined here since the factor (of $\sqrt{2}$) for thresholds would cancel. Current IAR Sensitivity of [PSR J0437$-$4715]{} {#sec:current_obs} =============================================== Here we describe the current observational status of [PSR J0437$-$4715]{}by the IAR. Observational and derived parameters for [PSR J0437$-$4715]{}are provided in Table \[table:J0437\]. @IAR demonstrated 88 high-precision timing measurements from A1 and 106 from A2 in its preliminary timing campaign in 2019. We will assume a standardized phase bin resolution of ${N_\phi}= 512$. This choice implies that comparisons of pulse S/N will vary between observing setups (NANOGrav uses a uniform 2048 but other groups use variable values) though the TOA uncertainty will still be constant [@NG9WN]. A uniform definition of pulse S/N, as in Eq. \[eq:TFerror\], will however allow us to project the sensitivity of the IAR observations more easily and therefore it is still useful to convert between the two quantities $\sigma_{\rm S/N}$ and $S$. Using the smoothed S/N-weighted average 20-cm template from the Parkes Pulsar Timing Array [PPTA; @Kerr+2020], we calculated the effective width ${W_{\rm eff}}$ of [PSR J0437$-$4715]{}to be 667.6 $\mu$s (see Table \[table:J0437\]). From the residuals in @IAR, we calculated the median template-fitting uncertainties to be 200 and 325 ns for A1 and A2, respectively. With ${N_\phi}= 512$, the equivalent median S/N for A1 is 148.8, while for A2 is 90.6. While different astrophysical and terrestrial effects will change the observed S/N of the pulses from epoch to epoch, we use these parameters to form the basis of our extrapolation to future observing configurations in the next section. Extrapolation of IAR Observations of [PSR J0437$-$4715]{}to Future Systems {#sec:future_obs} ========================================================================== In this section, we provide estimates of the IAR’s sensitivity with changes in the observing configuration, namely the use of different receiver ranges for the two antennas. We also compare with sensitivity estimates of other facilities observing [PSR J0437$-$4715]{}and show that the IAR’s substantial cadence, i.e., its time on the source, will provide it with a unique capability in observing this pulsar in the future. Upgraded Observing Configurations {#sec:upgraded} --------------------------------- To estimate the TOA uncertainties for each system, we combined the various noise components in the following manner. We took the A1 and A2 median S/Ns as estimated in §\[sec:current\_obs\] and scale by the ratio of the flux densities from 1400 MHz to the new center frequency, the ratio of the square root of the bandwidths and number of polarization channels, and inversely by the ratio of the receiver temperatures. With the new S/N, we calculated a new template-fitting error for the receiver. With the two center frequencies, we were able to calculate the TOA uncertainty due to DM estimation (Eq. \[eq:DMerr\]); we assumed that in all future timing analyses, DM will be estimated on a time-varying basis such that the rms timing fluctuations are negligible (e.g., from Eq. \[eq:sigma\_tau\]), otherwise the unmodeled DM variations will make any high-precision timing experiment impossible. Lastly, we added in 100 ns of uncertainty due to polarization miscalibration, comparable to the value calculated by @vS2006. Even though we approximated the observations across each band as single measurements, ignoring the varying spectral index with the band, changes in pulse width, etc., for simplicity we calculated the TOA uncertainties as if there are two “spot” measurements at each center frequency. While this is a simplification, it provided us with a [*conservative*]{} estimate since a full least-squares fit formalism will lead to reduced TOA uncertainties. In summary, our TOA uncertainties consisted of template-fitting errors, DM estimation errors, and polarization miscalibration errors, added together in quadrature. We considered several different configurations: both the current and future setups as described in @IAR, and setups where additional receivers are built. We restricted ourselves to discussing the configurations broadly, assuming that the receiver frontend is matched with a backend that will be able to adequately sample and coherently dedisperse the data. [@IAR] describe the future planned upgrades to the antennas, which include receivers reaching system temperatures $T_{\rm sys} < 50$ K, a 500 MHz bandwidth in a range between 1 and 2 GHz, and a new FPGA-based CASPER board backend capable of processing the increased bandwidth. For other configurations we list, we assumed similar matching as we primarily wanted to focus on estimating the TOA uncertainties given the assumed development of a specific system. We estimated these uncertainties under two scenarios with the current equipment, three in which one or more of these higher-bandwidth systems is deployed, and one optimistic scenario in which ultrawideband receivers are deployed on both antennas. The full list of assumptions is provided in Table \[table:configs\] along with the TOA uncertainties we estimate. The gains of the antennas in our analysis did not vary though some dish or efficiency improvements may be possible in the future. We describe the rationales for considering each configuration below. 1. Current. This configuration is as described in @IAR. 2. Optimized. Since the center frequencies are tunable, we allowed for the maximum separation in frequencies to minimize TOA uncertainties from DM misestimation. We also halved A1’s bandwidth coverage in favor of dual-polarization measurements, a requirement for high-precision timing. 3. Wideband. With a modest upgrade of two receivers and backends covering 500 MHz each as discussed in @IAR, the two antennas can cover the 1-2 GHz range. The target system temperature is $T_{\rm sys} < 50$ K. While some parts of the band will be lost due to RFI, no significant segments of the band are currently lost across this frequency range [@IAR]. 4. Low Frequency. Instead of two receivers covering the full L-band range, we instead selected one 500 MHz receiver to cover the top end of L band (1.5-2.0 GHz) for A1 and then for A2 considered a receiver from 400-450 MHz. This lower frequency range is used by NANOGrav at Arecibo for some pulsars. Its primary allocation in Argentina[^1] is for maritime communication and radionavigation (see @ITU for Region 2) which helps to limit fixed sources of RFI. In addition, scattering will minimally affect [PSR J0437$-$4715]{}given its low DM, and so the increased flux density (see Table \[table:J0437\]) can lead to high-S/N TOAs, also providing a significant frequency difference to estimate DM. For the L-band receiver, we used the higher end of the possible bandwidth range to minimize TOA uncertainties due to DM misestimation. 5. High Frequency. As a parallel to C4, we considered instead using a second receiver at higher frequencies, in the 2.5 GHz (S band) range which is also used by NANOGrav at Arecibo for some pulsars. This region of the frequency spectrum often contains significant RFI due to overlap with wireless communications. Nonetheless, we considered the potential for such a system. For the L-band receiver, we used the lower end of the possible bandwidth range to minimize TOA uncertainties due to DM misestimation. 6. Dual Ultrawideband. As an optimistic setup, we considered the receiver systems as described in @DSA2000 for the DSA-2000. This observatory will employ low-cost receiver systems from 0.4-2.0 GHz for each of its 2000 antennas. While their target system temperature is 25 K, we kept 50 K for use in this considered configuration. Since the receiver setup is identical for both antennas, for simplicity we assumed two bands centered at 700 and 1500 MHz with 600 and 1000 MHz of bandwidth, respectively, to cover the 400-2000 MHz range, and included a factor of $\sqrt{2}$ in our calculation.   As discussed in §\[sec:shorttimenoise\], one astrophysical cause of the S/N changing is due to diffractive scintillation. In considering wider bandwidths, the number of scintles will grow, thus causing the pulse S/N to tend more towards a mean value rather than cover an exponentially-distributed range. In practice, this means that for low-bandwidth setups as in the current configurations, there are epochs in which the template-fitting error in Eq. \[eq:TFerror\] breaks down [@NG9] and TOAs cannot be reliably estimated, resulting in a loss of usable data. In addition to the mean S/N $S_0$ increasing due to wider bandwidths, the change in the diffractive scintillation PDF means that the TOAs will tend towards higher mean/median values, without a loss of data. Thus, we considered the extrapolation from the median S/N in §\[sec:current\_obs\] to be more robust. For our analysis, we assumed that configurations C3 through C6 are operational by 2023, a rapid timeline, and assumed that the observing conditions are static into the future. As with all facilities, future upgrades can yield additional sensitivity. In addition, growing RFI will impact timing sensitivity as well. @IAR show that $\lesssim 10$% of observing times, with significantly less time at night. RFI affects both antennas differently given their differing proximities to the local IAR offices. The RFI tends to be narrowband and thus can be excised while retaining the majority of the band. In our work, we ignored the role of RFI, noting that it will play an important but small contribution in the assumptions we are making regardless. [l||cc|cc||cc|cc|cc||cc]{} Circular Polarizations ($N_{\rm pol}$) & 1 & 2 & 2 & 2 & 2 & 2 & 2 & 2 & 2 & 2 & 2 & 2\ Center Frequency ($\nu_0$, MHz) & 1400 & 1400 & 1572 & 1128 & 1750 & 1250 & 1750 & 425 & 2500 & 1250 & 1200 & 1200\ Bandwidth ($B$, MHz) & 112 & 56 & 56 & 56 & 500 & 500 & 500 & 50 & 500 & 500 & 1600 & 1600\ Receiver Temperature ($T_{\rm rcvr}$, K) & 100 & 110 & 100 & 110 & 50 & 50 & 50 & 50 & 50 & 50 & 50 & 50\ Gain ($G$, K/Jy) & 0.084 & 0.077 & 0.084 & 0.077 & 0.084 & 0.077 & 0.084 & 0.077 & 0.084 & 0.077 & 0.084 & 0.077\ & & & & & &\ \[table:configs\] Comparison with Other Observatories ----------------------------------- ![Plots of the timing precision metric $M$ (see Eq. \[eq:M\]) versus time. Increased $M$ implies improved sensitivity. The blue lines denote the various IAR configurations, where C4 and C6 start in 2023. The metric for Parkes is shown in orange and the two teal curves show the metric for the VLA assuming two different estimates for polarization noise.[]{data-label="fig:config_comp"}](comp_plot.pdf){width="52.00000%"} Using the single-pulsar-dependent terms in Eq. \[e:rho2(k)\] and the instrumental/telescope-dependent components of the TOA uncertainty, defined below, we constructed a metric that describes the overall timing precision of a pulsar in an array (and dropping the subscript $i$): M ()\^[1/2]{} 34.6 \^[-1]{} ()\^[1/2]{}()\^[1/2]{}()\^[-1]{}. \[eq:M\] When $T$ is the total observing baseline in years, $c$ is the observing cadence in years$^{-1}$, and $\sigma_{\mathrm{sto}}$ is the telescope-dependent [*stochastic*]{} TOA uncertainty in $\mu$s, then $M_i$ has units of $\mu$s$^{-1}$. We define $\sigma_{\mathrm{sto}}$ here as the combination (quadrature sum) of the short-timescale white noise, the DM estimation uncertainty due to the white noise (Eq. \[eq:DMerr\]), and polarization calibration uncertainty (taken to be 100 ns, see §\[sec:upgraded\]. The short-timescale white noise includes template-fitting, jitter, and scintillation-noise uncertainties. As discussed in §\[sec:sensitivity\], jitter becomes negligible for the duration of observations conducted by the IAR but not necessarily for other observatories. Scintillation noise is negligible given the small scattering timescale. The quantity $M$ acts like a signal-to-noise ratio, where larger $M$ implies a higher sensitivity, though we note that it is not a direct relationship. As the number of observations $Tc$ increases, then $M$ increases as the square root of that number, and as $\sigma_{\rm sto}$ decreases, then $M$ increases linearly just as many signal-to-noise ratios behave. Alternatively, this quantity can be viewed as proportional to the (square) inverse of the error on the mean of the residuals. Figure \[fig:config\_comp\] shows curves of $M$ as a function of time for a number of different telescopes and configurations. The three blue curves denote three different IAR configurations as listed in Table \[table:configs\], where for C2 we take the start time as mid-2020 and for all other upgraded configurations we assume an aggressive program start time of January 2023. Delays will shift the curves to the right. Nonetheless, it is quite clear how drastically instrumentation improvements can improve $M$. Comparisons to other observatories are described in the following subsections. ### The Parkes Telescope The PPTA has observed [PSR J0437$-$4715]{}since 1996 [@Hobbs2013], with a typical cadence of once every two weeks in three bands [@Manchester+2013; @Reardon+2016; @Kerr+2020] for approximately one hour in each of the 20 (1400 MHz) and 10/50 cm (3100/730 MHz) observations [@Hobbs2013]. The backend systems have improved over time, resulting in many different noise contributions to consider. We estimated the equivalent $\sigma_{\rm sto}$ as follows. Using the PPTA Second Data Release [@Kerr+2020], we used the empirically-derived noise parameters as an estimate for the telescope-dependent stochastic terms described previously. First, we took parameters that describe modifications to the template-fitting uncertainties, namely a noise added in quadrature (EQUAD) with the resultant quantity multiplied by a scaling factor (EFAC), to provide us with an estimate of the white noise for each TOA [@Reardon+2016]. Since the white noise parameters in @Kerr+2020 were derived from sub-banded TOAs, i.e., those taken from small frequency channels over a wider bandwidth, we then computed the epoch averaged error as in @optimalfreq (and also see references therein) to provide us with the amount of white noise per epoch per band. This amounts to computing the square root of $(\mathbf{U}^T \mathbf{C}^{-1} \mathbf{U})^{-1}$, where $\mathbf{U}$ is a column matrix of ones and $\mathbf{C}$ is the covariance matrix, which in this case is a diagonal matrix containing the squared of the modified TOA uncertainties. These epoch-averaged TOA uncertainties describe the measured white noise in the @Kerr+2020 data set. We also included the measured “band noise” terms, which describe additional noise as a function of specific frequency ranges. Such band noise can describe either additional chromatic propagation effects or terrestrial effects such as from RFI or polarization miscalibration [@Lentati+2016]. Since in §\[sec:sensitivity\] we argued that [PSR J0437$-$4715]{} likely has low levels of noise due to unmodeled propagation effects such as scattering, then we took the band noise to be due to terrestrial effects and as such consider it an important contribution to $\sigma_{\rm sto}$. Finally, to use a single number to project $M$ to, we computed the average “weight” of each uncertainty ($w_k = 1/\sigma_k^2$ for the $k$-th TOA), computed the mean weight, and then calculated the square root of reciprocal of this mean weight. We note that an ultrawideband receiver system has been deployed at Parkes [@Kerr+2020], which may result in reduced uncertainties in the future if RFI does not become sufficiently problematic. If so, this will improve future estimates of $M$ for Parkes as compared with what is shown in Figure \[fig:config\_comp\]. However, the majority of the amplitude of $M$ comes from the large observing baseline $T$ more so than the TOA uncertainty, so it is unclear by how much $M$ will be affected. ### The Very Large Array NANOGrav has also begun[^2] a program to observe [PSR J0437$-$4715]{}with the VLA monthly between 1 and 4 GHz, with 30 total minutes devoted to L band (1-2 GHz) and 30 minutes devoted to S band (2-4 GHz). While the full hour includes overhead time, for the purposes of our calculations, we will assume each observation is 30 minutes long. Without empirically-derived white-noise parameters as in the PPTA data release, while we could work out a full calculation of the white noise plus DM uncertainties, we will instead provide a heuristic argument to demonstrate that the observational setup with the VLA will result in a lower $M$. Let us assume that jitter is the only white-noise contribution to the TOA uncertainty. Using the values in @sod+14 (see also Table \[table:J0437\]), we have that the rms jitter for a 1 hour observation is 48 ns at L band and 41 ns at S band. As jitter scales as the square root of the number of pulses, the 30-minute rms jitter will be 68 and 58 ns for the two bands, respectively. For DM estimation errors, we will assume for this argument that the two spot frequencies are 1 and 4 GHz rather than center frequencies of 1.5 and 3 GHz, which will underestimate the uncertainty but will not matter for a comparison. Using our assumed numbers, calculating $\sigma_{{\ensuremath{\widehat{\mathrm{DM}}}}}$ using Eq. \[eq:DMerr\], and then finding the total infinite-frequency TOA uncertainty, we have 76 ns per observation as a [*minimum*]{} bound. With observations starting in 2016, and assuming 12 observations per year continued until 2033, we have $M$ between 123 and $170~\mu\mathrm{s}^{-1}$, assuming 50 and 100 ns of polarization calibration uncertainty, respectively. ### MeerKAT MeerKAT has begun pulsar-timing observations with the MeerTime project [@TPA; @MeerTime2020], with one goal of the project to observe MSPs. With the expected next generation of receivers extending the observatory’s bandwidth range, in a subarray mode, MeerKAT will be able to observe from 0.9-3.5 GHz [@MeerTime2018; @MeerTime2020]. If again we assume that the pulsar is jitter-dominated as with the VLA, then we expect comparable numbers as calculated above. ### Comparing $M$ for Different Observatories Figure \[fig:config\_comp\] shows $M$ for the VLA (again, comparable with MeerKAT), Parkes (PKS), and three configurations for the IAR: C2, C4, and C6. Given the assumptions presented above, we see that an aggressive timeline for improved instrumentation will drastically improve IAR’s sensitivity to [PSR J0437$-$4715]{} in comparison to other facilities. Delays in this timeline will shift the curve to the right but the conclusion remains the same: a modestly upgraded IAR facility (C4) will drastically aid in single-source GW science right in the era of the first CW sources, as well as the first observations by LISA. Using state-of-the-art yet low-cost instrumentation [e.g., C6 using similar instrumentation as described in @DSA2000] will provide the IAR with unparalleled sensitivity to [PSR J0437$-$4715]{}and thus a unique contributor to GW science with PTAs by the middle-to-end of the current decade. We note that other observatories like Parkes will still have an advantage over long-timescale ($\sim$decade) GW periods since its observations of [PSR J0437$-$4715]{} cover a sufficient number of cycles whereas observations by the IAR alone will not. GW Projections for [PSR J0437$-$4715]{} {#sec:projections} ======================================= In order to demonstrate the usefulness of [PSR J0437$-$4715]{}as an addition to the NANOGrav observing campaign, we projected the status of the PTA described in @NG11 into the next decade. Here we are only interested in how [PSR J0437$-$4715]{}contributes to the sensitivity for deterministic signals so we simply extended the baselines of the data set most recently used in these GW analyses, using the noise characteristics and sky locations of the current set of pulsars and those estimated for [PSR J0437$-$4715]{}at IAR. We only investigated single-source sensitivity because 1) the GWB is forecasted to be detected long before the dates of these projections, and 2) the short baseline of [PSR J0437$-$4715]{}when it is added to NANOGrav data sets will not be very useful for detection/characterization of the GWB until it has sensitivity at lower frequencies. Using the Python package [hasasia]{} [@hasasia] we simulated a PTA with the same characteristics as @NG11 but added 17 years of data to use as a baseline for our comparison, ending in 2033. We also simulated two versions of a PTA with IAR observations of [PSR J0437$-$4715]{} from 2023 onwards, i.e., 10 years of observations. The two versions correspond to using configurations C4 and C6. Figure \[fig:skymap\] shows the ratio of the sky sensitivity in source strain [see Eq. 80 in @Hazboun+2019] when adding observations of [PSR J0437$-$4715]{}versus not. We show the positions of the NANOGrav 11-year data set pulsars (white stars), [PSR J0437$-$4715]{}(red star), and nearby galaxy clusters and individual targets of interest [e.g., @Mingarelli+2017; @cz2018; @NG11CW; @3C66B]. We used a fiducial GW frequency of 10$^{-8}$ Hz (periods of $\sim$3 yr), approximately where the current most sensitive CW upper limits are [@NG11CW]. The sky maps are identical in structure but the overall sensitivity varies when using C4 versus C6. We see clear indications that observations with [PSR J0437$-$4715]{}versus without will increase the sensitivity of the NANOGrav PTA. In the case of C6, by 2033 the array’s sensitivity improves by a factor of $\approx$2-7 for 20% of the sky. Even with C4, the improvement ratio is $\approx$1.4-4. Since the observed GW strain is inversely proportional to the distance of a source, a factor of 2 increase in sensitivity leads to the array being able to search out a factor of 2 in distance, indicating that the volume NANOGrav will be able to probe in its current “blind spot” will be drastically increased. ![Ratio of the GW sensitivity when adding [PSR J0437$-$4715]{}to the PTA using configurations C4 and C6 versus otherwise. The sensitivity is calculated at 10$^{-8}$ Hz. NANOGrav 11-year data set pulsars are shown in white stars, [PSR J0437$-$4715]{}in the red star. Possible targets of interest for CW signals are shown in the black dots and labeled. Equatorial coordinates are shown, with right ascension increasing to the left every 2$^{\rm h}$ starting from the right and declination increasing upwards every 30$^\circ$.The structure of the sky map is the same for C4 and C6 but the corresponding colormap scale is shown at the bottom for each. We see clear significant increases to the sensitivity in NANOGrav’s current “blind spot” on the sky due to the addition of [PSR J0437$-$4715]{}. []{data-label="fig:skymap"}](skymap.pdf){width="48.00000%"} Predictions for Observations of Other MSPs {#sec:other_MSPs} ========================================== Our estimates and extrapolations for PSR J0437$-$4715 can be used to predict the general timing performance of other pulsars observable with the IAR. While PSR J0437$-$4715 is the brightest known MSP, other factors influence the overall timing precision. Table \[table:pulsars\] provides a list of some pulsars observed by the PPTA [@Kerr+2020], which we use as a representative list to quantify potential targets. For the same observational setup, we expect that the period-averaged signal-to-noise ratio $\bar{S}$ will be proportional to the period-averaged flux density $I_0$ at a fiducial frequency $\nu_0$, using the notation of @optimalfreq. Wide-bandwidth observations require that the spectral index of the pulsar flux be considered. In general, the minimum TOA uncertainty due to a finite S/N is given by Eq. \[eq:TFerror\], which uses the peak-to-off-pulse S/N, $S$. The conversion between the two is given by $\bar{S} = \bar{U}_{\rm obs}S$, where |[U]{}\_[obs]{} =\_[i=0]{}\^[[N\_]{}-1]{} [U\_[obs]{}]{}(\_i) is the mean value of the template shape ${U_{\rm obs}}$. A smaller value of $\bar{U}_{\rm obs}$ implies a sharper pulse profile which yields improved TOA uncertainties. Therefore, combining the pulse-shape-dependent factors with the pulsar fluxes, we can relate the ratio of the template-fitting uncertainties between two pulsars A and B as R = . \[eq:sigma\_ratio\] With Eq. \[eq:sigma\_ratio\], we were able to calculate the template-fitting errors for the pulsars provided in Table \[table:pulsars\] compared to PSR J0437$-$4715. In the table, the larger $R$ is, the better the template-fitting uncertainty will be. This analysis does not take into account other sources of uncertainty in TOA estimation. For example, sources with higher DM are more likely to be affected by pulse broadening due to interstellar scattering [@Bhat+2004]. We can more intuitively understand the impact of the pulse shape metrics compared to the flux densities in Eq. \[eq:sigma\_ratio\] by considering a Gaussian-shaped pulse. The effective width ${W_{\rm eff}}\propto (W_{50}P)^{1/2}$, the geometric mean of the pulse full-width-at-half-maximum $W_{50}$ and the pulse period $P$ (following from Eq. \[eq:Weff\], see also @NG9WN). In considering a Gaussian pulse, the mean value of the pulse will be the integral of the pulse shape divided by the pulse period, yielding the approximate relationship $\bar{U}_{\rm obs} \propto W_{50}/P$ for sufficiently narrow pulses. Combining, we have [W\_[eff]{}]{}|[U]{}\_[obs]{} (W\_[50]{}P)\^[1/2]{} = P ()\^[3/2]{}. \[eq:proportionalities\] We write the second equality in terms of the pulse duty cycle, $W_{50}/P$. We see from Eq. \[eq:proportionalities\] that decreases in either the pulse period or the duty cycle will lead to a lower $\sigma_{\rm S/N}$, which is to be expected. These improvements can help compensate for a difference between the flux densities of pulsars. A factor of two decrease in both the pulse period and duty cycle can then offset a factor of five decrease in the flux density of one pulsar compared to another and lead to equivalent $\sigma_{\rm S/N}$. Note that in comparison to [PSR J0437$-$4715]{}, decreases in $P$ or $W_{50}/P$ by more than a factor of $\sim$2 are generally not observed. Table \[table:pulsars\] gives the various pulse shape parameters for ten pulsars observed by the PPTA in the declination range of IAR and with $R > 0.01$. Pulse shape metrics were measured using the 20-cm Parkes Digital Filterbank System 4 (PDFB4) templates[^3] in @Kerr+2020. We used the median flux density at 1400 MHz reported by @Kerr+2020 in our analysis; the scintillation properties of the different pulsars will affect $R$ but we ignored this here and only considered a typical observation. From these measurements, we calculated $R$ according to Eq. \[eq:sigma\_ratio\]. Another pulsar with one of the lowest DMs, PSR J1744$-$1134, has been detected by the IAR in individual observations (L. Combi, private communication). As with [PSR J0437$-$4715]{}, we can then consider the impact of daily cadence using the IAR compared with $\sim$monthly observations available at other facilities. Let us naively consider [PSR J0437$-$4715]{}observed with $\sigma_{\rm W} \approx \sigma_{\rm S/N} = 200$ ns. While $R = 0.035$ for PSR J1744$-$1134, leading to many microseconds of uncertainty per epoch, the monthly-averaged uncertainty reduces by a factor of $\sqrt{30}$ to approximately 1 $\mu$s. This argument ignores differences in spectral index and other frequency-dependent profile changes but serves to demonstrate the potential of IAR’s unique capabilities. Observations of PSR J1744$-$1134 with daily cadence will be comparably as sensitive as other pulsars with microsecond timing observed with a monthly cadence, making observations of the pulsar of sufficient quality for PTA science. Several other possible targets, such as PSRs J1909$-$3744, J2241$-$5236, and possibly also J1600$-$3053, may be suitable candidates for daily monitoring. Concluding Remarks {#sec:conclusion} ================== The IAR’s unique capabilities to observe [PSR J0437$-$4715]{}will lead to a significant increase in CW sensitivity for PTAs. However, it is vital that the observatory receives the necessary upgrades to make this a possibility; without dual-polarization observations at frequencies spaced widely enough to accurately estimate DM, the IAR will not be able to meet the target sensitivities described here. This will require careful polarization calibration and the RFI environment around the observatory must remain in a clean state. In addition to direct CW sensitivity, high-cadence observations can contribute to other PTA-related science. For example, daily monitoring of DMs and scintillation which will feed back into understanding pulsar timing noise models is a planned goal of the IAR [@IAR]. Such observations will complement those of Northern-Hemisphere facilities like CHIME [$\delta >-20^\circ$ @Ng2018] which are beginning to provide unprecedented measurements of the interstellar medium on short timescales [@Ng+2020]. For example, given the proper motion of [PSR J0437$-$4715]{}of 141 mas/yr [@Kerr+2020] and distance of 156.8 pc, the pulsar’s transverse velocity is 105 km/s. Over 10 years of daily observations, the line of sight will probe length scales in the interstellar medium between 0.06 – 200 AU. Considering the relatively long tracks on [PSR J0437$-$4715]{}, breaking the observations into individual hour-long measurements would yield a factor of $\sim$2 reduction in the DM precision but the ability to probe scales down to 0.0025 AU, or half of a solar radius. International PTA efforts will also benefit from the inclusion of IAR observations. The International Pulsar Timing Array (IPTA) collaboration helps to coordinate these worldwide efforts and while the potential target pulsars listed here would not be new additions to any future combined data set, such a data set will have improved sensitivity to GWs of all kinds with additional high-precision data. Uniform observations of several pulsars by the IAR can additionally help constrain systematic variations between different frontend/backend systems for different telescopes as well as between telescopes, for example any timing offsets or calibration errors since a stable system can be compared against. Such observations can thus potentially reduce the overall noise of such a combined IPTA data set. We thank Luciano Combi and the Pulsar Monitoring in Argentina (PuMA) for useful discussions in the preparation of this work. The NANOGrav Project receives support from NSF Physics Frontiers Center award number 1430284. Aggarwal, K., Arzoumanian, Z., Baker, P. T., et al. 2019, , 880, 116 Aggarwal, K., Arzoumanian, Z., Baker, P. T., et al. 2020, , 889, 38 Alam, M., Arzoumanian, Z., Baker, P. T., et al. 2020a, arXiv e-prints, arXiv:2005.06490 Alam, M., Arzoumanian, Z., Baker, P. T., et al. 2020b, arXiv e-prints, arXiv:2005.06495 Anholm, M., Ballmer, S., Creighton, J. D. E., Price, L. R., & Siemens, X. 2009, , 79, 084030 Armstrong, J. W., Rickett, B. J., & Spangler, S. R. 1995, , 443, 209 Arzoumanian, Z., Brazier, A., Burke-Spolaor, S., et al. 2015, , 813, 65 Arzoumanian, Z., Brazier, A., Burke-Spolaor, S., et al. 2018a, , 235, 37 Arzoumanian, Z., Baker, P. T., Brazier, A., et al. 2018b, , 859, 47 Arzoumanian, Z., Baker, P. T., Brazier, A., et al. 2020, arXiv e-prints, arXiv:2005.07123 Babak, S., & Sesana, A. 2012, , 85, 044034 Bailes, M., Barr, E., Bhat, N. D. R., et al. 2018, arXiv e-prints, arXiv:1803.07424 Bailes, M., Jameson, A., Abbate, F., et al. 2020, arXiv e-prints, arXiv:2005.14366, accepted by PASA Bhat, N. D. R., Cordes, J. M., Camilo, F., Nice, D. J., & Lorimer, D. R. 2004, , 605, 759 Burt, B. J., Lommen, A. N., & Finn, L. S. 2011, , 730, 17 Chen, J.-W., & Zhang, Y. 2018, , 481, 2249 Cordes, J. M. & Rickett, B. J. 1998, , 507, 846 Cordes, J. M., Wolszczan, A., Dewey, R. J., Blaskiewicz, M., & Stinebring, D. R. 1990, , 349, 245 Cordes, J. M., & Chernoff, D. F. 1997, , 482, 971 Cordes, J. M., Kramer, M., Lazio, T. J. W., et al. 2004, , 48, 1413 Cordes, J. M., & Shannon, R. M. 2010, arXiv:1010.3785 Detweiler, S. 1979, , 234, 1100 Ellis, J. A., Jenet, F. A., & McLaughlin, M. A. 2012, , 753, 96 Ellis, J. A., Siemens, X., & Creighton, J. D. E. 2012, , 756, 175 Ente Nacional de Comunicaciones 2019, Cuadro de Atribución de Bandas de Frecuencias de la República Argentina (CABFRA; Table of Attribution of Frequency Bands of the Argentine Republic), (Buenos Aires, Argentina: ENACOM), <https://www.enacom.gob.ar/cuadro-de-atribucion-de-bandas-de-frecuencias-de-la-republica-argentina-cabfra-_p1588> Foster, R. S., & Backer, D. C. 1990, , 361, 300 Foster, R. S., & Cordes, J. M. 1990, , 364, 123 Gancio, G., Lousto, C. O., Combi, L., et al. 2019, , https://doi.org/10.1051/0004-6361/201936525 Gwinn, C. R., Hirano, C., & Boldyrev, S. 2006, , 453, 595 Hallinan, G., Ravi, V., Weinreb, S., et al. 2019, 51, 255 Hazboun, J. S., Romano, J. D., & Smith, T. L. 2019, , 100, 104028 Hazboun, J., Romano, J., & Smith, T. 2019, The Journal of Open Source Software, 4, 1775 Hellings, R. W., & Downs, G. S. 1983, , 265, L39 Hobbs, G. 2013, Classical and Quantum Gravity, 30, 224007 International Telecommunication Union Radiocommunication Sector (ITU-R) 2016, Radio Regulations, Volume I, Article 5 (Geneva, Switzerland:ITU), <http://handle.itu.int/11.1002/pub/80da2b36-en> Jenet, F. A., Hobbs, G. B., van Straten, W., et al. 2006, , 653, 1571 Johnston, S., Lorimer, D. R., Harrison, P. A., et al. 1993, , 361, 613 Johnston, S., Karastergiou, A., Keith, M. J., et al. 2020, , 493, 3608 Keith, M. J., Coles, W., Shannon, R. M., et al. 2013, , 429, 2161 Kelley, L. Z., Blecha, L., Hernquist, L., et al. 2018, , 477, 964 Kerr, M., Reardon, D. J., Hobbs, G., et al. 2020, , 37, e020 Lam, M. T., Cordes, J. M., Chatterjee, S., et al. 2016a, , 819, 155 Lam, M. T., Cordes, J. M., Chatterjee, S., et al. 2016b, , 821, 66 Lam, M. T., Cordes, J. M., Chatterjee, S., et al. 2017, , 834, 35 Lam, M. T., McLaughlin, M. A., Cordes, J. M., Chatterjee, S., & Lazio, T. J. W. 2018a, , 861, 12 Lam, M. T. 2018, , 868, 33 Lam, M. T., McLaughlin, M. A., Arzoumanian, Z., et al. 2019a, , 872, 193 Lam, M. T., Lazio, T. J. W., Dolch, T., et al. 2020, , 892, 89 Lentati, L., Shannon, R. M., Coles, W. A., et al. 2016, , 458, 2161 Levin, L., McLaughlin, M. A., Jones, G., et al. 2016, , 818, 166 Madison, D. R., Cordes, J. M., Arzoumanian, Z., et al. 2019, , 872, 150 Manchester, R. N., Hobbs, G. B., Teoh, A., & Hobbs, M. 2005, , 129, 1993 Manchester, R. N., Hobbs, G., Bailes, M., et al. 2013, , 30, e017 McLaughlin, M. A. 2013, Classical and Quantum Gravity, 30, 224008 Mingarelli, C. M. F., Lazio, T. J. W., Sesana, A., et al. 2017, Nature Astronomy, 1, 886 Ng, C. 2018, Pulsar Astrophysics the Next Fifty Years, 179 Ng, C., Pandhi, A., Naidu, A., et al. 2020, , doi:10.1093/mnras/staa1658 Perera, B. B. P., DeCesar, M. E., Demorest, P. B., et al. 2019, , 490, 4666 Phillips, J. A., & Wolszczan, A. 1991, , 382, L27 Ransom, S., Brazier, A., Chatterjee, S., et al. 2019, 51, 195 Reardon, D. J., Hobbs, G., Coles, W., et al. 2016, , 455, 1751 Rosado, P. A., Sesana, A., & Gair, J. 2015, , 451, 2417 Sazhin, M. V. 1978, , 22, 36 Shannon, R. M., & Cordes, J. M. 2010, , 725, 1607 Shannon, R. M., Os[ł]{}owski, S., Dai, S., et al. 2014, , 443, 1463 Siemens, X., Ellis, J., Jenet, F., & Romano, J. D. 2013, Classical and Quantum Gravity, 30, 224015 Taylor, J. H. 1992, Royal Society of London Philosophical Transactions Series A, 341, 117 Taylor, S., Ellis, J., & Gair, J. 2014, , 90, 104028 Taylor, S. R., Vallisneri, M., Ellis, J. A., et al. 2016, , 819, L6 Thrane, E., & Romano, J. D. 2013, , 88, 124032 Tiburzi, C., Verbiest, J. P. W., Shaifullah, G. M., et al. 2019, , 487, 394 van Straten, W. 2006, , 642, 1004 You, X. P., Hobbs, G. B., Coles, W. A., Manchester, R. N., & Han, J. L. 2007, , 671, 907 [^1]: Ente Nacional de Comunicaciones (ENACOM) follows ITU-R regulations, as described in @ENACOM. [^2]: Proposal IDs VLA/16B-240, VLA/18A-210, VLA/19A-356 [^3]: Accessible at <https://doi.org/10.25919/5db90a8bdeb59>.

ArXiv

--- abstract: 'This paper investigates the symmetric degrees of freedom (DoF) of multiple-input multiple-output (MIMO) cellular networks with $G$ cells and $K$ users per cell, having $N$ antennas at each base station (BS) and $M$ antennas at each user. In particular, we investigate achievability techniques based on either decomposition with asymptotic interference alignment or linear beamforming schemes, and show that there are distinct regimes of $(G,K,M,N)$ where one outperforms the other. We first note that both one-sided and two-sided decomposition with asymptotic interference alignment achieve the same degrees of freedom. We then establish specific antenna configurations under which the DoF achieved using decomposition based schemes is optimal by deriving a set of outer bounds on the symmetric DoF. Using these results we completely characterize the optimal DoF of any $G$-cell network with each user having a single antenna. For linear beamforming schemes, we first focus on small networks and propose a structured approach to linear beamforming based on a notion called packing ratios. Packing ratio describes the interference footprint or shadow cast by a set of transmit beamformers and enables us to identify the underlying structures for aligning interference. Such a structured beamforming design can be shown to achieve the optimal spatially normalized DoF (sDoF) of two-cell two-user/cell network and the two-cell three-user/cell network. For larger networks, we develop an unstructured approach to linear interference alignment, where transmit beamformers are designed to satisfy conditions for interference alignment without explicitly identifying the underlying structures for interference alignment. The main numerical insight of this paper is that such an approach appears to be capable of achieving the optimal sDoF for MIMO cellular networks in regimes where linear beamforming dominates asymptotic decomposition, and a significant portion of sDoF elsewhere. Remarkably, polynomial identity test appears to play a key role in identifying the boundary of the achievable sDoF region in the former case.' author: - 'Gokul Sridharan and Wei Yu[^1][^2]' bibliography: - 'IEEEabrv.bib' - 'ref\_file.bib' title: 'Degrees of Freedom of MIMO Cellular Networks: Decomposition and Linear Beamforming Design' --- Decomposition Based Schemes: Achievable DoF and Conditions for Optimality {#allAboutDecomp} ========================================================================= Linear Beamforming: Structured Design ===================================== Conclusion ========== In this paper we investigate the DoF of MIMO cellular networks. In particular we establish the achievable DoF through the decomposition based approach and linear beamforming schemes. Through a new set of outer bounds, we establish conditions for optimality of the decomposition based approach. Through these outer bounds it is apparent that the optimal DoF of a general $G$-cell, $K$-users/cell network exhibits two distinct regimes, one where decomposition based approach dominates over linear beamforming and vice-versa. With regard to linear beamforming, we develop a structured approach to linear beamforming that is DoF-optimal in small networks such as the two-cell two-users/cell network and the two-cell three-users/cell network. We also develop an unstructured approach to linear beamforming that is applicable to general MIMO cellular networks, and through numerical experiments, show that such an approach is capable of achieving the optimal-sDoF for a wide class of MIMO cellular networks. Although the structured design of linear beamformers takes a disciplined approach to constructing beamformers, the wide applicability of the unstructured approach and its apparent ability to achieve the optimal sDoF in regimes where the sDoF curve exhibits a piecewise-linear behavior renders it highly attractive. The remarkable effectiveness of the unstructured approach warrants a deeper investigation on the role of randomization and that of the polynomial identity test in designing aligned beamformers. DoF Outer Bound for the Two-Cell Three-Users/Cell Network When $\frac{5}{9} \leq \gamma < \frac{3}{4}$ {#genieouterboundB} ========================================================================================================= Achievability of the Optimal sDoF for the Two-Cell Two-Users/Cell Network and the Two-Cell Three-Users/Cell Network {#finerdetails} =================================================================================================================== [^1]: The authors are with the The Edward S. Rogers Sr. Department of Electrical and Computer Engineering, University of Toronto, Toronto, ON M5S4G4, Canada e-mail: (gsridharan@comm.utoronto.ca, weiyu@comm.utoronto.ca). [^2]: This work was supported by the Natural Science and Engineering Research Council (NSERC) of Canada. The material in this paper has been presented in part at Canadian Workshop Inf. Theory, Jun. 2013, IEEE Int. Symp. Inf. Theory, Jul. 2013, and IEEE Global Commun. Conf., Dec. 2013. Manuscript submitted to IEEE Transactions on Information Theory on December 9, 2013.

ArXiv

--- abstract: 'In the present paper we discuss the eigenvalue conjecture, suggested in 2012, in the particular case of $U_q(sl_2)$. The eigenvalue conjecture provides a certain symmetry for Racah coefficients and we prove that **the eigenvalue conjecture is provided by the Regge symmetry** for $U_q(sl_2)$, when three representations coincide. This in perspective provides us a kind of generalization of the Regge symmetry to arbitrary $U_q(sl_N)$.' author: - '[**Andrey Morozov$^{a,b,c}$**]{}, [**Alexey Sleptsov$^{a,b,c,d}$**]{}' date: --- IITP/TH-16/18\ ITEP/TH-27/18\ $^a$\ $^b$\ $^c$\ $^d$ Introduction ============ Racah coefficients (6-j symbols) are an important quantity in the theoretical and mathematical physics. They appear everywhere from quantum mechanics to the knot theory and integrable systems. Despite this there are a lot of open problems with Racah coefficients even for $U_q(sl_N)$. In the case of $U_q(sl_2)$ there is a general formula [@KirResh] for Racah coefficients. According to that formula 6-j symbols are described by the q-hypergeometric function $_4\Phi_3$. Using this fact the whole set of symmetries of $U_q(sl_2)$ Racah coefficients was described. The next step was done in [@NRZ1]-[@MMSracah] by extending the explicit Racah formula to the symmetric representations of arbitrary $U_q(sl_N)$. It covers only so-called [*exclusive*]{} 6-j symbols, which appear in the arborecent links [@arbor1; @arbor2] calculus coming from the Wess-Zumino-Novikov-Witten Confromal Field Theory consideration [@arborcalc1]-[@RTmod4]. Also exclusive 6-j symbols for the first non-symmetric representation $[2,1]$ was calculated in [@GuJ]. The modern version of the Reshetikhin-Turaev approach provide a general method to compute colored HOMFLY polynomials for arbitrary link [@RTmod]. Racah coefficients play the key role in this formalism, because they relate different $\mathcal{R}$-matrices. Some of the Racah coefficients appear in the Yang-Baxter equation together with $\mathcal{R}$-matrices. This led to the eigenvalue conjecture which relates the $\mathcal{R}$-matrix eigenvalues with the Racah coefficients. Eigenvalue conjecture, suggested in [@IMMMec], is a very important concept which has a very useful practical applications. This conjecture states that Racah coefficients are fully defined by the set of normalised eigenvalues of the corresponding $\mathcal{R}$-matrices. In its strong form the exact form for Racah coefficients from the $\mathcal{R}$-matrix eigenvalues is provided, and it is known for the matrices of the size less or equal to $5\times 5$ [@IMMMec; @Wenzl]. The explicit form for the Racah matrix of the size $6\times 6$ was proposed in [@MkrMM]. In its weak form it allows for example to calculate knot polynomials in any symmetric representation, expressing all the needed Racah coefficients through the known $U_q(sl_2)$ Racah coefficients [@DMMMRSS; @LL]. Eigenvalue conjecture provides a certain symmetry for the Racah coefficients. The eigenvalue conjecture gives rise to the question what happens with this symmetry in the well-known $U_q(sl_2)$ case. In the present paper we prove that: **for $U_q(sl_2)$ Racah coefficients the eigenvalue conjecture is provided by the Regge symmetry.** However, unlike the Regge symmetry, the eigenvalue conjecture is also formulated for arbitrary rank $U_q(sl_N)$. Thus one can hope that studies of the eigenvalue conjecture can give us some insights into the generalization of known symmetries from $U_q(sl_2)$ to $U_q(sl_N)$ case. Our paper is organized as follows. In section 2 we give some basic definitions, briefly review the Reshetikhin-Turaev approach to quantum link invariants and formulate the eigenvalue conjecture as the symmetry property on the Racah matrices. In section 3 we remind the symmetry properties of 6-j symbols. In section 4 we prove that the eigenvalue conjecture is true in the case of $U_q(sl_2)$ if the three incoming representations are same. This particular case of 6-j symbols is enough to calculate colored HOMFLY polynomials of 3-strand knots and 3-strand links, whose components colored by the same representations. We also show that the Regge symmetry provide the eigenvalue conjecture in this case. In section 5 we give some evidences that the eigenvalue conjecture correctly predicts the symmetry properties of the Racah matrices. Link invariants from quantum groups {#s.int} =================================== In this section we give a definiton of the colored HOMFLY polynomials of arbitrary links via Reshetikhin-Turaev approach based on the theory of quantum groups and quantum $\mathcal{R}$-matrix. $\mathcal{R}$-matrix and link invariants ---------------------------------------- First of all, let us define quantum $\mathcal{R}$-matrices, which are associated with multicolored braid. Suppose that we have a braid with $m$ strands. We associate a finite-dimensional representation $R_i$ of the quantized universal enveloping algebra $U_q(sl_N)$ with $i$-th strand. Since we assume that $q$ is a nonzero complex number, which is not a root of unity, all finite-dimensional representations are representations of highest weights and they can be enumerated by Young diagrams. Due to this fact for simplicity we identify Young diagram and representation and use the same notations for both when it can be done without ambiguity. $\bullet$ There exists universal R-matrix $$\check{\mathcal{R}} = q^{\sum\limits_{i,j}C_{ij}^{-1}H_i\otimes H_j} \prod_{\textrm{positive root }\beta} \exp_q [( 1-q^{-1}) E_\beta\otimes F_\beta]\,.$$ here $( C_{ij})$ is the Cartan matrix and $\{H_i,E_i,F_i\}$ are generators of $U_q(sl_N)$. $\bullet$ If we define invertible linear operators by $$\label{Rmat} \mathcal{R}_i = 1_{V_1}\otimes1_{V_2}\otimes\ldots\otimes P \check{\mathcal{R}}_{i,i+1} \otimes\ldots\otimes1_{V_m} \ \in \text{End}(V_1\otimes\ldots,\otimes V_m),$$ where $P(x\otimes y) = y\otimes x$ and $\check{\mathcal{R}}$ acts on two $U_q(sl_N)$-modules $V_i$ and $V_{i+1}$, then it is well known [@KirResh], [@RJ] that $\mathcal{R}_1,\ldots,\mathcal{R}_{m-1}$ define a representation of the Artin’s braid group $B_m$ on $m$ strands: $$\begin{array}{rcl} \pi: B_m &\rightarrow& \text{End}(V_1\otimes\ldots,\otimes V_m) \\ \pi(\sigma_i) &=& \mathcal{R}_i, \end{array}$$ where $\sigma_1,\ldots,\sigma_{m-1}$ are generators of the braid group $B_m$. Graphically we can represent $\mathcal{R}_i$ as follows: (850,125)(-250,-93) (-70,20)[(1,0)[90]{}]{} (-35,0) (-70,-20)[(1,0)[35]{}]{} (-70,-40)[(1,0)[35]{}]{} (-15,-20)[(1,0)[35]{}]{} (-15,-40)[(1,0)[35]{}]{} (-35,-60) (-70,-80)[(1,0)[90]{}]{} (-35,-20)[(1,-1)[20]{}]{} (-15,-20)(-10,-10)[2]{}[(-1,-1)[10]{}]{} (-130,-33) (-90,17) (-90,-23) (-90,-43) (-90,-83) Clearly, inverse crossing is given by $\mathcal{R}^{{-}1}_i$. Operators $\mathcal{R}_i$ satisfy relations of the braid group $B_m$: $$\begin{array}{lll} \textbf{far commutativity} & \mathcal{R}_i\mathcal{R}_j = \mathcal{R}_j\mathcal{R}_i, & \text{for} \ |i-j|\neq 1 \\ \textbf{braiding relation} & \mathcal{R}_i\mathcal{R}_{i+1}\mathcal{R}_{i} = \mathcal{R}_{i+1}\mathcal{R}_i\mathcal{R}_{i+1}, & \text{for} \ i=1,\ldots,m-2 \end{array}$$ Graphically, the braiding relation is nothing but the third Reidemeister move, while algebraically, it is a well-known *quantum Yang-Baxter equation* on quantum $\mathcal{R}$-matrix. $\bullet$ Alexander’s theorem states that any link ${\mathcal{L}}$ in $\mathbb{R}^3$ can be obtained by a closure of the corresponding braid. Let ${\mathcal{L}}$ be an oriented link with $L$ components ${\mathcal{K}}_1,\ldots,{\mathcal{K}}_L$ colored by representations $R_1,\ldots,R_L$ and $\beta_{{\mathcal{L}}} \in B_m$ is a some braid, which closure gives ${\mathcal{L}}$. Then according to Reshetikhin-Turaev approach [@RT0]-[@LiuPeng] the quantum group invariant, also known as colored HOMFLY polynomial, of the link ${\mathcal{L}}$ is defined as follows[^1]: $$\label{HMF1} H_{R_1,\dots,R_L}^{{\mathcal{L}}} = {} _q\text{tr}_{V_1\otimes\dots\otimes V_m}\left( \, \pi(\beta_{{\mathcal{L}}}) \, \right),$$ where $_q\text{tr}$ is a quantum trace. $\bullet$ Although a quantum trace is a standard notion in the theory of quantum groups [@Klimyk], we give here some details, which are useful in Reshetikhin-Turaev approach. Let $\rho$ denotes a half-sum of all positive roots of $sl_N$. There exists an element $K_{2\rho} \in U_q(sl_N)$ defined by $$\begin{array}{l} K_{2\rho} = K_1^{n_1} \, K_2^{n_2} \ldots K_{N-1}^{n_{N-1}}, \\ 2\rho = \sum\limits_{i=1}^{N-1} n_i\alpha_i, \ n_i \in \mathbb{N}_0, \\ K_i = q^{H_i} \ \forall i=1,...,N-1, \end{array}$$ where $\alpha_i$ are simple roots. Then for every $z \in \text{End}(V)$ one has the quantum trace $$\label{qtr} _q\text{tr}_V(z) = \text{tr}_V(zK_{2\rho}).$$ $\bullet$ Let us expand $V_1\otimes V_2\otimes \ldots \otimes V_m$ into a direct sum of irreducible representations: $$\label{irrdec} \bigotimes_i V_i = \bigoplus_{\mu} \mathcal{M}_\mu \otimes Q_{\mu},$$ where $Q_{\mu}$ is an irreducible representation and $\mathcal{M}_\mu$ is the subspace of highest weight vectors with highest weights[^2] corresponding to a Young diagram $\mu$. The dimension of the space $\mathcal{M}_\mu$ is called *the multiplicity* of representation $Q_{\mu}$. Let us evaluate quantum trace (\[qtr\]) on such decomposition (\[irrdec\]). Since $\mathcal{R}$-matrix commutes with any element from $U_q(sl_N)$, then $\mathcal{R}$-matrix gets a block structure corresponding to the decomposition on irreducible components (\[irrdec\]), i.e. it does not mix vectors from different representations. Futhermore, $\mathcal{R}$-matrix acts on $Q_{\mu}$ as an identity operator, while on $\mathcal{M}_\mu$ it acts nontrivially. The element $K_{2\rho}$ acts diagonally on $Q_{\mu}$, while on $\mathcal{M}_\mu$ it acts identically, because there are highest weight vectors with the same weight $\vec{\omega}(\mu)$. Therefore, the decomposition (\[irrdec\]) implies for the colored HOMFLY (\[HMF1\]) the following $$\begin{array}{ccc} H_{R_1,\dots,R_L}^{{\mathcal{L}}} \left(q,\,A=q^N \right) \ = \ \text{tr}_{V_1\otimes\dots\otimes V_m}\left( \, \pi(\beta_{{\mathcal{L}}}) \, K_{2\rho} \, \right) \ = \ \sum \limits_{\mu\, \vdash \sum |R_i|} \text{tr}_{\mathcal{M}_\mu}\left( \, \pi(\beta_{{\mathcal{L}}}) \, \right) \cdot \text{tr}_{Q_{\mu}}\left( \, K_{2\rho} \, \right) = \\ \\ = \sum\limits_{\mu\, \vdash \sum |R_i|} \text{tr}_{\mathcal{M}_\mu}\left( \, \pi(\beta_{{\mathcal{L}}}) \, \right) \cdot \text{qdim}_{\mu}, \label{linv} \end{array}$$ where $\text{qdim}_{\mu}$ is a quantum dimension of the representation $Q_{\mu}$ explicitly given in terms of Schur polynomials [@Lin]: $$\text{qdim}_{\mu} = s_{\mu}\left( x_1,\ldots,x_N \right){\Big|_{x_i=q^{N+1-2i}}} = s_{\mu}^{}(p_1,\dots,p_N)\Big|_{p_k=p_k^{*}} \equiv s_{\mu}^{*}(A,q),$$ where $p_k=\sum\limits_{i=1}^{N} x_i^k$ and $p_k^* = \frac{A^k-A^{-k}}{q^k-q^{-k}}$. Quantum Racah coefficients (6-j symbols) ---------------------------------------- In this subsection we define quantum Racah coefficients, also known as 6-j symbols. Consider three finite-dimensional irreducible representations $R_{i_1},\ R_{i_2}, \ R_{i_3}$ of highest weights of $U_q(sl_N)$. Since the tensor product of these representations is associative, one has a natural isomorphism: $$\label{asis} \left( R_{i_1}\otimes R_{i_2} \right)\otimes R_{i_3} \rightarrow R_{i_1} \otimes \left( R_{i_2}\otimes R_{i_3} \right).$$ Expand tensor product of two representations into irreducible components like in (\[irrdec\]): $$\begin{array}{l} R_{i_1}\otimes R_{i_2} = \bigoplus_{\mu} \mathcal{M}^{i_1,i_2}_\mu \otimes X_{\mu}, \\ R_{i_2}\otimes R_{i_3} = \bigoplus_{\nu} \mathcal{M}^{i_2,i_3}_\nu \otimes Y_{\nu}, \end{array}$$ and once again $$\begin{array}{l} \left( R_{i_1}\otimes R_{i_2} \right)\otimes R_{i_3} = \bigoplus_{\mu,\xi} \mathcal{M}^{i_1,i_2}_\mu \otimes \mathcal{M}^{\mu,i_3}_\xi \otimes R_{\xi}, \\ R_{i_1} \otimes \left( R_{i_2}\otimes R_{i_3} \right) = \bigoplus_{\nu,\xi} \mathcal{M}^{i_1,\nu}_\xi \otimes \mathcal{M}^{i_2,i_3}_\nu \otimes R_{\xi}. \end{array}$$ Then the associativity isomorphism (\[asis\]) implies $$\Phi^{i_1,i_2}_{i_3,\xi}: \ \bigoplus_{\mu} \mathcal{M}^{i_1,i_2}_\mu \otimes \mathcal{M}^{\mu,i_3}_\xi \ \rightarrow \ \bigoplus_{\nu} \mathcal{M}^{i_1,\nu}_\xi \otimes \mathcal{M}^{i_2,i_3}_\nu,$$ and the Racah coefficients are defined as the map components: $$U\left[\begin{array}{cc|c} i_1&i_2& \mu \\ i_3&\xi &\nu\end{array}\right] = \left( \Phi^{i_1,i_2}_{i_3,\xi} \right)_{\mu,\nu}.$$ Since this paper is devoted to Racah coefficients, then let us change some notations to make them more convenient and obvious. Instead of indices $i_1, \ i_2, i_3$ and $\xi$ we shall use representations $R_1, \ R_2, \ R_3$ and $R_4$, and instead of intermediate indices $\mu$ and $\nu$ we use $X \in R_1\otimes R_2$ and $Y \in R_2\otimes R_3$. Thus, we denote Racah matrix by $$U\left[\begin{array}{cc} R_1&R_2 \\ R_3&R_4 \end{array}\right]$$ and Racah coefficient by $$U\left[\begin{array}{cc|c} R_1&R_2& X \\ R_3&R_4 &Y\end{array}\right].$$ Usually quantum 6-j symbols have different normalisation than Racah coefficients, but for our purposes such normalisation is not important, hence we assume that they coincide: { [ccc]{} R\_1 & R\_2 & X\ R\_3 & R\_4 & Y } U Graphically Racah matrix can be represented as follows (240,110)(-100,-60) (0,0)[(1,0)[50]{}]{} (0,0)[(-1,1)[30]{}]{} (0,0)[(-1,-1)[30]{}]{} (50,0)[(1,1)[30]{}]{} (50,0)[(1,-1)[30]{}]{} (-45,-30) (-45,30) (85,-30) (85,30) (22,4) (130,0)[(1,0)[40]{}]{} (148,5) (250,0)[ (0,-20)[(0,1)[40]{}]{} (0,-20)[(-1,-1)[30]{}]{} (0,-20)[(1,-1)[30]{}]{} (0,20)[(1,1)[30]{}]{} (0,20)[(-1,1)[30]{}]{} (-45,-40) (-45,40) (35,40) (35,-40) (5,-4) ]{} Their explicit calculation in great details through highest weight vectors for different representation can be found in [@MMMS21]-[@China]. $\mathcal{R}$-matrices via Racah matrices ----------------------------------------- Now we describe how quantum Racah matrices relate different $\mathcal{R}$-matrices (\[Rmat\]) with each other [@KirResh]. $\bullet$ Let us choose the basis on $R_1\otimes\dots\otimes R_m$, which corresponds to the following order in the tensor product: $$B_{12,3,..,m} := \left(\ldots\Big(\,\left( R_1\otimes R_2 \right)\otimes R_3 \Big) \otimes \ldots \right) \otimes R_m.$$ Then the matrix ${\mathcal{R}}_1$ in this basis gets a block form on each space ${\mathcal{M}}_{\mu}$ from (\[irrdec\]). Different blocks corresponds to different representations $X_{\alpha}$ in the decomposition $$R_1 \otimes R_2 = \bigoplus_{\mu} \mathcal{M}^{1,2}_\alpha \otimes X_{\alpha}$$ as we discussed after formula (\[irrdec\]). The size of the block, which correspond to $X_{\alpha}$, is dim$\,\mathcal{M}^{1,2}_\alpha \times$ dim$\,\mathcal{M}^{1,2}_\alpha$. Therefore, if dim$\,\mathcal{M}^{1,2}_\alpha > 1$, then we need to rotate the basis additionally (those components, which correspond to $\alpha$) to diagonalize the matrix ${\mathcal{R}}_1$, but it is always possible to do. Thus, in the basis $B_{12,3,..,m}$ it is always possible to diagonalize the matrix ${\mathcal{R}}_1$. $\bullet$ In order to diagonalize the matrix $\mathcal{R}_2$ we have to repeat the same procedure but for the basis corresponding to $$B_{1,23,..,m} := \left(\ldots\Big( R_1\otimes \left(R_2 \otimes R_3 \right) \Big)\otimes \ldots \right)\otimes R_m.$$ Therefore, in order to diagonalize the matrix $\mathcal{R}_2$ one should make the basis transformation with the help of Racah matrix: $$\label{R2} \mathcal{R}_2 = U^{\dagger}\left[\begin{array}{cc} R_1&R_3 \\ R_2&R_4 \end{array}\right] \cdot \text{diag}\left( \lambda_{\mathcal{R}_2} \right) \cdot U\left[\begin{array}{cc} R_1&R_2 \\ R_3&R_4 \end{array}\right],$$ where $\lambda_{\mathcal{R}_2}$ mean eigenvalues of the matrix $\mathcal{R}_2$ and $U$ are corresponding Racah matrices. Similarly we can diagonalize all $\mathcal{R}$-matrices. So, in order to evaluate link invariant (\[linv\]) we need to find eigenvalues of $\mathcal{R}$-matrices and Racah coefficients. While eigenvalues of $\mathcal{R}$-matrices are known explicitly and we give answer for arbitrary representations below, Racah coefficients are known only for a few simple cases. Their calculation is puzzling and hard problem standing over 50 years both for quantum and classical groups, solved only for $U_q(sl_2)$ case. Eigenvalues of ${\mathcal{R}}$-matrix ------------------------------------- Now let us consider tensor product of two representations $R$ and $R$ and irreducible representations $X_{\alpha}$, which occur in their decomposition into irreducible components: $$R\otimes R = \bigoplus_{\alpha} X_{\alpha}.$$ In this sum we allow repeated summands. Then according to [@GZ; @Klimyk] eigenvalues of universal R-matrix are $$\label{evR} \begin{array}{l} \lambda_{X_{\alpha}}=\epsilon_{X_{\alpha}} q^{\varkappa(X_{\alpha})-4\varkappa({R})-|R|N} \end{array}$$ where $\varkappa(X_{\alpha})$ defined by $\sum_{(i,j)\in \alpha} \, (i-j)$, and $\epsilon_{X_{\alpha}}=\pm 1$ is a sign, which depends on whether highest weight vectors of the representations are *symmetric* or *antisymmetric* under permutation of two representations $R$ and $R$. These two types of representations $X_{\alpha}$ are said to belong to either symmetric or antisymmetric squares of the representation $R$. Finally we define [*normalised*]{} diagonal $\mathcal{R}$-matrix by the following conditions 1. $\prod_i {\lambda}_i = 1$, 2. $\log_q |\lambda_1| \geq \log_q |\lambda_2| \geq \ldots$, 3. if $\log_q |\lambda_i| = \log_q |\lambda_{i+1}|$, then we put eigenvalues with positive signs first, 4. $\text{sgn}(\lambda_1) > 0$. Eigenvalue conjecture {#evcj} --------------------- In [@IMMMec] the eigenvalue conjecture was suggested. It relates quantum Racah coefficients with eigenvalues of $\mathcal{R}$-matrices. Since in this paper we consider 3-strand braids only, then we give a formulation for this case only. In the simplest case of $3$-strand knots the eigenvalue conjecture states that the Racah matrices are fully-defined by the set of normalised eigenvalues of the corresponding $\mathcal{R}$-matrix. If we consider $3$-strand braid with representations $R_1=R_2=R_3=R$. Then there is the unique diagonal $\mathcal{R}$-matrices, say, $\mathcal{R}_{1}$ corresponding to basis $B_{12,3}$. The source of the eigenvalue conjecture is the Yang-Baxter equation, which for the 3-strand braid reads as \_1\_2\_1=\_2\_1\_2 Substituting formula in this equation we get \[req\] \_1U\^ \_1 U \_1 = U\^ \_1 U \_1 U\^ \_1 U. The latter equation is invariant under a scalar multiplication of the ${\mathcal{R}}_1$-matrix and permutations of diagonal elements. Therefore, if we have ${\mathcal{R}}_1$ and $\tilde{{\mathcal{R}}}_1$, which coincide after the normalisation procedure, then Racah matrices $U$ and $\tilde U$ satisfy same equation . A priori there is no reason this equation (in more exact terms it is a system on nonlinear equations) has always a unique solution corresponding to the quantum group. Nevertheless we formulate the following conjecture. **Conjecture:** *Let we have two sets $\{R,R,R,Q;\, {\mathcal{R}}_1, U\}$ and $\{\tilde R,\tilde R,\tilde R,\tilde Q;\,\tilde {\mathcal{R}}_1,\tilde U\}$, where $R,\tilde R,Q,\tilde Q$ are representations as follows $R^{\otimes 3}\rightarrow Q$, $\tilde R^{\otimes 3}\rightarrow \tilde Q$, matrices ${\mathcal{R}}_1$ and $\tilde R_1$ are the corresponding diagonal ${\mathcal{R}}$-matrices, $U$ and $\tilde U$ are the corresponding Racah matrices. Let normalised $R_1$ concides with normalised $\tilde {\mathcal{R}}_1$. Then $U=\tilde U$.* **Remark:** If the normalised $R_1$ concides with the normalised $\tilde {\mathcal{R}}_1^{-1}$, then we conjecture $U=\tilde U^{-1}$. The symmetry properties of 6-j symbols ====================================== In this section we briefly review the symmetry properties of quantum 6-j symbols. [$\bullet$ $\rm \bf U_q(sl_2)$]{} In this case all symmetries are known due to representation of 6-j symbols in terms of q-hypergeometric function. The symmetry group contains 144 elements, the full tetrahedral group $S_4$ is its subgroup. It describes *the tetrahedral* symmetry, which can be represented in a pictorial way by the associating a corresponding tetrahedron with the 6-j symbol: (850,160)(-250,-25) (-200,40) (-100,40)(30,-5.5)[6]{}[(6,-1)[14]{}]{} (-100,40)[(3,-2)[90]{}]{} (-10,-20)[(1,6)[25.4]{}]{} (-100,40)[(5,4)[115.5]{}]{} (-10,-20)[(5,2)[74.5]{}]{} (15.5,132.5)[(2,-5)[49]{}]{} (-55,90) (-65,0) (-12,60) (30,-15) (50,70) Other symmetries are given by the q-analog of *the Regge* symmetry. Let $p=\frac{1}{2}\left( r_1+r_2+r_3+r_4 \right)$, then the Regge symmetry holds: \[Regge\] { [ccc]{} r\_1 & r\_2 & r\_[12]{}\ r\_3 & r\_4 & r\_[23]{} } = { [ccc]{} p-r\_1 & p-r\_2 & r\_[12]{}\ p-r\_3 & p-r\_4 & r\_[23]{} } Alternatively, we can represent all symmetries in the following way. Let us take the 6-j symbol $$\left\{\begin{array}{ccc} r_1 & r_2 & r_{12} \\ r_3 & r_4 & r_{23} \end{array}\right\}$$ and reexpress it through a different set of parameters: $$\left[\begin{array}{ccc} A & B & C \\ \alpha & \beta & \gamma \end{array}\right] := \left\{\begin{array}{ccc} \frac{1}{2}(A+\alpha) & \frac{1}{2}(B+\beta) & \frac{1}{2}(C+\gamma) \phantom{\dfrac{1}{8}} \\ \frac{1}{2}(A-\alpha) & \frac{1}{2}(B-\beta) & \frac{1}{2}(C-\gamma) \end{array} \hspace{-2mm} \right\}$$ Then the 6-j symbol is invariant under separate permutations of $\{A,B,C\}$ alone, separate permutations of $\{\alpha,\beta,\gamma\}$ alone and separate change in sign of any pair of $\{\alpha,\beta,\gamma\}$. [$\bullet$ $\rm \bf U_q(sl_N)$]{} In this case we know only about the tetrahedral symmetry [@GuJ; @Lienert; @Pan]. The counterpart of the Regge symmetry is unknown at the present moment. Eigenvalue conjecture for $U_q(sl_2)$ ===================================== In this section we prove that the eigenvalue conjecture is true if $R_1=R_2=R_3=[r]$ for $U_q(sl_2)$. Moreover, we show that it follows from the Regge symmetry. Since $R_4 \in R_1\otimes R_2\otimes R_3$ the corresponding Young diagram according to the Littlewood-Richardson rule can take the form $[2m], \ m=0\ldots \frac{3}{2}r$. We consider two cases $R_4=[r+2k], \ k=0\ldots r$ and $R_4=[r-2k], \ k=0\ldots \lfloor\frac{r}{2}\rfloor$ separately. **1.** $R_4=[r+2k]$, where $k=0\ldots r$. To understand which eigenvalues correspond to the Racah matrix we should find all representations $X \in [r]\otimes [r]$, which contributes in this case. In general these representations can be denoted as $[2r-2k_1]$. Then there are the following restrictions: $$\begin{array}{lcl} \ [2r-2k_1]\in [r]\otimes [r] & & 0\leq k_1 \leq r \\ & \Rightarrow & \\ \ [r+2k]\in [2r-2k_1]\otimes [r] & & k_1\leq r-k \end{array}$$ Thus it comes to $0\leq k_1\leq r-k$. Therefore, the corresponding 6-j symbols are \[rac1\] { [ccc]{} r & r & 2r-2k\_1\ r & r+2k & 2r-2k\_2 },   k\_1,k\_2=0…r-k. The size of the Racah matrix is given by $d=r-k+1$. Since the eigenvalue conjecture implies to fix the size of the Racah matrices, then we must vary representation $r$ in in order to consider nontrivial cases. However the eigenvalue conjecture does not give us something nontrivial in this case, because there are no two different representations $r$ and $R$ such that the corresponding normalised eigenvalues coincide. It follows from a staraightforward calculation of a difference of the powers between two adjacent eigenvalues: (\[2r-2k\_1\]) - (\[2r-2k\_1-2\]) = 4r-4k\_1-3, k\_1=0…r-k-2. **2.** $R_4=[r-2k]$, where $k=0\ldots \lfloor\frac{r}{2}\rfloor$. Again let us look at the representations from the tensor square: $$\begin{array}{lcl} \ [2r-2k_1]\in r\otimes r & & 0\leq k_1 \leq r \\ & \Rightarrow & \\ \ [r-2k]\in [2r-2k_1]\otimes r & & |r-2k_1| \leq r-2k \leq |3r-2k_1| \end{array}$$ This gives $k\leq k_1\leq r-k$. Therefore, the corresponding Racah coefficients are $$\left\{\begin{array}{ccc} r & r & 2r-2k_1 \\ r & r-2k & 2r-2k_2 \end{array}\right\},\ \ \ k_1,k_2=k\ldots r-k$$ The size of the Racah matrix is given by $d=r-2k+1$. Similar reasoning, which we use above, shows that the eigenvalue conjecture says nothing about type 2 representations. **3.** Let us now consider a type 1 Racah matrix of the size $d\times d$ and a type 2 matrix of the same size and apply the eigenvalue conjecture. The corresponding sets of representations in the tensor square are $$\begin{array}{c} \ [2r],\ [2r-2], \, \ldots ,\ [2r-2d+2] \\ \ [R+d-1], \ [R+d-3], \, \ldots, \ [R-d+1] \end{array}$$ If $R+d-1=2r$ then eigenvalues coincide. So eigenvalue conjecture tells us that the following Racah coefficients should coincide. For a type 1 matrix and a representation $r$, substituting $k=r+1-d$, we have: $$\label{rrac} \boxed{ \left\{\begin{array}{ccc} r & r & 2r-2k_1 \\ r & 3r-2d+2 & 2r-2k_2 \end{array}\right\},\ \ \ k_1,k_2=0\ldots d-1 }$$ For a type 2 matrix and a representation $R$, substituting $R=2r-d+1$ and $k=\frac{R-d+1}{2}=r-d+1$: \[Rrac\] [l]{} { [ccc]{} 2r-d+1 & 2r-d+1 & 4r-2d+2-2k\_1\ 2r-d+1 & d-1 & 4r-2d+2-2k\_2 },   k\_1,k\_2=r-d+1…r =\ \ = Thus, the eigenvalue conjecture predicts that 6-j symbol coincides with . With the help of the Regge symmetry we can easily check that it is true. Indeed, the Regge symmetry applied to the 6-j symbols immediately gives 6-j symbols . Eigenvalue conjecture for $U_q(sl_N)$ ===================================== The eigenvalue conjecture also predicts the symmetry between different Racah matrices for the quantum group of the rank higher than 2. Such cases are more involved because they usually contain multiplicities. There is no analytical description of 6-j symbols like for $U_q(sl_2)$, consequently we can only perform some checks of the eigenvalue conjecture. $\bullet$ In the papers [@IMMMec; @Wenzl] there were found general nontrivial solutions of the Yang-Baxter equation if the size of the R-matrix not greater than 5 and all its eigenvalues are different. These solutions provide explicit formulas for matrix elements (i.e. for 6-j symbols) in terms of algebraic functions depending on eigenvalues. Therefore, these formulas provide the proof of the eigenvalue conjecture for this particular class of matrices. Note these formulas are valid for any rank $N$ of the algebra $U_q(sl_N)$. $\bullet$ If the size of the Racah matrix is greater than 5, then there are no explicit solutions of the Yang-Baxter equation. However we manage to present some particular examples of the Racah matrices, which confirm the eigenvalue conjecture. These examples we computed in the series of papers [@MMMS21]-[@China]: U&=& U, 66,  \ U&=& U, 66,  \ U&=& U, 99,  . Even these few examples give us a hope that further studies of the eigenvalue conjecture will provide some general symmetry of the Racah coefficients for any $U_q(sl_N)$ group. However at the moment some general expressions of the eigenvalue conjecture are lacking and this remains to be done. Based on the previous section we can provide a $U_q(sl_N)$ version of the same relation. The formulae (\[rrac\]) and (\[Rrac\]) become: $$\begin{array}{r} \left[\begin{array}{ccc} [2r-N+1] & [2r-N+1] & [3r-N+1-k_1,k_1+r-N+1] \\ {[}2r-N+1] & [3r-N+1,3r-2N+2] & [3r-N+1-k_2,k_2+r-N+1] \end{array}\right] = \\ = \left[\begin{array}{ccc} [r] & [r] & [2r-k_1,k_1] \\ {[}r] & [3r-N+1,N-1] & [2r-k_2,k_2] \end{array}\right] ,\ \ \ k_1,k_2=0\ldots N-1 \end{array}$$ We hope that this formula is just the beginning of the formulation of Racah coefficients symmetries in the $U_q(sl_N)$ case. Acknowledgements {#acknowledgements .unnumbered} ================ This work was funded by the Russian Science Foundation (Grant No.16-11-10291). [99]{} A. Kirillov and N. Reshetikhin, [*Representations of the Algebra $U_q(sl_2)$, q-Orthogonal Polynomials and Invariants of Links*]{}. New Developments in the Theory of Knots: pp. 202-256, World Scientific, Singapore (1990) S. Nawata, P. Ramadevi, Zodinmawia, Lett.Math.Phys. [**103**]{} (2013) 1389-1398, arXiv:1302.5143 S. Nawata, P. Ramadevi, Zodinmawia, J.Knot Theory and Its Ramifications, [**22**]{} (2013) 13, arXiv:1302.5144 A. Mironov, A. Morozov, A. Sleptsov, JHEP [**07**]{} (2015) 069, arXiv:1412.8432 A. Caudron, [*Classification des noeuds et des enlacements*]{}, Publ. Math. Orsay [**82-4**]{}, University of Paris XI, Orsay, 1982 F. Bonahon, L.C. Siebenmann, http://www-bcf.usc.edu/$\sim$fbonahon/Research/Preprints/BonSieb.pdf, [*New geometric splittings of classical knots and the classification and symmetries of arborescent knots*]{}, 2010 P. Ramadevi, T.R. Govindarajan, R.K. Kaul, Mod.Phys.Lett. [**A9**]{} (1994) 3205-3218, hep-th/9401095 D. Galakhov, D. Melnikov, A. Mironov, A. Morozov, A. Sleptsov, Phys.Lett. [**B743**]{} (2015) 71-74, arXiv:1412.2616 A. Mironov, A. Morozov, An. Morozov, P. Ramadevi, V.K. Singh, JHEP, [**1507**]{} (2015) 109, arXiv:1504.00371 R.K. Kaul, T.R. Govindarajan, Nucl.Phys. [**B380**]{} (1992) 293-336, hep-th/9111063 P. Ramadevi, T.R. Govindarajan, R.K. Kaul, Nucl.Phys. [**B402**]{} (1993) 548-566, hep-th/9212110 P. Ramadevi, T.R. Govindarajan, R.K. Kaul, Nucl.Phys. [**B422**]{} (1994) 291-306, hep-th/9312215 P. Ramadevi, T. Sarkar, Nucl.Phys. [**B600**]{} (2001) 487-511, hep-th/0009188 J. Gu, H. Jockers, Comm.Math.Phys., [**338**]{}(1) (2015) 393-456, arXiv: 1407.5643 A. Mironov, A. Morozov, An. Morozov, JHEP 03 (2012) 034, arXiv:1112.2654 H. Itoyama, A. Mironov, A. Morozov, An. Morozov, Int.J.Mod.Phys. [**A28**]{} (2013) 1340009, arXiv:1209.6304 I. Tuba, H. Wenzl, *Representations of the braid group $B_3$ and of SL(2,Z)* , Pacific Journal of Mathematics, Vol. 197 (2001), No. 2, 491510, arXiv:math/9912013 A.Mironov, A.Morozov, Physics Letters [**B 755**]{}, 10 April 2016, Pages 47-57 Saswati Dhara, A. Mironov, A. Morozov, An. Morozov, P. Ramadevi, Vivek Kumar Singh, A. Sleptsov, *Multi-Colored Links From 3-strand Braids Carrying Arbitrary Symmetric Representations*, arXiv:1805.03916 L. D. Landau and E. M. Lifshitz, [*Quantum Mechanics: Non-Relativistic Theory*]{}, (3rd ed.) Pergamon Press, 1977 M. Rosso, V. F. R. Jones, J. Knot Theory Ramifications, **2** (1993) 97-112 Guadagnini E., Martellini M., Mintchev M. (1990) *Chern-Simons field theory and quantum groups*. In: Doebner H.D., Hennig J.D. (eds) Quantum Groups. Lecture Notes in Physics, vol 370. Springer, Berlin, Heidelberg E. Guadagnini, M. Martellini, M. Mintchev, *Chern-Simons holonomies and the appearance of quantum groups*, Phys.Lett. [**B235**]{} (1990) 275 N.Yu. Reshetikhin, V.G. Turaev, *Ribbon graphs and their invariants derived from quantum groups*, Comm.Math.Phys. [**127**]{} (1990) 1-26 K. Liu, P. Peng, J. Differential Geom., Volume 85, Number 3 (2010), 479-525, arXiv:0704.1526 A. Klimyk, K. Schm$\rm \ddot{u}$dgen, *Quantum groups and their representations*, Springer Science $\&$ Business Media, 2012 X.-S. Lin, H. Zheng, Trans.Amer.Math.Soc. **362** (2010) 1-18 math/0601267 A. Mironov, A. Morozov, An. Morozov, A. Sleptsov, *Colored knot polynomials: HOMFLY in representation \[2, 1\]*, J.Mod.Phys. [**A30**]{} (2015) 1550169, arXiv:1508.02870 A. Mironov, A. Morozov, An. Morozov, A. Sleptsov, *HOMFLY polynomials in representation \[3, 1\] for 3-strand braids*, JHEP, [**2016**]{} (2016) 134, arXiv:1605.02313 A. Mironov, A. Morozov, An. Morozov, A. Sleptsov, *Quantum Racah matrices and 3-strand braids in irreps R with $|R|= 4$*, JETP Lett. [**104**]{} (2016) 56-61, arXiv:1605.03098 Sh. Shakirov, A. Sleptsov, *Quantum Racah matrices and 3-strand braids in representation \[3,3\]*, arXiv:1611.03797 C. Bai, J. Jiang, J. Liang, A. Mironov, A. Morozov, An. Morozov, A. Sleptsov, Journal of Geometry and Physics, Vol. 132, 2018, pp. 155-180, arXiv:1801.09363 M. Gould, Y. Zhang, *Quantum affine Lie algebras, Casimir invariants, and diagonalization of the braid generator*, Journal of Mathematical Physics, (1994), 35(12), 6757-6773, arXiv: hep-th/9311041 C. R. Lienert and P. H. Butler, *Racah-Wigner algebra for q-deformed algebras*, Journal of Physics A: Mathematical and General 25 (1992) 1223. F. Pan, *Racah coefficients of quantum group $U_q(n)$*, Journal of Physics A: Mathematical and General 26 (1993) 4621. [^1]: The usual framing factor [@LiuPeng] in front of the quantum trace, which provide the invariance under the first Reidemeister move, we incorporate in the ${\mathcal{R}}$-matrix by modifying its eigenvalues (\[evR\]). [^2]: Recall, that if $\mu$ is a Young diagram $\mu=\{\mu_1\geq\mu_2\geq\ldots,\mu_l>0\}$, then the highest weights $\vec{\omega}$ of the corrsponding representation are $\omega_i= \mu_i-\mu_{i+1} \ \forall \, i=1,\dots,l$, and vice versa $\mu_i = \sum_{k=i}^l \, \omega_k$.

ArXiv

--- abstract: 'Class imbalance is a long-standing problem relevant to a number of real-world applications of deep learning. Oversampling techniques, which are effective for handling class imbalance in classical learning systems, can not be directly applied to end-to-end deep learning systems. We propose a three-player adversarial game between a convex generator, a multi-class classifier network, and a real/fake discriminator to perform oversampling in deep learning systems. The convex generator generates new samples from the minority classes as convex combinations of existing instances, aiming to fool both the discriminator as well as the classifier into misclassifying the generated samples. Consequently, the artificial samples are generated at critical locations near the peripheries of the classes. This, in turn, adjusts the classifier induced boundaries in a way which is more likely to reduce misclassification from the minority classes. Extensive experiments on multiple class imbalanced image datasets establish the efficacy of our proposal.' author: - | Sankha Subhra Mullick\ Indian Statistical Institute\ Kolkata, India\ [sankha\_r@isical.ac.in]{} - | Shounak Datta\ Duke University\ Durham, NC, USA\ [shounak.jaduniv@gmail.com]{} - | Swagatam Das\ Indian Statistical Institute\ Kolkata, India\ [swagatam.das@isical.ac.in]{} bibliography: - 'paper.bib' title: Generative Adversarial Minority Oversampling --- Introduction {#sec:intro} ============ The problem of class imbalance occurs when all the classes present in a dataset do not have equal number of representative training instances [@He2009learningimb; @das2018handling]. Most of the existing learning algorithms produce inductive bias favoring the majority class in presence of class imbalance in the training set, resulting in poor performance on the minority class(es). This is a problem which routinely plagues many real-world applications such as fraud detection, dense object detection [@lin2017cost], medical diagnosis, etc. For example, in a medical diagnosis application, information about unfit patients is scarce compared to that of fit individuals. Hence, traditional classifiers may misclassify some unfit patients as being fit, having catastrophic implications [@mazurowski2008training]. 0.2in 0.03in 0.03in -0.4in Over the years, the machine learning community has devised many methods for tackling class imbalance [@krawczyk2016learning; @branco2016]. However, only a few of these techniques have been extended to deep learning even though class imbalance is fairly persistent is such networks, severely affecting both the feature extraction as well as the classification process [@xie2015holistically; @huang2016learning; @xie2017holistically; @buda2018systematic; @khan2018cost]. The existing solutions [@huang2016learning; @Chung2016c; @wang2016d; @lin2017cost; @rota2017cost] for handling class imbalance in deep neural networks mostly focus on cost tuning to assign suitably higher costs to minority instances. Another interesting class of approaches [@yan2015; @dong2018imbalanced] focuses on constructing balanced subsamples of the dataset. Wang et al. [@wang2017meta] proposed a novel meta-learning scheme for imbalanced classification. It is interesting to note that oversampling techniques like SMOTE [@chawla2002smote] have not received much attention in the context of deep learning, despite being very effective for classical systems [@fernandez2018smote]. This is because deep feature extraction and classification are performed in an end-to-end fashion, making it hard to incorporate oversampling which is typically done subsequent to feature extraction. An attempt to bridge this gap was made by Ando and Huang [@shin2017dos] in their proposed deep oversampling framework (DOS). However, DOS uniformly oversamples the entire minority class and is not capable of concentrating the artificial instances in difficult regions. Additionally, the performance of DOS depends on the choice of the class-wise neighborhood sizes, which must be determined by costly parameter tuning. Generative adversarial networks (GANs) are a powerful subclass of generative models that have been successfully applied to image generation. This is due to their capability to learn a mapping between a low-dimensional latent space and a complex distribution of interest, such as natural images [@goodfellow2014generative; @mirza2014conditional; @radford2015unsupervised; @odena2017conditional]. The approach is based on an adversarial game between a generator that tries to generate samples which are similar to real samples and a discriminator that tries to discriminate between real training samples and generated samples. The success of GANs as generative models has led Douzas and Bacao [@douzas2018effective] to investigate the utility of using GANs to oversample the minority class(es). However, attempting to oversample the minority class(es) using GANs can lead to boundary distortion [@santurkar2018classification], resulting in a worse performance on the majority class (as illustrated in Figure \[fig:exampleCGan\]). Moreover, the generated points are likely to lie near the mode(s) of the minority class(es) [@srivastava2017veegan], while new points around the class boundaries are required for learning reliable discriminative (classification) models [@hui2005bSmote; @He2008adasyn]. Hence, in this article, we propose (in Section \[sec:method\]) a novel end-to-end feature-extraction-classification framework called Generative Adversarial Minority Oversampling (GAMO) which employs adversarial oversampling of the minority class(es) to mitigate the effects of class imbalance. The contributions made in this article differ from the existing literature in the following ways: 1. Unlike existing deep oversampling schemes [@shin2017dos; @douzas2018effective], GAMO is characterized by a three-player adversarial game among a convex generator $G$, a classifier network $M$, and a discriminator $D$. 2. Our approach is fundamentally different from existing adversarial classification schemes (where the generator works in harmony with the classifier to fool the discriminator) [@salimans2016improved; @kumar2017semi; @springenberg2015unsupervised; @odena2017conditional], in that our convex generator $G$ attempts to fool both $M$ and $D$. 3. Unlike the generator employed in GAN [@goodfellow2014generative], we constrain $G$ to conjure points within the convex hull of the class of interest. Additionally, the discriminator $D$ further ensures that $G$ adheres to the class distribution for non-convex classes. Consequently, the adversarial contention with $M$ pushes the conditional distribution(s) learned by $G$ towards the periphery of the respective class(es), thus helping compensate for class imbalance effectively. 4. In contrast to methods like [@chawla2002smote; @douzas2018effective], $G$ can oversample different localities of the data distribution to different extents based on the gradients obtained from $M$. 5. For applications requiring a balanced training set of images, we also propose a technique called GAMO2pix (Section \[sec:imgGen\]) that can generate realistic images from the synthetic instances generated by GAMO in the distributed representation space. We undertake an ablation study as well as evaluate the performance of our method compared to the state-of-the-art in Section \[sec:exp\], and make concluding remarks in Section \[sec:concl\]. Related Works ============= The success of SMOTE [@chawla2002smote; @chawla2003smoteboost] has inspired several improvements. For example, [@hui2005bSmote; @chumphol2009sls] attempt to selectively oversample minority class points lying close to the class boundaries. Works like [@He2008adasyn; @Lin2013Dys; @barua2014mvmote], on the other hand, asymmetrically oversample the minority class such that more synthetic points are generated surrounding the instances which are difficult to classify. Although these methods achieved commendable improvement on classical classifiers, they can neither be extended to deep learning techniques nor be applied to images, respectively due to the end-to-end structure of deep learning algorithms and a lack of proper notion of distance between images. Extending GANs for semi-supervised learning, works like [@kumar2017semi; @salimans2016improved] fused a $c$-class classifier with the discriminator by introducing an extra output line to identify the fake samples. On the other hand, [@springenberg2015unsupervised] proposed a $c$-class discriminator which makes uncertain predictions for fake images. Additionally, [@odena2017conditional] proposed a shared discriminator-cum-classifier network which makes two separate sets of predictions using two different output layers. These approaches can loosely be considered to be related to GAMO as these also incorporate a classifier into the adversarial learning scheme. 0.2in -0.3in Proposed Method {#sec:method} =============== Let us consider a $c$-class classification problem with a training dataset $X \subset \mathbb{R}^{D}$ (of images vectorized either by flattening or by a convolutional feature extraction network $F$). Let the prior probability of the $i$-th class be $P_{i}$, where $i \in \mathcal{C}=\{1, 2, \cdots c\}$; $\mathcal{C}$ being the set of possible class labels. Without loss of generality, we consider the classes to be ordered such that $P_{1} \leq P_{2} \leq \cdots < P_{c}$. Moreover, let $X_{i}$ denote the set of all $n_{i}$ training points which belong to class $i \in \mathcal{C}$. We intend to train a classifier $M$ having $c$ output lines, where the $i$-th output $M_{i}(\mathbf{x})$ predicts the probability of any $\mathbf{x}\in X$ to be a member of the $i$-th class. Adversarial Oversampling ------------------------ Our method plays an adversarial game between a classifier that aims to correctly classify the data points and a generator attempting to spawn artificial points which will be misclassified by the classifier. The idea is that generating such difficult points near the fringes of the minority class(es) will help the classifier to learn class boundaries which are more robust to class imbalance. In other words, the performance of the classifier will adversarially guide the generator to generate new points at those regions where the minority class under concern is prone to misclassification. Moreover, the classifier will aid the generator to adaptively determine the concentration of artificial instances required to improve the classification performance in a region, thus relieving the user from tuning the amount of oversampling. Instead, we only need to fix the number of points to be generated to the difference between the number of points in the majority class and that of the (respective) minority class(es). {width="0.97\linewidth"} -0.1in Convex Generator ---------------- The generator tries to generate points which will be misclassified by the classifier. Hence, if left unchecked, the generator may eventually learn to generate points which do not coincide with the distribution of the intended minority class. This may help improve the performance on the concerned minority class but will lead to high misclassification from the other classes. To prevent this from happening, we generate the new points only as convex combinations of the existing points from the minority class in question. This will restrict the generated distribution within the convex hull of the real samples from the (respective) minority class(es). Since the generator attempts to conjure points that are difficult for the classifier, the points are generated near the peripheries of the minority class(es). Our convex generator $G$ comprises of two modules: a Conditional Transient Mapping Unit ($cTMU$) and a set of class-specific Instance Generation Units ($IGU$), which we propose to limit the model complexity. The $cTMU$ network learns a mapping $t$, conditioned on class $i$, from a $l$-dimensional latent space to an intermediate space. The $IGU_{i}$, on the other hand, learns a mapping $g_{i}$ from the $cTMU$ output space to a set of $n_{i}$ convex weights $g_{i}(t(\mathbf{z}|i)) \geq 0$, s.t. $\sum_{j=1}^{n_{i}}g_{i}(t(\mathbf{z}|i))=1$, using softmax activation. Hence, $G$ can generate a new $D$-dimensional sample for the $i$-th class as a convex combination of the data points in $X_{i}$, -0.12in $$G(\mathbf{z}|i)=\sum_{j=1}^{n_{i}}g_{i}(t(\mathbf{z}|i))\mathbf{x}_{j},$$ -0.04in where $\mathbf{z}$ is a latent variable drawn from a standard normal distribution and $\mathbf{x}_{j} \in X_{i}$. Formally, the adversarial game played by the proposed classifier-convex generator duo poses the following optimization problem, when cross entropy loss is considered: -0.2in $$\begin{aligned} \label{eqn:objective} & \quad \quad \quad \quad \quad \; \min_{G} \max_{M} J(G, M)= \sum_{i \in \mathcal{C}} J_{i}, \\ &\text{where \;} J_{i}=(J_{i1}+J_{i2}+J_{i3}+J_{i4}), \nonumber \\ &J_{i1}=P_{i}\mathbb{E}_{\mathbf{x} \sim p^{d}_{i}} [\log M_{i}(\mathbf{x})], \nonumber \\ &J_{i2}=\sum_{j \in \mathcal{C}\setminus \{i\}}P_{j}\mathbb{E}_{\mathbf{x} \sim p^{d}_{j}} [\log (1-M_{i}(\mathbf{x}))], \nonumber \\ &J_{i3}=(P_{c}-P_{i})\mathbb{E}_{G(\mathbf{z}|i) \sim p^{g}_{i}} [\log M_{i}(G(\mathbf{z}|i))], \text{and}, \nonumber \\ &J_{i4}=\sum_{j \in \mathcal{C}\setminus \{i\}}(P_{c}-P_{j})\mathbb{E}_{G(\mathbf{z}|j) \sim p^{g}_{j}} [\log (1-M_{i}(G(\mathbf{z}|j)))], \nonumber\end{aligned}$$ -0.05in while $p^{d}_{i}$ and $p^{g}_{i}$ respectively denote the real and generated class conditional probability distributions of the $i$-th class. The two-player minimax game formalized in (\[eqn:objective\]) is played between a classifier $M$ and a generator $G$. $M$ attempts to correctly classify all real as well as generated points belonging to all the classes. Whereas, $G$ strives to generate sample(s) which have a high probability of being classified by $M$ into all other classes. To demonstrate how such an adversarial game can aid $M$ to learn a better class boundary, we illustrate its chronological progression in a more explanatory manner in Figure \[fig:illustration\]. In Theorem \[thm:JS\], we show that the optimization problem in (\[eqn:objective\]) is equivalent to minimizing a sum of the Jensen-Shannon divergences. \[thm:JS\] Optimizing the objective function $J$ is equivalent to the problem of minimizing the following summation of Jensen-Shannon divergences: $$\sum_{i=1}^{c}JS\Big( \big( P_{i}p_{i}^{d}+(P_{c}-P_{i})p_{i}^{g} \big) \Big| \Big| \sum_{\substack{j \neq i \\ j=1}}^{c} \big( P_{j}p_{j}^{d}+(P_{c}-P_{j})p_{j}^{g} \big) \Big)$$ See the supplementary document. The behavior of the proposed approach can be understood by interpreting Theorem \[thm:JS\]. The optimization problem aims to bring the generated distribution, for a particular class, closer to the generated as well as real distributions for all other classes. Since the real distributions are static for a fixed dataset, the optimization problem in Theorem \[thm:JS\] essentially attempts to move the generated distributions for each class closer to the real distributions for all other classes. This is likely to result in the generation of ample points near the peripheries, which are critical to combating class imbalance. While doing so, the generated distributions for all classes also strive to come closer to each other. However, the generated distributions for the different classes do not generally collapses upon each other, being constrained to remain within the convex hulls of the respective classes. Additional Discriminator ------------------------ While the generator only generates points within the convex hull of the samples from the minority class(es), the generated points may still be placed at locations within the convex hull which do not correspond to the distribution of the intended class (recall Figure \[fig:exampleGamoNoDis\]). This is likely to happen if the intended minority class(es) are non-convex in shape. Moreover, we know from Theorem \[thm:JS\] that the generated distributions for different minority classes may come close to each other if the respective convex hulls overlap. To solve this problem, we introduce an additional conditional discriminator which ensures that the generated points do not fall outside the actual distribution of the intended minority class(es). Thus, the final adversarial learning system proposed by us consists of three players, viz. a multi-class classifier $M$, a conditional discriminator $D$ which given a class aims to distinguish between real and generated points, and a convex generator $G$ that attempts to generate points which, in addition to being difficult for $M$ to correctly classify, are also mistaken by $D$ to be real points sampled from the given dataset. The resulting three-player minimax game is formally presented in (\[eqn:objDis\]). $X$: training set, $l$: latent dimension, $b$: minibatch size, $u$, $v$: (hyperparameters, set to $\lceil \frac{n}{b} \rceil$ in our implementation).\ [**Output:**]{} A trained classification network $M$.\ [**Note:**]{} For flattened images there is no need to train $F$, i.e., $F(X)$ can be replaced by $X$. Sample $B_{d}=\{\mathbf{x}_{1}, \mathbf{x}_{2}, \cdots \mathbf{x}_{b}\}$ from $X$, with corresponding class labels $Y_{d}$. Update $F$ by gradient descent on $(M(F(B_{d})), Y_{d})$ keeping $M$ fixed. Sample $B_{d}=\{\mathbf{x}_{1}, \mathbf{x}_{2}, \cdots \mathbf{x}_{b}\}$ from $X$, with corresponding class labels $Y_{d}$. Sample $B_{n}=\{\mathbf{z}_{1}, \mathbf{z}_{2}, \cdots \mathbf{z}_{b}\}$ from $l$ dimensional standard normal distribution. Update $M$ and $D$ by respective gradient descent on $(M(F(B_{d})), Y_{d})$ and $(D(F(B_{d})|Y_{d}), \mathbf{1})$, keeping $F$ fixed. Generate labels $Y_{n}$ by assigning each $\mathbf{z}_{j} \in B_{n}$ to one of the $c-1$ minority classes, with probability $\propto$ $(P_{c}-P_{i})$; $\forall i \in \mathcal{C}\setminus\{c\}$. Update $M$ and $D$ by respective gradient descent on $(M(G(B_{n}|Y_{n})), Y_{n})$ and $(D(G(B_{n}|Y_{n})|Y_{n}), \mathbf{0})$, keeping $G$ fixed. Sample $B_{g}=\{\mathbf{z}_{1}, \mathbf{z}_{2}, \cdots \mathbf{z}_{b}\}$ from $l$ dimensional standard normal distribution. Generate labels $Y_{g}$ by assigning each $\mathbf{z}_{j} \in B_{g}$ to any of the $c-1$ minority classes with equal probability. Take ones’ complement of $Y_{g}$ as $\overline{Y_{g}}$. Update $G$ by gradient descent on $(M(G(B_{g}|Y_{g})), \overline{Y_{g}})$ keeping $M$ fixed. Update $G$ by gradient descent on $(D(G(B_{g}|Y_{g})|Y_{g}), \mathbf{1})$ keeping $D$ fixed. $$\begin{aligned} \label{eqn:objDis} & \min_{G} \max_{M} \max_{D} Q(G, M, D)= \sum_{i \in \mathcal{C}} Q_{i}, \\ & \text{where, } Q_{i}=(J_{i1}+J_{i2}+J_{i3}+J_{i4}+Q_{i1}+Q_{i2}), \nonumber \\ & Q_{i1}=P_{i}\mathbb{E}_{\mathbf{x} \sim p^{d}_{i}} [\log D(\mathbf{x}|i)], \text{and}, \nonumber \\ & Q_{i2}=(P_{c}-P_{i})\mathbb{E}_{G(\mathbf{z}|i) \sim p^{g}_{i}} [\log (1-D(G(\mathbf{z}|i)|i))]. \nonumber\end{aligned}$$ Least-Square Formulation ------------------------ Mao et al. [@mao2017least] showed that replacing the popular cross entropy loss in GAN with least square loss can not only produce better quality images but also can prevent the vanishing gradient problem to a greater extent. Therefore, we also propose a variant of GAMO using the least square loss, which poses the following optimization problem: $$\begin{aligned} & \min_{M} L_{M}=\sum_{i \in \mathcal{C}} (L_{i1}+L_{i2}+L_{i3}+L_{i4}), \label{eqn:lsObjM} \\ & \min_{D} L_{D}=\sum_{i \in \mathcal{C}} (L_{i5}+L_{i6}) \label{eqn:lsObjD}, \\ & \min_{G} L_{G}=\sum_{i \in \mathcal{C}\setminus \{c\}} (L_{i7}+L_{i8}+L_{i9}) \label{eqn:lsObjG}, \\ & \text{where, } L_{i1}=P_{i}\mathbb{E}_{\mathbf{x} \sim p^{d}_{i}} [(1-M_{i}(\mathbf{x}))^{2}], \nonumber \\ & L_{i2}=\sum_{j \in \mathcal{C}\setminus \{i\}}P_{j}\mathbb{E}_{\mathbf{x} \sim p^{d}_{j}} [(M_{i}(\mathbf{x}))^{2}], \nonumber \\ & L_{i3}=(P_{c}-P_{i})\mathbb{E}_{G(\mathbf{z}|i) \sim p^{g}_{i}} [(1-M_{i}(G(\mathbf{z}|i)))^{2}], \nonumber \\ & L_{i4}=\sum_{j \in \mathcal{C}\setminus \{i\}}(P_{c}-P_{j})\mathbb{E}_{G(\mathbf{z}|j) \sim p^{g}_{j}} [(M_{i}(G(\mathbf{z}|j)))^{2}], \nonumber \\ & L_{i5}=P_{i}\mathbb{E}_{\mathbf{x} \sim p^{d}_{i}} [(1-D(\mathbf{x}|i))^{2}], \nonumber \\ & L_{i6}=(P_{c}-P_{i})\mathbb{E}_{G(\mathbf{z}|i) \sim p^{g}_{i}} [(D(G(\mathbf{z}|i)|i))^{2}], \nonumber \\ & L_{i7}=\mathbb{E}_{G(\mathbf{z}|i) \sim p^{g}_{i}} [(M_{i}(G(\mathbf{z}|i)))^{2}], \nonumber \\ & L_{i8}=\sum_{j \in \mathcal{C}\setminus \{i, c\}}\mathbb{E}_{G(\mathbf{z}|j) \sim p^{g}_{j}} [(1-M_{i}(G(\mathbf{z}|j)))^{2}], \text{and}, \nonumber \\ & L_{i9}=\mathbb{E}_{G(\mathbf{z}|i) \sim p^{g}_{i}} [(1-D(G(\mathbf{z}|i)|i))^{2}]. \nonumber \end{aligned}$$ Putting it all together ----------------------- The model for the GAMO framework is detailed in Figure \[fig:gamoModel\], while the complete algorithm is described in Algorithm \[alg:gamo\]. To ensure an unbiased training for $M$ and $D$ we generate artificial points for the $i$-th class with probability $(P_{c} - P_{i})$ to compensate for the effect of imbalance. On the other hand, to also ensure unbiased training for $G$ we use samples from all classes with equal probability. Experiments {#sec:exp} =========== We evaluate the performance of a classifier in terms of two indices which are not biased toward any particular class [@Sokolova2009], namely Average Class Specific Accuracy (ACSA) [@huang2016learning; @wang2017meta] and Geometric Mean (GM) [@kubat1997; @branco2016]. All our experiments have been repeated 10 times to mitigate any bias generated due to randomization and the means and standard deviations of the index values are reported. Codes for the proposed methods are available at <https://github.com/SankhaSubhra/GAMO>. 0.15in ------------------ --------------- --------------- ------------------- ------------------- --------------- --------------- ------------------- ------------------- Algorithm ACSA GM ACSA GM ACSA GM ACSA GM Baseline CN 0.88$\pm$0.01 0.87$\pm$0.02 0.88$\pm$0.01 0.86$\pm$0.01 0.82$\pm$0.01 0.80$\pm$0.01 0.81$\pm$0.01 0.79$\pm$0.01 SMOTE+CN 0.88$\pm$0.02 0.87$\pm$0.03 0.89$\pm$0.01 0.89$\pm$0.01 - - - - Augment+CN - - - - 0.82$\pm$0.01 0.78$\pm$0.01 0.82$\pm$0.01 0.78$\pm$0.01 DOS - - - - 0.82$\pm$0.01 0.79$\pm$0.01 0.81$\pm$0.01 0.79$\pm$0.02 (cGAN/cDCGAN)+CN 0.88$\pm$0.01 0.87$\pm$0.01 0.89$\pm$0.01 0.88$\pm$0.01 0.81$\pm$0.02 0.78$\pm$0.01 0.82$\pm$0.01 0.80$\pm$0.01 cG+CN 0.86$\pm$0.03 0.85$\pm$0.02 0.86$\pm$0.03 0.85$\pm$0.03 0.79$\pm$0.02 0.77$\pm$0.02 0.80$\pm$0.01 0.77$\pm$0.02 cG+D+CN 0.85$\pm$0.02 0.83$\pm$0.01 0.85$\pm$0.02 0.82$\pm$0.02 0.79$\pm$0.02 0.78$\pm$0.01 0.79$\pm$0.01 0.78$\pm$0.02 GAMO\\D (Ours) 0.87$\pm$0.01 0.86$\pm$0.01 0.88$\pm$0.01 0.87$\pm$0.01 0.81$\pm$0.01 0.80$\pm$0.01 0.82$\pm$0.01 0.80$\pm$0.01 GAMO (Ours) 0.89$\pm$0.01 0.88$\pm$0.01 **0.91$\pm$0.01** **0.90$\pm$0.01** 0.82$\pm$0.01 0.80$\pm$0.01 **0.83$\pm$0.01** **0.81$\pm$0.01** ------------------ --------------- --------------- ------------------- ------------------- --------------- --------------- ------------------- ------------------- We have used a collection of 7 image datasets for our experiments, namely MNIST [@lecun1998mnist], Fashion-MNIST [@xiao2017fashion], CIFAR10 [@krizhevsky2009cifar], SVHN [@netzer2011svhn], LSUN [@yu2015lsun] and SUN397 [@xiao2010sun]. All the chosen datasets except SUN397 are not significantly imbalanced in nature, therefore we have created their imbalanced variants by randomly selecting a disparate number of samples from the different classes. Further, for all the datasets except SUN397, 100 points are selected from each class to form the test set. In the case of SUN397 (50 classes of which are used for our experiments) 20 points from each class are kept aside for testing. We refrain from using pre-trained networks for our experiments as the pre-learned weights may not reflect the imbalance between the classes. We, instead, train the models from scratch to emulate real-world situations where the data is imbalanced and there is no pre-trained network available that can be used as an appropriate starting point. We have obtained the optimal architectures and hyperparameters for each contending method in Section \[sec:exp\]-\[sec:imgGen\] using a grid search (see supplementary document). ![Ablation study on the MNIST dataset: SMOTE generates artificial samples from the minority class(es) as convex combinations of pairs of neighbors from the respective class(es). The oversampled dataset is then classified using a classifier network CN. SMOTE sometimes generates unrealistic “out-of-distribution" samples which are combinations of visually disparate images that happen to be Euclidean neighbors in the flattened image space. Using cGAN for generating new samples results in realistic images only from the more abundant minority classes. Training only a conditional Generator cG adversarially against CN, to generate images which will be misclassified by CN, results in new samples which all resemble the majority class ‘0’. Introducing a discriminator D (to ensure that cG adheres to class distributions) into the mix results in new samples which are somewhat in keeping with the class identities, but still unrealistic in appearance. Employing our proposed convex generator G to generate new samples by training it adversarially with CN (the GAMO\\D formulation) results in samples which are in keeping with the class identities, but often “out-of-distribution" as the classes are non-convex. Finally, introducing D into this framework results in the complete GAMO model which can generate realistic samples which are also in keeping with the class identities.[]{data-label="fig:mnist_resmod"}](mnist-results_model.pdf){width="0.98\linewidth"} -0.35in 0.15in Dataset Algorithm ACSA GM --------- ---------------- ------------------- ------------------- Baseline CN 0.45$\pm$0.01 0.37$\pm$0.01 Augment+CN 0.47$\pm$0.01 0.39$\pm$0.02 cDCGAN+CN 0.42$\pm$0.02 0.32$\pm$0.03 DOS 0.46$\pm$0.02 0.37$\pm$0.01 GAMO\\D (Ours) 0.47$\pm$0.01 0.40$\pm$0.01 GAMO (Ours) **0.49$\pm$0.01** **0.43$\pm$0.02** Baseline CN 0.74$\pm$0.01 0.73$\pm$0.01 Augment+CN 0.69$\pm$0.01 0.63$\pm$0.01 cDCGAN+CN 0.69$\pm$0.01 0.66$\pm$0.02 DOS 0.71$\pm$0.02 0.68$\pm$0.01 GAMO\\D (Ours) 0.75$\pm$0.01 0.75$\pm$0.02 GAMO (Ours) **0.76$\pm$0.01** **0.75$\pm$0.02** : Comparison of classification performance on CIFAR10 and SVHN datasets.[]{data-label="tab:cifarSvhnRes"} -0.2in 0.15in ------------- --------------- --------------- ------------------- ------------------- --------------- --------------- ------------------- ------------------- Algorithm ACSA GM ACSA GM ACSA GM ACSA GM Baseline CN 0.91$\pm$0.01 0.91$\pm$0.01 0.59$\pm$0.01 0.45$\pm$0.04 0.93$\pm$0.01 0.92$\pm$0.01 0.71$\pm$0.01 0.60$\pm$0.03 Augment+CN 0.74$\pm$0.06 0.70$\pm$0.09 0.62$\pm$0.05 0.47$\pm$0.08 0.82$\pm$0.01 0.79$\pm$0.01 0.72$\pm$0.01 0.66$\pm$0.02 cDCGAN+CN 0.86$\pm$0.01 0.84$\pm$0.01 0.59$\pm$0.01 0.36$\pm$0.02 0.87$\pm$0.01 0.86$\pm$0.01 0.67$\pm$0.01 0.58$\pm$0.02 DOS 0.82$\pm$0.03 0.80$\pm$0.02 0.61$\pm$0.01 0.48$\pm$0.02 0.84$\pm$0.01 0.83$\pm$0.02 0.72$\pm$0.01 0.64$\pm$0.02 GAMO (Ours) 0.92$\pm$0.01 0.91$\pm$0.01 **0.66$\pm$0.01** **0.54$\pm$0.02** 0.91$\pm$0.01 0.91$\pm$0.01 **0.75$\pm$0.01** **0.70$\pm$0.02** ACSA GM ACSA GM ACSA GM ACSA GM Baseline CN 0.90$\pm$0.01 0.89$\pm$0.01 0.50$\pm$0.01 0.28$\pm$0.05 0.87$\pm$0.01 0.87$\pm$0.01 0.61$\pm$0.02 0.54$\pm$0.03 Augment+CN 0.67$\pm$0.06 0.64$\pm$0.09 0.54$\pm$0.03 0.45$\pm$0.07 0.70$\pm$0.03 0.65$\pm$0.03 0.64$\pm$0.02 0.58$\pm$0.03 cDCGAN+CN 0.80$\pm$0.02 0.79$\pm$0.02 0.53$\pm$0.02 0.43$\pm$0.03 0.81$\pm$0.02 0.80$\pm$0.02 0.60$\pm$0.02 0.53$\pm$0.03 DOS 0.78$\pm$0.03 0.76$\pm$0.02 0.54$\pm$0.02 0.44$\pm$0.02 0.79$\pm$0.02 0.77$\pm$0.02 0.63$\pm$0.02 0.61$\pm$0.03 GAMO (Ours) 0.93$\pm$0.01 0.93$\pm$0.01 **0.57$\pm$0.01** **0.50$\pm$0.02** 0.80$\pm$0.01 0.80$\pm$0.01 **0.70$\pm$0.02** **0.68$\pm$0.03** ------------- --------------- --------------- ------------------- ------------------- --------------- --------------- ------------------- ------------------- -0.1in MNIST and Fashion-MNIST ----------------------- The experiments in this section are conducted using imbalanced subsets of the MNIST and Fashion-MNIST datasets. In case of both the datasets, we have sampled $\{4000, 2000, 1000, 750, 500, 350, 200, 100, 60, 40\}$ points from classes in order of their index. Thus, the datasets have an Imbalance Ratio (IR: ratio of the number of representatives from the largest class to that of the smallest class) of 100. We begin by establishing the effectiveness of our proposed framework. We also compare between the two variants of GAMO which use Cross Entropy (CE) and Least Square (LS) losses, respectively. We undertake an ablation study on MNIST using flattened images to facilitate straightforward visualization of the oversampled instances. Convolutional features are used for Fashion-MNIST. For MNIST, we have compared GAMO, against baseline classifier network (CN), SMOTE+CN (training set is oversampled by SMOTE), cGAN+CN (training set oversampled using cGAN, which is then used to train CN), and also traced the evolution of the philosophy behind GAMO, through cG+CN (conditional generator cG adversarially trained against CN, in contrast to cGAN+CN where CN does not play any part in training cGAN), cG+D+CN (cG+CN network coupled with a discriminator D), and GAMO\\D (GAMO without a discriminator) on the MNIST dataset. SMOTE+CN and cGAN+CN are respectively replaced by Augment+CN (data augmentation is used to create new images for balancing the training sets), and cDCGAN+CN (oversampled using conditional deep convolutional GAN) for Fashion-MNIST. GAMO is also compared with DOS on Fashion-MNIST. The ablation study is shown visually in Figure \[fig:mnist\_resmod\] and the results for both datasets are tabulated in Table \[tab:mnistFmnistRes\]. Overall, GAMO is observed to perform better than all other methods on both datasets. Interestingly, GAMO\\D performs much worse than GAMO on MNIST but improves significantly on Fashion-MNIST. This may be due to the fact that the convolutional feature extraction for Fashion-MNIST results in distributed representations where the classes are almost convex with little overlap between classes, enabling the convex generator to always generate data points which reside inside the class distributions. Since we observe from Table \[tab:mnistFmnistRes\] that the LS variants of the classifiers mostly perform better than their CE based counterparts (which according to [@mao2017least] is contributed by the more stable and better decision boundary learned in LS), all the experiments in the subsequent sections are reported using the LS formulation for all the contending algorithms. CIFAR10 and SVHN ---------------- In case of CIFAR10 and SVHN the classes are subsampled (4500, 2000, 1000, 800, 600, 500, 400, 250, 150, and 80 points are selected in order of the class labels) to achieve an IR of 56.25. From Table \[tab:cifarSvhnRes\] we can see that GAMO performs better than others on both of these datasets, closely followed by GAMO\\D, further confirming the additional advantage of convolutional feature extraction in the GAMO framework. Interestingly, Augment+CN performs much worse than the other methods on the SVHN dataset. This may be due to the nature of the images in the SVHN dataset, which may contain multiple digits. In such cases, attempting to augment the images may result in a shift of focus from one digit to its adjacent digit, giving rise to a discrepancy with the class labels. 0.15in Algorithm ACSA GM ------------- ------------------- ------------------- Baseline CN 0.26$\pm$0.04 0.19$\pm$0.05 Augment+CN 0.30$\pm$0.04 0.21$\pm$0.04 cDCGAN+CN 0.20$\pm$0.05 0.00$\pm$0.00 DOS 0.28$\pm$0.04 0.20$\pm$0.05 GAMO (Ours) **0.32$\pm$0.04** **0.24$\pm$0.03** : Comparison of classification performance on SUN397.[]{data-label="tab:sun50Res"} -0.2in CelebA and LSUN --------------- The experiment on CelebA and LSUN are undertaken to evaluate the performance of GAMO on images of higher resolution, as well as to assess the effects of an increase in the number of instances from the different classes. In case of CelebA the images are scaled to $64 \times 64$ size, while for LSUN the same is done on a central patch of resolution $224 \times 224$ extracted from each image. In the case of CelebA we have created two 5 class datasets by selecting samples from non-overlapping classes of hair colors, namely *blonde, black, bald, brown*, and *gray*. The first dataset is the smaller one (having 15000, 1500, 750, 300, and 150 points in the respective classes) with an IR of 100, while the second one is larger (having 28000, 4000, 3000, 1500, and 750 points in the respective classes) with an IR of 37.33. Similarly, in the case of LSUN we select 5 classes namely *classroom, church outdoor, conference room, dining room*, and *tower*, and two datasets are created. The smaller one (with 15000, 1500, 750, 300, and 150 points from the respective classes) has an IR of 100, while the larger one (with 50000, 5000, 3000, 1500, and 750 points) has an IR of 66.67. In Table \[tab:celebaLsunRes\], we present the ACSA and GM over both the training and test set for the small and large variants of the two datasets. We can observe that all the algorithms manage to close the gap between their respective training and testing performances as the size of the dataset increases. Moreover, while Augment+CN seems to have the lowest tendency to overfit (smallest difference between training and testing performances), GAMO exhibits a greater ability to retain good performance on the test dataset. 0.2in -0.3in SUN397 ------ We have randomly selected 50 classes from SUN397 to construct a dataset containing $64 \times 64$ sized images (depending on the image size either a $512 \times 512$ or a $224 \times 224$ center patch is extracted, which is then scaled to $64 \times 64$) with an IR of 14.21. The experiment on SUN397 is performed to evaluate the performance of GAMO over a large number of classes. A scrutiny of the result tabulated in Table \[tab:sun50Res\] reveals that despite all four contending techniques being severely affected by the complexity of the classes and the scarcity of data samples from many of the classes, GAMO is able to retain overall better performance than its competitors. GAMO2pix {#sec:imgGen} ======== GAMO results ultimately in a classifier trained to properly classify samples from all the classes. However, some application may require that actual samples be generated by oversampling to form an artificially balanced dataset. While GAMO directly generates images if flattened images are used, it only generates vectors in the distributed representation space (mapped by the convolutional layers) for the convolutional variant. Therefore, we also propose the GAMO2pix mechanism to obtain images from the GAMO-generated vectors in the distributed representation space. 0.05in Dataset GAMO2pix (Ours) cDCGAN --------------- -------------------- ---------------- Fashion-MNIST **0.75$\pm$0.03** 5.57$\pm$0.03 SVHN **0.17$\pm$0.02** 0.59$\pm$0.04 CIFAR10 **1.59$\pm$0.03** 2.96$\pm$0.03 CelebA-Small **11.13$\pm$0.04** 15.12$\pm$0.05 : Comparison of FID of cDCGAN and GAMO2pix.[]{data-label="tab:fidScores"} Our network for generating images (as illustrated in Figure \[fig:gamo2pix\]) from the GAMO-generated vectors is inspired by the Variational Autoencoder (VAE) [@kingma2013auto; @rezende2014auto]. VAE, unlike regular autoencoders, is a generative model which attempts to map the encoder output to a standard normal distribution in the latent space, while the decoder is trained to map samples from the latent normal distribution to images. We follow the design of a standard VAE in GAMO2pix, only replacing the encoder part with the convolutional feature extractor $F$ trained by GAMO. The GAMO2pix network is trained separately for each class while keeping the encoder part fixed. Such a setting should learn the inverse map from the $D$-dimensional feature space induced by $F$ to the original image space and consequently be able to generate realistic images of the concerned class given GAMO-generated vectors for that class as input. We present the images respectively generated by cDCGAN and GAMO2pix on CIFAR10, Fashion-MNIST, SVHN and CelebA-Small in Figures \[fig:cdcganim\]-\[fig:gamoim\]. We can see that GAMO2pix can indeed generate more realistic and diverse images, compared to cDCGAN which also suffers from mode collapse for minority classes. This is further confirmed by the lower Fr[é]{}chet Inception Distance (FID) [@martin2017fid] (calculated between real and artificial images from each class and averaged over classes) achieved by GAMO2pix, as shown in Table \[tab:fidScores\]. Conclusions and Future Work {#sec:concl} =========================== The proposed GAMO is an effective end-to-end oversampling technique for handling class imbalance in deep learning frameworks. Moreover, it is also an important step towards training robust discriminative models using adversarial learning. We have observed from our experiments that the convolutional variant of GAMO is more effective due to the distributed representations learned by the convolutional layers. We also found that the LS loss variant of GAMO generally performs better than the CE loss variant. An interesting area of future investigation is to improve the quality of the images generated by GAMO2pix by employing a different architecture such as BEGAN [@berthelot2017began]. To reduce the tendency of GAMO to overfit as well as to potentially improve its performance, one may consider hybridization with improved GAN variants [@gurumurthy2017deliganFuture] which can achieve good performance even with less number of training samples. Further, one may explore the efficacy of GAMO to learn new classes by taking inspiration from Memory Replay GAN [@wu2018memoryFuture], or study the usefulness of the proposed convex generator for handling boundary distortion in GANs.

ArXiv

--- abstract: | Although fundamental for astrophysics, the processes that produce massive stars are not well understood. Large distances, high extinction, and short timescales of critical evolutionary phases make observations of these processes challenging. Lacking good observational guidance, theoretical models have remained controversial. This review offers a basic description of the collapse of a massive molecular core and a critical discussion of the three competing concepts of massive star formation: - monolithic collapse in isolated cores - competitive accretion in a protocluster environment - stellar collisions and mergers in very dense systems We also review the observed outflows, multiplicity, and clustering properties of massive stars, the upper initial mass function and the upper mass limit. We conclude that high-mass star formation is not merely a scaled-up version of low-mass star formation with higher accretion rates, but partly a mechanism of its own, primarily owing to the role of stellar mass and radiation pressure in controlling the dynamics. --- to appear in Annual Reviews of Astronomy and Astrophysics, Vol. 45, 2007 [**Toward Understanding\ Massive Star Formation**]{} [Hans Zinnecker[^1] Harold W. Yorke[^2]]{} [**Key Words**]{} accretion, circumstellar disks, HII regions, massive stars, protostars, star formation INTRODUCTION ============ Basic Issues ------------ Massive stars play a key role in the evolution of the Universe. They are the principal source of heavy elements and UV radiation. Through a combination of winds, massive outflows, expanding HII regions, and supernova explosions they provide an important source of mixing and turbulence in the interstellar medium (ISM) of galaxies. Turbulence in combination with differential rotation drives galactic dynamos. Galactic magnetic fields are generated, interacting with supernova shock fronts that accelerate cosmic rays. Cosmic rays, UV radiation, and dissipation of turbulence are the principal sources of heating in the ISM, whereas heavy elements found in dust, molecules, and in atomic/ionic form ultimately are responsible for its cooling. Massive stars thus profoundly affect the star- and planet-formation process (Bally, Moeckel & Throop 2005) as well as the physical, chemical, and morphological structure of galaxies (e.g., Kennicutt 1998, 2005). In spite of the dominant role that massive stars play in shaping galactic structure and evolution, our understanding of their formation and early evolution is still sketchy. There are many reasons. High dust extinction makes it difficult to observe high-mass stars during critical early formation phases. They are rare. They evolve quickly and important evolutionary phases are short-lived. The theoretical problem is extremely complex. Finally, massive stars are seldomly (if at all) formed in isolation; the proximity of other high-mass stars compounds the complex influence of the forming star on its local environment via gravitational interactions, powerful outflows and winds, ionizing radiation, and supernovae. The low number statistics of young or forming high-mass stars is only partially offset by their higher luminosities, which allow us to study them at greater distances than their low-mass counterparts. However, insufficient spatial resolution is an issue – an entire OB-star cluster is often contained in a single observing pixel (e.g., Henning & Stecklum 2002). Definitions ----------- Star formation typically starts with a collapsing gas condensation (core) inside a larger subunit (clump) of a molecular cloud (cf. Williams, Blitz & McKee 2000). A protostar forms that increases its mass by accretion (accumulation) of neighboring gas, while at the same time some mass loss occurs through a bipolar outflow and/or a collimated jet. Let us define some of the terminology adopted here. One of the most misused terms in papers dealing with star formation is protostar, which is considered the Holy Grail (Wynn-Williams 1982) of IR astronomy. Here, we reserve the term protostar or protostellar object for a gaseous object in hydrostatic equilibrium (gas pressure forces balance the gas’ self-gravity), which has not yet begun hydrogen burning but which will, given time, burn hydrogen. At the point hydrogen burning commences, we shall speak of a zero-age main-sequence (ZAMS) star; as long as hydrogen burning occurs in the center, we shall speak of a main-sequence star. Note that the size scale of a protostar is at most a few tens of solar radii. We use the terms massive star and high-mass star interchangeably to denote an OB star sufficiently massive to produce a type II supernova (${\rm M}_*/{\rm M}_\odot \ga$8 for solar abundances). With these definitions in mind, the term high-mass protostar denotes a $\ga$8M$_\odot$ hydrostatic object that has not yet begun hydrogen burning. As we shall see in the following, such objects exist only briefly during a transitory stage between “accreting intermediate-mass protostar” and “accreting high-mass star.” Because it will be impossible to distinguish observationally when an accreting object begins burning hydrogen, we suggest that the terms massive protostar and high-mass protostar generally be avoided. In [**Table 1**]{}, we give a crude classification of massive stars in terms of logarithmic mass intervals and the corresponding main sequence spectral types. [**Mass**]{} [**Designation**]{} [**Sp. type**]{} ----------------- -- ---------------------------- -- ------------------ -- -- 8–16M$_\odot$ Early B-type massive stars B3V to B0V 16–32M$_\odot$ Late O-type massive stars O9V to O6V 32–64M$_\odot$ Early O-type massive stars 05V to O2V$^a$ 64–128M$_\odot$ O/WR-type massive stars WNL-H$^b$ : Main Sequence massive star definition (logarithmic mass ranges) \[tab:defs\] -- ---------------------------------------------------------------------------- $^a$O2V main sequence stars have been identified by Walborn et al. (2002). $^b$WNL-H: N-rich late-type Wolf-Rayet (WR) stars, still on the Main Sequence (H-burning) – see Crowther (2007). -- ---------------------------------------------------------------------------- : Main Sequence massive star definition (logarithmic mass ranges) We reserve the terms very massive star (VMS) and supermassive star (SMS) for stars in the mass ranges of $100 \la {\rm M}_*/{\rm M}_\odot \la 1000$ and $10^4 \la {\rm M}_*/{\rm M}_\odot \la 10^8$ , respectively, and introduce the term ultramassive star (UMS) for stars in the mass range of $10^3 \la {\rm M}_*/{\rm M}_\odot \la 10^4$. SMSs are equilibrium configurations that are dominated by radiation pressure — baryons and electron-positron pairs provide only a minor contribution to the equation of state. At some point during their evolution SMSs collapse owing to a general relativistic gravitational instability. Whereas in the present epoch VMSs, UMSs, and SMSs are unlikely to be formed except under very special conditions, stars with masses in excess of 100M$_\odot$ are expected during the first epoch of star formation (Bromm & Larson 2004; Abel, Bryan & Norman 2000). VMSs, UMSs, and SMSs are not discussed in this review. A recent discussion of the formation and evolution of VMSs is given in Portegies Zwart et al. (2006) and in Belkus, van Bever & Vanbeveren (2007). What is accretion? The term accretion is used in a variety of senses. Measured accretion rates often refer to the rate of mass inflow toward star-forming sites – not the rate at which a star or protostar gains mass. Originating from a 0.1pc scale, this material cannot possibly fall into a sub-10$^{-6}$pc region without carrying significant angular momentum. Instead, it either forms a disk or hits and is mixed with prior existing disk material (see [**Figure 1**]{}). Thus, we distinguish between the accretion of cloud core material onto a disk (${\rm \dot M_{D-acc}}$) and the accretion onto a (proto-)star (${\rm \dot M_{S-acc}}$). Analogous to accretion, mass loss is used in a variety of senses. Measured mass loss rates from jets and outflows do not necessarily reflect the mass loss from an isolated young star. We thus distinguish between the mass loss from the (proto-)star via a wind (${\rm \dot M_{S-wind}}$), the mass loss launched from the accretion disk (${\rm \dot M_{D-wind}}$), which never reaches the (proto-)star, and the material swept up into the outflow from the surrounding molecular cloud (${\rm \dot M_{load}}$). The measured outflow could have contributions from several stars and several disks. The interrelation between disk accretion and disk winds is a fascinating aspect of massive star formation (see, e.g., the recent magneto-hydrodynamic models of Banerjee & Pudritz 2007) – and is at the focus of the frequently asked question: Is high-mass star formation a scaled-up version of low-mass star formation? The answer to this question will be better defined by the time we reach the end of this review. Recommended Reading ------------------- The study of the origin of massive stars is a relatively new field of astrophysical research. There is no comprehensive monograph on the subject, but there are several conference proceedings over the past few years dedicated to the topic, of which we recommend the following: - [*Massive Stars: Their Lives in the interstellar Medium*]{} (Cassinelli & Churchwell 1993) - [*Hot Star Workshop III: The Earliest Stages of Massive Star Birth*]{} (Crowther 2002) - [*Massive Star Birth: A Crossroads of Astrophysics*]{} (Cesaroni et al. 2005a) The reviews on [*Environment and Formation of Massive Stars*]{} (Garay & Lizano 1999), [*Control of Star Formation by Supersonic Turbulence*]{} (Mac Low & Klessen 2004), [*The Formation of the First Stars in the Universe*]{} (Glover 2005), [*The Birth of Massive Stars and Star Clusters*]{} (Tan 2005), [*High Mass Star Formation by Gravitational Collapse of Massive Cores*]{} (Krumholz 2007), and [*The Critical Role of Disks in the Formation of High-Mass Stars*]{} (Cesaroni et al. 2007) are also recommended. Among the reviews in the Proceedings of Protostars and Planets V, we particularly recommend [*The Formation of Massive Stars*]{} (Beuther et al. 2007). There are also a few related Annual Reviews articles: - [*Compact HII Regions and OB Star Formation*]{} (Habing & Israel 1979) - [*The Search for Infrared Protostars*]{} (Wynn-Williams 1982) - [*The Dynamic Evolution of HII Regions – Recent Theoretical Developments*]{} (Yorke 1986) - [*The Orion Molecular Cloud and Star-Forming Region*]{} (Genzel & Stutzki 1989) - [*Physical Conditions in Regions of Star Formation*]{} (Evans 1999) - [*Ultra-Compact HII Regions and Massive Star Formation*]{} (Churchwell 2002) - [*Massive Stars in the Local Group: Implications for Stellar Evolution and Star Formation*]{} (Massey 2003) - [*The First Stars*]{} (Bromm & Larson 2004) Except for the last, none of these earlier reviews had its focus on the formation aspect of massive stars but rather provided a descriptive observational summary of the properties of young OB stars and their HII regions. The present Annual Reviews article will be accompanied in the same volume by an article on ’Physical properties of Wolf-Rayet stars’ (Crowther 2007) and an article on ’Theory of Star Formation’ (McKee & Ostriker 2007) which mostly addresses low-mass star formation, but also includes an important section on high-mass star formation. The Focus of This Review ------------------------ The major questions we wish to address in this review are: 1. What is the sequence of observable states leading from molecular clouds to young high-mass stars? 2. What are the initial conditions of massive star formation (gas densities, temperatures, clump masses, etc.) and how do they come about? 3. Do massive stars always form in dense stellar clusters or can they form in isolation? What special conditions are necessary to allow coalescence, i.e., mergers of stars? 4. Which clues to the origin can be gleaned from multiplicity observations? How do we explain the very tight massive spectroscopic binaries and OB runaway stars? 5. How does the forming massive star influence its immediate surroundings, possibly limiting its final mass and/or the final mass of its neighbors? 6. How do young massive stars influence their global environment, either by inhibiting or by triggering further star formation? How do we get a starburst? To tackle these questions, we first discuss some key observations related to massive star formation. MASSIVE STAR FORMATION:\ KEY OBSERVATIONS ======================== Observable Stages ----------------- The optically visible main-sequence life of OB-type stars is preceded by an embedded phase that lasts about 15% of their lifetime (Churchwell 2002). As summarized by Menten, Pillai & Wyrowski (2005) and van der Tak & Menten (2005), observations at mid-IR through radio wavelengths have shown that this embedded phase can be subdivided into several groups of objects: - IR dark clouds (Perault et al. 1996, ISOCAM; Egan et al. 1998, MSX; Benjamin et al. 2003, [*Spitzer*]{}). Their internal density maxima and temperature minima likely represent the initial conditions of high-mass star formation; a compilation of several dozens of such high-mass starless cores has been given by Sridharan, Williams & Fuller (2005). Some of these cores probably contain low-mass and intermediate-mass accreting protostars, which are faint and hard to detect. Protostellar outflow activity has been detected in one of them (Beuther, Sridharan & Saito 2005). - Hot molecular cores (Kurtz et al. 2000, Cesaroni 2005). These have large masses of warm and dense gas, and large abundances of complex organic molecules evaporated off dust grains; they are signposted by methanol maser emission (Menten 1991, Walsh et al. 1998, Hill et al. 2005); ground-based detectability on the Wien part of the spectral energy distribution with sufficient spatial resolution is difficult (Stecklum et al. 2002), but comes into reach with dedicated 8-m class telescope observations (De Buizer & Minier 2005, Linz et al. 2005). - Hypercompact and ultracompact HII regions (Kurtz 2005, Hoare et al. 2007). In these regions, small but growing pockets of ionized gas have developed that stay confined to the stellar vicinity. Whereas hypercompact HII regions probably represent individual photoevaporating disks (Keto 2007; see also the example in Nielbock et al. 2007), ultracompact HII regions probably represent disk-less stars photoionizing their own cocoons and massive envelopes. - Compact and classical HII regions (Mezger et al. 1967, Yorke 1986). Their gas is ionized globally, often by several ionizing sources. It expands hydrodynamically as a whole and disrupts the parent molecular cloud, revealing both the embedded high-mass and lower mass stellar population for optical and near-IR observations (Carpenter et al. 1993; Zinnecker, McCaughrean & Wilking 1993). Initial Conditions ------------------ Massive star formation occurs inside dense, compact clumps in giant molecular clouds (H$_2$ column densities are 10$^{23}$–10$^{24}$cm$^{-2}$). Smaller mass clumps with lower peak H$_2$ column densities do not form massive stars. Several types of molecular cloud surveys, predominantly near HII regions, have been carried out: CS-molecule surveys for dense molecular gas, 1.2-mm dust continuum surveys for massive cold dust (and hence gas) condensations, as well as OH, H$_2$O, and methanol maser emission surveys for shock-excited compact regions as signposts for massive star formation. (Note that methanol maser and OH maser emission is exclusively associated with high-mass star formation, whereas H$_2$O masers may also be found in low-mass star-forming regions. This is because methanol and OH masers are radiatively pumped and need an intense far-IR source in their vicinity; H$_2$O masers, in contrast, are collisionally pumped in gas shocked by outflows.) These gas and dust surveys have revealed dense cold clumps (molecular hydrogen density n$_{\rm H_2}$=10$^5$cm$^{-3}$, gas temperature T=10–20K, diameter $\sim$0.5pc) with gas masses ranging from a few hundred to a few thousand solar masses (Plume et al. 1997, Shirley et al. 2003, Garay et al. 2004, Motte et al. 2005, Evans 2005). The methanol maser surveys at 6.7GHz point to hot molecular cores with internal heat sources and outflows, as well as protoclusters (Burton et al. 2005, De Buizer 2003, Minier et al. 2005). As mentioned before, large-scale observations with the ISO, MSX, and [*Spitzer*]{} satellites have revealed a new class of clouds, the so-called IR dark clouds or IRDCs, and Simon et al. (2006) have identified more than 10,000 such IRDCs from the MSX data base. Many of these appear to be located in the 4–5kpc Galactic molecular ring (Jackson, private communication; see also Bronfman et al. 2000). The IRDCs are dense clouds seen in absorption against mid-IR background emission. They are mostly filamentary structures that contain condensations of cold massive cores where massive stars or even star clusters seem to form (Rathborne, Jackson & Simon 2006). Recent mid-IR and millimeter-continuum observations show different evolutionary stages of massive star formation in adjacent cores: dense millimeter-continuum sources with and without mid-IR emission (Garay et al. 2004, their figure 4). The origin of these structures appears to derive from supersonic turbulence in giant molecular clouds, that is, shock compression from convergent turbulent gas streams. Depending on the direction of the compression with respect to the direction of the magnetic field lines, the magnetic field will be boosted through flux freezing and hence the resulting clump will be stabilized by magnetic forces against gravitational collapse (subcritical compression). If not, the compressed clump is quickly set up for collapse (supercritical compression), in fact so quickly that the set-up time is shorter than the free-fall time. Magnetically stabilized clumps take much longer for collapse to begin, and their internal turbulent structure may make the clump prone to subfragmentation. If this is true, only supercritical compression leads to massive star formation (cf. Shu, Adams & Lizano 1987). An interesting speculation would be that clouds without sufficiently strong magnetic fields can form lots of massive stars quasi-simultaneously, giving rise to gigantic starbursts. Endproducts ----------- ### OB Clusters The endproducts of massive star formation are either dense gravitationally bound OB star clusters or loose unbound OB associations (Lada & Lada 2003, Briceño et al. 2007). Classical examples of OB star clusters include the Orion Nebula Cluster (ONC), the dense compact cluster associated with the giant galactic HII region NGC 3603, and the R136 cluster in the 30 Dor region in the Large Magellanic Cloud (LMC; see [**Figure 2**]{}). These clusters roughly define a richness sequence in powers of 10: they contain 1, 10 (21), and $\sim$100 massive O-type stars per cluster, with estimated total cluster masses of 10$^3$, 10$^4$, and 10$^5$M$_\odot$, respectively (e.g., ONC: Hillenbrand 1997, Hillenbrand & Hartmann 1998; NGC 3603: Moffat, Drissen & Shara 1994, Drissen et al. 1995; R136: Parker & Garmany 1993, Massey & Hunter 1998). Although R136 can probably be considered a small young globular cluster (M. Andersen et al., submitted), there are more massive young clusters in the nearby universe, such as the very massive embedded super star cluster in the center of the NGC 5253 dwarf galaxy (Turner et al. 2003). This cluster has an ionizing flux equivalent to the presence of 4000–6000 O7V stars that comes from a very compact region about 1pc in size, measured with the VLA. How can the formation of so many massive stars in such a small volume be possible? This question is at the heart of understanding the origin of globular clusters. ### OB Associations The classical examples of OB associations include the nearby Scorpius OB2 and Orion OB1 associations (Blaauw 1964, 1991). Another fine example of what may ultimately become an OB association is the Carina star formation complex at 2.3kpc (Smith & Brooks 2007). One of the best studied extragalactic OB associations is NGC 604 in M33 (Maíz-Apellániz, Pérez & Mas-Hesse 2004). In all these cases, the OB stars are spread over the whole face of the parent giant molecular cloud and are not densely packed at all, with distances between massive stars ranging from 1 to 10pc. This then appears to be a completely different mode of massive star formation, although it must be noted that OB associations often contain dense clusters, too (e.g., the Carina complex harbors the well-known Trumpler 14 and 16 clusters). The question is whether OB associations are superpositions of expanded young clusters (e.g., Kroupa, Aarseth & Hurley 2001, Bastian & Goodwin 2006). ### Field OB Stars Do massive stars always occur in young star clusters or OB associations, or can massive stars also be found outside these regions, i.e., in the field? The answer is they can. It has long been realized that there exists a class of massive stars, the so-called runaway OB stars (Blaauw 1961, Poveda, Ruiz & Allen 1967, Gies & Bolton 1986) that are ejected from their birthplaces – clusters and associations – with velocities in excess of 40kms$^{-1}$. About 10–25% of all O stars and about 2% of all B stars belong to this class of massive field stars. The question whether – these runaway stars aside – other massive stars occur in the field (implying that they would be formed in isolation) has been studied by de Wit et al. (2004, 2005), following Mason et al. (1998). These authors established that 43 among the 227 O stars brighter than eighth V-magnitude are in the field. Of these, about half can be traced back to a cluster or association origin, but about 10–20 stars (i.e., $\sim$5–10%) could not be assigned to any group of origin and therefore might be true field stars, born outside clusters and associations, an issue already raised in a pioneering paper by Roberts (1957). A case in point is HD93521, a high-latitude O9.5V star, more than 1kpc above the Galactic plane, which must have formed locally in the halo (Irvine 1989)! These examples suggest that the question of the birthplaces of massive stars is not yet completely settled. The forthcoming [*Spitzer*]{} 8-micron imaging survey of the Magellanic Clouds (see Meixner et al. 2006) can shed new light upon the question, and in particular can pinpoint isolated massive stars in the LMC/SMC, should these objects indeed exist. Clues from multiplicity ----------------------- The multiplicity of massive stars is believed to be higher than that of young low-mass premain-sequence stars (Preibisch, Weigelt & Zinnecker 2001, Duchêne et al. 2001). This means massive stars have more physical companions than low-mass stars on average. For reference, the multiplicity or, more precisely, the companion star fraction (csf) of a stellar population has been defined by Reipurth & Zinnecker (1993) to be $${\rm csf\,=\,(B\,+\,2T\,+\,3Q\,+\,...)/(S\,+\,B\,+\,T\,+\,Q\,+\,...)},$$ where S is the number of single, B the number of binary, T the number of triple, and Q the number of quadruple systems, etc. (i.e., triple systems contribute two companions, quadruple systems three companions, etc.). For example, the multiplicity of the four OB stars in the Trapezium Cluster in Orion is as follows: the most massive star $\theta^1$ C is double, the next massive star $\theta^1$ A is triple (a hierarchical system with a close spectroscopic binary and a wider companion), $\theta^1$ B is at least quadruple (kind of a Trapezium system within the Trapezium Cluster), and $\theta^1$ D is apparently single. \[Kraus et al. (2007) find some indications that $\theta^1$ D appears extended in their speckle images. There is a fifth star in the Trapezium Cluster, $\theta^1$ E, which has recently been found to be a double-lined intermediate-mass spetroscopic binary (Herbig & Griffin 2006, Costero et al. 2006).\] Putting the number of companions of $\theta^1$ A, B, C, D into the above formula, we get csf=1.5 at face value (probably a lower limit). This should be compared to the multiplicity of the low-mass stellar members in the Orion Nebula cluster, which is csf=0.5 (Padgett, Strom & Ghez 1997; B. Reipurth et al., submitted), i.e., significantly lower. The multiplicity statistics of other OB clusters, including clusters rich in O stars (N$>$5) and poor in O stars (N$<$3) has been studied by Mermilliod & García (2001) and García & Mermilliod (2001), with interesting results: The spectroscopic binary frequency in O-star rich clusters can vary enormously in different clusters, from 15% to 80%, with no apparent correlation. If anything, there is an anticorrelation of the binary frequency and the cluster density, but this needs to be reinvestigated and confirmed. The above statistics often rely on relatively poor data; and the sampling is not complete. Recent work on higher quality data, but on a more limited number of clusters, tends to obtain a lower binary fraction in the range of 20% to 60% (Sana, Rauw & Gosset 2005). The most dramatic example is the IC 1805 cluster where the binary frequency went down from 80% to 20% based on better data (De Becker et al. 2006). In the O-star poor clusters almost all O stars are spectroscopic binaries, often double-lined and even eclipsing. These massive binaries are usually members of hierarchical triple or quadruple systems, or of trapezia, and are often located at the cluster center. The exciting star of the Orion Nebula Cluster, $\theta^1$ Ori C belongs in this category, although $\theta^1$ Ori C is not a massive close spectroscopic binary but a very eccentric visual binary, with masses of 34M$_\odot$ (O5.5V) and 15M$_\odot$ (O9.5V) and an orbital period of about 11yr (Kraus et al. 2007). The orbital periods of the spectroscopic binaries in the O-star rich clusters are concentrated in the range of 4–5 days, whereas in the O-star poor clusters there is a pile-up of orbital periods around 3$\pm$1 days. In NGC 6231, for example, according to Sana (private communication), 10 out of the 16 O stars are double-lined spectroscopic binaries: 6 with periods under 10 days (4 below 5 days); 2 with periods between 3 and 9 months, and 2 with periods of the order of a year or greater (in addition, 1 star is probably a triple-lined spectroscopic binary). The luminosity ratios are all in the range 1–10, otherwise one would not detect them as double-lined spectroscopic binaries. This implies that the secondaries are probably early B stars, and the primary-to-secondary mass ratios must be about 3 (at maximum) or lower. It is also worth mentioning that there is no very short-period highly eccentric O+O binary known at this time (Sana, private communication). These fascinating and surprising facts challenge our views of massive star formation and provide clues to their origin, clues too complex to fully decipher yet but hinting at gravitational dynamics playing a role – beyond mere disk or filament fragmentation (see [**Section 5**]{}, where the formation of binary and multiple systems is extensively discussed; see also Zinnecker 2003). Of course, it is equally important to study the multiplicity of massive stars in OB associations where the stellar density is much lower than in OB clusters, and dynamical interactions between the forming massive stars should be less of an issue. For example, the Orion OB1 association contains $\sim$70 massive stars in its three subgroups 1a,b,c (subgroup 1d is the Orion Nebula Cluster). Of these 70 OB stars, 20% are spectroscopic binaries with periods less than 10 days (Morrell & Levato 1991). The three subgroups show variations in their spectroscopic binary fractions: subgroup 1a is average, subgroup 1b is a factor 1.5 above the average, and subgroup 1c is a factor 1.5 below the average. The situation in the nearby Scorpius-Centaurus OB2 association is as follows: Among the 48 early B stars (B05 to B3V) there are 25 binaries, and 20 of them are spectroscopic binaries with known periods in the range of 0.9–34.2 days, with a median of 5.7 days (Brown 2001). In conclusion, it seems that the spectroscopic binary fraction among massive stars in OB associations is surprisingly similar to that in OB clusters rich in O stars (about 40% on average), and the fraction of very close spectroscopic and eclipsing massive binaries in OB associations with orbital periods below 5–10 days (about 20%, judging from the Orion and Scorpius-Centaurus regions) is a factor of two lower than in OB clusters. Thus OB clusters appear to contain more of a population of very tight (hard) binaries, possibly an effect owing to dynamical encounters after birth; this is an effect that is absent in OB associations. The future of spectroscopic massive binary research lies in the near-IR and in multiepoch radial velocity surveys of embedded massive stars. First results (and successes) have been reported by Apai et al. (2007), indicating that massive close binaries indeed form at a very early stage. Upper Initial Mass Function and Upper Mass Limit ------------------------------------------------ The mass distribution function of massive stars at birth (the so-called Initial Mass Function or IMF for short) is a complicated matter, because ([*a*]{}) massive stars quickly lose some of their initial mass through stellar winds, ([*b*]{}) many of these massive stars are unresolved binaries, and ([*c*]{}) massive stars tend to be born in the centers of OB clusters or tend to sink preferentially toward the cluster center, leaving behind their lower mass siblings that live in the cluster outskirts. This introduces a bias into the mass distribution, flattening a power-law slope. The upper IMF from about 10 to 100M$_\odot$ is usually found to be a universal power law, with logarithmic slope -1.35, first found by Salpeter (1955) for a range of masses below 10M$_\odot$. We refer here to the early observational work of Garmany, Conti & Chiosi (1982) and the summary of Massey (1998). The implication of a Salpeter slope or other similar slopes of the IMF for the number of stars born in different mass intervals (for convenience spaced by a factor of two) can be seen in [**Table 2**]{}, which has been normalized to contain exactly one object in the highest mass interval. This is instructive, because it shows dramatically how rare the O/WR-type massive stars (interval 64–128M$_\odot$) are compared with the early B-type massive stars (8–16M$_\odot$), or with the solar-type low-mass stars (1–2M$_\odot$); see Zinnecker (1996). [**Mass range**]{} [ Logarithmic slope]{} -------------------- -- ----- ------------------------ ------- x=1 x=1.35 x=1.7 0.5–1M$_\odot$ 128 700 3822 1–2M$_\odot$ 64 275 1176 2–4M$_\odot$ 32 108 362 4–8M$_\odot$ 16 42 111 8–16M$_\odot$ 8 16.6 34.3 16–32M$_\odot$ 4 6.5 10.6 32–64M$_\odot$ 2 2.55 3.25 64–128M$_\odot$ 1 1 1 : Initial Mass Function (dN/dlogM$\sim$M$^{\rm -x}$) examples The IMF of massive stars in the aforementioned OB clusters (Orion Nebula Cluster, NGC 3603, and R136 in 30 Dor) is discussed by Pflamm-Altenburg & Kroupa (2006), Stolte et al. (2006), and Massey & Hunter (1998), respectively. \[Other studies of the Orion Nebula Cluster include Zinnecker, McCaughrean & Wilking (1993); Hillenbrand & Hartmann (1998). The stellar content in the biggest galactic HII region NGC 3603 was also investigated by Drissen et al. (1995); Hofmann, Seggewiss & Weigelt (1995); Eisenhauer et al. (1998); Brandl et al. (1999); Moffat et al. (2004); and Sung & Bessell (2004). An important investigation of the IMF of R136 is that of Sirianni et al. (2000).\] Whereas Pflamm-Altenburg & Kroupa (2006) find a deficit of high-mass stars in the Orion cluster, Stolte et al. (2006) derive an excess of massive stars in the NGC 3603 cluster core, reflected by a power-law slope of -0.9 (probably due to mass segregation). It is only in R136 that the power-law slope is almost exactly the same as the Salpeter value. It is worth noting here that the slope of the mass function of high-mass stars in the range of 8–40M$_\odot$ in the wider field in the 30 Dor region is apparently the same as in the R136/NGC 2070 cluster (Selman & Melnick 2005). The latter authors do not find the much steeper slope of the IMF (for the range of 25M$_\odot$ to 120M$_\odot$) derived by Massey (2002) for the global OB field population in the LMC. They suspect that selective incompleteness at V=12 owing to detector saturation and Be star contamination lies at the origin of this discrepancy. The question of the IMF of massive stars in OB associations was discussed long ago by Garmany, Conti & Massey (1980) and later by Massey, Johnson & Degioia-Eastwood (1995). They concluded that the IMF is normal, i.e., consistent with a Salpeter power-law. A fine discussion of upper IMF slopes in various young clusters and associations including all the caveats and selection effects was given by Scalo (1998). He noted that individual realizations of IMF slopes can vary, but the average slope is indeed close to the Salpeter value -1.35. The question of whether there is a physical (rather than statistical) upper mass end to the IMF is of great interest to anyone interested in population synthesis, galactic evolution, and cosmology. For example, if the stellar upper mass limit were 120M$_\odot$, pair instability supernovae requiring stellar masses at the time of explosion between 140 and 260M$_\odot$ (Heger et al. 2003) could not happen. Larson (1982) originally asked the observational question whether there was a correlation between the mass of a molecular cloud and the maximum mass of a star that could form in it. His result was that indeed the maximum stellar mass scaled with the mass of the parent cloud, roughly with its square root. It takes a giant molecular cloud of 10$^5$M$_\odot$ to form a 50M$_\odot$ star; a cloud of 10$^3$M$_\odot$ can only spawn a maximum stellar mass of 8M$_\odot$. The implication is that massive stars form in clouds of mass between 10$^3$ and 10$^5$M$_\odot$ or more, probably because only these have sufficiently massive substructure (clumps). Weidner & Kroupa (2004) and Figer (2005) discussed the upper limit to the masses of stars, based on observations of the R136 cluster in the LMC and the Arches cluster near the Galactic Center, respectively. They pointed out that these clusters are so massive that given a Salpeter IMF one would expect to find stars as massive as 750M$_\odot$ and 500M$_\odot$, respectively, whereas the most massive stars seen do not exceed 140M$_\odot$ and 130M$_\odot$, respectively. This suggests a firm upper mass limit of 150M$_\odot$. Otherwise, a sharp down-turn of the IMF near 150M$_\odot$ would be required (see the extensive discussion in Elmegreen 2000). Or the very massive stars have already exploded/imploded during the dust-obscured, hidden, early evolutionary stages – an unlikely scenario. Oey & Clarke (2005) also gave a statistical confirmation of a stellar upper mass limit around 120–200M$_\odot$, if the IMF is Salpeter-like. Koen (2006) further analyzed the upper IMF in the R136 cluster with two different statistical techniques and suggested an upper mass limit of 140–160M$_\odot$. Thus all four studies agree on the existence of a physical upper limit in the stellar mass distribution. Feedback and Triggering ----------------------- This topic deserves its own review. The question we ask here is the following: What does the energy and momentum input of massive stars in terms of expanding HII regions, stellar winds, or supernova shock waves do to the parent clouds? Is the cloud primarily disrupted or is new star formation triggered? Which of the above agents (HII regions, stellar winds, or supernova shock waves) provides the best trigger for new OB star formation and for new low-mass star formation? We are only beginning to answer these questions. A key observation in this context is the fact that the high-mass and low-mass stellar populations in the subgroups of OB associations appear to be coeval, i.e., the nuclear age of the massive stars is the same as that of the lower mass premain-sequence objects (Preibisch & Zinnecker 1999, 2007; Briceño et al. 2007). This would appear to require a fast, coherent trigger, such as a supernova shock wave. Indeed, in the Scorpius-Centaurus association there is evidence that the shock wave of a supernova in one OB subgroup triggered the formation of another subgroup (de Geus 1992). However, there is also other evidence that radiation from massive stars \[by a process called radiative implosion (e.g., Kessel-Deynet & Burkert 2003)\] can only trigger the formation of low- and intermediate-mass objects (Lee & Chen 2007 and references therein). The classic theory of triggered massive star formation is the one by Elmegreen & Lada (1977). In their theory, an ionization shock front provides the pressure on an adjacent layer of molecular gas to compress it and heat it, thus stimulating gravitational instability of massive gas layers, hence the formation of massive stars – and perhaps only massive stars. Low-mass star formation may not be triggered with this mechanism. If true, this would produce an anomalous IMF with only massive stars (known as bimodal star formation, cf. Güsten & Mezger 1982). It is worth noting that such characteristics – small groups of massive stars only – may be needed for the decay of small N-body systems that give rise to the dynamical ejection of runaway OB stars (Clarke & Pringle 1992). MASSIVE STAR FORMATION:\ BASIC THEORY ======================== Sequence of Events ------------------ Starting from a pre-existing giant molecular cloud, the sequence of events is likely as follows: 1. Formation of cold dense molecular cores or filaments, induced by gravo-turbulent cloud fragmentation (Mac Low & Klessen 2004). This means that supersonic turbulence rapidly produces localized compressed pockets of gas, some of which remain gravitationally bound and provide the initial conditions for collapse (Padoan & Nordlund 2002; Klessen et al. 2005). A characteristic density of about 10$^5$cm$^{-3}$ and temperature of 10–15K results from the equation of state of dusty molecular gas of solar abundance (Jappsen et al. 2005, Larson 2005). 2. Nonhomologous gravitational collapse of portions of the cores into optically thick, pressure-supported protostellar embryos with initial masses of the order of 10$^{-3}$M$_\odot$ (Larson 1969, Bate 2000). The term nonhomologous collapse refers to the fact that the relative distribution of material changes, as opposed to a homologous or self-similar collapse. (This is different from Shu’s (1977) inside-out collapse of a self-similar isothermal sphere). 3. Accretion of material onto protostellar objects as they evolve toward the main sequence. For low-mass objects the accretion stops well before hydrogen burning commences. These premain-sequence objects of fixed mass then slowly and quasi-hydrostatically contract to the main sequence (Palla & Stahler 1993, Baraffe et al. 2002). However, high-mass objects eventually start burning hydrogen and develop radiation-driven winds as they continue to accrete and evolve up the main sequence to hotter and more luminous states (Kudritzki 2002). 4. Disruption of the birth cloud, as the first high-mass stars strongly influence their environment by their winds, outflows, and UV radiation, and eventually become supernovae. The most massive stars go supernova after $\sim$3Myr. When the remnant molecular cloud has been dissipated, the result is mostly a cluster of OB stars or an OB association, with an associated cospatial population of lower mass stars (Zinnecker, McCaughrean & Wilking 1993). Often, several evolutionary stages of star formation can be found side by side; the 30 Dor region in the LMC is a good example (Walborn et al. 1999). We shall denote these four phases as the compression, collapse, accretion, and disruption phases of high-mass star formation, respectively. All phases can occur simultaneously and side-by-side in a molecular cloud. Below, we give a fairly detailed yet still greatly simplified discussion of these phases, including protostellar luminosity and ionization evolution as a function of growing stellar mass. The Compression Phase --------------------- This first step toward high-mass star formation is either a starless core ($\sim$100M$_\odot$) or a starless clump ($\sim$1000$_\odot$) of molecular gas in a giant molecular cloud. McKee & Tan (2003) envisage that these cores are molecular condensations in a turbulence-supported quasi-equilibrium that ultimately form single or gravitationally bound multiple massive protostars. As these authors argue, turbulent and pressurized clouds permit sufficient material to be available in the cores of giant molecular clouds for high-mass star formation. Mechanical energy must be continuously injected into the clump in order to maintain this quasi-equilibrium between turbulence and gravity. The assumption is that this energy is either injected from within the cores from the kinetic energy of outflows and accretion shocks, or it comes from the outside and cascades down to smaller scale sizes. It is furthermore assumed that (small-scale) turbulence acts as an isotropic pressure. In an alternate scenario by Bonnell et al. (1997, 2001a), the compression phase is more of a transient phase due to the random motions in the self-gravitating cloud. Smoothed particle hydrodynamics (SPH, see the sidebar Smoothed Particle Hydrodynamics with Sink Particles) simulations of this phase show sheets, filaments, and cores forming (see [**Figure 3**]{}). Some cores collapse and fragment, marking the beginning of the collapse phase for these objects, while in other parts of the cloud compression is still occurring. The philosophical difference between the approach of McKee & Tan (2003) and that of Bonnell et al. (1997, 2001a) can be summarized as “monolithic collapse” versus “competitive accretion.” For the former, the mass necessary for massive star formation is intimately associated with the final product. If there are bulk motions of the embryo star, its protostellar core participates in those motions. The only competition for the infalling material is between close members of a multiple system. For competitive accretion, the material that makes up a particular star can come from various parts of the parent cloud. Protostars move relative to the molecular gas; the only gas that is intimately associated with any particular protostar is in its circumstellar disk and envelope. Because multiple protostars are often formed together at the same time, each protostar competes for the available molecular material (see [**Section 4.2**]{}). Crutcher’s (1999, 2005) summary of the available Zeeman measurements of magnetic field strengths in molecular clouds suggests that magnetic fields likely play an important role in molecular cloud dynamics. Vázquez-Semadeni et al. (2005) studied the evolution of clumps and cores formed as turbulent density fluctuations in nearly isothermal molecular clouds, considering both the magnetic and nonmagnetic cases with driven turbulence. In the nonmagnetic case the authors find that the cores are unlikely to reach a hydrostatic state – necessary for monolithic collapse – if the molecular clouds have an effective polytropic exponent less than 4/3. In this case, cores are transient, either proceeding directly to collapse or re-expanding on a dynamical timescale. The magnetically subcritical clouds simulated by Vázquez-Semadeni et al. (2005) do not produce magnetostatic clumps, but rather a few marginally bound clumps that are subsequently dispersed. Ambipolar diffusion – had it been included in the simulations – could have increased the clumps’ likelihood to become bound and subsequently to collapse. For clouds with weaker magnetic fields a few cores form and collapse on a timescale slightly larger than the cloud’s free-fall timescale. In their most supercritical simulation, fewer clumps and cores form than in their nonmagnetic counterpart, and these cores reexpand because they are not Jeans unstable. The authors thus conclude that not all cores observed in molecular clouds will necessarily form stars and that magnetic fields may help reduce the star-formation efficiency by reducing core formation rather than by delaying or inhibiting the collapse of individual cores. The assumption of driven as opposed to decaying turbulence is important in these and similar simulations (cf. Heitsch, Mac Low & Klessen 2001, Li & Nakamura 2006). Indeed, Li & Nakamura (2006) using the best 3D MHD simulation technique to date have shown that initial turbulence in a cluster forming region is quickly replaced by motions generated by protostellar outflows. This protostellar outflow-driven turbulence can keep a protocluster clump close to virial equilibrium long after the initial turbulence has decayed away. This may explain the early molecular cloud observations of Bertoldi & McKee (1992) and lend support to the recent equilibrium cluster formation models of Tan, Krumholz & McKee (2006). It could even imply that the stellar IMF is regulated by outflow feedback (Silk 1995, Adams & Fatuzzo 1996). However, the ultimate test of these predictions has not yet been made and requires much more observational work. At the same time, competing, more violent cluster formation models have been proposed in the literature (e.g. Bonnell, Bate & Vine 2003), and the jury is still out. Essentially, the question is whether star formation is slow or fast (Ballesteros-Paredes, Hartmann & Vázquez-Semadeni 1999, Glover & Mac Low 2007). The Collapse Phase ------------------ Gravity plays the dominant role in star formation. To form a star, gravity must overcome pressure, magnetic forces, internal turbulence, and rotation. In the simplest case of gravity versus gas pressure, one defines the Jeans mass, $${\rm M_{Jeans}} \simeq 1.1\;{\rm M}_\odot \left[ { {\rm T}_{\rm gas} \over 10\,{\rm K}} \right]^{3/2} \left[ { \rho \over 10^{-19}\,{\rm g\; cm}^{-3}} \right]^{-1/2}$$ as the smallest mass for which gravity can become dominant. The normalization is consistent with typical initial conditions. Turbulence as a repulsive force will exceed gas pressure if motions are supersonic. Unless continually replenished, however, supersonic turbulence dies out on a dynamical timescale (see, e.g., Stone, Ostriker & Gammie 1998; Clark & Bonnell 2005; Kritsuk, Norman & Padoan 2006; for a summary of earlier work, see Mac Low & Klessen 2004). As Shu, Adams & Lizano (1987) point out, once gravity dominates pressure and magnetic forces in an optically thin gas capable of radiating compressional heat, it remains dominant and the gas collapses on a free-fall timescale $${\rm t_{ff}} \simeq 2.1 \times 10^5\, {\rm yr} \left[ { \rho \over 10^{-19}\,{\rm g\; cm}^{-3}} \right]^{-1/2} \; .$$ That is, the densest parts collapse the fastest, and the Jeans mass decreases during collapse. The gas collapses nonhomologously until the densest parts become optically thick, allowing the gas to heat up adiabatically and to increase the gas pressure dramatically. Rotational (centrifugal) forces increase during gravitational collapse owing to conservation of angular momentum, so flattened structures and accretion disks (e.g., Black & Bodenheimer 1975; Terebey, Shu & Cassen 1984; Yorke & Bodenheimer 1999) are expected phenomena of gravitational collapse. The Accretion Phase ------------------- ### Formation of the Hydrostatic Core The formation of low-mass stars is explained by the nonhomologous collapse of a slowly rotating fragment of molecular material; the collapse is stopped in the central regions when the object becomes optically thick (Larson 1969, Woodward 1978, Winkler & Newman 1980, Masunaga, Miyama & Inutsuka 1998; for a more recent review, see Klein, Fisher & McKee 2004). There is a second, inside-out collapse in the center for ${\rm T_{gas}} \simeq 2000$K when molecular hydrogen dissociates. When the second core is optically thick and thermally ionized, pressure forces are able to balance gravity on a dynamical timescale, and one speaks of an accreting quasi-hydrostatic core as more material rains down. When it forms, the second hydrostatic core contains somewhat more than a Jupiter mass and has a radius on the order of (3–5)R$_\odot$ (Winkler & Newman 1980, Masunaga & Inutsuka 2000). As long as material continues to flow onto the quasi-hydrostatic core, the core grows in mass. Simultaneously, it contracts on a thermal Kelvin-Helmholtz timescale ${\rm \tau_{KH} \sim GM_*^2/R_*L_*}$ toward hydrogen-burning densities and temperatures. [**Figure 4**]{} shows the Kelvin-Helmholtz timescale to the ZAMS as a function of stellar mass. Except for the growth of mass and angular momentum, the interior regions of the hydrostatic core are generally ignorant of and not influenced by the details of the accretion process. This allows us to separate the problem of star formation into several distinct parts: ([*a*]{}) the evolution of the central core, ([*b*]{}) the details of transporting material from the disk onto the core, ([*c*]{}) transporting material inward within the disk, and ([*d*]{}) accretion onto the disk. Fundamental differences between low-mass and high-mass star formation can be attributed to differences in the above processes and, in particular, to significant differences in the timescales involved and in the local radiative environment. ### Evolution of Accreting Cores Accretion onto the (proto)star will affect the core’s outer (atmosphere) regions and the overall spectral appearance. We estimate the thickness ${\rm \Delta R_{dyn}}$ of the outer stellar regions that are dynamically affected by accretion to be: $${\rm \frac{\Delta R_{dyn}}{R_*} \sim \frac{\dot M_{S-acc} t_{S-dyn}}{M_*} = \frac{t_{S-dyn}}{t_{S-acc}}}$$ where ${\rm t_{S-dyn} \simeq R_*/v_{esc}}$, the stellar dynamical timescale, is the time for an acoustic wave to cross through the core or the orbital period of a body just above the core’s surface (generally less than a day). Because ${\rm t_{S-dyn} <<< t_{S-acc} =}$ ${\rm M_*/\dot M_{S-acc}}$ (the accretion timescale), ${\rm \Delta R_{dyn}/R_*}$ is extremely small ($<$10$^{-5}$ for accretion rates ${\rm \dot M_{S-acc} < 10^{-2}}$M$_\odot$yr$^{-1}$). It is not easy to estimate the size of the region thermally affected by the accretion flow onto the (proto)star, because the radiation transfer in an optically thick plasma depends on details of the complex accretion geometry. Exactly how material flows from the disk onto the (proto)star remains an unsolved theoretical problem. One can expect a (proto)star’s location in the Hertzsprung-Russell (HR) diagram to be strongly affected by accretion whenever the Kelvin-Helmholtz timescale ${\rm \tau_{KH}}$ exceeds the accretion timescale ${\rm \tau_{acc}}$. However, even when this criterion is not fulfilled, the Kelvin-Helmholtz timescale for the outer regions of a star can locally exceed the accretion timescale, and accretion can affect the radius, effective temperature, and luminosity of the star (Kippenhahn & Meyer-Hofmeister 1975). Kippenhahn & Meyer-Hofmeister (1975) consider the case of mass transfer within a close binary system. At high accretion rates ${\rm \dot M_{S-acc} \ga 10^{-3}}$M$_\odot$yr$^{-1}$, accreting main-sequence stars of mass M$_* \ga 5$M$_\odot$ bloat up to radii exceeding 10 times their ZAMS values. Although not strictly applicable to the case of accreting premain-sequence stars, one can expect the same qualitative effect of high accretion rates in this mass range. These results demonstrate the importance of accretion for the appearance of the central hydrostatic cores. For low-mass stars, it is generally accepted that material is transported onto the central core through an accretion disk (e.g., Shu, Adams & Lizano 1987). The net gain per unit time of gravitational potential energy of the accreted material ${\rm G M_*\dot M_{S-acc}/R_*}$ is partly converted into rotational energy of the core and disk ($\sim$1/4) and partially converted into heat ($\sim$3/4), which is radiated away (Yorke & Krügel 1977). Part of the gravitational energy is converted into heat in a series of disk accretion shocks (Yorke & Bodenheimer 1999); part of it is converted into heat within the disk by the same viscous processes that transport angular momentum outward and allow the radial flow of material inward; and part of it is converted into heat in the accretion flow and shocks/relaxation zones on the (proto)stellar surface. As mentioned above, the details of how the gas is ultimately transported from the disk onto the core are still unclear. For low-mass stars it has been postulated that magnetically focused flows (Blandford & Payne 1982) and/or accretion columns (Königl 1991, Edwards et al. 1993) and/or X-winds (Shu et al. 1994, 1995) are involved. Because 50% ($\simeq$3/8${\rm G M_* \dot M_{S-acc}/R_*}$) of the total gravitational energy is converted into heat and radiated away within 1R$_*$ of the core, one can expect higher temperatures and luminosities as either M$_*$ or ${\rm \dot M_{S-acc}}$ increases or as the core contracts. The temperature of this hot shocked gas $${\rm T \simeq \frac{ G M_*/R_*}{k/\mu m_H} \simeq 10^6\,K\ \left[ \frac{M_*}{1\,M_\odot} \right] \left[ \frac{R_*}{10\,R_\odot} \right]^{-1}}$$ is sufficiently high to produce X-rays even for low-mass stars. The above argument is only slightly modified when the core and disk produce a wind or outflow. We then speak of a net mass accretion of ${\rm \dot M_{S-acc} - \dot M_{S-wind}}$. A portion of the core’s rotational energy and angular momentum can be converted into kinetic energy and angular momentum of the wind. A necessary condition to accrete sufficient material to produce a massive star is thus, $${\rm M_* = \int_0^t \left[ \dot M_{S-acc}(t') - \dot M_{S-wind}(t') \right]\, dt' \ga 8\,M_\odot} . $$ That is, the accretion rate ${\rm \dot M_{S-acc}}$ onto an embryo object must exceed the outflow rate ${\rm \dot M_{S-wind}}$ during a significant proportion of the formation process. For this to occur, the acceleration owing to gravity must exceed the outward directed radiative acceleration of the accreting core. Whereas gravity ${\rm G M_*/r^2}$ at each radial point in an envelope increases linearly with core mass, the radiative acceleration of dusty material ${\rm \kappa L/4\pi r^2 c}$ is proportional to the core’s luminosity. This increases as a high power of stellar mass. A lower limit to the core’s luminosity is the ZAMS luminosity (see [**Figure 4**]{}). Thus, to allow infall, we require as a necessary condition $${\rm \kappa_{eff} L / 4 \pi r^2 c < G M_* / r^2} $$ with L=L$_*$+L$_{\rm acc}$, which translates into $${\rm \kappa_{eff} < 130\, cm^2 g^{-1} \left[ \frac {M_*}{10\,M_\odot} \right] \left[ \frac {L}{1000\,L_\odot} \right]^{-1}} , \label{eq:critopacity}$$ where the effective opacity for radiative acceleration of accretable material is defined as $${\rm \kappa_{eff} = {\int_0^\infty} \kappa_\nu^{\rm rad} F_\nu\,d\nu / F} , $$ here $\kappa_\nu^{\rm rad}$ is the frequency-dependent gram-opacity (cgs-units: cm$^2$g$^{-1}$) of the material subject to radiative acceleration and ${\rm F = \int_0^\infty F_\nu\,d\nu}$ is the radiative flux. (For a more complete discussion of the definition of $\kappa_\nu^{\rm rad}$, including effects of nonisotropic scattering and photoejection of particles from the dust, see Yorke 1988.) The (proto)star’s luminosity is given by the sum of its intrinsic luminosity L$_*$ and the luminosity ${\rm L_{acc}}$ emitted by the dissipation of kinetic energy of the material being accreted. For dusty gas, $\kappa_\nu^{\rm rad}$ is strongly frequency-dependent. Depending on the hardness of radiation emitted by the central source and the accretion shocks, dusty gas could be repelled from the star. The ISM extinction in the J-band, for instance, corresponds to ${\rm \kappa}$=130cm$^2$g$^{-1}$. In [**Figure 5**]{} we indicate values of $\kappa_{\rm eff}$ for dusty gas at solar abundances ($\kappa_\nu^{\rm rad}$ taken from the Preibisch et al. 1993 dust model with and without ice-coated grains) illuminated by 600K and 6000K black bodies \[i.e., ${\rm F_\nu \propto B_\nu (T_*)}$\], respectively, and for a fully ionized dustless hydrogen plasma. [**Figure 6**]{} displays the mean effective opacity of dusty gas as defined by [**Equation 4**]{}, using the Preibisch et al. (1993) dust model with and without ice-coated grains ([*blue*]{} and [*red*]{} curves, respectively) and assuming blackbody radiation at the temperature T$_{\rm rad}$. Using the mass scale on the right of [**Figure 6**]{} we note that the net force on dusty gas surrounding a deeply embedded 100M$_\odot$ main-sequence star is directed toward the star as long as the star appears to the dust to be a $\la$50-K source. By contrast, for an unobscured 3M$_\odot$ main-sequence star with T$_{\rm eff} \ga 10^4$K the net force on dusty gas would be directed away from the star. Overcoming Radiative Acceleration --------------------------------- Next we discuss how [**Equation 2**]{} can be satisfied, i.e., accretional growth of an already existing stellar embryo can be enabled. At least one of the following conditions must be met: ([*a*]{}) ${\rm \kappa_{eff}}$ must be sufficiently low, i.e., significantly lower than its ISM value for optical/UV radiation; ([*b*]{}) the total luminosity L must be reduced; or ([*c*]{}) gravity (i.e., the stellar mass M$_*$ ) must be increased. Below, we discuss each of these three possibilities. ### Reduce ${\rm \kappa_{eff}}$ Owing to the strong frequency dependence of dust opacity, ${\rm \kappa_{eff}}$ can be significantly lower than its ISM value if the radiation field seen by the accreting material is that of a cold, embedded object. When embedded the protostellar radiation field is shifted from the optical/UV – where the dust absorbs the photons – into the far-IR where the dust reemits the absorbed energy. Alternatively, ${\rm \kappa_{eff}}$ can also be reduced if the average size of dust grains increases (but remains compact rather than becoming fractal) or if most of the dust is destroyed. In their pioneering efforts, Kahn (1974) and Wolfire & Cassinelli (1987) studied the 1D, spherically symmetric accretion problem for massive star formation with an emphasis on the dust opacity. Kahn concluded that a 40-M$_\odot$ star could be formed by spherically symmetric accretion, but his assumed dust destruction temperature T$_{\rm sub}$=3600K is too high by a factor of about two (implying for the outer regions greater protection from radiation pressure). Furthermore, his assumed grain opacities, which can be written in the form ${\rm \kappa \simeq 100 cm^2 g^{-1} (T_{dust}/9000 K)^2}$, were somewhat too low. In a more careful treatment of the dust (but still assuming steady-state spherically symmetric infall), Wolfire & Cassinelli (1987) concluded that very massive stars can form only if the dust has been significantly modified. By contrast, Yorke & Krügel (1977) showed in a hydrodynamical simulation that spherically symmetric accretion must be nonsteady for the high-mass case. They were able to produce stars of masses 17M$_\odot$ and 36M$_\odot$ from clouds of masses 50M$_\odot$ and 150M$_\odot$, respectively, in a highly variable accretion flow. These early attempts to explain high-mass star formation suffer significantly from the fact that spherically symmetric infall was assumed. The effective opacity of the accreting material can also be reduced by density inhomogeneities resulting from the photon bubble instability. The radiation escapes readily through the gaps between the shocks that are driven by disturbances in the radiation flux (Turner, Quataert & Yorke 2007). Another possibile explanation for reducing the effective opacity is the accretion of optically thick blobs or fingers. These can be expected to form via Rayleigh-Taylor instabilities in the radiation-inflated cavities produced by luminous (proto)stars (Krumholz, McKee & Klein 2005a; Yorke 2002). In this case, $${\rm \kappa_{eff} = \pi R^2_{blob} / M_{blob}}.$$ As a particular subset of this family of solutions, Bonnell, Bate & Zinnecker (1998) considered building up massive stars by coalescence of intermediate-mass stars within a very dense protostellar cluster (see also Bonnell & Bate 2002, Zinnecker & Bate 2002). We return to this issue in more detail below. Opacity modifications owing to coagulation of dust and dust destruction processes during the collapse phase were calculated by Suttner & Yorke (2001) for three different detailed dust models: compact spherical particles, fractal BPCA (ballistic particle-cluster agglomeration) grains, and fractal BCCA (ballastic cluster-cluster agglomeration) grains. (BPCA dust consists of grains with widely different grains sizes, whereas BCCA dust is formed by coagulation of similar-sized grains). Assuming axial symmetry, Suttner & Yorke followed the dynamics of gas and 30 individual dust components. They found that even during the early collapse and the first $\sim$10$^4$yr of dynamical disk evolution, the initial dust size distribution is strongly modified. Close to the disk’s midplane, coagulation produces dust particles of several tens of $\mu$m in size (for compact spherical grains) up to several millimeters in size (for fluffy BCCA grains). In contrast, in the vicinity of the accretion shock front located several density-scale heights above the disk, large velocity differences inhibit coagulation. Dust particles larger than about 1$\mu$m segregate from smaller grains behind the accretion shock. Owing to the combined effects of coagulation and grain segregation the IR dust emission is modified. Within the accretion disk, an interstellar medium dust size distribution (Mathis, Rumpl & Nordsieck 1977) provides a poor description of the general dust properties. Nevertheless, the radiative force acting on material infalling from the envelope to the disk is hardly affected by coagulation (see also Weidenschilling & Ruzmaikina 1994). ### Reduce the Effective Luminosity Nakano, Hasegawa & Norman (1995) and Jijina & Adams (1996) pointed out that, because we expect accretion to proceed through an accretion disk, radiation pressure could blow away the tenuous polar regions but not the massive disk. Yorke & Bodenheimer (1999) and Yorke & Sonnhalter (2002) studied this effect quantitatively and substantiated this claim through numerical simulations. They found that, whereas the central object may emit radiation isotropically, the radiation field quickly becomes anisotropic farther from the center. For an outside observer, and in particular, for a dust grain attempting to accrete onto an existing protostellar disk, the radiative flux close to the equatorial plane can be much smaller than the component parallel to the rotation axis. This so-called flashlight effect (beaming of radiation in the polar direction) occurs whenever a circumstellar disk forms. Yorke & Bodenheimer (1999) considered the evolution of a 10-M$_\odot$ molecular fragment that produced a 8.2-M$_\odot$ star (their case F). [**Figure 7**]{} depicts an intermediate stage of the evolution, after 7M$_\odot$ of material have accreted onto the central protostar, while 2.8M$_\odot$ have accumulated in a circumstellar disk and 0.2M$_\odot$ still reside in an larger extended infalling envelope. Inner and outer disk accretion shocks (density and temperature discontinuities) are visible above and below the disk. The dark red regions with an opening angle of about 45$^\circ$ show preferential heating of polar regions by the protostar. The light red regions within 45$^\circ$ of the equator beyond $\sim$1000AU indicate shadowing by the disk. Yorke & Bodenheimer (1999) estimated that the edge-on and pole-on bolometric fluxes can differ by more than a factor of 30 after about one-half of the mass of a 2-M$_\odot$ collapsing protostellar clump has accreted onto the protostar. The difference in radiative acceleration is much greater than this factor of $\ga$30, however, because the edge-on flux is dominated by the far-IR, which is far less effective at radiatively accelerating dusty gas than mid- and near-IR light seen pole-on. Yorke & Sonnhalter (2002) showed by calculating frequency-dependent radiation transfer that the flashlight effect is further enhanced by the fact that the central star’s optical and UV radiation blows out material in the polar direction, reducing back-scattering of radiation toward the disk. A similar effect has been reported by Krumholz, McKee & Klein (2005a) from polar cavities blown free by outflows originating close to the star. Thus, photons emitted or scattered into these directions will not hinder accretion of material within the disk or material in the disk’s shadow regions. Although the flashlight effect allows dusty material to come close to the central source via a circumstellar disk, the material to be accreted eventually encounters optical and UV radiation from the central source. For this material to be accreted rather than blown out by radiation, the dust must be largely destroyed or it must have coagulated into larger particles so that the opacity is dominated by the gaseous component. Even though no massive disk has yet been directly observed around a main-sequence O star, there is much indirect evidence that such disks exist (see the review by Zhang 2005 and our more detailed discussion in [**Section 6**]{}). A compelling argument that disks exist during the early phases of massive star formation is the observation of massive bipolar outflows. Such massive outflows are probably powered by disk accretion, and, as in their low-mass counterparts, the flow energetics appear to scale with the luminosity of the source (see Shepherd & Churchwell 1996a,b; Richer et al. 2000; Henning et al. 2000; Wu et al. 2005). If the primary source of the massive star’s material is accretion from the surrounding molecular core, then a circumstellar disk should be the natural consequence of the star-formation process even in the high-mass case. However, it should be difficult to observe disks around massive stars. The high far UV and extreme UV fluxes associated with high-mass stars will begin to photoevaporate the disks on timescales of $\sim$10$^5$yr (Hollenbach, Yorke & Johnstone 2000). The results will be observable as deeply embedded ultracompact HII-regions with comparable lifetimes (Richling & Yorke 1997). The fact that disks around O stars photoevaporate so quickly provides negative feedback for disk accretion. This limits the build-up of more massive stellar objects and may even imply an upper mass limit for star formation. ### Increase Gravity For completeness we mention the fact that the gravitational acceleration is enhanced with respect to radiative acceleration when massive stars form within a dense cluster of not-so-brightly-radiating objects. In this scenario, one requires a density-peaked cluster of low-mass objects embedded within a molecular cloud, with ${\rm \rho_{objects} \gg \rho_{gas}}$. The effective gravity near the cluster’s center is enhanced relative to an isolated molecular cloud without the cluster and relative to off-center regions of the molecular cloud (cf. Keto 2002). If this were the only way to form massive stars, isolated massive stars would exist only in very exceptional cases. Stellar and Protostellar Luminosity Evolution --------------------------------------------- Because the luminosity is so critical during accretion up to high stellar masses, one must also consider the luminosity evolution of the accreting object. As discussed by Behrend & Maeder (2001) and Yorke (2002), [**Figure 8**]{} shows that protostars do not evolve along premain-sequence tracks until they land on and remain at a unique spot on the main sequence where hydrogen burning starts. Rather, they reach the main sequence – that is, the hydrogen-burning state – well before they have finished accreting mass. After that, they continue growing in mass and evolve up the main sequence until they run out of material to accrete. This means that an initially low-mass object that gains mass through accretion evolves substantially differently in the HR diagram than would a nonaccreting premain-sequence star of the same final mass. [**Figure 8**]{} shows a number of protostellar evolutionary tracks in the HR diagram, assuming a sequence of conceivable accretion rates increasing by factors of 10, but each constant in time. These accreting tracks were calculated by Yorke (2002) and are qualitatively similar to the more detailed stellar evolution calculations by Norberg & Maeder (2000), Behrend & Maeder (2001), and – for the lower accretion rates – Palla & Stahler (1992). Differences to the two former investigations can be attributed to Yorke’s (2002) starting mass, 0.1M$_\odot$ instead of 1M$_\odot$, and the differing accretion rates. In all cases published to date, not only do the tracks of accreting objects consistently lie slightly below the equilibrium deuterium-burning birthline, but the qualitative effect of more rapid accretion is to shift the tracks away from the birthline. (In this context, the birthline is defined as the point on the Hayashi track at which equilibrium deuterium burning begins in a nonaccreting star. Palla & Stahler (1993), however, use the concept of birthline as the evolutionary track of accreting premain-sequence stars in the HR diagram. Deuterium burning of newly accreted material keeps these tracks well above the main sequence until accretion stops.) The tracks of accreting stars eventually converge to the main sequence and follow along the ZAMS as more material is added. For example, at an accretion rate of 10$^{-3}$M$_\odot$yr$^{-1}$, hydrogen burning begins at t$\simeq$1.3$\times$10$^4$yr, after $\sim$13M$_\odot$ have been accreted. The tracks discussed above do not display the degree of bloating seen by Kippenhahn & Meyer-Hofmeister (1975) for accreting main-sequence stars M$_*$$\sim$5–10M$_\odot$. Yorke’s (2002) simplifying assumption of thermally adjusted (pre)stellar objects is not strictly fulfilled, especially for the ${\rm \dot M_{S-acc}}$=10$^{-3}$M$_\odot$yr$^{-1}$ case. Norberg & Maeder (2000) did not consider accretion rates ${\rm \dot M_{S-acc}}$$>$10$^{-4}$M$_\odot$yr$^{-1}$ and Behrend & Maeder (2001) attained high accretion rates ${\rm \dot M_{S-acc}}$$\ga$10$^{-3}$M$_\odot$yr$^{-1}$ for high-mass objects M$_*$$\ga$25M$_\odot$ only. Palla & Stahler (1992, 1993) found significant bloating for the highest accretion rate $10^{-4}$M$_\odot$yr$^{-1}$ they considered, attributable to shell deuterium burning. We expect that (proto)stars accreting at extremely high rates are not fully thermally adjusted and thus bloat up. The degree of bloating depends on the accretion rate and mass of the accreting object (Kippenhahn & Meyer-Hofmeister 1975). We remind the reader that these theoretical tracks in the HR diagram do not reflect the actual observable bolometric luminosities of accreting protostars nor do they reflect the observable effective temperatures. Plotted in the HR diagram are the intrinsic (proto)stellar properties, from which the observed properties must be derived by multidimensional radiative transfer calculations (e.g. Indebetouw et al. 2006). Note that much of the accretion luminosity, ${\rm L_{acc}}$, is indistinguishable from the intrinsic luminosity, L$_*$, of the star. Also, owing to the existence of shock fronts and their postshock relaxation zones, strong deviations from a stellar spectral energy distribution can be expected. The importance of accretion luminosity, ${\rm L_{acc} \sim G M_*\dot M/R_*}$, is seen in [**Figure 5**]{}. At high accretion rates the L/M ratio can attain high values even for low-mass cores, so that radiative acceleration can clear out part of the infalling material in the polar regions. What order of magnitude of mass accretion rate can be expected? In order to produce a star of mass M$_*$ within, say 200,000yr, an average accretion rate of ${\rm 5 \times 10^{-6}\,M_\odot\,yr^{-1} [M_*/M_\odot]}$ is necessary. Assuming this average accretion rate, we note that during the main accretion phase the luminosity of low-mass stars is dominated by accretion luminosity, whereas for high-mass stars the luminosity is initially determined by accretion but is eventually dominated by the intrinsic stellar luminosity. Of course, the actual accretion rate may vary strongly from this average value. The maximum sub-Eddington accretion rate possible onto a core hydrogen-burning star, assuming electron scattering and the effects of both the intrinsic stellar luminosity and accretion luminosity, can be inferred from [**Figure 5**]{}. Whereas the dotted curve corresponding to an accretion rate of ${\rm 10^{-3}\,M_\odot\,yr^{-1}}$ lies below the value permitted by electron scattering, an accretion rate ten times higher would clearly lie above this value everywhere. Thus, accretion of ionized material onto a stellar core at a rate $\ga {\rm 10^{-2}\,M_\odot\,yr^{-1}}$ implies super-Eddington accretion. Stellar Evolution Beyond the Zero Age Main Sequence --------------------------------------------------- As hydrogen burning proceeds, the stars begin to evolve away from the ZAMS even as they accrete material through a disk and lose material through a stellar wind. The star’s evolution strongly depends on its mass loss rate and on the internal mixing that is induced by rotation. These effects must be taken into account when modeling these stars and their effect on their environment. If the stars attain their final mass and become optically visible on a timescale much shorter than their main-sequence lifetime, they will not have evolved far from the ZAMS. If, however, the stars first become optically visible after significant stellar evolution, the traditional concept of the ZAMS as the starting point for newly formed massive stars is flawed. To illustrate this point we compare in [**Figure 9**]{} the current best estimate of the locations in the HR diagram of spectral types O3 through O9.5 for luminosity classes V, III, and I with theoretical isochrones and evolutionary tracks. It is not surprising that the O dwarfs do not fall on the ZAMS, because we are dealing with average properties over a range of ages. A quantitative explanation of the magnitude and systematic variation of the offset for early to late O dwarfs, for example a shift of average ages from $\sim$1 to 5Myr, is still lacking (Martins, Schaerer & Hillier 2005). Such a shift could occur, because stars of later spectral types have weaker winds and lower UV fluxes and therefore remain embedded in their parental cloud longer. [**SpT**]{} [**Luminosity class**]{} ------------- -- ------ -------------------------- ------ V III I B0 – – – O9.5 16.5 21.0 30.4 O9 18.0 23.1 32.0 O8 22.0 26.9 36.8 O7 26.5 31.2 40.9 O6.5 29.0 33.7 43.1 O6 31.7 36.4 45.8 O5 37.3 41.5 50.9 O4 46.2 48.8 58.0 O3 58.3 58.6 66.9 O2.5 – – – : O-stars spectral type versus mass for different luminosity classes (Martins, Schaerer & Hillier 2005) -- ----------------------------- Note: masses in solar units -- ----------------------------- : O-stars spectral type versus mass for different luminosity classes (Martins, Schaerer & Hillier 2005) We find it useful to end this stellar evolution section with a compilation ([**Table 3**]{}) of the masses of massive stars as a function of spectral type (from O9.5 through O3) and luminosity class (V, III, and I). Ionization Evolution and Cloud Disruption ----------------------------------------- Because massive stars are built by accretion, whether from a monolithic collapse or competitively, they will produce an increasing number of hydrogen ionizing photons even while they grow as they become hotter and more luminous. Thus, in spite of accretion, hydrogen-ionizing and helium-ionizing fluxes similar to those expected from ZAMS stars are likely (see [**Figure 10**]{}). Indeed, the process of accretion itself is likely to produce hard ionizing radiation in a series of accretion shocks close to the stellar surface. Moreover, powerful winds interacting with surrounding material will produce strong shocks and hard radiation. Thus, material in the immediate vicinity of the accreting massive star, in particular material in the circumstellar disk and in any nearby star-disk systems, can be ionized. In the case of circumstellar disks, a thin layer of ionized gas on the surface of the disk results. Close to the star, where the escape velocity is much greater than the sound speed, ionized gas remains bound. Ionized gas outside a radius r$_{\rm evap}$, that is, the radius where the sound speed exceeds the escape velocity, expands outward. The situation depicted in [**Figure 11**]{} is likely to result. The photoionization of circumstellar disks produces an outflowing disk wind of ionized gas that interacts with the stellar wind. Depending on its strength (and the photoionization rate), the stellar wind will be collimated by the ionized disk wind (Richling & Yorke 1997). The stronger the stellar wind (or the weaker the ionizing photon flux and hence the ionised disk wind), the less collimated the stellar wind will be. This is different from the case of collimated jets and outflows due to X-winds and disk winds in young low-mass stars, where magnetic fields and forces are involved (Shu et al. 1995). In the high-mass case, although magneto-centrifugal forces cannot be dismissed from influencing the collimation process, we can have rather collimated outflows without magnetic fields. This being said, we suspect that magnetic fields do play a role not only in low-mass but also in high-mass star formation, not least because there are by now at least two cases where strong kilo-Gauss stellar magnetic fields have been inferred from Zeeman spectro-polarimetry, including the oblique magnetic rotator $\theta^1$ Ori C (Stahl et al. 1996, Donati et al. 2002) and the Of?p star HD 191612 (Walborn et al. 2003, Donati et al. 2006). The mismatch in $\theta^1$ Ori C between the magnetic field direction and the stellar spin axis could be an indication that this well-known massive star formed in a collisional process from a merger of two lower mass stars. Finally, we mention the detection of hard X-ray emission from the W3 massive star-forming region, an indication that embedded massive stars must be magnetically active (Hofner et al. 2002). Hoare et al. (2007) examined the suggestion first made by Hollenbach et al. (1994) and later calculated numerically by Yorke & Welz (1996) and Richling & Yorke (1997) that photoevaporating disks can explain the existence of unresolved ultracompact HII regions (sizes $\sim$0.1pc). Many of these have since been resolved with long-baseline radio interferometry and have been reclassifed as hypercompact HII regions (sizes $\sim$0.01pc; see Kurtz 2005). The latter most likely represent the individual photoevaporating disks associated with individual deeply embedded O stars (Keto 2007). The phase of photoevaporating disks is the first evolutionary stage after the ionizing flux has turned on. It remains to be seen whether the thermally evaporating flow interacting with a stellar wind can produce the high velocities seen in the broad H92$\alpha$ recombination line seen toward hypercompact HII regions (Sewilo et al. 2004). Hoare et al. (2007) concluded that in a later evolutionary phase most ultracompact HII regions need to be interpreted as external photoevaporation of a molecular clump. The longevity of these HII regions is explained by a relative motion of the ionizing source and its stellar wind into the clump, producing a combination of bow shock and unipolar flow (a so-called champagne flow, see the review by Yorke 1986). If this is the case, then the transition from hypercompact to ultracompact HII regions may mark the end of the existence of the accretion disk. The ultracompact HII region phase then marks the beginning of the dissipation of the molecular cloud as the stellar wind and ionizing radiation of the newly formed massive star are able to interact with lower density, more poorly gravitationally bound molecular material. An important point to make in the evolution of an ultracompact HII region is that the gravitational force of the star(s) responsible for the HII region should be included (Keto 2002). The gravitational attraction of the star(s) can maintain the accretion flow within the ionized gas and prevent the HII region from expanding hydrodynamically. This is true as long as the radius of the ionization equilibrium is smaller than the radius where the sound speed of the ionized gas, about 10kms$^{-1}$, approximates the escape velocity (concept of the trapped HII region). Indeed, observations of the H66$\alpha$ recombination line from the ultracompact HII region G10.6-0.4 powered by a compact and very luminous ($\sim$10$^6$L$_\odot$) cluster of newly formed massive stars surprisingly show inward motion (Keto 2002). This implies that, despite the high luminosity and ionizing radiation of several O stars, neither radiation pressure nor thermal pressure has reversed the accretion flow (Keto & Wood 2006). The ram pressure of the observed accretion flow, with a rate of M$_*$=10$^{-3}$M$_\odot$yr$^{-1}$, can confine the bubbling HII region and allow the massive stars to continue growing in mass – from ionized gas! One may speculate whether the ionization surrounding the moderately massive stars helps rather than hinders their further growth by accretion (Keto & Wood 2006), the reason being that ionized gas can couple strongly with any magnetic fields. Magnetic fields can transport angular momentum outward, thereby allowing a high rate of infall to smaller radii to be maintained. Eventually, of course, with the stellar masses steadily increasing, the ionizing flux will increase enough (roughly $\propto$M$_*^4$, see [**Figure 10**]{}) so that the radius of ionization equilibrium grows beyond the critical radius within which the gas is gravitational bound. At this point the HII region will burst free and start to disrupt the dense cloud, preferentially unidirectionally in a champagne flow, revealing the initially deeply embedded cluster (both high-mass and low-mass stars). MASSIVE STAR FORMATION: COMPETING CONCEPTS AND CALCULATIONS =========================================================== Three different concepts describing the origin of massive stars have been discussed in the recent literature, each of which may occur in nature, depending on the initial and environmental conditions for the parent molecular clouds (e.g., the Mach number of supersonic turbulence in the clouds and the external pressure in the ISM). These are (1) monolithic collapse and disk accretion, (2) competitive accretion and runaway growth, and (3) stellar collisions and mergers. Whether the parent molecular cloud clumps are magnetically subcritical or supercritical (Shu, Adams & Lizano 1987, Crutcher & Troland 2007) may play a crucial role in determining whether massive stars form (1) in isolation or (2) are strongly clustered. Furthermore, in extreme star-forming environments (such as in massive protoglobular cluster clouds), the initial gas densities may have been so high that (3) stellar collisions become an unescapable ingredient in massive star formation. We discuss all these routes toward massive star formation in this section. In addition, we briefly discuss competing, but not mutually exclusive scenarios related to massive star formation stimulated by the pressure of expanding HII regions versus a supernova blast wave. Similarly, we highlight the two competing, but not mutually exclusive processes accounting for runaway OB stars. Finally, the important concept of mass segregation (the fact that massive stars are often concentrated in the centers of star clusters) is reviewed in terms of nature (a birthmark) or nurture (subsequent dynamical evolution). Monolithic Collapse and Disk Accretion -------------------------------------- Yorke & Sonnhalter (2002) consider the collapse of isolated, rotating, nonmagnetic, massive molecular cores of masses 30M$_\odot$, 60M$_\odot$, and 120M$_\odot$ using a frequency-dependent radiation hydrodynamics code. The flashlight effect discussed in [**Section 3.5**]{} allows material to enter into the central regions through a disk. For massive stars, it is important to take into account the frequency-dependent nature of the opacity and the flux within the disk rather than assuming either Rosseland or Planck gray opacities. For their 60M$_\odot$ case, Yorke & Sonnhalter find that 33.6M$_\odot$ are accreted in the central regions as opposed to 20.7M$_\odot$ in a comparison gray calculation. Because these simulations cannot spatially resolve the innermost regions of the molecular core, they cannot distinguish between the formation of a dense central cluster, a multiple-star system, or a single massive object. They also cannot exclude significant mass loss from the central object(s) that may interact with the inflow into the central grid cell. With the basic assumption that all material in the innermost grid cell accretes onto a single object, they are only able to provide an upper limit to the mass of stars that could possibly be formed for the cases considered. Note that the $\sim$43M$_\odot$ star formed during the collapse of the 120M$_\odot$ molecular clump (Yorke & Sonnhalter 2002) does not represent an upper limit to the stellar mass that is enabled by the flashlight effect. Larger initial masses or a more focused flow along a filament could conceivably lead to even more massive stars. In an ongoing simulation of the collapse of 100- and 200-M$_\odot$ clouds, Krumholz, Klein & McKee (2005) report that $\ga$27M$_\odot$ ($\ga$25M$_\odot$ for the turbulent case) accreted onto a stellar core. As in the simulations of Yorke & Sonnhalter (2002), a disk formed around the accreting (proto)stars. One can speculate on the effect outflows have on the accretion through an accretion disk. The inner part of the accretion disk could well look like the configuration shown in [**Figure 11**]{}. Radiation and the stellar wind from the central star (presumably already hydrogen burning) evacuate a cavity in the polar direction. At the interface between the supersonic outflowing stellar wind and the denser subsonic HII disk atmosphere, some disk material will be removed, but this cannot prevent inward flow of disk material. Inward radial flow of dusty molecular gas is allowed in the equatorial plane of the disk as angular momentum is transfered outward. In cases of magnetized disks with high radiation densities, photon bubbles can lead to more efficient vertical transport of radiation in the disk (Gammie 1998; Begelman 2001; Blaes & Socrates 2003; Turner et al. 2005; Turner, Quataert & Yorke 2007). This can explain super-Eddington accretion in a variety of luminous systems, including accreting compact objects and very massive stars. Efficient angular momentum transfer could result from weak magnetic fields in the disk (Balbus & Hawley 1991, Hawley & Balbus 1991, Balbus 2003), from turbulence and/or spiral density waves (Bodenheimer 1995) excited by gravitational instabilities in nonmagnetized (Laughlin & Bodenheimer 1994) or magnetized (Fromang et al. 2004) disks, or from the tidal effects of nearby stars (Terquem 2001). Indeed, rapid accretion through a disk may be a direct consequence of having nearby companions. This may explain why massive stars are generally members of multiple systems. Once the disk material crosses r$_{\rm dust} \sim$25AU\[M$_*$/30M$_\odot]^{1.6}$, the radius of dust destruction, its opacity decreases and it is not easily stopped by radiation. It is, however, still unclear how the disk material ultimately flows onto the star. Surely, the disk puffs up close to the star, in analogy to the accretion disks in active galactic nuclei. Beyond ${\rm r_{evap} \sim 130 AU [M_*/30 M_\odot]}$, where the escape velocity is less than 10kms$^{-1}$, the disk loses material via photoevaporation on a timescale of $\sim$10$^5$yr. This is of the same order as the accretion timescale. These competing effects (accretion and photoevaporation) will determine the final mass of the star and perhaps even the upper mass limit. Competitive Accretion and Runaway Growth ---------------------------------------- Bonnell et al. (1997, 2001a) present the first 3D numerical simulations of the growth of stellar masses by competitive accretion in small young star clusters. The best way to visualize the idea of competitive accretion is to compare it to an economic model based on two complementary concepts; the real estate concept – “location, location, location” – and the capitalistic concept – “the rich get richer.” The former reminds us that environmental influences can be very important, such that being in a fortunate location can significantly promote growth, whereas the latter simply means that the gravitational attraction increases with success – in this case, with the increasing mass of the star. A protostar’s ability to grow depends on the size of its accretion domain, i.e., the region from which gas can be gathered. A location in the center of a protostellar cluster is beneficial, as gas flowing down to the center of the cluster increases the gas reservoir available to an individual star. The early birth of a protostar may also give it an unfair advantage in the competition to end up large and massive. Imagine a large, dense molecular gas cloud with a number of protostellar seeds distributed inside the cloud that have initially condensed from some denser portions of the cloud. These condensations or cores subsequently have the chance to grow in mass by accumulation (accretion) of lower density cloud gas from their individual accretion domains (Larson 1978, Zinnecker 1982). These accretion domains are systematically larger for higher mass seeds. With increasing mass, the gravitational spheres of influence keep growing. Equally important, the amount of material that enters an individual accretion domain depends on the external environment. In isolation, it would simply be proportional to the mean gas density of the region. By contrast, in a larger scale potential such as that provided by a protocluster cloud, the gas density can become significantly larger in the center of the protocluster, as the gas settles into the deepest part of the potential, there to be accreted by the growing protomassive star. The accretion domain of an off-center protostar is tidally limited by the total mass in the inner part of the cloud, whereas the accretion domain of a protostar in the center of the cloud is the whole cloud. Because the gas reservoir is limited (the cloud has a finite mass), the protostellar masses will eventually compete for cloud gas, especially after the accretion domains start to overlap. The action of the cluster to gather matter from larger distances and focus it toward the accreting stars, combined with the increasing accretion radii of these stars owing to their increasing mass, is what makes competitive accretion such a powerful mechanism. This is also true in a molecular cloud with hierarchical substructure, in particular for subcluster clumps of gas in a bigger protocluster cloud. The implication is that each subcluster clump is likely to have one most massive protostellar object in its center surrounded by a hierarchy of lower mass objects (Bonnell, Bate & Vine 2003). When and if those subcluster clumps merge to one big protocluster (see [**Figure 12**]{}), the complex accretion history of each protostar is no longer related to a single local cloud core (Schmeja & Klessen 2004). In this scenario, the massive stars had accretion histories that were priviledged at every stage – from Jeans-instability protostellar birth to location and density in the gas cloud, all factors affecting accretion were more favorable than the average. This is why massive stars are rare. The rarest and most massive of them probably formed in the most favorable conditions by runaway accretion until their gas reservoir was exhausted or dissipated (e.g., by ionization feedback, see von Hoerner 1968 and recently Clarke, Edgar & Dale 2005). The above model of competitive accretion has been critized by Krumholz, McKee & Klein (2005b) on the grounds that Bonnell’s SPH simulations start from very strongly gravitationally bound protocluster clouds, while observationally such clouds appear to be supported by turbulent motions. In other words, the simulations use a virial parameter ${\rm \alpha}$=E$_{\rm turb}$/E$_{\rm grav}$$\ll$1 while molecular observations suggest ${\rm \alpha}$$\sim$1. Krumholz, McKee & Klein (2005b) argue from analytical considerations that protostellar masses cannot grow in such a turbulent medium, not even by a factor of two. The Bondi-Hoyle accretion rate is far too low owing to the high relative velocities between the accreting stars and the turbulent gas. In addition, radiative feedback from the incipient massive star may prevent Bondi-Hoyle accretion altogether (Edgar & Clarke 2004). In response to this criticism, Bonnell & Bate (2006) argue that global initial collapse versus quasi-equilibrium support is not the issue, as their more recent SPH calculations start with initial conditions where the turbulent kinetic energy is close to the gravitational energy. In fact, the scaling laws of supersonic turbulence (Larson 1981) imply small turbulent velocity differences between protostars and their neighboring gas. This then allows for significant growth in stellar mass while more distant high velocity turbulent gas cannot be accreted. In other words, protostars are swept along with their neigboring gas for some time in a similar global motion before they hit more distant and thus less-correlated gas. The above Bondi-Hoyle accretion problem is also not as serious as suspected, considering that the most massive star forms in the center of the clump with little motion relative to the surrounding gas. Future detailed observations of the motions in massive ($\sim$1000M$_\odot$) star-forming clumps will discriminate between these different dynamical scenarios (small versus large gas motions relative to the protostars, turbulent cloud support, or overall cloud collapse). A first step in this direction was made by Peretto, André & Belloche (2006) in their dust continuum and molecular study of the NGC 2264 clumps, with the conclusion that the observations are most consistent with “a picture of massive star formation intermediate between the scenario of stellar mergers of Bonnell et al. (1998) and the massive turbulent core model of McKee & Tan (2003), whereby a turbulent, massive ultra-dense core is formed by the gravitational merger of two or more Class 0 protostellar cores at the center of a collapsing protocluster”. Stellar Collisions and Mergers ------------------------------ Historically, the original reason for proposing stellar collisions as a formation process for massive stars was twofold. ([*a*]{}) At the time, radiation pressure on dust was considered a severe hindrance to gas accretion (then assumed to occur spherically symmetrically). Today this concern has gone away (see [**Section 4.1**]{}). ([*b*]{}) The packing of massive stars in dense clusters was too tight, and so there was concern that a sufficiently large gas reservoir for monolithic collapse was not available. This is still a concern, although not all massive stars form in densely packed clusters – many form in widely spread OB associations. In any case, it is possible that a collisional build-up of high-mass stars can occur, especially for the most massive stars in very dense clusters. The problem is the very high stellar density of already massive or at least intermediate-mass stellar objects that is required to get the process going. We can estimate the threshold stellar number density n$_{\rm star}$ or, equivalently, the average star-star separation (s=n$_{\rm star}^{-1/3}$) in a dense cluster for stellar collisions (encounters) to be important. Because of the fundamental role of gravitational focusing coupled with a distribution of velocities (i.e., a finite probability of low relative velocities at infinity v$_\infty$), the cross section ${\rm \sigma_{grav}}$, where $${\rm \sigma_{grav} = \pi R^2_{min} \left(1 + \frac{2 G M_*}{v_\infty^2 R_{min}}\right)}$$ for two stars of mass M$_*$ passing each other at periastron within a minimum distance R$_{\rm min}$ is vastly enhanced over the geometrical cross section. Gravitational focusing can enhance the effective cross section (by factors of $\sim$10$^4$), thus rendering close stellar encounters in very young stellar clusters realistic. Of course, the condition to meet for massive star growth through mergers is that the collision (close encounter) time t$_{\rm coll}$ for the stars to collide must be shorter than the timescale for stellar evolution of the most massive star in the cluster ($\sim$3Myr). Using the formula for the stellar collision time per star (Binney & Tremaine 1987, Dale & Davies 2006), i.e., $${\rm \tau_{coll} = \frac{1}{n_{star} \sigma_{grav} v_{rms}} = \left[4 \sqrt{\pi} n_{star} v_{rms} \left(R_{min}\right)^2 \left(1 + \frac{2 G M_*} {R_{min} v^2_{rms}} \right)\right]^{-1}}$$ we obtain, when gravitational focusing dominates $${\rm \tau_{coll} = 7 \times 10^7 \left[\frac{n_{star}}{10^6\,pc^{-3}} \right]^{-1} \left[\frac{M_*}{10\,M_\odot} \right]^{-1} \left[\frac{R_{min}}{1\,R_\odot} \right]^{-1} \times \left[\frac{v_{rms}}{10\,km\,s^{-1}} \right] yr}.$$ Here we have normalized the expression with reasonable numbers for the number density, mass, size, and velocity dispersion of the stellar collision partners, assumed to be of equal mass. We ignore binary stars for the time being. Note that the velocity dispersion is the 1D-velocity dispersion, which is 2.3kms$^{-1}$ in the Orion Nebula cluster (van Altena et al. 1988), but is $\sim$5kms$^{-1}$ in denser and more massive cluster cores such as NGC 3603 and R136, or even larger in young globular clusters such as those in the Antennae (Mengel et al. 2002). [**Equation 5**]{} is based on equal mass collision partners; however, recently the formula has been extended to cover nonequal mass encounters (Moeckel & Bally 2006, 2007), as well as larger cross sections due to circumstellar disks and binary components (Davies et al. 2006). This can decrease the threshold stellar number density for collisions from 10$^7$ or 10$^8$pc$^{-3}$ down to about 10$^6$pc$^{-3}$ (Bonnell & Bate 2005). In [**Figure 13**]{} we show 3D SPH numerical simulations of the collision process of two pairs of stars, one pair of equal mass and one pair of unequal mass. Triggered OB Star Formation --------------------------- The classical model of triggered OB star formation goes back to Elmegreen & Lada (1977). The idea is that the ionization shock front of one group of massive stars provides the external pressure to compress adjacent molecular cloud layers, thereby inducing the formation of a new group of massive stars, which in turn, by the same process, induces the formation of another generation of massive stars and so on. The Elmegreen & Lada (1977) model was developed to explain the sequence of spatially distinct OB subgroups in nearby OB associations such as Orion OB1 or Sco-OB2 (Blaauw 1964, 1991). It is important to note that the Elmegreen & Lada model did not predict the formation of the observed coeval, low-mass T Tauri star population in the subgroups of OB associations (Preibisch & Zinnecker 2007, Briceño et al. 2007). The low-mass population was assumed to form independently and in many locations spread out over the cloud. Hence the Elmegreen & Lada (1977) process of sequential triggered star formation, contrary to accepted wisdom, may not be the main mechanism that accounts for the existence of OB subgroups. Rather, supernova triggering could be at work, assuming that a supernova blast wave can trigger both high-mass and low-mass stars at the same time (which would then explain the coevality of the high-mass and low-mass stars). Numerical simulations along these lines are just beginning (e.g., Melioli et al. 2006). However, there are several clear cases known in our Galaxy where expanding HII regions have swept up molecular gas at their periphery and in which new massive stars have formed or are about to form. The latter is indicated by luminous IR sources (e.g., Sh 104 and RCW 79, Zavagno et al. 2005; RCW 108, Comerón, Schneider & Russeil 2005). This is indeed reminscent of the Elmegreen & Lada (1977) model of triggered sequential star formation (the so-called collect and collapse scenario, Elmegreen 1998; see also Whitworth et al. 1994 and Dale, Bonnell & Whitworth 2007). However, according to Elmegreen & Lada (1977), it may be expected that in a two-stage starburst (cf. Walborn & Parker 1992) the second generation protocluster hosts only a small group of high-mass OB stars without the concomitant multitude of low-mass stars. Such a small N-body group, much like the Orion Trapezium system, would then be highly dynamically unstable and could help explain the occurrence of runaway OB field stars (Clarke & Pringle 1992, Allen, Poveda & Hernández-Alcántara 2004). An important implication of this scenario would be that the HII region of a big star cluster (e.g., NGC 3603) is not able to trigger the formation of a similarly massive second generation star cluster, but only a smaller mass cluster (e.g., IRS9 in NGC 3603; see Nürnberger 2003), potentially with a top-heavy stellar IMF; similarly, the R136 cluster in 30 Dor does not seem to trigger the formation of a new massive cluster but just a few small groups of embedded protostars (Rubio et al. 1998, Walborn et al. 1999, Brandner et al. 2001, Walborn, Maíz-Apellániz & Barbá 2002). A very interesting case is N81, one of the most compact HII regions in the Small Magellanic Cloud (SMC). N81 is isolated in an area of low extinction in Shapley’s wing, 1.2kpc away from the main body of the SMC. In contrast to other compact HII regions, typically located within or at the edge of giant HII regions, it appears that SMC N81 has been formed in isolation. The study of its stellar inventory shows that this ’high-excitation blob’ is ionized by at least eight near-ZAMS O stars in an instantaneous burst (Heydari-Malayeri et al. 2002, 2003). The question here is, Where is the trigger? How did these massive stars form? Back to the Galaxy, another region of note is the Carina nebula powered by the most extreme grouping of massive stars in the southern Milky Way (Smith & Brooks 2007). Here $\sim$65 O stars (including many in the clusters Tr14 and Tr16, but excluding $\eta$ Car) provide a total of $\sim$10$^{52}$erg of kinetic energy and a Lyman continuum luminosity of $\sim$10$^{51}$ photonss$^{-1}$, creating a giant superbubble. Most of the bubble (seen as polycyclic aromatic hydrocarbons in [*Spitzer*]{}/MSX images) resides as atomic gas in the photodissociation regions and not in dense molecular clouds. The synchronized star formation around the periphery of Carina strongly suggests that star formation was triggered by stellar winds. The second-generation population appears to involve a cascade toward preferentially intermediate-mass and low-mass stars (Smith & Brooks 2007), but the situation may change soon when $\eta$ Car and its siblings explode as supernovae, rejuvenating massive star formation. The idea that it is supernovae rather than HII regions and stellar winds that stimulate wide-spread massive star formation is not new (see Herbst & Assousa 1977 for the Canis Majoris star formation region, see also Gerola & Seiden 1978 for galactic spiral arms, and the pioneering paper by Öpik 1953). However, what is new is the evidence from the Upper Scorpius OB subgroup data that supernovae can trigger both high-mass and low-mass star formation in an OB subgroup at the same time (Preibisch & Zinnecker 2007). Supernovae need a critical distance from a molecular cloud – not too close and not too distant – to be an effective trigger for star formation because of momentum transfer of the blast wave onto prestellar cores (Vanhala & Cameron 1998, Vanhala et al. 1998). Other triggering mechanisms, like radiatively driven implosion of globules, also operate, but seem to be secondary processes, forming only small stellar groups rather than whole OB subgroups with thousands of stars (for a review of a whole variety of triggering mechanisms, see Elmegreen 1998). The Carina star-forming region (diameter 150pc, age 3Myr, total IR-luminosity $\sim$10$^7$L$_\odot$) may be the galactic analog of giant extragalactic HII regions such as NGC 604 in M33 (size 40arcsec or 140pc), a very large OB association (see Maíz-Apellániz, Pérez & Mas-Hesse 2004). Closer to home, the Orion Nebula cluster is likely an example of triggered massive star formation. The appearance of the Orion A and B molecular clouds suggest an interaction caused by energy input from the Orion OB1ab association subgroups (Blaauw 1991). Furthermore, the formation of the nearest massive star in the Orion BN/KL region, which hosts a very bright IR source, has probably been triggered by the Orion Trapezium star’s HII region pushing against the Orion Molecular Cloud right behind the Trapezium cluster (see Bally 2002). Recent [*Spitzer*]{}/IRAC Galactic plane observations (Churchwell et al. 2006) have revealed dozens of parsec-sized bubbles, formed by hot young stars in massive star forming regions. Among the 80 or so ring-like structures produced by young OB stars, several show secondary bubbles on the rim of the primary bubbles, suggestive of triggered star formation. However, is this morphological evidence conclusive, i.e., is triggering necessary to explain such spatial correlations? In an interesting paper, Dale, Clark & Bonnell (2007) attempt to address this and similar issues. They conduct SPH simulations to examine the difference between triggered and revealed star formation. They study the impact of irradiation by an external source of ionizing photons on a turbulent massive molecular cloud and compare the results (the number and type of stars formed) with a control simulation where the turbulent cloud evolves without the impact of the irradiation. They find that, although the external ionization has a dramatic effect on the morphology of the model cloud, its impact in terms of extra star formation is surprisingly minor; the feedback effects can be both positive and negative, accelerating the formation of some objects and delaying the formation of others. Only a few objects form that would otherwise not have formed, and the effect of induced star formation on the overall star formation efficiency is less than a factor of two. Dynamical Evolution: Mass Segregation and Runaway OB Stars ---------------------------------------------------------- ### Mass Segregation Massive stars are often found near the centers of star clusters but not exclusively. Examples where massive stars are preferentially located in and near the cluster center include the Trapazium in the Orion Nebula cluster (Hillenbrand 1997) or the WR and O stars in the NGC 3603 cluster (Drissen et al. 1995). The issue is whether this is a birthmark (we call this effect prompt mass segregation) or whether this is an N-body evolutionary effect after birth (in which case, we call it dynamical mass segregation). The way to decide between these two possibilities is to estimate the dynamical time for massive stars to sink to the center of the gravitational potential from the half-mass radius of the cluster. If this timescale turns out to be too long, i.e., longer than the age of the cluster, dynamical mass segregation is ruled out and prompt mass segregation is indicated (e.g., Bonnell & Davies 1998, who infer that mass segregation in the Orion Nebula cluster must have been prompt and hence an important constraint for understanding massive star formation). ### Runaway OB Stars Runaway OB stars were defined by Blaauw (1961) as massive stars with radial velocities in excess of 40kms$^{-1}$. Their high space velocities prompt the question of how they were accelerated to such kinetic energies. The original proposal involved an asymmetric supernova in a massive binary system that provided the kick to eject the companion from the system (Blaauw 1961). A strong prediction of this model is that the runaway massive stars should themselves be single. More recently, an alternative suggestion for the runaway phenomenon was put forward (Poveda, Ruiz & Allen 1967, Gies 1987). Here the idea is that massive stars in dense cluster cores (which is their preferential location) undergo dynamical three-body encounters, especially if a massive binary is involved. In such gravitational interactions, potential and kinetic energy are exchanged: The binary orbit shrinks, and the intruder extracts the potential energy and converts it into the corresponding kinetic energy. It can also happen that an exchange reaction occurs so that one of the binary components (usually the lighter one, rather than the intruder) is ejected. Conservation of momentum leads to a recoil for each of the two system components, so they fly apart in roughly opposite directions. In this scenario, the runaway stars need not be single but can, in principle, be binaries, at least in 50% of the cases. Hoogerwerf, de Bruijne & de Zeeuw (2001) give evidence that both processes for runaway OB stars occur in nature: while $\zeta$ Ophiuchi is a very high proper motion massive O9.5 star without an antipode in the opposite direction, and hence an example for the supernova scenario, another famous case, that of $\mu$ Columbae and AE Aurigae, argues for the dynamical interaction scenario, where two hard binary systems likely had a close encounter in the 5Myr old NGC 1980 Orion cluster, in which two of the binary components were set off in opposite directions with speeds of around 100kms$^{-1}$. In their wake is $\iota$ Ori, a tight, eccentric, massive binary (Gualandris, Portegies Zwart & Eggleton 2004), known to be classic, colliding-wind X-ray binary. Other famous examples of massive runaway stars include $\lambda$ Cep, $\xi$ Per, and $\alpha$ Cam (astronomy picture of the day on 24 Nov 2006). The dynamical ejection model for runaway stars would predict not only very fast ejection speeds, but also slower ones, as milder interaction events with wider, i.e., softer, binary pairs can occur (Kroupa 2000). This must be kept in mind when discussing the origin of massive field stars: with an escape speed of only 5–10kms$^{-1}$, a massive O star with an age of around 3–5Myr could still travel 15–50pc from its cluster birth place. Some 40 nearby field O stars are known and at least half of them can be associated with a cluster origin (de Wit et al. 2004); for the rest the situation is unclear. It is interesting to recall that the spectroscopic and visual binary frequency among these runaway O stars is very low (Mason et al. 1998). MASSIVE STAR FORMATION:\ BINARY AND MULTIPLE SYSTEMS =========================== Some of the most important clues toward understanding the formation of massive stars, that have been neglected in previous reviews (e.g., Massey 2003), come from their high frequency of binary and multiple systems, together with an analysis of their properties (period distributions, mass ratios, orbital eccentricities); see [**Section 2.4**]{} and the discussion later in [**Section 6.2**]{}. Massive stars often come in hierarchical triples with an almost equal-mass close massive binary and a third more distant companion. One example includes $\theta^2$ Ori A in Orion’s Bar (see Preibisch, Weigelt & Zinnecker 2001). Another is $\sigma$ Ori in the center of the namesake cluster, whose hierarchical configuration is even more complex (see Sanz-Forcada, Franciosini & Pallavicini 2004). These close massive binaries likely play an important role in the evolution and age dating of starburst galaxies, as pointed out by van Bever & Vanbeveren (1998), and an equally important role in estimating supernova rates and the rates of high- and low-mass X-ray binaries (see Verbunt 1993; see also H.A. Kobulnicky, C.L. Fryer & D.C. Kiminki, submitted). Massive close binaries also bias and obfuscate measurements of the velocity dispersion in dense starburst clusters (see Bosch et al. 2001). It seems worthwhile and appropriate to summarize the various formation processes of massive binary and multiple systems. The following five processes could be relevant. Disk or Filament Fragmentation ------------------------------ This is the most obvious mechanism; it is similar to suggestions of how to form low-mass T Tauri binaries from filament fragmentation (Zinnecker 1991, Bonnell & Bastien 1992, Monin et al. 2007). A filamentary geometry may play a key role in the fragmentation process, because the isothermal case is a critical one for the collapse of a cylinder: the collapse and fragmentation of a cylinder can continue freely as long as the temperature continues to decrease, but not if it begins to increase (Larson 2005). As for disk fragmentation, gas thermal physics controls the non-linear outcome of gravitational instability in low-mass circumstellar disks (Durisen 2001). The fragmentation of massive disks has recently been studied analytically by Kratter & Matzner (2006). They found that these disks are unstable to fragment if they are cold enough, and catastrophically so when ${\rm \dot M}$$>$10$^{-3}$M$_\odot$yr$^{-1}$. This mechanism accounts only for initially wide binaries. Cold rotating Keplerian disks produce circular orbits; sub-Keplerian disk rotation or cold filaments tumbling end over end produce highly eccentric orbits. $\theta^1$ Ori C may be an example of the latter (Kraus et al. 2007). Note that fragmentation is different from fission, the splitting up of a hydrostatic, rapidly rotating body. Fission does not work for a compressible fluid (Tohline & Durisen 2001); it only leads to the ejection of spiral arms and torques that slow down the rapid stellar rotation. Accretion onto a Low-Mass, Wide Binary -------------------------------------- This is a less obvious but potentially key mechanism to form tight massive binaries. Speculated upon by Maeder & Behrend (2002), it was worked out by Bonnell & Bate (2005). These authors realized that the orbital separation of the components of close binaries is much smaller than the Jeans length at reasonable gas densities and temperatures. Therefore, they investigated the physical idea that the orbital separation of an initially wide binary would shrink by letting both components accrete and grow in mass; the final separation then depends on the orbital angular momentum of the accreted material, as follows: Let us consider the angular momentum A=Mv$\times$R of a binary system. Because of v$\propto$(M/R)$^{1/2}$ (v is the Keplerian orbital speed), $${\rm A \propto M^{3/2} R^{1/2}}$$ where M is the instantaneous total mass of the binary system and R is the separation of the binary components. If the accreted material has zero net angular momentum, as would be expected if the infall were spherically symmetric, then A remains constant (A=const), implying that R$\propto$M$^{-3}$, i.e., the binary separation should be a strong function of the binary mass. If, instead, the accreted material has constant specific angular momentum (A/M=const), then the total angular momentum will scale with the mass of the binary (A$\propto$M), implying that R$\propto$M$^{-1}$. Thus it is easy to see that accretion onto a binary system can significantly decrease its separation at the same time that it increases its mass (Bate 2000). In the numerical SPH simulations, the early evolution is well parameterized by the relation R$\propto$M$^{-2}$, intermediate between the above two cases, indicating that the binary is accreting some angular momentum with the mass but that the net specific angular momentum of the accreting gas is decreasing with time. The reason for this is that random velocities of gas at large radii tend to cancel each other out, resulting in accretion of a lower net angular momentum. With an R$\propto$M$^{-2}$ relation, we infer that a low-mass (e.g., 3M$_\odot$) wide pair (e.g., separation of order 100AU) will become a binary system with a semimajor axis of around 1AU by the time the system has accreted up to 30M$_\odot$. That is, accretional growth in mass by a factor of 10 will make the separation decrease by a factor of 100, or more if the growth in mass is more than a factor of ten. This may explain the observed very tight O-star spectroscopic binaries (see [**Section 2.4**]{}). The component masses, even if unequal in the beginning, would tend to become equal later when higher angular momentum cloud material falls preferentially on the secondary component (Bate & Bonnell 1997), as the lower mass secondary is, by definition, further away from the center of mass and would thus carry the higher angular momentum in the system initially. Another very interesting new model to explain the origin of close massive ’twins’ is proposed by Krumholz & Thompson (2006). They invoke mass transfer in close, rapidly accreting protobinaries; this always pushes the initial binaries toward mass ratio unity. Their model is superficially similar to the model of Bonnell & Bate (2005), but in fact the physical details are quite different, involving the swelling of protostars undergoing deuterium shell burning. Failed Mergers in Stellar Collisions ------------------------------------ If massive stars can form through stellar mergers, then near misses might sometimes form tight and eccentric massive binaries. A necessary condition is that the kinetic energy of the quasi-parabolic encounter can be dissipated as tidal energy (Fabian, Pringle & Rees 1975) in a gravitationally focused grazing fly-by (Zinnecker & Bate 2002, Dale & Davies 2006). Apart from the very high stellar number densities required for this process to be significant and frequent, the main problem is the extreme fine tuning of the collision impact parameter. Indeed, the cross section is a very small concentric annulus in impact parameter space that is much smaller than the central area inside the annulus relevant for massive mergers (F. Rasio, private communication). One way out is disk formation during the tidal disruption of a lower density star by a higher density star in a non-head-on collision; see the example of a grazing collision between a 3M$_\odot$ premain-sequence star and a 10M$_\odot$ main-sequence star in [**Figure 13**]{}. This can lead to subsequent disk-assisted capture of a companion, to be discussed extensively in the next subsection (the so-called ’shred and add’ scenario, Davies et al. 2006). A variant of the failed merger scenario is the off-center collision of two protostellar cores with extended accretion envelopes (Stahler, Palla & Ho 2000). These protostellar envelopes offer a much higher interaction cross section and can provide the necessary orbital drag for collisional massive binary formation (cf. Silk 1978). Disk-Assisted Capture --------------------- If massive stars form via disk accretion, then the large disk radii increase the interaction cross section considerably. This suggests that disk interactions with neighboring stars could assist in capturing binary companions (Bally & Zinnecker 2005). Although this mechanism has been found insufficient for solar mass stars with disks (Heller 1995, Boffin et al. 1998), recent SPH/N-body simulations by Moeckel & Bally (2006, 2007) convincingly showed that disk-assisted capture is much more efficient in a regime suited to massive stars (ca. 20M$_\odot$) with large disks (ca. 500AU). We note that this process works particularly well in providing massive stars with lower mass companions in rather wide orbits (separation ca. 100AU). The same authors also discuss the consequences of a mass-dependent velocity dispersion and of an initial mass segregation for the capture rates. Furthermore, they considered the long-term survival of these binaries in a dense cluster. The point is that massive binaries with different separations and mass ratios form by different processes. Although accretion seems to be the only way to form the observed tight equal-mass binaries in young clusters, fragmentation and disk-assisted capture can form the wider, unequal-mass binaries. N-Body Dynamical Evolution -------------------------- Small-N groups of massive and intermediate-mass stars are seen forming in 3D SPH simulations of the collapse and fragmentation of gas-rich protoclusters (e.g., Bonnell, Bate & Vine 2003). Binary formation in these groups is common and occurs through dynamical three-body capture (different from tidal capture). To begin with, a massive star typically has a lower mass wide companion. With time, an exchange interaction with a more massive third object occurs, thus forming an almost equal mass but still wide binary. The wide binary then shrinks (’hardens’) as it takes up most of the binding energy of the small group when the lower mass group members get kicked and escape. The typical final outcome is a tight equal-mass massive binary, with a lower mass wide companion (Bate, Bonnell & Bromm 2002). This scenario explains many aspects of massive binaries, including the average companion star fraction of 1.5 and the prevalence of short-period spectroscopic systems. Also, the timescales for the N-body dynamics are short enough for the whole process to occur on or before the ZAMS (van Albada 1968, Aarseth 2003). However, it would seem that the frequency of massive close binaries should correlate with the overall stellar number density in young clusters, which is not what is observed. In fact, at face value the opposite – an anticorrelation – is observed (Mermilliod & García 2001). We have no real explanation as to how this environmental effect fits into the N-body picture, unless most of the most massive binaries in the densest young clusters have managed to merge into a single object. Stellar mergers are thought to occur in old globular clusters (e.g., Dale & Davies 2006), and this is one possible explanation for the presence of blue stragglers (younger stars with masses higher than the cluster turnoff mass). Stellar mergers may therefore be even more important in dense, massive, young globular clusters, both now and in the past. The Origin of Trapezium Systems ------------------------------- Finally, we address the problem of how Trapezium systems of massive stars, such as the one in the center of the Orion Nebula Cluster ($\theta^1$ Ori), likely came into being. In brief, the idea is as follows: numerical SPH simulations of supersonic gravo-turbulent fragmentation of a protocluster cloud (1000M$_\odot$) suggest that a collapsing cloud develops a few subclusters (star+gas systems), which subsequently merge into a single cluster entity. Each subcluster carries one most massive star (likely already part of a multiple). Hence the merging of subclusters will result in a central Trapezium-type system (see [**Figure 14**]{}), as observed in the core of the Orion Nebula Cluster (see [**Figure 15**]{}). Note that components A1 and B1 of the Orion Trapezium are spatially unresolved eclipsing spectroscopic binaries; for a summary of the parameters of the Trapezium multiple stars we refer to Preibisch et al. (1999) and Schertl et al. (2003). Note also that component B is itself a Trapezium-like system, indicating the hierarchical nature of massive star formation. The dynamical evolution of Trapezium systems, including stellar ejections, is discussed by Allen, Poveda & Hernández-Alcántara (2004), taking into account their multiplicity substructure. Future studies of Trapezium-type systems will likely concentrate on embedded systems, such as W3-IRS5 (Megeath, Wilson & Corbin 2005). MASSIVE STAR FORMATION: DISCUSSION ================================== Disks and Outflows ------------------ Disks and outflows are a general phenomenon in low-mass star formation that is explained by the accretion-ejection connection (Camenzind 1990, Ferreira & Pelletier 1995). That is, disk accretion energy powers either a disk wind (Pudritz & Norman 1983, Pudritz et al. 2007) or an X-wind (Shu et al. 1994, 1995). These in turn are magnetically collimated into a molecular jet. The jet then runs into the parent molecular cloud and local ISM, thereby accelerating entrained ambient molecular gas into a wider momentum-driven molecular flow. Textbook examples include the famous Hubble Space Telescope (HST) source HH30 (Burrows et al. 1996) and the most beautiful IR jets HH211 and HH212 (McCaughrean, Rayner & Zinnecker 1994; Gueth & Guilloteau 1999; Zinnecker, McCaughrean & Rayner 1998; Lee et al. 2006). Similar but usually less collimated outflows have been observed around more luminous and, hence, more massive young stars (Shepherd & Churchwell 1996a, b; see the compilation by Beuther & Shepherd 2005). This has led to the notion that high-mass star formation is a scaled-up version of low-mass star formation. However, as noted by Arce et al. (2007), no highly collimated outflow has been observed for high-mass star-forming regions exceeding 10$^5$L$_\odot$ (corresponding to $\sim$25M$_\odot$). Likewise, circumstellar disks have been detected around young B-type massive stars (Zhang 2005), but disks around the progenitors of O-type massive stars in excess of 20M$_\odot$ have eluded detection (Cesaroni et al. 2007). This suggests that simple-minded claims of exact similarity between high-mass star formation and low-mass star formation are partly based on a casual use of the term ’high-mass star’ and partly on wishful thinking. Yes, the rotating disks around IRAS 20216+4104 (Cesaroni et al. 1997, 1999; Cesaroni et al. 2005b), G192.16–3.82 (Shepherd, Claussen & Kurtz 2002), Cep A (Patel et al. 2005), or GL 490 (Schreyer et al. 2006) are wonderful examples of millimeter-interferometric observations. But the central sources are at best early B-type stars (below 20M$_\odot$) rather than massive O stars where radiation pressure and photoevaporation play a role in the formation process. The same seems to be true for the large silhouette disk in M17 (Chini et al. 2004), which is associated with a molecular hydrogen jet (Nürnberger et al. 2007). Let there be no misunderstanding. It is likely that the majority of massive stars forms by disk accretion. But the issue is really whether stars of 50–100M$_\odot$ can also form by some other, more dramatic process like stellar or protostellar collisions (Bonnell, Bate & Zinnecker 1998; Stahler, Palla & Ho 2000). One could imagine that the reason why no disks and collimated outflows have been found for the most massive stars in statu nascendi is because their disks are being photoeroded at the same time the stellar mass grows by disk accretion. The combination of a powerful stellar wind and radiative acceleration clear the polar regions of material and magnetic fields, thus leading to poorer collimation for their molecular outflows. Maybe the outflow from AFGL 2591 ([**Figure 16**]{}) is such a case. The most messy outflow from a massive star-forming region is from the BN/KL region in Orion. The OMC1 outflow (mass 10M$_\odot$, velocity 30–100kms$^{-1}$) and the H$_2$ fingers (Allen & Burton 1993, McCaughrean & Mac Low 1997, Kaifu et al. 2000) have been interpreted as the result of a powerful explosion that occurred in the center of OMC1 within the last 1000yr. It has been proposed that a dynamical interaction 500yr ago, possibly leading to a merger, may have produced the OMC1 outflow and H$_2$ finger system (Bally & Zinnecker 2005). In this scenario, interactions with surrounding gas have decelerated the impulsive outflow powered by the stellar collision by about a factor of two, thereby reconciling the 1000-yr dynamical timescale with the 500-yr timescale for the interaction. The detection of oppositely directed motions in radio sources IRc2-I and BN provides support for models in which the OMC1 outflow was powered by a dynamical interaction between high-mass or intermediate-mass stars in an ultradense environment (Rodríguez et al. 2005a; see also Tan 2004). Another issue that needs to be appreciated in the context of massive outflows is the question: to what extent do the observed massive outflows trace the combination of stellar wind and intrinsic disk wind? The latter is able to entrain a lot of ambient gas, and care must be taken not to mistake this gas mass for the intrinsic wind mass. This is important because the accretion rate is estimated from the outflow rate, so if the outflow rate is overestimated, the accretion rate is overestimated too. Somewhat analogously, one must also be careful when deriving the stellar mass from the luminosity of the source (e.g., using IRAS and submm data) when part of this luminosity is nonstellar and comes from accretion onto the disk and onto the (proto-)star (see [**Section 3.4.2**]{} and [**Figure 5**]{}). Accretion occurs as long as there is outflow activity, and the highest accretion rate translates into the most violent outflow. Spatially unresolved binary and multiple systems compound the issue (see the next section). A final issue is the inferred nonlinear mass dependence of the disk accretion rate on the instantaneous central stellar mass (Norberg & Maeder 2000, Behrend & Maeder 2001). This is a consequence of a misunderstanding perpetuated in the literature. It dates back to the observed relation that the rate of mass outflow increases roughly linearly with the source luminosity (for a recent plot see Wu et al. 2005). If the outflow rate scales with the disk accretion rate and the luminosity varies with some power of the stellar mass (e.g., square), then we obtain ${\rm \dot{M}_* \sim M_*^2}$ (Zhang 2005). However, it is a leap of faith to interpret this relation as one of instantaneous quantities rather than one regarding average quantities (the latter simply means that the formation of more massive stars requires higher average accretion rates, whereas the former would imply that the accretion rate keeps increasing with increasing central stellar mass). By completely ignoring the contribution of the accretion luminosity in this scenario, one implicitly assumes that a high luminosity implies a high core mass, as opposed to a high luminosity likely indicating that the accretion rate onto a lower mass (proto)star is high! A problem with an ever increasing accretion rate is not only how to suddenly stop the accretion, but also the observation that outflow activity is strongest during the earliest phases of star formation rather than during the final phases. In fact, the same problem exists for the turbulent core scenario of massive star formation (McKee & Tan 2003) where the accretion rate is predicted to increase linearly with time, implying that the most vigorous outflows should occur toward the end of accretion rather than at the beginning. Binary Statistics and the Most Massive Binary Systems ----------------------------------------------------- The first spectroscopic binary survey among O stars (V$\le$7) was conducted by Garmany, Conti & Massey (1980). They found that 24 out of their 67 O stars are spectroscopic binaries (36%). Some 15 are double-lined and 4 are single-lined, mostly with 1–10 day periods; the remaining 5 O stars are binary candidates. A later important work is the optical speckle survey of O stars carried out by Mason et al. (1998). This is a magnitude-limited sample (V$<$8) of 227 bright O stars, observed at 0.1arcsec spatial resolution (15 new detections). In addition, the previously known spectroscopic binaries were included in a comprehensive analysis. We summarize their results as follows. Besides the 50 known spectroscopic binary systems with periods less than 30 days and typical mass ratios of 0.5–1.0, there is a similar fraction of wide visual binaries (42 systems with probable orbital motions, with periods larger than 100–1000yr, and typical mass ratios from 0.5 to 0.1). Clearly the distribution of mass ratios is different for wide binaries and for close binaries. However there is a huge gap in orbital periods between close and wide binaries. This is a selection effect: Such binaries have been hard to detect with past techniques. Interferometry will be able to find such binaries in the near future. Mason et al. (1998) estimate that the total O star (spectroscopic + visual) binary frequency is about 60%, but they suggest that when the above gap is filled, the frequency may well be 100%. They also note that most binaries occur in clusters and associations and that binaries are less common among field stars and especially among runaway stars. 10–25% of all O stars are runaway stars, whereas among the B stars this fraction is only about 2% (Gies & Bolton 1986). The fact that the O-star runaways are 10 times more common than the B-star runaways implies that the secondaries of O-star primaries (which are the ones that get ejected, Leonhard & Duncan 1990) should also preferentially be O stars. That is, many O–O binaries must have mass ratios skewed to unity (Clarke & Pringle 1992), in accordance with the observed high frequency of double-lined spectroscopic binaries among O-type stars in young clusters. However, the binary frequency for O-type runaways is generally lower than that in young clusters (Mason et al. 1998). A possible explanation could be that the runaway recoil process produces a tight binary and a single massive star ejected in almost opposite directions, thus predicting a spectroscopic binary frequency for runaway O stars that is half that of young O-star clusters. Wolf-Rayet (WR) stars are evolved massive stars and, strictly speaking, do not belong to our main-sequence statistics. On the other hand, their progenitors were massive stars that lost much of their mass because of strong stellar winds. The mass loss rate and terminal wind speed depend on metallicity; more metal-rich stars lose more mass and lose it more quickly. The initial masses of WR stars cannot easily be determined from observations alone. Many new WR star binaries as well as mixed WR-OB star pairs have recently been discovered. The important question is whether the binary frequency of WR binaries and WR-OB star pairs depends on metallicity. We leave a discussion of WR-star binary statistics and related topics to Crowther (2007, in this volume). The O3 stars HD 93129A and HD 93205 in the Carina young star clusters Trumpler 14 and 16 were once thought to be the most massive single stars in the Galaxy (Taresch et al. 1997, Antokhina et al. 2000). In the case of HD 93129A this claim is based on the detailed non-local thermodynamic equilibrium analysis of the UV and hydrogen and helium optical spectrum of this O-type supergiant (T$_{\rm eff}$=52,000K). This analysis implies a very high bolometric luminosity, 2.5$\times$10$^6$L$_\odot$. The extreme dynamical stellar wind properties ($\dot {\rm M}$=-20M$_\odot$Myr$^{-1}$, v=3200kms$^{-1}$) can be used to infer a stellar mass of 130$\pm$20M$_\odot$ (Taresch et al. 1997). However, Walborn (2003) and Nelan et al. (2004), using the Fine Guidance Sensor on HST, recently resolved HD 93129A into two components with a separation of 60mas or 165AU and a visual magnitude difference of 0.5mag. This corresponds to component masses of about 80 and 50M$_\odot$. It is unclear how the new stellar parameters can be consistent with the high luminosity and the extreme wind properties, but this example shows strikingly how we can be misled in our conclusions if the binary nature of massive stars is unrecognized. As for HD 93205, this object consists of an O3V and an O8V noneclipsing pair with an orbital period of about 6 days. Because the inclination angle of the orbit is unknown, we only know that M$_1 \cdot \sin^3{\rm i}$=29M$_\odot$ and M$_2\cdot\sin^3{\rm i}$=13M$_\odot$ (Antokhina et al. 2000). An O8V star should correspond to about 20–25M$_\odot$, so one can tentatively infer a primary mass M$_1$=50–60M$_\odot$. Morrell et al. (2001) suggest a spectral type O3.5 V and a mass of 48M$_\odot$ for the primary of HD 93205, making it one of the earliest main-sequence stars in the Galaxy at this time. There is at least one other massive double-lined spectroscopic binary in the Trumpler 16 cluster (Tr 16-110 with P=3.5 days), as well as two massive eclipsing binaries (Tr 16-1 and Tr 16-104) with P=2.2 days (Rauw, personal communication). Also in the same cluster, there is one of the most massive WR stars in the Galaxy (WR 22, HD 92740), which is the primary of a spectroscopic binary with P=80 days. Its current mass is 45M$_\odot$ (Schweickhardt et al. 1999), corresponding to a ZAMS mass $\sim$80M$_\odot$ (Rauw, personal communication). We refer to Walborn et al. (2002) for further discussion of the most massive stars known (O2V stars). These include Cyg OB2-22A and Pismis 24-1, with masses in excess of 100M$_\odot$ and, up until recently, believed to be single; however, the latter has now been resolved by HST as a visual triple system (see Maíz-Apellániz et al. 2007). A similar case was LBV 1806–20: Eikenberry et al. (2004) determined its mass to be close to 200M$_\odot$, but Figer, Najarro & Kudritzki (2004) dethroned it, finding that it is most likely a spectroscopic binary. Therefore, at present, the ’Pistol Star’ with an estimated initial mass of 200–250M$_\odot$ (Figer et al. 1998) in the Quintuplet Cluster near the Galactic Center is still considered to be the most massive star in the Galaxy. However, the most massive stars with dynamically measured masses reside in the double-lined spectroscopic, eclipsing binary WR20a: the primary and secondary component masses are 83M$_\odot$ and 82M$_\odot$, respectively, with an error bar of 5M$_\odot$. Rauw et al. (2004) originally discovered the spectroscopic binary nature of the object with an orbital period of about 3.7 days. Bonanos et al. (2004) obtained an eclipse light curve in the I-band and so could determine the inclination angle and hence the masses. Both objects are slightly evolved (SpT WN6ha), and still undergoing core hydrogen burning, but on their way from early O-type to WR stars – with no previous phase of Roche lobe overflow. The binary system lies in the Westerlund 2 cluster associated with the RCW 49 HII region, but interestingly not near the cluster center. Speculation has it that this most heavy close binary system was ejected from the cluster core due to a dynamical interaction. A spectacular [*Spitzer*]{}/IRAC image of the RCW 49 HII region and cluster can be found in the Infrared Legacy Gallery at IPAC, courtesy of E. Churchwell and NASA/JPL-Caltech. A glimpse of some other spectacular [*Spitzer*]{}/IRAC images of high-mass star forming regions is given in Brandl et al. (2005), including the embedded proto-OB association W49A. Additional striking examples of multiplicity among very massive stars can be found in the 30 Dor cluster in the LMC, where Massey, Penny & Vukovich (2002) have identified four tight, double-lined spectroscopic binaries, three of which are eclipsing systems (R136–38 is the most massive system, with component masses of 57 and 23M$_\odot$, while R136–42 is a close second with component masses of 40 and 33M$_\odot$ (the latter is actually a physical pair of O3V stars with a period of 2.89 days). Other massive stars in 30 Dor, including Melnick 34, are being monitored for radial velocity variations. Unlike the center of NGC 3603, which is dominated by a massive WR binary system, a recent, still unpublished near-IR integral field spectroscopic survey of the R136 central regions with the ESO-VLT has not identified any WR spectroscopic binaries (T. Moffat & O. Schnurr, personal communication). This result is contrary to the physical intuition that would lead us to expect the heaviest objects, i.e., massive binaries, to be located in the very center (mass segregation due to dynamical friction). The center of the R136 cluster that was once considered to host a supermassive star (Feitzinger et al. 1980; Cassinelli, Mathis & Savage 1981) actually consists of a group of eight massive stars within a projected radius of 0.4arcsec or equivalently 0.1pc (Weigelt & Baier 1985; Pehlemann, Hofmann & Weigelt 1992), corresponding to a central mass density in massive stars of 10$^5$M$_\odot$ per cubic parsec. We do not know how much mass there is in low-mass stars in the center, as two-body dynamical mass segregation could have led to a depletion of low-mass stars there. The density of massive stars alone may not be high enough at this time to enable a runaway stellar collision process, leading to the formation of an intermediate mass black hole, as envisaged by Portegies Zwart & McMillan (2002). It is enough, though, for the occasional collision between two massive stars, leading to the formation of a rapidly rotating massive object, possibly the progenitor for a gamma-ray burster (Fryer & Heger 2005, Zinnecker 2006b). Although the R136 cluster is quite dense in its center, more massive young globular clusters such as those seen in the Antennae merging galaxies may be denser still, and a runaway collision process cannot be excluded. The Universality of the Upper Initial Mass Function --------------------------------------------------- How can we understand the universality of the upper IMF? Wherever we look, the data seem to be compatible with a Salpeter power-law with a logarithmic slope of -1.35. Even population studies at moderate to high redshift are consistent with the assumption of a Salpeter power law (e.g., Baldry & Glazebrook 2003). In essence, two completely different schools of thought have attempted to explain the robustness of the Salpeter IMF despite various kinds of environmental factors that could be expected to change this distribution (such as metallicity, gas pressure, or the density of stellar systems, to name but a few). These are competitive accretion and random sampling of fractal clouds. As for competitive accretion, this has been discussed extensively in the review of Bonnell, Larson & Zinnecker (2007) and will not be repeated here. Suffice it to say that competitive accretion of protostars for cluster gas can explain the observed Salpeter stellar mass distribution for massive stars, if most massive stars form in dense clusters (Bonnell et al. 2001b, Klessen 2001a,b, Klessen & Burkert 2001; for a discussion of the onset and richness of clustering as a function of the most massive cluster star see Testi et al. 1997 and Testi, Palla & Natta 1999). If, however, a substantial fraction of massive stars forms in less dense OB associations (Garmany 1994; Clark, Bonnell & Zinnecker 2005 ) or in isolated places in the field (for example, where was Betelgeuse – a $\sim$20M$_\odot$ red supergiant – born?), then competitive accretion cannot be invoked, and another idea is needed (e.g. Lamers et al. 2002; Li, Klessen & Mac Low 2003). Elmegreen (1997) presented such an idea, essentially a geometrical model, with a physical icing on the (geometrical) cake. This model is based on a random sampling of mass in turbulent fractal interstellar clouds (see [**Figure 17**]{}). A star was assumed to get a fixed fraction of the gas mass of the cloud piece in which it formed. The mass distribution of the pieces in any hierarchy has a logarithmic slope of -1 (i.e., equal mass in equal logarithmic intervals), so the IMF would seem to end up with this slope, but the sampling rate for pieces was assumed to be proportional to the square root of the local gas density to mimic the dynamical processes that are involved. This local density depends on the level in the hierarchy according to the fractal scaling of density with size (because mass scales with size to a power equal to the fractal dimension, density scales with size to a power equal to the fractal dimension minus three). This density dependence means that lower mass regions are sampled more frequently, and the mass function slope steepens from -1 to -1.35, which is the Salpeter slope. The net result of this sampling is a mass function for model stars that is indistinguishable from the observed IMF for young clusters and OB associations. Another interesting attempt to connect star formation to fractal cloud structure was made by Henriksen (1991, see his figure 1), but the details differ from Elmegreen’s (1997) model. An important variant of the fractal molecular cloud IMF model is the concept in which the stellar mass function derives directly from the mass distribution of cloud cores. Recent observations show a surprising similarity between the scaled cloud core mass distribution and the stellar IMF (Alves, Lombardi & Lada 2007). Yet another alternative for explaining a power-law IMF is due to Basu & Jones (2004). They note that the power-law tail in the mass function of protostellar condensations and stars arises from the accretion of ambient cloud material on to the condensation, coupled with a nonuniform (exponential) distribution of accretion lifetimes (cf. radioactive decay). Thus, this model assumes that not all condensations accrete for the same time. If we start with protostars with a log-normal mass distribution (possibly a reasonable assumption expected from the central limit theorem, see Zinnecker 1985), this log-normal distribution develops a power-law tail at high masses if the accretion rate is directly proportional to the instantaneous mass of the accreting object and if the probability of stopping accretion is constant in time. The latter implies an exponential probability distribution of accretion timescales with a constant death rate. How general this random accretion model is and how it relates to the competitive accretion model above needs to be further explored (cf. Bate & Bonnell 2005). Needless to say that the characteristic mass scale of fragmentation and the IMF (a few tenths of a solar mass) is beyond the scope of this review, and we refer to Larson (1985, 2005) and also Whitworth, Boffin & Francis (1998) for an in-depth discussion of this issue. Finally, we refer to Klessen, Spaans & Jappsen (2007) for the first numerical hydrodynamical calculations of the characteristic stellar mass in starburst regions, predicting a top-heavy IMF. This is naturally explained as a consequence of the elevated thermal Jeans mass in the warmer and dustier starburst environment (including the Galactic Center). The Number of Accreting Massive Protostars in the Galaxy -------------------------------------------------------- In this subsection, we derive an estimate of high-mass accreting stellar objects in the Galaxy (note that we are trying to avoid the term protostars). To this end, we first calculate how much stellar mass is formed in the Galaxy over a time-interval t$_{\rm acc}$ corresponding to the accretion phase (200,000yr). Using a total Galactic star-formation rate (SFR) of 5M$_\odot$yr$^{-1}$ (Smith, Mezger & Biermann 1978, Diehl et al. 2006) and multiplying by 200,000yr, we obtain 10$^6$M$_\odot$. Of these, we assume a fraction f$\approx$0.10 ends up in massive stars (above 10M$_\odot$, say). The rest goes into intermediate- and low-mass stars according to a Salpeter (1955) field star IMF, with an effective lower mass limit of 0.1M$_\odot$ and upper mass limit of 100M$_\odot$ (see Smith, Mezger & Biermann 1978, their table 2). A more realistic low-mass field star IMF (Kroupa 2002, Chabrier 2003) would increase the fraction f by about a factor of 2, yet would decrease the SFR derived above by about the same factor, thus leaving the product f$\times$SFR approximately constant. The number N$_*$($>$M$_*$) of stellar progenitors in the Galaxy to be found in the accretion phase, which eventually attain a final mass above a given value M$_*$, can then be derived from the two equations (the first solving for N$_0$), $${\rm N_0 \int^{100}_{0.1} m^{-x} dm = f \times SFR \times \Delta t_{acc}}$$ $${\rm N_* (> m_*) = N_0 \int^\infty_{m_*} m^{-x} d log m}$$ Here m$_*$=M$_*$/M$_\odot$ is the dimensionless stellar mass and dN/dlog(m)=N$_0$m$^{\rm -x}$ is the power-law upper stellar IMF with a slope x (x=1.35 for a Salpeter IMF). Solving this equation results in the numbers N$_*$($>$M$_*$) given in [**Table 4**]{}. ------------------------------ --------------- --------------- --------------- [**M$_*$**]{} [**N$_*$**]{} [**N$_*$**]{} [**N$_*$**]{} $\left[{\rm M}_\odot\right]$ x=1.3 x=1.35 x=1.7 $> 10$ 6300 5400 1600 $> 20$ 2400 2000 480 $> 30$ 1300 1000 220 $> 50$ 480 390 74 ------------------------------ --------------- --------------- --------------- : Expected number N of stellar progenitors in the Galaxy to be found in the accretion phase$^a$, which eventually attain a final mass above M$_*$, for various logarithmic slopes x of the IMF$^b$ -- --------------------------------------------------------------------------- $^a$Assumed to last 200,000yr. $^b$Assuming a total Galactic star formation rate of 5M$_\odot$yr$^{-1}$. -- --------------------------------------------------------------------------- : Expected number N of stellar progenitors in the Galaxy to be found in the accretion phase$^a$, which eventually attain a final mass above M$_*$, for various logarithmic slopes x of the IMF$^b$ The table shows that we can expect about 1000 accreting massive stars in the Galaxy with masses in excess of 30M$_\odot$ (i.e., spectral types earlier than O6.5V, see [**Table 3**]{}). Assuming 10kpc as the reference radius of the Milky Way, we see that the average surface density of accreting early O-type massive stars is about 3objects/kpc$^2$. This demonstrates the rareness of massive protostars, but is in agreement with the few luminous and totally embedded objects known, such as BN/KL and Cep-A (see [**Table 5**]{} in the Appendix which provides a list of massive star forming regions within 1kpc from the Sun and some properties of the exciting stars). Is There a Maximum Stellar Mass Set by Star Formation? ------------------------------------------------------ In [**Section 2.5**]{} we reviewed the observational evidence for an upper stellar mass limit and concluded there was such a limiting mass somewhere in the range between 100 and 200M$_\odot$, at least in Population I systems. Here we present a few thoughts on the possible physics of the stellar upper mass limit. The first question is whether this limit is set by stellar stability considerations or star-formation theory. Ledoux (1941), and later Schwarzschild & Härm (1959), concluded from linear stability analysis of radial stellar pulsations that there is an upper mass limit of the order of 100M$_\odot$, corresponding to the existence of a vibrational instability owing to nuclear reactions. This refers to massive stars of Population I chemical composition, whereas for lower chemical abundances the unstable mass is higher. Appenzeller (1970) and Ziebarth (1970) concluded, from nonlinear calculations, that this vibrational instability (also known as $\epsilon$-instability) does not have a global disrupting effect, except perhaps in stars with masses greatly exceeding the critical mass. A more recent discussion can be found in Stothers & Chin (1993). These authors concluded that, for a metallicity equal to or greater than 0.02, no nuclear-induced pulsational instability developed for masses up to at least 150M$_\odot$. For a metallicity equal to 0.004, corresponding to the value of the Small Magellanic Cloud, the critical mass for such an instability was still very high, around 140M$_\odot$. Larson & Starrfield (1971) were the first to stress a possible upper mass limit resulting from star-formation theory. They came up with a value of 50–100M$_\odot$ as a larger and larger fraction of the growing mass is thrown out by radiation before the star reaches stellar conditions. Today, we realize that the opacity of dusty gas does not determine the upper mass limit, whereas the opacity of ionized gas (electron scattering and UV line blanketing) certainly may. The constraint imposed by dust opacity (see [**Figure 6**]{}) shows that a luminous object with an L/M ratio consistent with a main-sequence star M$_*$$>$100M$_\odot$ can still accrete material as long as the object appears to be cool ($<$50K). This allows material to fall onto the disk. Within the disk much of the radial radiation flux is diverted into the polar direction (flashlight effect). Provided angular momentum transfer is adequate, dusty material can thus flow radially inward until at some point the dust is destroyed (sublimation radius). Dust-free neutral gas has a lower opacity and it can continue to flow inward. Eventually, as the disk gas gets closer to the accreting star, the gas is ionized and the opacity increases again sharply. Disk models currently do not allow an accurate estimate of the opacity of the ionized disk gas close to the star; the lower limit to the opacity provided by electron scattering limits the maximum L/M to $\sim$6$\times$10$^4$, well in excess of that expected from a main-sequence star of M$_*$=200M$_\odot$. Because very massive stars also possess strong radiation-driven winds, it is clear that the opacity of ionized stellar gas is sufficiently high that radiative acceleration exceeds surface gravity. The upper mass limit could thus result from the fact that mass loss from the star matches or exceeds accretion. As the line-driven wind mass loss from massive stars is metallicity-dependent (Kudritzki 2002) the upper mass limit would be metallicity-dependent, too. However, if mass loss is by a continuum-driven instability, as suggested for $\eta$ Car and other Luminous Blue Variables (Smith & Owocki 2006), the upper mass limit would be independent of metallicity. It is unclear if stellar collisions and mergers in dense stellar systems can beat the opacity-limited accretion discussed above and form stars with masses beyond 200M$_\odot$. The merging of a binary star with two 100-M$_\odot$ components is conceivable (Zinnecker 1986, Bonnell & Bate 2005), doubling the single star upper mass limit. It is likely that a coalescence process would deplete the number density of stars to collide with near the center of a deep cluster potential, thus imposing a density-dependent limiting mass, especially for young clusters undergoing core collapse (Portegies Zwart et al. 2006). Interestingly, in this case, too, the upper mass limit would be independent of the heavy element abundance. We refer to Omukai & Palla (2003) for a theoretical prediction of the increased upper mass limit ($\sim$600M$_\odot$) in zero-metallicity, Population III conditions, albeit based on 1D models. Evolutionary Sequence --------------------- We can characterize the formation of massive stars by the following crude four-stage evolutionary sequence: CDMC $\longrightarrow$ HDMC $\longrightarrow$ DAMS $\longrightarrow$ FIMS. Here CDMC means cold dense massive core, HDMC means hot dense massive core, DAMS stands for disk-accreting main-sequence star, and FIMS denotes the final main-sequence star. The FIMS star is more evolved than the theoretical concept of the ZAMS star, which assumes no prior hydrogen consumption. Of course, the sequence needs a more detailed description. The CDMCs are starless turbulent gravitationally bound condensations; they have near-virial equilibrium and are either on the verge of collapsing or, indeed, already collapsing. In the center an intermediate-mass protostar forms that can heat up the dense massive core, turning it into a HDMC. At this stage, collimated jets and outflows first appear, traced by H$_2$O and later by methanol maser emission. The central star grows in mass primarily by disk accretion and quickly becomes a DAMS object powered more by hydrogen burning than by disk accretion. The collimated outflows become less collimated with widening opening angles. At this stage, the accretion disk starts to get photoionized and partly photoevaporated, giving rise to a gravitationally confined hypercompact HII region (HCHII) with broad hydrogen recombination lines. Eventually the star has accreted its final mass and most of the disk has been dissipated. At this point the ionizing radiation is no longer quenched by the accretion flow and can expand freely into the vicinity of the star, blending together with ionized bubbles from other similarly massive stars. This is when we speak of an ultracompact HII region (UCHII), often formed by a group of ionizing stars within a volume of (0.1pc)$^3$ (e.g., W3-IRS5). UCHII regions frequently exhibit OH maser 18-cm radio emission \[e.g., W3(OH)\], which can be used to obtain a magnetic field strength by measuring the Zeeman splitting of the 1667MHz line. Subsequently, the UCHII region keeps expanding and evolves into a compact radio HII region and finally into a normal diffuse optical HII region like the Orion Nebula. How this simple-minded picture changes when we deal with a massive starburst protocluster, like the progenitor of R136 with dozens of massive stars born in a small volume, is anybody’s guess. It is conceivable that competitive accretion takes place in a gravitationally bound dense HII region. It is even conceivable that competitive accretion turns into cooperative accretion (E. Keto, private communication) which means the ionized flow pulled in by the collective gravity of the cluster is redirected onto the less massive members of the cluster. Instead of being greedy and competing for additional mass with their siblings, the near-Eddington-limit massive stars start to deflect the inflow onto neighboring stars. By doing so, they aid the further growth of their lower mass neighbors. Thus, accretion becomes cooperative! Another possibility is the hierarchical merging of subclusters which naturally leads to prompt initial mass segregation in the resulting final cluster (McMillan, Vesperini & Portegies Zwart 2007). In this scenario, many but not all subclusters (gas+stars) merge with each other and sink to the center of gravity owing to the loss of kinetic energy in these highly inelastic collisions. There they form a dense stellar cluster of massive stars, which will expel the residual cluster gas. Some subclusters miss and won’t collide; these will form massive stars in the cluster outskirts. Such a core-halo structure is observed in many young clusters, including R136 (e.g., Moffat, Drissen & Shara 1994). Super OB associations like NGC 604 in M33 have more widely spread subclusters that are unable to merge, thus avoiding dense cluster formation (cf. Hunter 1995). WHY HIGH-MASS STAR FORMATION IS NOT A SCALED-UP VERSION OF LOW-MASS STAR FORMATION ================================================================================== In this section, we try to convince the reader that high-mass star formation is indeed different from low-mass star formation, and not merely a scaled-up version of star formation by disk accretion for higher accretion rates. Many complex new physical processes enter the scene for high-mass stars. To start with, radiative forces on gas and dust play little or no role in the build-up of low-mass, solar-type stars, whereas a substantial fraction of the luminosity of high-mass stars is emitted in ionizing radiation, which introduces new effects such as the photoevaporation of the star’s accretion disk and protostellar envelope. This dramatically limits late accretion and the final stellar mass. In addition, the ionizing photons can photoevaporate the disks of the neighboring lower mass stars (cf. the proplyds in the Orion Nebula). The nonionizing far-UV radiation will influence the massive stars’ molecular cloud environment by dissociating H$_2$ and CO molecules. This requires photons of about 11.2eV; even early-type B stars can produce these photons, but low-mass and intermediate-mass stars cannot. The radiative acceleration of dusty and gaseous matter also leads to radiation-driven bipolar winds, and ionizing radiation can escape through these wind-blown cavities (flashlight effect); however, the bipolar outflows from low-mass stars are generated by magneto-centrifugal forces. The second big difference between the formation of low-mass and high-mass stars is the fact that massive stars are practically born on the main sequence, whereas low-mass stars spend a considerable part of their youth as contracting premain-sequence objects (30Myr for a solar-mass star). A massive star, forming by accretional growth from an initially low-mass star with an accretion rate of 10$^{-4}$M$_\odot$yr$^{-1}$ begins central hydrogen burning after about 9M$_\odot$ have accumulated – 13M$_\odot$ for an accretion rate of 10$^{-3}$M$_\odot$yr$^{-1}$. For low-mass stars, circumstellar disk evolution proceeds during the whole extended premain-sequence phase, whereas for massive stars the disk lifetime is very short (less than 1Myr). When we see massive O stars close to the ZAMS (Walborn 2007), their disks have been dissipated (but see Kastner et al. 2006). It is only in the embedded phase that circumstellar CO-bandhead emission and hydrogen emission lines are detected (Bik et al. 2005, Blum 2005), indicating a dense neutral remnant disk with an ionized upper layer. Bipolar outflows are another strong indication of the existence of disks. A further significant difference between low- and high-mass star formation is the role of competitive accretion in protoclusters. Competitive accretion is far more important for high-mass stars than for low-mass stars. The latter can form directly by Jeans-type gravitational instability and turbulence-induced cloud fragmentation. The former must accrete large amounts of protocluster gas. Gravitational dynamics (N-body interactions) also have a much greater effect when massive stars are involved. This can best be seen when considering the dynamical ejection of members of massive multiple systems producing runaway OB stars. This phenomenon is largely absent for low-mass stars. It appears that massive stars, when they form in their own local subclusters, are always accompanied by a small group of lower mass stars, some of which stay bound to their big parent star even after subclusters merge. Thus the higher companion star fraction observed for massive stars compared to solar-type stars testifies to a more dynamic scenario for the heavy objects thanks to their higher than average gravitational attraction. Finally, massive stars have a much bigger influence on triggering new star formation in adjacent regions than low-mass stars. Massive stars provide external pressure in the form of expanding HII regions, stellar winds, and supernovae explosions. They are capable of sustaining sequential and self-propagating star formation (Gerola & Seiden 1978), a process that low-mass stars are incapable of. Through runaway OB stars, massive star formation can trigger further massive star formation over large (kpc) distances, an important feature in sustaining large-scale nuclear starbursts. We speculate that, if massive star formation is massively triggered, the individual collapse of massive cores is outside-in, instead of inside-out (Banerjee, Pudritz & Anderson 2006), and always magnetically supercritical (Shu, Adams & Lizano 1987). In other words, magnetic fields likely play a more passive role in massive star formation, whereas in low-mass stars this is opposite. In defense of the notion of a scaled-up formation picture, some arguments have been raised. Foremost is the observation that collimated outflows occur both in high-mass and low-mass young stars (Beuther et al. 2002, Davis et al. 2004). However, a closer look reveals that collimated outflows and jets do not occur in the most luminous sources, but only in sources with total luminosities up to 10$^4$L$_\odot$ (Shepherd 2005), with one or two exceptions (Garay et al. 2003, Rodríguez et al. 2005b). This implies that massive star formation can be a scaled-up version of low-mass star formation, but only up to early B stars. The outflow morphology does not scale further for O stars, which generate powerful, wide-angle, ionized winds, calling into question the relationship between outflow and accretion (Shepherd 2005). Another point that has been made relates to the IMF. It has been suggested (see the discussion in Zinnecker 2004) that there should be a feature (knee) in the IMF, at a critical mass where the intermediate-mass and high-mass star-formation processes diverge. However, this feature has not been seen and the upper IMF is a power-law with a constant slope (Kroupa 2002). It is unclear how to refute this argument, but the constant slope of the upper IMF may simply imply that stellar collisions and mergers are not a dominant factor in massive star formation, except for the most massive stars (Bonnell & Bate 2002) where statistical fluctuations can hide a slope change. Collisional growth of massive stars (Bonnell, Bate & Zinnecker 1998; Bally & Zinnecker 2005) also seems to be ruled out by recent observations of stellar rotation: There is a continuous behavior of the specific stellar spin angular momentum over the full range of stellar masses, i.e., J/M$\sim$M$^{0.3}$ between 0.2M$_\odot$ and 50M$_\odot$, indicative of a single stellar formation and angular momentum regulation mechanism (Wolff et al. 2006). However, a collisional process and stellar mergers are not ruled out for the most massive early O-type stars (M$>$50M$_\odot$). OUTLOOK: RELEVANT FUTURE OBSERVATIONS ===================================== Here we suggest a number of future observations that would help promote?advance our understanding of massive star formation. The list, however, is by no means complete. 1. IR-observations with an Extremely Large Telescope (ELT)\ (diameter 30–42m)\ As described in Zinnecker (2006a), such a powerful telescope could penetrate the dust extinction of ultracompact HII regions (A$_{\rm V}$=100–200mag) in the near-IR (A$_{\rm K}$=10–20) and see the stellar photospheres of massive stars, resolving very tight embedded clusters, such as W3-IRS5 (cf. Megeath, Wilson & Corbin 2005). There is hope that an ELT might have sharp enough vision to test the prediction (Zinnecker 2006b) that gamma-ray bursts occur in the centers of young massive protoglobular clusters. 2. Submillimeter observations with the [*Atacama Large Millimeter Array*]{}\ The spatial resolution of the [*Atacama Large Millimeter Array*]{} ([*ALMA*]{}) at 350 microns ($\sim$0.01arcsec, i.e., 50AU at 5kpc) would allow us to map dense molecular clumps, where massive stars are believed to form, in the dust continuum and see whether fragmentation is occurring (Dobbs, Bonnell & Clark 2005) or not (Krumholz 2006). [*ALMA*]{} should also be able to measure rotation curves of circumstellar disk gas around massive stars and thus measure their enclosed stellar mass (M. Krumholz, R.I. Klein & C.F. McKee, submitted). 3. Mid-infrared observations with the [*James Webb Space Telescope*]{}\ The [*James Webb Space Telescope*]{} will have unparalleled background-limited sensitivity to reveal very embedded populations around young massive stars in the making in near- and far-galactic star forming regions as well as regions like 30 Dor/R136 in the Magellanic Clouds. Progress may also be expected for understanding star formation in the Galactic Center Regions, including resolution of the Arches and Quintuplet clusters and the OB populations near Sgr A$^*$ (Figer 2003, Kim et al. 2006, Paumard et al. 2006, Martins et al. 2006). 4. Proper motion observations with [*Gaia*]{}\ Precise proper motion data of young clusters and OB associations are eagerly awaited, surpassing the limited milli-arcsec precision of Hipparcos (de Zeeuw et al. 1999). This will be crucial to understanding the dynamical nature of OB associations, such as the question of whether an OB association is just an expanding, dissolving massive star cluster or something of its own (Blaauw 1983, Briceño et al. 2007). Hence [*Gaia*]{} will tell us if the stars of an association expand from a common center or from several centers or from no center at all. [*Gaia*]{} will also provide us with a complete sample of field O stars and their parallaxes within a few kiloparsecs. 5. Far-IR observations with [*Herschel*]{}\ The immediate future will see the launch of ESA’s [*Herschel*]{} space observatory, whose Photodetector Array Camera & Spectrometer (PACS) instrument will make it possible to follow up on [*Spitzer*]{}/MIPS observations and to detect highly embedded massive protostellar sources down to a spatial resolution of 3 arcsec at 60 microns. Surveys with the Spectral and Photometric Imaging Receiver (SPIRE) will allow us to identify and characterize pre-collapse cores at an early evolutionary stage (Boss & Yorke 1995). PACS together with SPIRE can be used to better constrain the ’collect and collapse’ model of triggered massive star formation (Zavagno et al. 2005), discriminating between true triggering versus the compression of pre-existing clumps. 6. Interferometric near-IR observations with VLTI and LBTI\ Very high angular resolution observations with closure phases will give us interferometric images (and not just visibilities) of ultracompact HII regions and the stellar content of hot molecular cores. This will allow us to obtain a clearer picture of the maternities of massive stars (Garay 2005), including images of disks and jets as well as astrometric information, such as the stellar velocity dispersion, dynamics of the stars in protocluster centers, and the orbital motion of short-period embedded binary and multiple systems. 7. Long-term variability surveys The formation and evolution of massive stars is likely to involve a variety of cataclysmic or eruptive phenomena on short timescales (days to years), such as episodic FU Orionis-like accretion (Chini, private communication) and stellar mergers (Bally & Zinnecker 2005), or $\eta$ Car-like LBV eruptions (Smith & Owocki 2006) and even infrared supernovae in luminous starburst galaxies (Maiolino et al. 2002). Without continuous monitoring, these (embedded) events may go unnoticed. This calls for long-term variability compaigns, in the near-infrared, for example from a medium-sized survey telescope at Dome C in Antarctica. 8. “Observations” with a Numerical Telescope\ As numerical simulations of the star formation process become more refined in 3D resolution and the microphysics involved (magnetic fields, chemistry, ionization, heating, cooling, dust physics, etc.) the ability to accurately “observe” the evolution at various stages with respect to dust continuum features, line intensities and line profiles increases in importance. In other words, theoretical modeling will have to keep pace with the rapid development of observing capabilities as described above. FINAL SUMMARY ============= Massive star formation is not a simple scaled-up version of low-mass star formation, particularly when it comes to ZAMS O stars ($>$20M$_\odot$). The formation of early-type B stars, however, may still be considered a continuation of low-mass star formation. In terms of formation processes, the monolithic collapse and disk accretion model and the competitive accretion scenario are the two opposite ends of a continuum of cases in the accretion theory of O-type massive star formation. The coalescence process, introduced to circumvent the obstacle of radiation pressure, is no longer generally necessary but may still occur in exceptional circumstances, especially for very massive stars in the centers of very dense protoclusters or subclusters with strong initial mass segregation. The observed multiplicity and clustering of massive stars suggests complex N-body and gas dynamical interactions (tidal, drag, and capture effects) among the youngest stars or protostars that would contribute to the argument against a monolithic collapse of isolated massive protostellar cores. Massive stars are rare and correspond to the tail of a power-law stellar mass distribution. For a Salpeter IMF, the number of all newborn ZAMS OB stars ($>$8M$_\odot$) is a mere 10% of all stars in the mass range of 1–2M$_\odot$. At present, competitive accretion (i.e., runaway growth of a few objects from cluster gas) provides the best physical explanation of the high-mass IMF and its surprisingly universal Salpeter slope. The maximum stellar mass (around 150M$_\odot$), long believed to be a result of stellar instability (Eddington-Ledoux limit), might be due to the negative feedback caused by the increasingly destructive radiative erosion of massive accretion disks at increasingly higher stellar masses. This self-limiting star formation can, however, be beaten by stellar collisions under very special circumstances. Rapid external shock compression (i.e., supersonic gas motions) generating high column densities in less than a local free-fall time rather than slow quasi-static build-up of massive cores may be the recipe to set up the initial conditions for local and global bursts of massive star formation. Such bursts can both enhance or quench further massive star formation (positive and negative feedback), depending on environment (gas density and pressure), but such a discussion is beyond the scope of this review. The question of how feedback from massive stars can influence proto galaxy evolution and morphological types (cf. Sandage 1986, Elmegreen 1999, Silk 1997, 2005) should be the topic of a future Annual Reviews article. APPENDIX – NEARBY ($<$1kpc)\ MASSIVE STAR-FORMATION REGIONS ============================== In this Appendix, we attempt to give a complete table of star formation regions within 1kpc that contain at least one massive star (see [**Table 5**]{}). Some of these nearby regions of massive star formation are described in more detail by Bally et al. (2005). Moreover, the massive stellar content of nearby (distance $<$650pc) OB associations are disscussed in the classic Hipparcos paper by de Zeeuw et al. (1999), including Sco OB2, Vel OB2, Per OB2, and Cep OB2 (see their figures 1 and 29). We also should mention the Galactic O-star catalog of Maíz-Apellániz, Pérez & Mas-Hesse (2004), which includes spectral classification, optical-NIR photometry, multiplicity, and astrometric information, as well as cluster?association membership. (See also the living O-star Web-catalogue http://www-int.stsci.edu/ jmaiz/GOS/GOSmain.html). O stars in open cluster regions can also be selected from the combined photometric and astrometric membership analysis of 520 previously known (Kharchenko et al. 2004) and 130 newly detected (Kharchenko et al. 2005) open clusters. As a final highlight, we show here Westerlund 1, the most massive young star cluster in the Galaxy ([**Figure 18**]{}). [**Region**]{} [**Sources**]{} [**Distance**]{} [**Age**]{} [**L/SpT**]{} [**References**]{} ---------------- ----------------- ------------------ ------------- ----------------- --------------------------- OMC-1 BN/KL 450pc emb$^a$ 10$^5$L$_\odot$ Menten & Reid 1995 Greenhill et al. 2004 OMC-1S FIR4 450pc emb 10$^4$L$_\odot$ Schmid-Burgk et al. 1990 Smith et al. 2004 NGC1977 HD 37018 450pc 5Myr B1V Makinen et al. 1985 Howe et al. 1991 NGC1980 $\iota$ Ori 550pc 5Myr O9III Johnstone & Bally 2006 Piskunov et al. 2006 NGC2023 HD 37903 475pc $\sim$5Myr B1.5V Howe et al. 1991 Wyrowski et al. 2000 NGC2024 IRS2b 360pc emb O8V Bik et al. 2003 Lenorzer et al. 2004 $\sigma$ Ori $\sigma$ Ori AB 350pc $\sim$3Myr O9.5V Sanz-Forcada et al. 2004 Hernandez et al. 2007 Coll69 $\lambda$ Ori 400pc $\sim$6Myr O8III Dolan & Mathieu 2002 Barrado y Navascués 2005 S106 IRS4 600pc emb O9V Hodapp & Rayner 1991 Furuya et al. 1999 W40 OS1a, 2a, 3a 600pc obsc$^b$ O9V Smith et al. 1985 Vall[é]{}e & MacLeod 1994 : Star formation regions (within 1kpc) with massive stars (B2 and earlier) \[tab:regions\] [**Region**]{} [**Sources**]{} [**Distance**]{} [**Age**]{} [**L/SpT**]{} [**References**]{} ---------------- ----------------- ------------------ ------------- -------------------------- ------------------------- NGC1579 LkH$\alpha$101 700pc obsc B0.5V Barsony et al. 1991 Herbig et al. 2004 NGC2264 IRS1 800pc emb B2V Schwartz et al. 1985 Schreyer et al. 2003 Mon R2 IRS3 830pc emb B1V Carpenter et al. 1997 Preibisch et al. 2002 GGD 12-15 IRAS06084 830pc emb B0.5V Gomez et al. 1998 Gutermuth et al. 2005 S140 IRS1 900pc emb B2V Lester et al. 1986 Preibisch et al. 2001 IC 1396 HD 206267 800pc $\sim$3Myr O6.5V Schulz et al. 1997 de Zeeuw et al. 1999 Cep-OB3 Cep-A/HW2 725pc emb 2$\times$10$^4$L$_\odot$ Garay et al. 1996 Hartigan et al. 2000 Cyg-X AFGL 2591 1000pc emb 2$\times$10$^4$L$_\odot$ Tamura & Yamashita 1992 van der Tak et al. 2006 $^a$emb stands for embedded, i.e., a very young phase ($<$1Myr); sources optically invisible\ $^b$obsc stands for obscured, i.e., a young phase ($\sim$1Myr); sources optically visible but heavily dust extincted\ Note: AFGL 2591 may be at a distance of 1700pc (Schneider et al. 2006) and part of the Cyg OB2 association, one of the richest regions of massive stars in the Galaxy with $\sim$100 O stars, including a hidden very massive OB cluster, shown in [**Figure 18**]{}. Note that this massive star forming region is a source of significant diffuse $\gamma$ ray emission, i.e. the 1.8MeV line from the radioactive decay of $^{26}$Al (Plüschke et al. 2002) GLOSSARY ======== - Accretion\ gas accumulation of a star or protostar, increasing the mass of the object - Competitive accretion\ gas initially unbound to a star or protostar and moving with relative speed, v$_{\rm {rel}}$, w.r.t. to the object is added to it, i.e., to its gravitational sphere or column of influence (defined by an impact parameter 2GM/v$^2_{\rm {rel}}$) - Coalescence\ growth of the mass of a stellar or protostellar object by a physical collision and merger with another object - Merger\ inelastic collision between two stars or protostars leading to the amalgamation of the two bodies - Massive star\ star more massive than about 8M$_{\odot}$ that ends its life with a type II supernova (unless it is in a close binary system with mass transfer); a massive star on the main sequence has spectral type B3 or earlier. - OB star\ star of spectral type O ($>$16M$_{\odot}$) or type B (B$>$4M$_{\odot}$) - Ultracompact HII region\ small blob (typical density $10^4$cm$^{-3}$, size 0.1pc) of ionized gas emitting radio continuum radiation of high emission measure (Ln$_{\rm e}^{2}$=10$^7$pccm$^{-6}$) - Hypercompact HII region\ very small blob (density 10$^6$cm$^{-3}$, typical size 0.01pc) of ionized gas emitting radio continuum radiation of very high emission measure (Ln$_{\rm e}^{2}$=10$^{10}$pccm$^{-6}$) - Protostar\ object on its way to become a star, with more than half of its final mass still to be accumulated - Infrared Dark Cloud (IRDC)\ dense interstellar cloud (often filamentary) seen in absorption against background thermal IR emission; detected with [*ISO*]{}, [*MSX*]{}, and [*Spitzer*]{} - Core\ basic molecular cloud unit (small dense gas fragment, mass 10–100M$_\odot$, size $\sim$0.1pc) to form one or a few stars - Clump\ basic molecular cloud unit (large dense gas fragment, mass $\sim$1000–5000M$_\odot$, size $\sim$0.5pc) to form a young cluster of stars - Hot Molecular Core (HMC)\ dense warm (T$>$100K) molecular gas in emission (CO, NH$_3$, CH$_3$OH) heated from a protostar inside; a compact region (size $<$0.1pc) in a star forming molecular clump such as the BN/KL region in Orion - Orion Nebula Cluster (ONC)\ elongated 1$\times$2pc cluster of young stars, centered on the high-mass Trapezium stars in Orion; about 4000 cluster members with a total stellar mass of $\sim$1800M$_{\odot}$ - Trapezium Cluster\ The 4–5 stars in the center of the Orion nebula, $\theta^1$ C, A, B, D, E in order of decreasing luminosity, the main exciting/ionising source being $\theta^1$ C (O5.5V, ca. 35M$_{\odot}$) - 30 Doradus (30 Dor)\ prime example of a giant (200pc) extragalactic HII region, powered by $\sim$100 massive O stars stars in the NGC 2070 cluster (cluster mass some 10$^4$M$_{\odot}$, age $<$3Myr, cluster radius about 1arcmin or 15pc) - R136\ nominally the 1arcsec core of the 30 Dor nebula and HII region (Radcliffe object No. 136), but R136 also often designates the central dense core of 1pc (4arcsec) radius of the exciting star cluster NGC 2070 - Jeans mass\ critical mass that must be exceeded for a gas cloud to collapse dynamically, with the self-gravity of the gas cloud being opposed only by thermal pressure; if squeezed by an external pressure, the critical mass for the onset of collapse is called the Bonnor-Ebert mass. Other anisotropic forces such as rotation or magnetic fields can oppose the cloud’s self-gravity and thus impede star formation. - Free-fall timescale\ timescale it takes for a pressure-free cloud to collapse dynamically to a very small size (nominally a point) under its own self-gravity; the timescale depends on the average cloud gas density only (for number density n=10$^5$cm$^{-3}$, t$_{\rm {ff}}$=10$^5$yr). - Kelvin-Helmholtz (KH) timescale\ time it takes for a young stellar object or stellar core to radiate its thermal energy content, typically 10$^4$yr for young high-mass stars but 10$^7$yr for solar-type stars (t$_{\rm {KH}}$=GM$^2$/RL). Nuclear energy generation eventually offsets the thermal energy loss. - Premain-Sequence evolution\ While high-mass stars are born on the Main-Sequence (i.e. they reach hydrogen burning while still accreting matter onto their stellar surface during their deeply embedded phase), the collapse of a low-mass star does not immediately lead to a hydrogen-burning star, rather to a quasi-static object that keeps contracting in radius by factors of a few before hydrogen burning ignites in the center. During this contraction, the young stellar object has a time-dependent radius, effective temperature, and luminosity that can be calculated by pre-MS evolutionary models. - ZAMS\ zero-age Main Sequence: a star that has contracted enough to start hydrogen burning in its central region - Initial Mass Function (IMF)\ mass distribution of (single) stars at birth introduced by Salpeter in 1955 <!-- --> - [*Spitzer*]{}\ NASA mid-infrared space telescope of 85cm diameter launched in 2003 - [*MSX*]{}\ mid-infrared satellite explorer with an aperture of 20cm diameter - [*Chandra*]{}\ 0.5–10keV NASA X-ray telescope in space with particularly good imaging quality comparable to optical seeing (angular resolution $\sim$1arcsec) - [*HST*]{}\ [*Hubble Space Telescope*]{}, 2.4m aperture, workhorse for high angular resolution space astronomy in UV, optical, and near-IR; in orbit since 1990 [**Acknowledgments**]{} The present review benefited greatly from insightful discussions with John Bally, Henrik Beuther, Adriaan Blaauw, Ian Bonnell, Bernhard Brandl, Wolfgang Brandner, Anthony Brown, Ed Churchwell, Paul Crowther, Melvyn Davies, Lise Deharveng, Bruce Elmegreen, Neal Evans, Simon Glover, Wolf-Rainer Hamann, Thomas Henning, George Herbig, David Hollenbach, Jim Jackson, Eric Keto, Spyros Kitsionas, Ralf Klessen, Pavel Kroupa, Mark Krumholz, Richard Larson, Hendrik Linz, Andre Maeder, Eric Mamajek, Jorge Melnick, Karl Menten, Vincent Minier, Tony Moffat, Dieter Nürnberger, Tolya Piskunov, Simon Portegies Zwart, Thomas Preibisch, Gregor Rauw, Hugues Sana, Fernando Selman, Joe Silk, Nathan Smith, Bringfried Stecklum, Jonathan Tan, Neal Turner, Virpi Niemela (deceased), Nolan Walborn, and Gerd Weigelt. We thank Geoffrey Burbidge and Allan Sandage for their vote of confidence, and Ewine van Dishoeck and particularly John Kormendy for their critical reading of the manuscript. Portions of this research was conducted at the Jet Propulsion Laboratory, California Institute of Technology, which is supported by the National Aeronautics and Space Administration (NASA). The assistance of U. Hanschur at AIP Potsdam was indispensible in completing this review and is gratefully acknowledged. H.Z. would like to thank Andrea Lagarini and the Magalhaes family in Rio de Janeiro and ESO in Santiago de Chile for their hospitality during the very final stages of writing this review. This collaboration has a long history, dating back to 1990–1995 when the authors were colocated at the University of Würzburg. Aarseth SJ. 2003. [*Gravitational N-Body Simulations.*]{} Cambridge, UK: Cambridge Univ. Press Abel T, Bryan GL, Norman ML. 2000. [*Ap. J.*]{} 540:39–44 Adams FC, Fatuzzo M. 1996. [*Ap. J.*]{} 464:256–71 Allen DA, Burton MG. 1993. [*Nature.*]{} 363:54–56 Allen C, Poveda A, Hernández-Alcántara A. 2004. [*Rev. Mex. Astron. Astrofis. Ser. Conf.*]{} 21:195–99 Alves J, Lombardi M, Lada CJ. 2007. [*Astron. Astrophys.Lett.*]{} 462:L17–21 Antokhina EA, Moffat AFJ, Antokhin II, Bertrand J-F, Lamontagne R. 2000. [*Ap. J.*]{} 529:463–76 Apai D, Bik A, Kaper L, Henning T, Zinnecker H. 2007. [*Ap. J.*]{} 655:484–91 Appenzeller I. 1970. [*Astron. Astrophys.*]{} 9:216–20 Arce HG, Shepherd D, Gueth F, Lee C-F, Bachiller R, et al. 2007. See Reipurth et al. 2007, pp. 245–60 Balbus SA. 2003. [*Annu. Rev. Astron. Astrophys.*]{} 41:555–97 Balbus SA, Hawley JF. 1991. [*Ap. J.*]{} 376:214–22 Baldry IK, Glazebrook K. 2003. [*Ap. J.*]{} 593:258–71 Ballesteros-Paredes J, Hartmann L, Vázquez-Semadeni E. 1999. [*Ap. J.*]{} 527:285–97 Bally J. 2002. See Crowther 2002, pp. 219–33 Bally J, Moeckel N, Throop H. 2005. In [*Chondrites and the Protoplanetary Disk,*]{} ed. AN Krot, ERD Scott, B Reipurth. [*ASP Conf. Ser.*]{} 341:81–106 Bally J, Cunningham N, Moeckel N, Smith N. 2005. See Cesaroni et al. 2005a, pp. 12–22 Bally J, Zinnecker H. 2005. [*Astron. J.*]{} 129:2281–93 Banerjee R, Pudritz RE, Anderson DW. 2006. [*MNRAS*]{} 373:1091–06 Banerjee R, Pudritz RE. 2007. [*Ap. J.*]{} 660:479–88 Baraffe I, Chabrier G, Allard F, Hauschildt PH. 2002. [*Astron. Astrophys.*]{} 382:563–72 Barrado y Navascués D. 2005. [*Rev. Mex. Astron. Astrofis. Ser. Conf.*]{} 24:217–18 Barsony M, Schombert J, Kis-Halas K. 1991. [*Ap. J.*]{} 379:221–31 Bastian N, Goodwin SP. 2006. [*MNRAS Lett.*]{} 369:L9–13 Basu S, Jones CE. 2004. [*MNRAS Lett.*]{} 347:L47–51 Bate MR. 2000. [*MNRAS*]{} 314:33–53 Bate MR, Bonnell IA, Price NM. 1995. [*MNRAS*]{} 277:362–76 Bate MR, Burkert A. 1997. [*MNRAS*]{} 288:1060–72 Bate MR, Bonnell IA. 1997. [*MNRAS*]{} 285:33–48 Bate MR, Bonnell IA, Bromm V. 2002. [*MNRAS*]{} 336:705–13 Bate MR, Bonnell IA. 2005. [*MNRAS*]{} 356:1201–21 Begelman MC. 2001. [*Ap. J.*]{} 551:897–906 Behrend R, Maeder A. 2001. [*Astron. Astrophys.*]{} 373:190–98 Belkus H, Van Bever J, Vanbeveren D. 2007. [*Ap. J.*]{} 659:1576–81 Benjamin RA, Churchwell E, Babler BL, Bania TM, Clemens DP, et al. 2003. [*Publ. Astron. Soc. Pac.*]{} 115:953–64 Bertoldi F, McKee CF. 1992. [*Ap. J.*]{} 395:140–57 Beuther H, Schilke P, Sridharan TK, Menten KM, Walmsley CM, Wyrowski F. 2002. [*Astron. Astrophys.*]{} 383:892–904 Beuther H, Shepherd DS. 2005. In [*Cores to Clusters: Star Formation with Next Generation Telescopes,*]{} ed. MS Nanda Kumar, M Tafalla, P Caselli. [*Astrophys. Space Sci. Libr.*]{} 324:105–19 Beuther H, Sridharan TK, Saito M. 2005. [*Ap. J. Lett.*]{} 634:L185–88 Beuther H, Churchwell E, McKee CF, Tan JC. 2007. See Reipurth et al. 2007, pp. 165–80 Bik A, Lenorzer A, Kaper L, Comerón F, Waters LBFM. 2003. [*Astron. Astrophys.*]{} 404:249–54 Bik A, Kaper L, Hanson MM, Smits M. 2005. [*Astron. Astrophys.*]{} 440:121–37 Binney J, Tremaine S. 1987. [*Galactic Dynamics.*]{} Princeton, NJ: Princeton Univ. Press. 755 pp. Blaauw A. 1961. [*Bull. Astron. Inst. Neth.*]{} 15:265–90 Blaauw A. 1964. [*Annu. Rev. Astron. Astrophys.*]{} 2:213–46 Blaauw A. 1983. [*Irish Astron. J.*]{} 16:141–47 Blaauw A. 1991. In [*The Physics of Star Formation and Early Stellar Evolution,*]{} ed. CJ Lada, ND Kylafis. [*NATO ASI Ser. C,*]{} 342:125–54 Dordrecht: Kluwer Black DC, Bodenheimer P. 1975. [*Ap. J.*]{} 199:619–32 Blaes O, Socrates A. 2003. [*Ap. J.*]{} 595:509–37 Blandford RD, Payne DG. 1982. [*MNRAS*]{} 199:883–903 Blum RD. 2005. See Cesaroni et al. 2005a, pp. 216–24 Bodenheimer P. 1995. [*Annu. Rev. Astron. Astrophys.*]{} 33:199–238 Bodenheimer P, Laughlin GP, Różyczka M, Yorke HW. 2007. [*Numerical Methods in Astrophysics. An Introduction.*]{} New York: Taylor & Francis Boffin HMJ, Watkins SJ, Bhattal AS, Francis N, Whitworth AP. 1998. [*MNRAS*]{} 300:1189–204 Bonanos AZ. 2007. [*Astron. J.*]{} 133:2696–708 Bonanos AZ, Stanek KZ, Udalski A, Wyrzykowski L, Żebruń K, et al. 2004. [*Ap. J. Lett.*]{} 611:L33–36 Bonnell I, Bastien P. 1992. [*Ap. J.*]{} 401:654–66 Bonnell IA, Bate MR, Clarke CJ, Pringle JE. 1997. [*MNRAS*]{} 285:201–08 Bonnell IA, Bate MR, Zinnecker H. 1998. [*MNRAS*]{} 298:93–102 Bonnell IA, Davies MB. 1998. [*MNRAS*]{} 295:691–98 Bonnell IA, Bate MR, Clarke CJ, Pringle JE. 2001a. [*MNRAS*]{} 323:785–94 Bonnell IA, Clarke CJ, Bate MR, Pringle JE. 2001b. [*MNRAS*]{} 324:573–79 Bonnell IA, Bate MR. 2002. [*MNRAS*]{} 336:659–69 Bonnell IA, Bate MR, Vine SG. 2003. [*MNRAS*]{} 343:413–18 Bonnell IA, Vine SG, Bate MR. 2004. [*MNRAS*]{} 349:735–41 Bonnell IA, Bate MR. 2005. [*MNRAS*]{} 362:915–20 Bonnell IA, Bate MR. 2006. [*MNRAS*]{} 370:488–94 Bonnell IA, Larson RB, Zinnecker H. 2007. See Reipurth et al. 2007, pp. 149–64 Bosch G, Selman F, Melnick J, Terlevich R. 2001. [*Astron. Astrophys.*]{} 380:137–41 Boss AP, Yorke HW. 1995. [*Ap. J. Lett.*]{} 439:L55–58 Brandl B, Sams BJ, Bertoldi F, Eckart A, Genzel R, et al. 1996. [*Ap. J.*]{} 466:254–73 Brandl B, Brandner W, Eisenhauer F, Moffat AFJ, Palla F, Zinnecker H. 1999. [*Astron. Astrophys. Lett.*]{} 352:L69–72 Brandl BR, Townsley LK, Churchwell E, Carey S, Zinnecker H, et al. 2005. See Cesaroni et al. 2005a, pp. 311–17 Brandner W, Grebel EK, Barbá RH, Walborn NR, Moneti A. 2001. [*Astron. J.*]{} 122:858–65 Brandner W, Clark JS, Stolte A, Waters R, Neguerela I, Goodwin SP. 2007. [*Astron. Astrophys.*]{} In press Briceño C, Preibisch T, Sherry WH, Mamajek EA, Mathieu RD, et al. 2007. See Reipurth et al. 2007, pp. 345–60 Bromm V, Larson RB. 2004. [*Annu. Rev. Astron. Astrophys.*]{} 42:79–118 Bronfman L, Casassus S, May J, Nyman L-. 2000. [*Astron. Astrophys.*]{} 358:521–34 Brown A. 2001. [*Astron. Nachr.*]{} 322:43–46 Burrows CJ, Stapelfeldt KR, Watson AM, Krist JE, Ballester GE, et al. 1996. [*Ap. J.*]{} 473:437–51 Burton MG, Hill T, Longmore SN, Purcell CR, Walsh AJ. 2005. See Cesaroni et al. 2005a, pp. 157–62 Camenzind M. 1990. [*Rev. Mod. Astron.*]{} 3:234–65 Carpenter JM, Snell RL, Schloerb FP, Skrutskie MF. 1993. [*Ap. J.*]{} 407:657–79 Carpenter JM, Meyer MR, Dougados C, Strom SE, Hillenbrand LA. 1997. [*Astron. J.*]{} 114:198–221 Cassinelli JP, Mathis JS, Savage, BD. 1981. [*Science*]{} 212:1497–501 Cassinelli JP, Churchwell EB, ed. 1993. [*Massive Stars: Their Lives in the Interstellar Medium, ASP Conf. Ser.,*]{} Vol. 35. San Francisco: ASP Cesaroni R. 2005. [*Ap. Space Sci.*]{} 295:5–17 Cesaroni R, Felli M, Testi L, Walmsley CM, Olmi L. 1997. [*Astron. Astrophys.*]{} 325:725–44 Cesaroni R, Felli M, Jenness T, Neri R, Olmi L, et al. 1999. [*Astron. Astrophys.*]{} 345:949–64 Cesaroni R, Felli M, Churchwell E, Walmsley M, ed. 2005a. [*Massive Star Birth: A Crossroads of Astrophysics, IAU Symp. 227.*]{} Cambridge: Cambridge University Press Cesaroni R, Neri R, Olmi L, Testi L, Walmsley CM, Hofner P. 2005b. [*Astron. Astrophys.*]{} 434:1039–54 Cesaroni R, Galli D, Lodato G, Walmsley CM, Zhang Q. 2007. See Reipurth et al. 2007, pp. 197–212 Chabrier G. 2003. [*Publ. Astron. Soc. Pac.*]{} 115:763–95 Chini R, Hoffmeister V. Kimeswenger S, Nielbock M, Nürnberger D, et al. 2004. [*Nature*]{} 429:155–57 Churchwell E. 2002. [*Annu. Rev. Astron. Astrophys.*]{} 40:27–62 Churchwell E, Povich MS, Allen D, Taylor MG, Meade MR, et al. 2006. [*Ap. J.*]{} 649:759–78 Clark JS, Negueruela I, Crowther PA, Goodwin SP. 2005. [*Astron. Astrophys.*]{} 434:949–69 Clark PC, Bonnell IA. 2005. [*MNRAS*]{} 361:2–16 Clark PC, Bonnell IA, Zinnecker H. 2005. [*MNRAS*]{} 359:809–18 Clarke CJ, Pringle JE. 1992. [*MNRAS*]{} 255:423–30 Clarke CJ, Edgar RG, Dale JE. 2005. See Corbelli et al. 2005, pp. 449–54 Comerón F, Schneider N, Russeil D. 2005. [*Astron. Astrophys.*]{} 433:955–77 Comerón F, Pasquali A. 2007. [*Astron. Astrophys. Lett.*]{} 467:L23–27 Corbelli E, Palla F, Zinnecker H. ed. 2005. [*The Initial Mass Function 50 Years Later, Astrophys. Space Sci. Libr. 327.*]{} Dordrecht: Springer Costero R, Echevarria J, Richer MG, Poveda A, Li W. 2006. [*IAU Circ.*]{} 8669 Crowther PA, ed. 2002. [*Hot Star Workshop III: The Earliest Stages of Massive Star Birth, ASP Conf. Proc.,*]{} Vol. 267. San Francisco: ASP Crowther PA. 2007. [*Annu. Rev. Astron. Astrophys.*]{} 45:In press Crutcher RM. 1999. [*Ap. J.*]{} 520:706–13 Crutcher RM. 2005. See Cesaroni et al. 2005a, pp. 98–107 Crutcher RM, Troland TH. 2007. See Elmegreen, Palous. 2007, pp. 141–47 Dale JE, Davies MB. 2006. [*MNRAS*]{} 366:1424–36 Dale JE, Bonnell IA, Whitworth AP. 2007. [*MNRAS*]{} 375:1291–98 Dale JE, Clark PC, Bonnell IA. 2007. [*MNRAS*]{} 377:535–44 Davies MB, Bate MR, Bonnell IA, Bailey VC, Tout CA. 2006. [*MNRAS*]{} 370:2038–46 Davis CJ, Varricatt WP, Todd SP, Ramsay Howat SK. 2004. [*Astron. Astrophys.*]{} 425:981–95 De Becker M, Rauw G, Manfroid J, Eenens P. 2006. [*Astron. Astrophys.*]{} 456:1121–30 De Buizer JM. 2003. [*MNRAS*]{} 341:277–98 De Buizer JM, Minier V. 2005. [*Ap. J. Lett.*]{} 628:L151–54 de Geus EJ. 1992. [*Astron. Astrophys.*]{} 262:258–70 de Wit WJ, Testi L, Palla F, Vanzi L, Zinnecker H. 2004. [*Astron. Astrophys.*]{} 425:937–48 de Wit WJ, Testi L, Palla F, Zinnecker H. 2005. [*Astron. Astrophys.*]{} 437:247–55 de Zeeuw PT, Hoogerwerf R, de Bruijne JHJ, Brown AGA, Blaauw A. 1999. [*Astron. J.*]{} 117:354–99 Diehl R, Halloin H, Kretschmer K, Lichti GG, Schönfelder V, et al. 2006. [*Nature*]{} 439:45–47 Dobbs CL, Bonnell IA, Clark PC. 2005. [*MNRAS*]{} 360:2–8 Dolan CJ, Mathieu RD. 2002. [*Astron. J.*]{} 123:387–403 Donati J-F, Babel J, Harries TJ, Howarth ID, Petit P, Semel M. 2002. [*MNRAS*]{} 333:55–70 Donati J-F, Howarth ID, Bouret J-C, Petit P, Catala C, Landstreet J. 2006. [*MNRAS Lett.*]{} 365:L6–10 Drissen L, Moffat AFJ, Walborn NR, Shara MM. 1995. [*Astron. J.*]{} 110:2235–41 Duchêne G, Simon T, Eisloffel J, Bouvier J. 2001. [*Astron. Astrophys.*]{} 379:147–61 Durisen RH. 2001. See Zinnecker, Mathieu. 2001, pp. 381–90 Edgar R, Clarke C. 2004. [*MNRAS*]{} 349:678–86 Edwards S, Strom SE, Hartigan P, Strom KM, Hillenbrand LA, et al. 1993. [*Astron. J.*]{} 106:372–82 Egan MP, Shipman RF, Price SD, Carey SJ, Clark FO, Cohen M. 1998. [*Ap. J. Lett.*]{} 494:L199–202 Eikenberry SS, Matthews K, LaVine JL, Garske MA, Hu D, et al. 2004. [*Ap. J.*]{} 616:506–18 Eisenhauer F, Quirrenbach A, Zinnecker H, Genzel R. 1998. [*Ap. J.*]{} 498:278–92 Elmegreen BG. 1997. [*Ap. J.*]{} 486:944–54 Elmegreen BG. 1998. In [*The Origin of Stars and Planetary Systems,*]{} ed. CE Woodward, JM Shull, HA Thronson Jr. [*ASP Conf. Ser.,*]{} 148:150–83 Elmegreen BG. 1999. [*Ap. J.*]{} 517:103–07 Elmegreen BG. 2000. [*Ap. J.*]{} 539:342–51 Elmegreen BG, Lada CJ. 1977. [*Ap. J.*]{} 214:725–41 Elmegreen BG, Palous J. ed. 2007. [*Triggered Star Formation in a Turbulent ISM, IAU Symp. 237.*]{} Cambridge: Cambridge University Press Evans NJ. 1999. [*Annu. Rev. Astron. Astrophys.*]{} 37:311–62 Evans NJ. 2005. See Cesaroni et al. 2005a, pp. 443–48 Fabian AC, Pringle JE, Rees MJ. 1975. [*MNRAS*]{} 172:15–18 Feitzinger JV, Schlosser W, Schmidt-Kaler T, Winkler C. 1980. [*Astron. Astrophys.*]{} 84:50–59 Ferreira J, Pelletier G. 1995. [*Astron. Astrophys.*]{} 295:807–32 Figer DF. 2003. In [*A Massive Star Odyssey: From Main Sequence to Supernova,*]{} ed. K van der Hucht, A Herrero, E César. [*IAU Symp.*]{} 212:487–96. San Francisco: ASP Figer DF. 2005. [*Nature*]{} 434:192–94 Figer DF, Najarro F, Morris M, McLean IS, Geballe TR, et al. 1998. [*Ap. J.*]{} 506:384–404 Figer DF, Najarro F, Kudritzki RP. 2004. [*Ap. J. Lett.*]{} 610:L109–12 Fromang S, Balbus S, Terquem C, de Villiers JP. 2004. [*Ap. J.*]{} 616:364–75 Fryer CL, Heger A. 2005. [*Ap. J.*]{} 623:302–13 Furuya RS, Kitamura Y, Saito M, Kawabe R, Wootten HA. 1999. [*Ap. J.*]{} 525:821–31 Gammie CF. 1998. [*MNRAS*]{} 297:929–35 Garay G. 2005. See Cesaroni et al. 2005a, pp. 86–91 Garay G, Ramirez S, Rodríguez LF, Curiel S, Torrelles JM. 1996. [*Ap. J.*]{} 459:193–208 Garay G, Lizano S. 1999. [*Publ. Astron. Soc. Pac.*]{} 111:1049–87 Garay G, Brooks KJ, Mardones D, Norris RP. 2003. [*Ap. J.*]{} 587:739–47 Garay G, Faúndez S, Mardones D, Bronfman L, Chini R, Nyman L-. 2004. [*Ap. J.*]{} 610:313–19 García B, Mermilliod JC. 2001. [*Astron. Astrophys.*]{} 368:122–36 Garmany CD. 1994. [*Publ. Astron. Soc. Pac.*]{} 106:25–37 Garmany CD, Conti PS, Massey P. 1980. [*Ap. J.*]{} 242:1063–76 Garmany CD, Conti PS, Chiosi C. 1982. [*Ap. J.*]{} 263:777–90 Genzel R, Stutzki J. 1989. [*Annu. Rev. Astron. Astrophys.*]{} 27:41–85 Gerola H, Seiden PE. 1978. [*Ap. J.*]{} 223:129–35 Gies DR. 1987. [*Ap. J. Suppl.*]{} 64:545–63 Gies DR, Bolton CT. 1986. [*Ap. J. Suppl.*]{} 61:419–54 Glover SCO. 2005. [*Space Sci. Rev.*]{} 117:445–508 Glover SCO, Mac Low M-M. 2007. [*Ap. J.*]{} 659:1317–37 Gomez Y, Lebron M, Rodríguez LF, Garay G, Lizano S, et al. 1998. [*Ap. J.*]{} 503:297–306 Greenhill LJ, Gezari DY, Danchi WC, Najita J, Monnier JD, Tuthill PG. 2004. [*Ap. J. Lett.*]{} 605:L57–60 Gualandris A, Portegies Zwart S, Eggleton PP. 2004. [*MNRAS*]{} 350:615–26 Güsten R, Mezger PG. 1982. [*Vistas Astron.*]{} 26:159–224 Gueth F, Guilloteau S. 1999. [*Astron. Astrophys.*]{} 343:571–84 Gutermuth RA, Megeath ST, Pipher JL, Williams JP, Allen LE, et al. 2005. [*Ap. J.*]{} 632:397–420 Habing HJ, Israel FP. 1979. [*Annu. Rev. Astron. Astrophys.*]{} 17:345–85 Hartigan P, Morse J, Bally J. 2000. [*Astron. J.*]{} 120:1436–48 Hawley JF, Balbus SA. 1991. [*Ap. J.*]{} 376:223–33 Heger A, Fryer CL, Woosley SE, Langer N, Hartmann DH. 2003. [*Ap. J.*]{} 591:288–300 Heitsch F, Mac Low M-M, Klessen RS. 2001. [*Ap. J.*]{} 547:280–91 Heller CH. 1995. [*Ap. J.*]{} 455:252–59 Henning T, Schreyer K, Launhardt R, Burkert A. 2000. [*Astron. Astrophys.*]{} 353:211–26 Henning T, Stecklum B. 2002. In [*Modes of Star Formation and the Origin of Field Populations,*]{} ed. EK Grebel, W Brandner. [*ASP Conf. Proc.*]{} 285:40–48 Henriksen RN. 1991. [*Ap. J.*]{} 377:500–09 Herbig GH, Andrews SM, Dahm SE. 2004. [*Astron. J.*]{} 128:1233–53 Herbig GH, Griffin RF. 2006. [*Astron. J.*]{} 132:1763–67 Herbst W, Assousa GE. 1977. [*Ap. J.*]{} 217:473–75 Hernandez J, Hartmann L, Megeath T, Gutermuth R, Muzerolle J, et al. 2007. [*Ap. J.*]{} In press (astro-ph/0701476) Heydari-Malayeri M, Rosa MR, Schaerer D, Martins F, Charmandaris V. 2002. [*Astron. Astrophys.*]{} 381:951–58 Heydari-Malayeri M, Meynadier F, Charmandaris V, Deharveng L, Le Bertre T, et al. 2003. [*Astron. Astrophys.*]{} 411:427–35 Hill T, Burton MG, Minier V, Thompson MA, Walsh AJ, et al. 2005. [*MNRAS*]{} 363:405–51 Hillenbrand LA. 1997. [*Astron. J.*]{} 113:1733–68 Hillenbrand LA, Hartmann LW. 1998. [*Ap. J.*]{} 492:540–53 Hoare MG, Kurtz SE, Lizano S, Keto E, Hofner P. 2007. See Reipurth et al. 2007, pp. 181–96 Hodapp K-W, Rayner J. 1991. [*Astron. J.*]{} 102:1108–17 Hofmann K-H, Seggewiss W, Weigelt G. 1995. [*Astron. Astrophys.*]{} 300:403–14 Hofner P, Delgado H, Whitney B, Churchwell E, Linz H. 2002. [*Ap. J. Lett.*]{} 579:L95–98 Hollenbach D, Johnstone D, Lizano S, Shu F. 1994. [*Ap. J.*]{} 428:654–69 Hollenbach D, Yorke HW, Johnstone D. 2000. See Mannings et al. 2000, pp. 401–28 Hoogerwerf R, de Bruijne JHJ, de Zeeuw PT. 2001. [*Astron. Astrophys.*]{} 365:49–77 Howe JE, Jaffe DT, Genzel R, Stacey GJ. 1991. [*Ap. J.*]{} 373:158–68 Hunter DA. 1995. In [*Gaseous Nebulae and Star Formation. Rev. Mex. Astron. Astrofis. Ser. Conf.*]{} 3:1–7 Indebetouw R, Whitney BA, Johnson KE, Wood K. 2006. [*Ap. J.*]{} 636:362–80 Irvine NJ. 1989. [*Ap. J. Lett.*]{} 337:L33–35 Jappsen A-K, Klessen RS, Larson RB, Li Y, Mac Low M-M. 2005. [*Astron. Astrophys.*]{} 435:611–23 Jijina J, Adams FC. 1996. [*Ap. J.*]{} 462:874–87 Johnstone, Doug; Bally, John. 2006. [*Ap. J.*]{} 653:383–97 Kahn FD. 1974. [*Astron. Astrophys.*]{} 37:149–62 Kaifu N, Usuda T, Hayashi SS, Itoh Y, Akiyama M, et al. 2000. [*Publ. Astron. Soc. Jpn.*]{} 52:1–8 Kastner JH, Buchanan CL, Sargent B, Forrest WJ. 2006. [*Ap. J. Lett.*]{} 638:L29–32 Kennicutt RC. 1998. [*Annu. Rev. Astron. Astrophys.*]{} 36:189–232 Kennicutt RC. 2005. See Cesaroni et al. 2005a, pp. 3–11 Kessel-Deynet O, Burkert A. 2003. [*MNRAS*]{} 338:545–54 Keto E. 2002. [*Ap. J.*]{} 580:980–86 Keto E. 2007. [*Ap. J.*]{} In press (astro-ph/0603856) Keto E, Wood K. 2006. [*Ap. J.*]{} 637:850–59 Kharchenko NV, Piskunov AE, Röser S, Schilbach E, Scholz R-D. 2004. [*Astron. Nachr.*]{} 325:740–48 Kharchenko NV, Piskunov AE, Röser S, Schilbach E, Scholz R-D. 2005. [*Astron. Astrophys.*]{} 440:403–08 Kim SS, Figer DF, Kudritzki RP, Najarro F. 2006. [*Ap. J. Lett.*]{} 653:L113–16 Kippenhahn R, Meyer-Hofmeister E. 1975. [*Astron. Astrophys.*]{} 54:539–42 Kitsionas S, Whitworth AP. 2002. [*MNRAS*]{} 330:129–36 Klein RI, Fisher R, McKee CF. 2004. In [*Gravitational Collapse: From Massive Stars to Planets,*]{} ed. G Garcia-Segura, G Tenorio-Tagle, J Franco, HW Yorke. [*Rev. Mex. Astron. Astrophys. Ser. Conf.,*]{} 22:3–7 Klessen RS. 2001a. [*Ap. J. Lett.*]{} 550:L77–80 Klessen RS. 2001b. [*Ap. J.*]{} 556:837–46 Klessen RS, Burkert A. 2001. [*Ap. J.*]{} 549:386–401 Klessen RS, Ballesteros-Paredes J, Vázquez-Semadeni E, Durán-Rojas C. 2005. [*Ap. J.*]{} 620:786–94 Klessen RS, Spaans M, Jappsen A-K. 2007. [*MNRAS Lett.*]{} 374:L29–33 Knodlseder J. 2000. [*Astron. Astrophys.*]{} 360:539–48 Koen C. 2006. [*MNRAS*]{} 365:590–94 Konigl A. 1991. [*Ap. J. Lett.*]{} 370:L39–43 Kratter KM, Matzner CD. 2006. [*MNRAS.*]{} 373:1563–76 Kraus S, Balega YY, Berger J-P, Hofmann K-H, Millan-Gabet R, et al. 2007. [*Astron. Astrophys.*]{} 466:649–59 Kritsuk AG, Norman ML, Padoan P. 2006. [*Ap. J.*]{} 638:L25–28 Kroupa P. 2000. In [*Massive Stellar Clusters,*]{} ed. A Lancon, C Boily. [*ASP Conf. Ser.,*]{} 211:233–40 Kroupa P. 2002. [*Science*]{} 295:82–91 Kroupa P, Aarseth S, Hurley J. 2001, [*MNRAS*]{} 321:699–712 Krumholz MR. 2006. [*Ap. J. Lett.*]{} 641:L45–48 Krumholz MR. 2007. In [*Massive Stars: From Pop III and GRBs to the Milky Way,*]{} ed. M Livio, E Villaver. In press (astro-ph/0607429) Krumholz MR, McKee CF, Klein RI. 2004. [*Ap. J.*]{} 611:399–412 Krumholz MR, Klein RI, McKee CF. 2005. See Cesaroni et al. 2005a, pp. 231–36 Krumholz MR, McKee CF, Klein RI. 2005a. [*Ap. J.*]{} 618:L33–36 Krumholz MR, McKee CF, Klein RI. 2005b. [*Nature*]{} 438:332–34 Krumholz MR, Klein RI, McKee CF. 2007. [*Ap. J.*]{} 656:959–79 Krumholz MR, Thompson TA. 2007. [*Ap. J.*]{} In press (astro-ph/0611822) Kudritzki R. 2002. [*Ap. J.*]{} 577:389–408 Kurtz S. 2005. See Cesaroni et al. 2005a, pp. 111–19 Kurtz S, Cesaroni R, Churchwell E, Hofner P, Walmsley CM. 2000. See Mannings et al. 2000, pp. 299–326 Lada CJ, Thronson HA Jr., Smith HA, Schwartz PR, Glaccum W. 1984. [*Ap. J.*]{} 286:302–09 Lada CJ, Lada EA. 2003. [*Annu. Rev. Astron. Astrophys.*]{} 41:57–115 Lamers HJGLM, Panagia N, Scuderi S, Romaniello M, Spaans M, et al. 2002. [*Ap. J.*]{} 566:818–32 Larson RB. 1969. [*MNRAS*]{} 145:271–95 Larson RB. 1978. [*MNRAS*]{} 184:69–85 Larson RB. 1981. [*MNRAS*]{} 194:809–26 Larson RB. 1982. [*MNRAS*]{} 200:159–74 Larson RB. 1985. [*MNRAS*]{} 214:379–98 Larson RB. 2005. [*MNRAS*]{} 359:211–22 Larson RB, Starrfield S. 1971. [*Astron. Astrophys.*]{} 13:190–97 Laughlin G, Bodenheimer P. 1994. [*Ap. J.*]{} 436:335–54 Ledoux P. 1941. [*Ap. J.*]{} 94:537–38 Lee C-F, Ho PTP, Beuther H, Bourke TL, Zhang Q, et al. 2006. [*Ap. J.*]{} 639:292–302 Lee H-T, Chen WP. 2007. [*Ap. J.*]{} 657:884–96 Lenorzer A, Bik A, de Koter A, Kurtz SE, Waters LBFM, et al. 2004. [*Astron. Astrophys.*]{} 414:245–59 Leonard PJT, Duncan MJ. 1990. [*Astron. J.*]{} 99:608–16 Lester DF, Harvey PM, Joy M, Ellis HB Jr. 1986. [*Ap. J.*]{} 309:80–89 Li Y. Klessen RS, Mac Low M-M. 2003. [*Ap. J.*]{} 592:975–85 Li Z-Y, Nakamura F. 2006. [*Ap. J. Lett.*]{} 640:L187–90 Linz H, Stecklum B, Henning T, Hofner P, Brandl B. 2005. [*Astron. Astrophys.*]{} 429:903–21 Mac Low M-M, Klessen RS. 2004. [*Rev. Mod. Phys.*]{} 76:125–94 Maeder A, Behrend R. 2002. See Crowther 2002, pp. 179–92 Maiolino R, Vanzi L, Mannucci F, Cresci G, Ghinassi F, Della Valle M. 2002. [*Astron. Astrophys.*]{} 389:84–92 Maíz-Apellániz J, Pérez E, Mas-Hesse JM. 2004. [*Astron. J.*]{} 128:1196–218 Maíz-Apellániz J, Walborn NR, Morrell NI, Niemela VS, Nelan EP. 2007. [*Ap. J.*]{} In press (astro-ph/0612012) Makinen P, Harvey PM, Wilking BA, Evans NJ II. 1985. [*Ap. J.*]{} 299:341–50 Mannings I, Boss AP, Russell SS, ed. 2000. [*Protostars and Planets IV.*]{} Tucson: Univ. Ariz. Press Martel H, Evans NJ II, Shapiro PR. 2006 [*Ap. J. Suppl.*]{} 163:122–44 Martins F, Schaerer D, Hillier DJ. 2005. [*Astron. Astrophys.*]{} 436:1049–65 Martins F, Trippe S, Paumard T, Ott T, Genzel R, et al. 2006. [*Ap. J. Lett.*]{} 649 :L103–06 Mason BD, Gies DR, Hartkopf WI, Bagnuolo WG Jr., ten Brummelaar T, McAlister HA. 1998. [*Astron. J.*]{} 115:821–47 Massey P. 1998. In [*The Stellar Initial Mass Function,*]{} ed. G Gilmore, D Howell. [*ASP Conf. Ser.,*]{} 142:17–44 Massey P. 2002. [*Ap. J. Suppl.*]{} 141:81–122 Massey P. 2003. [*Annu. Rev. Astron. Astrophys.*]{} 41:15–56 Massey P, Johnson KE, Degioia-Eastwood K. 1995. [*Ap. J.*]{} 454:151–71 Massey P, Hunter DA. 1998. [*Ap. J.*]{} 493:180–94 Massey P, Penny LR, Vukovich J. 2002. [*Ap. J.*]{} 565:982–93 Masunaga H, Miyama SM, Inutsuka S-I. 1998. [*Ap. J.*]{} 495:346–69 Masunaga H, Inutsuka S-I. 2000. [*Ap. J.*]{} 531:350–65 Mathis JS, Rumpl W, Nordsieck KH. 1977. [*Ap. J.*]{} 217:425–33 McCaughrean MJ. 2001. In [*From Darkness to Light: Origin and Evolution of Young Stellar Clusters,*]{} ed. T Montmerle, P André. [*ASP Conf. Ser.,*]{} 243:449–60 McCaughrean MJ, Rayner JT, Zinnecker H. 1994. [*Ap. J. Lett.*]{} 436:L189–92 McCaughrean MJ, Mac Low M-M. 1997. [*Astron. J.*]{} 113:391–400 McKee CF, Tan JC. 2003. [*Ap. J.*]{} 585:850–71 McKee CF, Ostriker EC. 2007. [*Annu. Rev. Astron. Astrophys.*]{} 45:In press McMillan SLW, Vesperini E, Portegies Zwart SF. 2007. [*Ap. J. Lett.*]{} 655:L45–49 Megeath ST, Wilson TL, Corbin MR. 2005. [*Ap. J. Lett.*]{} 622:L141–44 Meixner M, Gordon KD, Indebetouw R, Hora JL, Whitney B, et al. 2006. [*Astron. J.*]{} 132:2268–88 Melioli C, de Gouveia Dal Pino EM, de La Reza R, Raga A. 2006. [*MNRAS*]{} 373:811–18 Mengel S. Lehnert MD, Thatte N, Genzel R. 2002. [*Astron. Astrophys.*]{} 383:137–52 Mengel S, Tacconi-Garman LE. 2007. [*Astron. Astrophys.*]{} In press (astro-ph/0701415) Menten KM. 1991. [*Ap. J. Lett.*]{} 380:L75–78 Menten KM, Reid MJ. 1995. [*Ap. J. Lett.*]{} 445:L157–60 Menten KM, Pillai T, Wyrowski F. 2005. See Cesaroni et al. 2005a, pp. 23–34 Mermilliod J-C, García B. 2001. See Zinnecker, Mathieu. 2001, pp. 191–98 Meynet G, Maeder A. 2003. [*Astron. Astrophys.*]{} 404:975–90 Meynet G, Maeder A. 2005. [*Astron. Astrophys.*]{} 429:581–98 Meynet G, Maeder A, Schaller G, Schaerer D, Charbonnel C. 1994. [*Astron. Astrophys. Suppl.*]{} 103:97–105 Mezger PG, Altenhoff W, Schraml J, Burke BF, Reifenstein EC III, Wilson, TL.1967. [*Ap. J.*]{} 150:L157–66 Minier V, Burton MG, Hill T, Pestalozzi MR, Purcell CR, et al. 2005. [*Astron. Astrophys.*]{} 429:945–60 Moeckel N, Bally J. 2006. [*Ap. J.*]{} 653:437–46 Moeckel N, Bally J. 2007. [*Ap. J.*]{} 656:275–86 Moffat AFJ, Drissen L, Shara MM. 1994. [*Ap. J.*]{} 436:183–93 Moffat AFJ, Poitras V, Marchenko SV, Shara MM, Zurek DR, et al. 2004. [*Astron. J.*]{} 128:2854–61 Monaghan JJ. 1992. [*Annu. Rev. Astron. Astrophys.*]{} 30:543–74 Monaghan JJ. 2005. [*Rep. Prog. Phys.*]{} 68:1703–59 Monin J-L, Clarke CJ, Prato L, McCabe C. 2007. See Reipurth et al. 2007, pp. 395–409 Morrell NI, Levato H. 1991. [*Ap. J. Suppl.*]{} 75:965–85 Morrell NI, Barbá RH, Niemela VS, Corti MA, Albacete Colombo JF, et al. 2001. [*MNRAS*]{} 326:85–94 Motte F, Bontemps S, Schilke P, Lis DC, Schneider N, Menten KM. 2005. See Cesaroni et al. 2005a, pp. 151–56 Muench AA, Lada EA, Lada CJ, Alves J. 2002. [*Ap. J.*]{} 573:366–93 Nakano T, Hasegawa T, Norman C. 1995. [*Ap. J.*]{} 450:183–95 Nelan EP, Walborn NR, Wallace DJ, Moffat AFJ, Makidon RB, et al. 2004. [*Astron. J.*]{} 128:323–29 Nielbock M, Chini R, Hoffmeister VH, Scheyda CM, Steinacker J, et al. 2007. [*Ap. J. Lett.*]{} 656:L81–84 Norberg P, Maeder A. 2000. [*Astron. Astrophys.*]{} 359:1025–34 Nürnberger DEA. 2003. [*Astron. Astrophys.*]{} 404:255–65 Nürnberger DEA, Chini R, Eisenhauer F, Kissler-Patig M, Modigliani A, et al. 2007. [*Astron. Astrophys.*]{} 465:931–36 Oey MS, Clarke CJ. 2005. [*Ap. J.*]{} Lett. 620:L43–46 Omukai K, Palla F. 2003. [*Ap. J.*]{} 589:677–87 Öpik EJ. 1953. [*Irish Astron. J.*]{} 2:219–33 Padgett DL, Strom SE, Ghez A. 1997. [*Ap. J.*]{} 477:705–10 Padoan P, Nordlund . 2002. [*Ap. J.*]{} 576:870–79 Palla F, Stahler SW. 1992. [*Ap. J.*]{} 392:667–77 Palla F, Stahler SW. 1993. [*Ap. J.*]{} 418:414–25 Parker JW, Garmany CD. 1993. [*Astron. J.*]{} 106:1471–83 Patel NA, Curiel S, Sridharan TK, Zhang Q, Hunter TR, et al. 2005. [*Nature*]{} 437:109–11 Paumard T, Genzel R, Martins F, Nayakshin S, Beloborodov AM, et al. 2006. [*Ap. J.*]{} 643:1011–35 Pehlemann E, Hofmann K-H, Weigelt G. 1992. [*Astron. Astrophys.*]{} 256:701–14 Perault M, Omont A, Simon G, Seguin P, Ojha D, et al. 1996. [*Astron. Astrophys. Lett.*]{} 315:L165–68 Peretto N, André P, Belloche A. 2006. [*Astron. Astrophys.*]{} 445:979–98 Pflamm-Altenburg J, Kroupa P. 2006. [*MNRAS*]{} 373:295–304 Pietrinferni A, Cassisi S, Salaris M, Castelli F. 2004. [*Ap. J.*]{} 612:168–90 Piskunov AE, Kharchenko NV, Röser S, Schilbach E, Scholz RD. 2006. [*Astron. Astrophys.*]{} 445:545–65 Plume R, Jaffe DT, Evans NJ II, Martin-Pintado J, Gomez-Gonzalez J. 1997. [*Ap. J.*]{} 476:730–49 Plüschke S, Cervi[ñ]{}o M, Diehl R, Kretschmer K, Hartmann DH, Knödlseder J. 2002. [*New Astron. Rev.*]{} 46:535–39 Poetzel R, Mundt R, Ray TP. 1992. [*Astron. Astrophys.*]{} 262:229–47 Portegies Zwart SF, McMillan SLW. 2002. [*Ap. J.*]{} 576:899–907 Portegies Zwart SF, Baumgardt H, McMillan SLW, Makino J, Hut P, Ebisuzaki T. 2006. [*Ap. J.*]{} 641:319–26 Poveda A, Ruiz J, Allen C. 1967. [*Bol. Obs. Tonantzintla Tacubaya*]{} 4:86–90 Preibisch T, Ossenkopf V, Yorke HW, Henning T. 1993. [*Astron. Astrophys.*]{} 279:577–88 Preibisch T, Balega Y, Hofmann K-H, Weigelt G, Zinnecker H. 1999. [*New Astron.*]{} 4:531–42 Preibisch T, Zinnecker H. 1999. [*Astron. J.*]{} 117:2381–97 Preibisch T, Weigelt G, Zinnecker H. 2001. See Zinnecker & Mathieu 2001, p. 69–78 Preibisch T, Balega YY, Schertl D, Smith MD, Weigelt G. 2001. [*Astron. Astrophys.*]{} 378:539–45 Preibisch T, Balega YY, Schertl D, Weigelt G. 2002. [*Astron. Astrophys.*]{} 392:945–54 Preibisch T, Balega YY, Schertl D, Weigelt G. 2003. [*Astron. Astrophys.*]{} 412:735–43 Preibisch T, Zinnecker H. 2007. See Elmegreen, Palous. 2007, pp. 270–77 Pudritz RE, Norman CA. 1983. [*Ap. J.*]{} 274:677–97 Pudritz RE, Ouyed R, Fendt C, Brandenburg A. 2007. See Reipurth et al. 2007, pp. 277–94 Rathborne JM, Jackson JM, Simon R. 2006. [*Ap. J.*]{} 641:389–405 Reipurth B, Zinnecker H. 1993. [*Astron. Astrophys.*]{} 278:81–108 Rauw G, De Becker M, Nazé Y, Crowther PA, Gosset E, et al. 2004. [*Astron. Astrophys. Lett.*]{} 420:L9–13 Reipurth B, Jewitt D, Keil K, ed. 2007. [*Protostars and Planets V.*]{} Tucson: Univ. Ariz. Press Richer JS, Shepherd DS, Cabrit S, Bachiller R, Churchwell E. 2000. See Mannings et al. 2000, pp. 867–96 Richling S, Yorke HW. 1997. [*Astron. Astrophys.*]{} 327:317–24 Roberts MS. 1957. [*Publ. Astron. Soc. Pac.*]{} 69:59–64 Rodríguez LF, Poveda A, Lizano S, Allen C. 2005a. [*Ap. J. Lett.*]{} 627:L65–68 Rodríguez LF, Garay G, Brooks KJ, Mardones D. 2005b. [*Ap. J.*]{} 626:953–58 Rubio M, Barbá RH, Walborn NR, Probst RG, García J, Roth MR. 1998. [*Astron. J.*]{} 116:1708–18 Salpeter EE. 1955. [*Ap. J.*]{} 121:161–67 Sana H, Rauw G, Gosset E. 2005. In [*Massive Stars and High-Energy Emission in OB Associations, Proc. Workshop JENAM,*]{} ed. G Rauw, Y Nazé, R Blomme, E Gosset, pp. 107–10 Sandage A. 1986. [*Astron. Astrophys.*]{} 161:89–101 Sanz-Forcada J, Franciosini E, Pallavicini R. 2004. [*Astron. Astrophys.*]{} 421:715–27 Scalo JM. 1998. In [*The Stellar Initial Mass Function,*]{} ed. G Gilmore, D Howell. [*ASP Conf. Ser.*]{} 142:201–36 Schaerer D, de Koter A. 1997. [*Astron. Astrophys.*]{} 322:598–614 Schertl D, Balega YY, Preibisch T, Weigelt G. 2003. [*Astron. Astrophys.*]{} 402:267–75 Schmeja S, Klessen RS. 2004. [*Astron. Astrophys.*]{} 419:405–17 Schmid-Burgk J, Guesten R, Mauersberger R, Schulz A, Wilson TL. 1990. [*Ap. J. Lett.*]{} 362:L25–28 Schneider N, Bontemps S, Simon R, Jakob H, Motte F, et al. 2006. [*Astron. Astrophys.*]{} 458:855–71 Schreyer K, Stecklum B, Linz H, Henning T. 2003. [*Ap. J.*]{} 599:335–41 Schreyer K, Semenov D, Henning T, Forbrich J. 2006. [*Ap. J. Lett.*]{} 637:L129–32 Schulz NS, Berghoefer TW, Zinnecker H. 1997. [*Astron. Astrophys.*]{} 325:1001–12 Schwartz PR, Thronson HA Jr., Odenwald SF, Glaccum W, Loewenstein RF, Wolf G. 1985. [*Ap. J.*]{} 292:231–37 Schwarzschild M, Härm R. 1959. [*Ap. J.*]{} 129:637–46 Schweickhardt J, Schmutz W, Stahl O, Szeifert T, Wolf B. 1999. [*Astron. Astrophys.*]{} 347:127–36 Selman FJ, Melnick J. 2005. [*Astron. Astrophys.*]{} 443:851–61 Sewilo M, Churchwell E, Kurtz S, Goss WM, Hofner P. 2004. [*Ap. J.*]{} 605:285–99 Shepherd D. 2005. See Cesaroni et al. 2005a, pp. 237–46 Shepherd DS, Churchwell E. 1996a. [*Ap. J.*]{} 472:225–39 Shepherd DS, Churchwell E. 1996b. [*Ap. J.*]{} 457:267–76 Shepherd DS, Claussen MJ, Kurtz SE. 2002. See Crowther 2002, pp. 415–16 Shirley YL, Evans NJ II, Young KE, Knez C, Jaffe DT. 2003. [*Ap. J. Suppl.*]{} 149:375–403 Shu FH. 1977. [*Ap. J.*]{} 214:488–97 Shu FH, Adams FC, Lizano S. 1987. [*Annu. Rev. Astron. Astrophys.*]{} 25:23–81 Shu FH, Najita J, Ostriker E, Wilkin F, Ruden S, Lizano S. 1994. [*Ap. J.*]{} 429:781–96 Shu FH, Najita J, Ostriker EC, Shang H. 1995. [*Ap. J. Lett.*]{} 459:L155–8 Silk J. 1978. In [*Protostars and Planets I,*]{} ed. T Gehrels, p. 172. Tucson: Univ. Ariz. Press Silk J. 1995. [*Ap. J. Lett.*]{} 438:L41–44 Silk J. 1997. [*Ap. J.*]{} 481:703–09 Silk J. 2005. See Corbelli et al. 2005, pp. 439–47 Simon R, Jackson JM, Rathborne JM, Chambers ET. 2006. [*Ap. J.*]{} 639:227–36 Sirianni M, Nota A, Leitherer C, De Marchi G, Clampin M. 2000. [*Ap. J.*]{} 533:203–14 Smith J, Bentley A, Castelaz M, Gehrz RD, Grasdalen GL, Hackwell JA. 1985. [*Ap. J.*]{} 291:571–80 Smith LF, Mezger PG, Biermann P. 1978. [*Astron. Astrophys.*]{} 66:65–76 Smith N, Bally J, Shuping RY, Morris M, Hayward TL. 2004. [*Ap. J. Lett.*]{} 610:L117–20 Smith N, Owocki SP. 2006. [*Ap. J. Lett.*]{} 645:L45–48 Smith N, Brooks KJ. 2007. [*MNRAS.*]{} In press (arXiv:0705.3053, astro-ph) Sridharan TK, Williams SJ, Fuller GA. 2005. [*Ap. J. Lett.*]{} 631:L73–76 Stahl O, Kaufer A, Rivinius T, Szeifert T, Wolf B, et al. 1996. [*Astron. Astrophys.*]{} 312:539–48 Stahler SW, Palla F, Ho PTP. 2000. See Mannings et al. 2000, pp. 327–51 Stecklum B, Brandl B, Henning T, Pascucci I, Hayward TL, Wilson JC. 2002. [*Astron. Astrophys.*]{} 392:1025–29 Stolte A, Brandner W, Brandl B, Zinnecker H. 2006. [*Astron. J.*]{} 132:253–70 Stone JM, Ostriker EC, Gammie CF. 1998. [*Ap. J. Lett.*]{} 508:L99–102 Stothers RB, Chin C-W. 1993. [*Ap. J. Lett.*]{} 408:L85–88 Sung H, Bessell MS. 2004. [*Astron. J.*]{} 127:1014–28 Suttner G, Yorke HW. 2001. [*Ap. J.*]{} 551:461–77 Tamura M, Yamashita T. 1992. [*Ap. J.*]{} 391:710–18 Tan JC. 2004. [*Ap. J. Lett.*]{} 607:L47–50 Tan JC. 2005. See Cesaroni et al. 2005a, pp. 318–27 Tan JC, Krumholz MR, McKee CF. 2006. [*Ap. J. Lett.*]{} 641:L121–24 Taresch G, Kudritzki RP, Hurwitz M, Bowyer S, Pauldrach AWA, et al. 1997. [*Astron. Astrophys.*]{} 321:531–48 Terebey S, Shu FH, Cassen P. 1984. [*Ap. J.*]{} 286:529–51 Terquem CEJMLJ. 2001. See Zinnecker & Mathieu 2001, pp. 406–09 Testi L, Palla F, Prusti T, Natta A, Maltagliati S. 1997. [*Astron. Astrophys.*]{} 320:159–66 Testi L, Palla F, Natta A. 1999. [*Astron. Astrophys.*]{} 342:515–23 Tohline JE, Durisen RH. 2001. See Zinnecker & Mathieu 2001, pp. 40–44 Trinidad MA, Curiel S, Cantó J, D’Alessio P, Rodríguez LF. 2003. [*Ap. J.*]{} 589:386–96 Truelove JK, Klein RI, McKee CF, Holliman JH II, Howell LH, Greenough JA. 1997. [*Ap. J. Lett.*]{} 489:L179–83 Turner JL, Beck SC, Crosthwaite LP, Larkin JE, McLean IS, Meier DS. 2003. [*Nature*]{} 423:621–23 Turner NJ, Blaes OM, Socrates A, Begelman MC, Davis SW. 2005. [*Ap. J.*]{} 624:267–88 Turner NJ, Quataert E, Yorke HW. 2007. [*Ap. J.*]{} In press (astro-ph/0701800) Vallée JP, MacLeod JM. 1994. [*Astron. J.*]{} 108:998–1001 van Albada TS. 1968. [*Bull. Astron. Inst. Neth.*]{} 20:57–68 van Altena WF, Lee JT, Lee J-F, Lu PK, Upgren AR. 1988. [*Astron. J.*]{} 95:1744–54 van Bever J, Vanbeveren D. 1998. [*Astron. Astrophys.*]{} 334:21–28 van der Tak FFS, Menten KM. 2005. [*Astron. Astrophys.*]{} 437:947–56 van der Tak FFS, Walmsley CM, Herpin F, Ceccarelli C. 2006. [*Astron. Astrophys.*]{} 447:1011–25 Vanhala HAT, Boss AP, Cameron AGW, Foster PN. 1998. [*29th Annu. Lunar Planet. Sci. Conf., March 16-20, 1998,*]{} Houston. Abstr. 1470 Vanhala HAT, Cameron AGW. 1998. [*Ap. J.*]{} 508:291–307 Vázquez-Semadeni E, Kim J, Shadmehri M, Ballesteros-Paredes J. 2005. [*Ap. J.*]{} 618:344–59 Verbunt F. 1993. [*Annu. Rev. Astron. Astrophys.*]{} 31:93–127 von Hoerner S. 1968. In [*Interstellar Ionized hydrogen, Proc. Symp. HII Regions,*]{} ed. Y Terzian, pp. 101–69. New York: Benjamin Walborn NR. 2003. In [*A Massive Star Odyssey: From Main Sequence to Supernova,*]{} ed. K van der Hucht, A Herrero, E César. [*IAU Symp.*]{} 212:13–21. San Francisco: ASP Walborn NR. 2007. In [*Massive Stars: From Pop III and GRBs to the Milky Way,*]{} ed. M Livio, E Villaver. In press (astro-ph/0701573) Walborn NR, Parker JW. 1992. [*Ap. J. Lett.*]{} 399:L87–89 Walborn NR, Drissen L, Parker JW, Saha A, MacKenty JW, White RL. 1999. [*Astron. J.*]{} 118:1684–99 Walborn NR, Howarth ID, Lennon DJ, Massey P, Oey MS, et al. 2002. [*Astron. J.*]{} 123:2754–71 Walborn NR, Maíz-Apellániz J, Barbá RH. 2002. [*Astron. J.*]{} 124:1601–24 Walborn NR, Howarth ID, Herrero A, Lennon DJ. 2003. [*Ap. J.*]{} 588:1025–38 Walsh AJ, Burton MG, Hyland AR, Robinson G. 1998. [*MNRAS*]{} 301:640–98 Weidenschilling SJ, Ruzmaikina TV. 1994. [*Ap. J.*]{} 430:713–26 Weidner C, Kroupa P. 2004. [*MNRAS*]{} 348:187–91 Weigelt G, Baier G. 1985. [*Astron. Astrophys. Lett.*]{} 150:L18–20 Whitworth AP, Bhattal AS, Chapman SJ, Disney MJ, Turner JA. 1994. [*MNRAS*]{} 268:291–98 Whitworth AP, Boffin HMJ, Francis N. 1998. [*MNRAS*]{} 299:554–61 Williams JP, Blitz L, McKee CF. 2000. See Mannings et al. 2000, pp. 97–120 Winkler K-H, Newman M. 1980. [*Ap. J.*]{} 236:201–11 Wolff SC, Strom SE, Dror D, Lanz L, Venn K. 2006. [*Astron. J.*]{} 132:749–55 Wolfire MG, Cassinelli JP. 1987. [*Ap. J.*]{} 319:850–67 Woodward PR. 1978. [*Annu. Rev. Astron. Astrophys.*]{} 16:555–84 Wu Y, Zhang Q, Chen H, Yang C, Wei Y, Ho PTP. 2005. [*Astron. J.*]{} 129:330–47 Wynn-Williams CG. 1982. [*Annu. Rev. Astron. Astrophys.*]{} 20:587–618 Wyrowski F, Walmsley CM, Goss WM, Tielens AGGM. 2000. [*Ap. J.*]{} 543:245–56 Yorke HW. 1986. [*Annu. Rev. Astron. Astrophys.*]{} 24:48–87 Yorke HW. 1988. In [*Dust in the Universe, Proc. Conf.,*]{} ed. ME Bailey, DA Williams, pp. 355–72. Cambridge, UK: Cambridge Univ. Press Yorke HW. 2002. See Crowther 2002, pp. 165–78 Yorke HW, Krügel E. 1977. [*Astron. Astrophys.*]{} 54:183–94 Yorke HW, Welz A. 1996. [*Astron. Astrophys.*]{} 315:555–64 Yorke HW, Bodenheimer P. 1999. [*Ap. J.*]{} 525:330–42 Yorke HW, Sonnhalter C. 2002. [*Ap. J.*]{} 569:846–62 Zavagno A, Deharveng L, Brand J, Massi F, Caplan J, et al. 2005. See Cesaroni et al. 2005a, pp. 346–51 Zhang Q. 2005. See Cesaroni et al. 2005a, pp. 135–44 Ziebarth K. 1970. [*Ap. J.*]{} 162:947–62 Zinnecker H. 1982. [*NY Acad. Sci.*]{} 395:226–35 Zinnecker H. 1985. In [*Birth and Infancy of Stars. Proc. Les Houches Summer School,*]{} ed. R Lucas, A Omont, R Stora, pp. 473–75. NATO: Knudsen. Zinnecker H. 1986. In [*Luminous Stars & Associations in Galaxies,*]{} ed. CWH de Loore, AJ Willis, P Laskarides, [*IAU Symp.*]{} 116:271–73. Dordrecht: Kluwer Academic Publishers Zinnecker H. 1991. In [*Fragmentation of Molecular Clouds and Star Formation,*]{} ed. E Falgarone, F Boulanger, G Duvert, [*IAU Symp.*]{} 147:526–32. Dordrecht: Kluwer Academic Publishers Zinnecker H. 1996. In [*The Interplay Between Massive Star Formation, the ISM and Galaxy Evolution. Proc. 11th IAP Astrophys. Meet.,*]{} ed. D Kunth, B Guiderdoni, M Heydari-Malayeri, TX Thuan, pp. 249–58. Gif-sur-Yvette: Editions Frontieres. Zinnecker H. 2002. In [*The Origin of Stars and Planets: The VLT View,*]{} ed. JF Alves, MJ McCaughrean, pp. 179–86. Berlin/Heidelberg: Springer-Verlag Zinnecker H. 2003. In [*A Massive Star Odyssey: From Main Sequence to Supernova,*]{} ed. K van der Hucht, A Herrero, E César. [*IAU Symp.*]{} 212:80–90. San Francisco: ASP Zinnecker H. 2004. In [*The Formation and Evolution of Massive Young Star Clusters,*]{} ed. HJGLM Lamers, LJ Smith, A Nota, [*ASP Conf. Ser.*]{} 322:349–58 Zinnecker H. 2006a. In [*The Scientific Requirements for Extremely Large Telescopes,*]{} ed. PA Whitelock, M Dennefeld, B Leibundgut. [*IAU Symp.*]{} 232:324–27. Cambridge: Cambridge University Press Zinnecker H. 2006b. In [*Stellar Evolution at Low Metallicity: Mass Loss, Explosions, Cosmology,*]{} ed. HJGLM Lamers, N Langer, T Nugis, K Annuk, [*ASP Conf. Ser.*]{} 353:339–47 Zinnecker H, McCaughrean MJ, Wilking BA. 1993. In [*Protostars and Planets III,*]{} ed. EH Levy, JI Lunine, pp. 429–95. Tucson: Univ. Ariz. Press Zinnecker H, McCaughrean MJ, Rayner JT. 1998. [*Nature*]{} 394:862–65 Zinnecker H, Mathieu RD, ed. 2001. [*The Formation of Binary Stars, IAU Symp. 200.*]{} San Francisco: ASP Zinnecker H, Bate MR. 2002. See Crowther 2002, pp. 209–18 {width="12.0cm"} ![ Hubble Space Telescope optical/IR image of the dense massive young cluster R136/30 Dor (courtesy of M.J. McCaughrean; FOV $\sim$30arcsec$\times$30arcsec or 7.5pc$\times$7.5pc). Dozens of massive O stars are found crowded within the half-light radius of 2pc (Brandl et al. 1996). ([*a*]{}) A VLT image of NGC 3603 (Brandl et al. 1999) and ([*b*]{}) a VLT image of the Trapezium Cluster in Orion (McCaughrean 2001) are shown, as these two galactic clusters would be seen if they were located at the distance of R136 in the Large Magellanic Cloud (50kpc) and imaged with similar angular resolution (see Zinnecker 2002).](araa_fig02.jpg){width="12.0cm"} {width="12.0cm"} {width="12.0cm"} ![ ZAMS luminosity to mass ratio (left scale) and corresponding critical effective opacity (right scale) as defined by [**Equation 3**]{} as a function of stellar mass. The solid grey line depicts the ZAMS models shown in [**Figure 4**]{}. Dotted grey lines indicate the total luminosity (including accretion luminosity) of stars accreting at the indicated constant rate as discussed in the text (cf. [**Figure 8**]{}). The triangle denotes the position of an O5V star (see [**Figure 9**]{}). Dashed lines denote the opacities of dusty gas in the light of black body sources at the temperature indicated (for two different grain types, see [**Figure 6**]{}) and the contribution from electron scattering in a fully ionized plasma. Because of the combined effect of UV lines (UV line blanketing), the actual opacity of a hot plasma can be greater than that from electron scattering alone. This may be relevant for the existence of an upper stellar mass limit.](araa_fig05.jpg){width="12.0cm"} ![ Planck-weighted mean effective opacity of dusty gas, using the (Preibisch et al. 1993) dust model with ice-coated grains (blue curve) and grains without ice mantels (red curve), assuming black-body radiation at the temperature T$_{\rm rad}$ and solar abundances. The right-hand scale (grey) is based on the ZAMS models shown in [**Figure 5**]{} (solid grey line).](araa_fig06.jpg){width="12.0cm"} {width="12.0cm"} ![ Evolutionary tracks in the HR diagram for (proto-)stars accreting at a constant rate (colored lines) are contrasted to the tracks of non-accreting stars (black lines). All accreting tracks are assumed to begin at the birthline of an equilibrium deuterium burning 0.1M$_\odot$ pre-main-sequence star. Non-accreting tracks up to H burning were calculated by Yorke for M$_*$/M$_\odot$=0.1, 0.3, 1, 3, 10, and 30, using the computer code supplied with the book by Bodenheimer et al. (2007). \[Adapted from Yorke (2002)\]](araa_fig08.jpg){width="12.0cm"} ![ Luminosity and effective temperature calibration of luminosity classes I (blue-filled circles), III (green squares), and V (red triangles) for stars of spectral classes O3, O4, O5, O5.5, O6, O6.5, O7, O7.5, O8, O8.5, O9, and O9.5. For comparison, theoretical evolutionary tracks from Meynet & Maeder (2003) for non-rotating (dashed grey) and for rotating stars (continuous grey) with mass loss are labeled by their starting ZAMS mass. Meynet et al. (1994) isochrones for non-rotating stars are also plotted for 0 (solid black line labeled ZAMS), 1, 3, and 5Myr (dotted black lines). \[Adapted from Figure 14 of Martins, Schaerer & Hillier (2005)\]](araa_fig09.jpg){width="12.0cm"} ![ Flux of radiation that can ionize hydrogen (blue lines) and neutral helium (red dashed lines) as a function of age (counted from the moment of arrival on the ZAMS), for stars of ZAMS masses (top to bottom) of M$_*$/M$_\odot$=120, 85, 60, 40, 25, and 20. Mass loss is included in the evolution. \[Based on models from Schaerer & de Koter (1997)\]](araa_fig10.jpg){width="12.0cm"} {width="12.0cm"} ![ Time-dependent evolution of a turbulent, self-gravitating 1000-M$_\odot$ cloud showing the formation of sheets, filaments, and cores, the latter of which become gravitationally unstable. Newly formed stars are shown in dark blue, the gas is shown in blue-green. Stars tend to cluster, they continue to accrete material in competition with other stars. Times shown are 1.0, 1.4, 1.8, and 2.4 initial free-fall times (t$_{\rm ff}$), from left to right and top to bottom. \[Adapted from Bonnell, Bate & Vine (2003)\].](araa_fig12.jpg){width="12.0cm"} ![([*a*]{}) A grazing encounter between two 3M$_\odot$ premain-sequence stars that results in the formation of a binary (r$_{\rm min}$=25.8 solar radii). ([*b*]{}) A detached encounter between a 3M$_\odot$ premain-sequence star and a 10M$_\odot$ zero age main-sequence star with the same minimum periastron distance as in the (a) encounter. The stars have radii of 12.9 and 3.92 solar radii, respectively. The greater density of the 10M$_\odot$ star results in tidal disruption of the low-density 3M$_\odot$ star to form a disk around the massive star. The encounters have zero relative velocity at infinity (i.e., they are parabolic encounters). \[ Adapted from Zinnecker & Bate (2002), Davies et al. (2006)\]. The cross section for a subsequent collision with (say) another 10M$_\odot$ star is significantly increased, and a runaway collisional growth in mass is possible. This is the so-called ’shred and add’ process.](araa_fig13.jpg){width="12.0cm"} ![ Massive stars (dark blue circles) are formed in the center of individual subclusters of low-mass stars (light circles) because of competitive accretion. These subclusters evolve by merging with the other subclusters. The final state of the simulation is a single, centrally condensed cluster with little substructure but with 4 massive stars, one from each subcluster. This then is a model for the origin of Trapezium-type systems. \[Adapted from Bonnell, Vine & Bate (2004)\]](araa_fig14.jpg){width="12.0cm"} ![ The multiple star systems of the Orion Trapezium as revealed by bispectrum speckle interferometry. \[Courtesy of G. Weigelt & Th. Preibisch; inserts adapted from Schertl et al. (2003) and Kraus et al. (2007)\]](araa_fig15.jpg){width="12.0cm"} {width="12.0cm"} {width="12.0cm"} {width="12.0cm"} [^1]: Astrophysikalisches Institut Potsdam, An der Sternwarte 16, D-14482 Potsdam, Germany; email: hzinnecker@aip.de [^2]: Jet Propulsion Laboratory, California Institute of Technology, 4800 Oak Grove Drive, Pasadena, CA 91109 USA; email: Harold.Yorke@jpl.nasa.gov

ArXiv

--- abstract: 'We combine a complete sample of 113 pointed observations taken with the [*Rossi X-ray Timing Explorer*]{} between 1996–1999, monitoring observations taken with the Ryle telescope and the Green Bank Interferometer, and selected observations with the Very Large Array to study the radio and X-ray properties of [GRS 1915$+$105]{} when its X-ray emission is hard and steady. We establish that radio emission always accompanies the hard-steady state of [GRS 1915$+$105]{}, but that the radio flux density at 15.2 GHz and the X-ray flux between 2–200 keV are not correlated. Therefore we study the X-ray spectral and timing properties of [GRS 1915$+$105]{} using three approaches: first, by describing in detail the properties of three characteristic observations, then by displaying the time evolution of the timing properties during periods of both faint and bright radio emission, and lastly by plotting the timing properties as a function of the the radio flux density. We find that as the radio emission becomes brighter and more optically thick, 1) the frequency of a ubiquitous 0.5–10 Hz QPO decreases, 2) the Fourier phase lags between hard (11.5–60 keV) and soft (2–4.3 keV) in the frequency range of 0.01–10 Hz change sign from negative to positive, 3) the coherence between hard and soft photons at low frequencies decreases, and 4) the relative amount of low frequency power in hard photons compared to soft photons decreases. We discuss how these results reflect upon basic models from the literature describing the accretion flow around black holes and the possible connection between Comptonizing electrons and compact radio jets.' author: - 'Michael P. Muno, Ronald A. Remillard, Edward H. Morgan,' - 'Elizabeth B. Waltman' - 'Vivek Dhawan, Robert M. Hjellming' - Guy Pooley title: 'Radio Emission and the Timing Properties of the Hard X-ray State of [GRS 1915$+$105]{}' --- Introduction ============ The microquasar [GRS 1915$+$105]{} is one of the most interesting galactic sources of radio, infrared, and X-ray emission. It was first discovered as a transient X-ray source by [*Granat*]{} in 1992 [@ct92], and later was observed to emit highly relativistic radio jets with intrinsic velocities greater than 0.9c [@mr94; @fen99b]. High optical extinction has prevented studies of the binary companion of [GRS 1915$+$105]{}, and there is no known orbital period or mass function for the system. However, the X-ray luminosity of [GRS 1915$+$105]{} greatly exceeds the Eddington luminosity for a neutron star (Greiner, Morgan, and Remillard 1996), and the other galactic sources of highly relativistic ($> 0.9$c) jets, [GRO J1655$-$40]{} [@ob97; @sha00] and V4641 Sgr [@oro01], have mass functions greater than the maximum mass for a neutron star, 3 [M$_{\odot}$]{}. [GRS 1915$+$105]{} is therefore thought to be a black hole binary system. The properties of microquasars in the galaxy are reviewed by @mr99. Although [GRS 1915$+$105]{} is often noted for exhibiting X-ray, radio, and infrared emission which is variable on time scales of seconds to minutes [@gmr97; @pf97; @mir98; @eik98], the source also exhibits steady emission which is reminiscent of black hole candidates such as [Cyg X$-$1]{} and [GX 339$-$4]{}. Both the steady and variable emission from [GRS 1915$+$105]{} exhibit two (Muno, Morgan, & Remillard 1999; Rao, Yadav, & Paul 2000) or three [@bel00] basic modes of X-ray emission. In the two-state description, the hard state resembles the very high state of canonical black hole binaries (Morgan, Remillard, & Greiner 1997), as it exhibits X-ray flux with a power-law spectrum above 40 keV, prominent thermal emission from the accretion disk below $\sim 5$ keV, and a strong (up to 15% RMS) 0.5–10 Hz quasi-periodic oscillation (QPO). Conversely, the soft state of [GRS 1915$+$105]{} exhibits little X-ray emission above 40 keV and no 0.5–10 Hz QPO, although a stationary 67 Hz QPO and several low frequency ($< 0.1$ Hz) QPOs are often observed during this state. Long-term radio and X-ray monitoring of [GRS 1915$+$105]{}  [@rod95; @fos96; @har97; @pf97; @ban98] has revealed that periods of hard, steady X-ray emission are often accompanied by bright, optically thick radio emission. This has been dubbed the “plateau state” by Foster et al. (1996; see also Fender 2001). The radio plateau emission in [GRS 1915$+$105]{} resembles steady, optically thick radio emission observed coincident with hard X-ray emission in the black hole candidates [Cyg X$-$1]{} [@hgo75; @bro99], [GX 339$-$4]{} [@cor00], and [Cyg X$-$3]{} [@wal95; @mcc99]. The connection between the X-ray and radio emission in [GRS 1915$+$105]{} is illustrated in Figure \[asm\], where we plot the intensity as a function of time as observed in soft X-rays by the All-Sky Monitor (ASM) aboard the [*Rossi X-ray Timing Explorer*]{} ([[*RXTE*]{}]{}; first panel), and at radio wavelengths by the Ryle telescope (15.2 GHz; third panel). We also plot the the ASM hardness ratio HR2 (5-12 keV / 3-5 keV; second panel), and the radio spectral index $\alpha$ taken from monitoring data at 2.25 and 8.3 GHz from the Green Bank Interferometer (i.e. $S_{\nu} \propto \nu^{\alpha}$; only observations where $S_\nu$ at both wavelengths are greater than 20 mJy are plotted). Bright radio plateau emission and hard-steady X-ray emission is evident on several occasions (e.g. MJD 50730–50750), yet the radio emission is sometimes a factor of 10 fainter during other hard X-ray states (e.g. MJD 50450–50550). Observations with the Very Long Baseline Array during the plateau state reveal that the optically thick radio emission originates in a compact jet on order 10 AU across (Dhawan, Mirabel, & Rodríguez 2000). The radio plateau state is of further interest because it is often preceded and/or followed by optically thin radio flares in [GRS 1915$+$105]{} [@fen99b], as are similar hard-steady states with optically thick radio emission in [GX 339$-$4]{} [@cor00] and [Cyg X$-$3]{} [@mcc99]. The optically thin emission in [GRS 1915$+$105]{} originates from relativistic ejections of synchrotron-emitting blobs detached from the central source [@mr94; @fen99b; @dmr00]. In Figure \[asm\], optically thin radio flares precede and terminate the radio plateau emission which occurred between MJD 50730–50750, while weaker flares with shorter duration terminate the long period of weaker radio emission from MJD 50450–50550. These events are associated with flares in the ASM light curve, which suggests that the most powerful jets are associated with instabilities in the inner accretion disk. Moreover, there have been several occasions in which weaker radio and infrared flares recur at $\sim$30 minute intervals, and are preceded by distinctive dip and flare patterns in X-rays [@pf97; @eik98; @mir98]. During the plateau state of [GRS 1915$+$105]{}, the radio luminosity is low compared to the large optically thin flares, but the power in the radio jets may still represent a significant fraction of the accretion energy inferred from the X-ray emission [@fb99; @fen01]. Since this emission originates from a compact jet within 10 AU of the central source and is present only during the hard X-ray state, it seems very likely that the compact jet is influenced by the properties of the inner accretion flow, and vice versa. This has led @fen99a to suggest that the hard X-rays originate from inverse Compton scattering of seed photons at the base of the compact jet of relativistic electrons. This idea motivates further efforts to compare the properties of the radio emission to the X-ray spectral and timing information during the hard state of [GRS 1915$+$105]{}. The timing properties of [GRS 1915$+$105]{} provide a promising means of probing the accretion flow near the black hole. In particular, the 0.5–10 Hz QPO observed during the hard state appears to provide a link between the optically thick accretion disk and the Comptonizing electrons, because (1) the frequency of the QPO increases as the thermal flux from the disk increases (Chen, Taam & Swank 1997; Trudolyubov, Churazov, & Gilfanov 1999; Markwardt, Swank, & Taam 1999), (2) the spectrum of the QPO amplitude is hard, indicating an origin in the power law component of the spectrum [@mrg97], and (3) the QPO is only present when the power-law component of the X-ray spectrum is strong, while it is absent when the power-law component is weak [@mmr99]. Additional timing characteristics, such as the Fourier phase lags and the coherence function, may further constrain the relationship between hard and soft X-ray components. @rei00 have found that the signs of the phase lags change from positive to negative as the frequency of the 0.5–10 Hz QPO increases, which has provided impetus for others [e.g. @nob00] to develop models describing the structure of the Comptonizing corona and how it is related to the optically thick accretion disk. In this paper, we use a complete sample of “hard-steady” observations of [GRS 1915$+$105]{} (referred to as the $\chi$ state by Belloni et al. 2000) to examine in detail the connection between the hard, steady X-ray emission and the steady radio emission. First, we establish that the hard-steady X-ray states of [GRS 1915$+$105]{} always exhibit detectable radio emission. However, the intensities of the 2–200 keV X-ray and 15.2 GHz radio flux are not correlated, so we proceed to examine whether there are more subtle relationships between the spectrum and timing properties of the X-ray emission and the intensity and spectrum of the radio emission. We investigate these correlations in three different manners: first, by describing in detail the properties of three characteristic observations, then by displaying the time evolution of the timing properties during periods of both faint and bright radio emission, and lastly by examining the timing properties as a function of the radio flux density. Observations and Data Analysis ============================== Ryle Monitoring Observations ---------------------------- The Ryle Telescope at the Mullard Radio Astronomy Observatory, Cambridge, is primarily used for microwave-background measurements, but during gaps in that program it has been used to monitor variable galactic and extra-galactic sources. It operates at 15.2 GHz, and uses a single set of linearly-polarized feeds. In good observing conditions, the RMS noise in a 1 minute integration is about 3.5 mJy. Since the measurements are unbiased, the noise level decreases as the square root of the observing time. Details of the observing and analysis routine are given in @pf97. We have averaged the radio measurements in two ways for the purposes of this paper. We have used the averaged daily flux density measured with the Ryle Telescope when displaying monitoring light curves of [GRS 1915$+$105]{} (e.g. Figure \[asm\]). When comparing the Ryle measurements to the properties of pointed [[*RXTE*]{}]{} observations, we have averaged data taken within 0.5 days before and after the [[*RXTE*]{}]{} observation in order to sample the radio properties which most closely correspond to the X-ray observation (see Table \[rxte\]). Green Bank Interferometer Monitoring Observations ------------------------------------------------- Monitoring with the Green Bank Interferometer has been performed simultaneously at 2.25 and 8.3 GHz on a 2.4 km baseline with 35 MHz bandwidth and dual circular polarization. Most scans are 10-15 minute integrations, and the calibration procedure is described by @fos96. Random (one sigma) errors in the GBI data are flux dependent: 4 mJy (2 GHz) or 6 mJy (8 GHz) for fluxes $< 100$ mJy, 15 mJy (2 GHz) or 50 mJy (8 GHz) for fluxes $\simeq 1$ Jy. In addition, systematic errors are introduced by atmospheric and hardware effects, which we estimate may approach 10% at 2 GHz and 20% at 8 GHz, occasionally higher for extreme local hour angles. Flux density measurements below 20 mJy approach the noise level of the instruments, and therefore we do not include measurements below this level in our analysis. As with the Ryle data, daily GBI measurements have been averaged when producing monitoring light curves, and GBI data taken within 0.5 days of a pointed [[*RXTE*]{}]{} observation have been averaged when comparing individual observations (Table \[rxte\]). In both cases, we also have computed the spectral index $\alpha$ from the average flux in the 2.25 and 8.3 GHz bands. Very Large Array Observations ----------------------------- The VLA is a multi-frequency, multi-configuration aperture synthesis imaging instrument, consisting of 27 antennas of 25 m diameter. The receivers at 1.42, 5.0, 8.45, and 15.2 GHz, have been used on various occasions, with 2 adjacent bands of 50 MHz nominal bandwidth processed in continuum mode. The corresponding 1-sigma sensitivities in 10 minutes are 0.06, 0.06, 0.05, and 0.2 mJy respectively. Switching between multiple frequencies takes about 30 s. The array configuration is varied every 4 months to cycle between 4 sets, A, B, C, and D, with maximum baselines of about 36, 11, 3.4, and 1 km. Observations of [GRS 1915$+$105]{} while in its radio faint state have been obtained from the public archive in order to constrain the radio spectrum. The observations we have found are listed in Table \[vla\], along with mean fluxes from Ryle and GBI observations taken within a half day of the VLA observations. For all of the observations reported here, [GRS 1915$+$105]{} was unresolved by the synthesized beam (see Table \[vla\]) in any configuration at any wavelength. Other sources in the field of view (mainly at 1.4 GHz) are well separated in the images, so there is no confusion with [GRS 1915$+$105]{}. Calibration and imaging were carried out with standard tasks in the NRAO AIPS (Astronomical Image Processing System) package. For all of these observations, the phase calibrator was 1925$+$211. The primary flux density calibrator was 3C286 (1328$+$307) for all observations except 1996 November 14, when 3C48 (0134$+$329) was used. In practice, the flux density errors are not set by the RMS receiver (thermal) noise stated above, but by errors in the flux density scale, estimated to be 3-5% of the measurement, and/or source variability, depending on the occasion. More details of the VLA are given in the Observational Status Summary (<http://www.aoc.nrao.edu/vla/obstatus/vlas/vlas.html>), and the data reduction is treated in formal detail in Taylor, Carilli, & Perley (1999) and in the VLA cookbook (<http://www.cv.nrao.edu/aips/cook.html>). Pointed [[*RXTE*]{}]{} Observations ----------------------------------- The [*Rossi X-ray Timing Explorer*]{} ([[*RXTE*]{}]{}) consists of three instruments: the All-Sky Monitor [ASM; @lev96], the Proportional Counter Array [PCA; @jah96], and the High Energy Timing Experiment [HEXTE; @roth98]. Figure \[asm\] shows the ASM light curve of [GRS 1915$+$105]{} along with daily radio measurements from Ryle and the spectral index from the GBI. There is no apparent correlation between the radio emission and the ASM count rate, although there is some correlation with the ASM hardness ratio HR2 (5-12 keV / 3-5 keV). In addition, there has been a long campaign of weekly observations with with the PCA and HEXTE, which we use to study the detailed timing and spectral properties of the hard emission from [GRS 1915$+$105]{}. We define intervals as “hard-steady” if the ratio of the mean PCA count rates in the 12–60 keV to the 2–12 keV bands exceeds 0.05 and the mean variability in the count rate (2–60 keV) at 1s time intervals is less than 12% for an [[*RXTE*]{}]{} orbit (typically 3000 s). In Figure \[asm\] we have marked the pointed [[*RXTE*]{}]{} observations of the hard-steady state with vertical bars above the ASM light curve, and we have listed them in Table \[rxte\]. The small dots represent the remainder of the pointed observations. We have analyzed all 113 observations of [GRS 1915$+$105]{} in the hard steady state taken with [[*RXTE*]{}]{} between 1996 April 15 (MJD 50188) through 1999 March 22 (MJD 51259), during the time interval for the PCA instrument gain setting known as “epoch 3”. In order to improve the signal-to-noise of the spectral and timing measurements whenever possible, if the variability in the mean count rate in 16 s bins is less than 5% during a single day, we have analyzed all of the data for that day together. Otherwise, we have grouped the data while remaining within the 5% variability limit (16 s bins), with the smallest unit being a single [[*RXTE*]{}]{} orbit. For each time unit, we have integrated 128-channel energy spectra (“Standard2”) data from PCUs 0 and 1 of the PCA, and 64 channel (archive mode) spectra from both clusters of HEXTE. In order to calculate the flux from GRS 1915+105, we have fit the spectra with the phenomenological model described in @mmr99. The model consists of the sum of a multi-color blackbody, a power law (with a high-energy cutoff when necessary), and a Gaussian line (representing iron emission between 5–7 keV). Interstellar absorption is taken into account assuming a column density equivalent to $6\times10^{22}$ cm$^2$. A normalization factor is allowed to account for differences in the effective area of the PCUs and the HEXTE clusters which are not taken into account by the response matrices. The standard FTOOLS 4.2 background subtraction algorithm for bright sources has been applied to the data, and a systematic uncertainty of 1% has been added to each spectral bin. We have computed an X-ray flux for each observation, by summing the bolometric flux implied by multi-temperature disk model ($F_{bb} = 1.08\times 10^{-11} N_{\rm app} \sigma T_{\rm app}^4$), the flux in the power law between 2–200 keV ($\int_2^{200} 1.60 \times 10^{-9} N_\Gamma E^{-\Gamma+1}\exp[-E/E_{\rm cut}] dE$), and the flux in the iron line ($N_{\rm Gauss}/(\sqrt{2\pi}W_{\rm Gauss})\int_0^\infty \exp[ 0.5(E - E_{\rm Gauss})^2/W_{\rm Gauss}^2 ] dE$). We also have created power density spectra (PDS) and cross spectra (CS) for each 256 s interval of data with $2^{-7}$ s time resolution and four energy channels, using combinations of the binned and event modes for each observation. The four energy channels used are as follows: 2–4.3 keV (henceforth referred to as the “low energy” band), 4.3–7.8 keV, 7.8–11.5 keV, and 11.5–60 keV (referred to as the “high energy” band). The PDS and CS have been averaged for the entire time unit (an [[*RXTE*]{}]{} orbit or an observation, as defined above) and logarithmically re-binned. The PDS have been corrected for dead-time effects and Poisson noise [@mrg97], and have been normalized to the fractional RMS squared per Hz. To quantify the features of the PDS, we have fit PDS from each energy band with a model consisting of a power law with two breaks and Lorentzians for any QPOs present in the PDS. We have searched for the 67 Hz QPO during each of these hard-steady observations using data with $2^{-13}$ s time resolution in a single energy channel, but we did not detect it. Typical upper limits range from 0.1–0.5% (1 $\sigma$). The coherence function and phase lags have been calculated from the CS in the manner described in @vn97. We have used the standard convention that “hard” or “positive” lags indicate that higher energy photons lag behind lower energy photons. The uncertainties have been calculated for the case of high signal power and high measured coherence, using an estimate of the dead-time-corrected Poisson noise [@mrg97]. If the signal power is smaller than the estimated noise, or if the coherence is less than the noise divided by the signal, we do not plot either the phase lags or the coherence functions. Guided by the long-term monitoring light curves in Figure \[asm\], we have defined subsets of the hard-steady observations based upon the radio emission from [GRS 1915$+$105]{} in order to facilitate the comparison between the radio and X-ray emission. We have defined “radio plateau” hard-steady conditions for observations for which [GRS 1915$+$105]{} is brighter than 20 mJy at 15.2 GHz, and exhibits a flat radio spectrum with index (measured from the daily average of the flux) $\alpha > -0.2$. We also have defined “radio steep” conditions for observations when [GRS 1915$+$105]{} is brighter than 20 mJy at 15.2 GHz, and exhibits a optically thin radio spectrum with index $\alpha < -0.2$. Figure \[asm\] demonstrates that the radio steep conditions typically represent the transition into and out of periods of radio plateau emission, and occur during the decays of large optically thin radio flares [see also @fen99b; @dmr00]. Our discussion of radio steep observations will be limited, as the optically thin emission probably results from radio ejecta that could evolve independently of the instantaneous X-ray conditions on 0.5 day time scales. Finally, we have defined “radio-faint” conditions for those hard-steady X-ray observations for which the daily average of the radio flux at 15.2 GHz is less than 20 mJy, because for these observations the radio flux at the GBI frequencies is too low to reliably measure a spectral index. The radio faint conditions occur with two variants— those observations for which the X-ray emission is bright, and those for which it is faint. The behavior of the timing properties with respect to the radio emission is similar for both variants, although the shape of the X-ray spectrum and the values of the timing properties (e.g. the QPO frequency and the phase lags) are distinctive, and warrant a separate discussion. Results ======= We first investigate 1) whether all of the hard-steady observations are coincident with radio emission, 2) whether the X-ray and radio flux are correlated, and 3) how the radio conditions (plateau, steep, and faint) are distributed in a plot of radio vs. X-ray flux. Figure \[rvx\] displays a plot of the radio flux density at 15.2 GHz (log scale) versus the 2–200 keV X-ray flux for each of the hard-steady observations. It is clear that all hard steady observations have radio flux densities greater than 2 mJy. Panel 3 of Figure \[asm\] shows that the radio flux from [GRS 1915$+$105]{} does drop below 1 mJy (e.g. MJD 50825–50875), but such intervals are associated with the soft X-ray state. Thus, [*radio emission always accompanies the hard-steady X-ray state*]{}, although a note of caution is appropriate for those observations associated with variable, optically thin radio emission, as noted above. However, Figure 2 also shows the radio and X-ray fluxes are not correlated in the hard-steady state. Likewise, we have found that the total 50-100 keV flux, the power law flux, and the thermal (multi-color disk) flux are not correlated with the strength of the radio emission (not shown). The radio conditions are represented in Figure \[rvx\] as follows: radio plateau observations are plotted as triangles, radio steep as circles, and radio faint as $\times$’s. Observations for which the GBI spectral index has not been measured are plotted only with error bars. The plot can be roughly split into four quadrants. The radio plateau observations are clearly localized in the upper left quadrant, with $S_{15.2} > 20$ mJy and $F_{\rm X} < 4\times 10^{-8}$ [erg cm$^{-2}$ s$^{-1}$]{}. Some radio steep observations are also present in this quadrant. The upper right quadrant contains radio steep observations and observations for which no spectral index is available from the GBI. These latter observations were taken between MJD 50313–50320 (before the current GBI monitoring program), a time interval which is included in the analysis of @ban98. During this time span a large radio flare occurred terminating a plateau state, and the radio flux density at 1.4 GHz was consistently larger than that at 3.3 GHz, which suggests that these points should be classified as radio steep. Therefore, the upper right quadrant ($S_{15.2} > 20$ mJy and $F_{\rm X} > 4\times 10^{-8}$ [erg cm$^{-2}$ s$^{-1}$]{}) appears to include only radio steep observations, during which the radio and X-ray emission may vary independently since the radio measurements may be dominated by the evolution of an expanding jet detached from the central source. For this reason, we choose to ignore the upper right quadrant. The lower quadrants, with $S_{15.2} < 20$ mJy, contain radio faint observations by construction. After examining the temporal clustering of these radio-faint observations, we feel that it is useful also to consider them in two quadrants: an X-ray faint quadrant with $F_{\rm X} < 4\times 10^{-8}$ [erg cm$^{-2}$ s$^{-1}$]{}, and an X-ray bright quadrant with $F_{\rm X} > 4\times 10^{-8}$ [erg cm$^{-2}$ s$^{-1}$]{}. This separation is further motivated by their differing energy spectral and timing properties, which are described in the next section. Characteristic Observations --------------------------- From each of the three quadrants under consideration from the radio versus X-ray flux plane (Figure \[rvx\]), we have selected a characteristic observation in order to illustrate the range of X-ray spectral and timing properties within the hard-steady X-ray state. In Figures \[spec\], \[pdsetc\], and \[vsspec\], the left panel is a radio faint, X-ray faint observation (MJD 50488; 1997 February 9); the center panel is a radio plateau, X-ray faint observation (MJD 50737; 1997 October 16); and the right panel is a radio faint, X-ray bright observation (MJD 50708; 1997 September 17). The radio and X-ray fluxes from these three observations are indicated in Figure \[rvx\]. ### X-ray Energy Spectra Figure \[spec\] illustrates the results of the spectral fits for the three representative observations with differing values of the radio and X-ray flux. The spectrum in the left panel is from a radio faint observation when the X-ray emission was also relatively faint ($2.2\times10^{-8}$[erg cm$^{-2}$ s$^{-1}$]{}). The soft emission during this observation can be modeled ($\chi^2_\nu \sim 1$) using a cool multi-temperature blackbody with an apparent inner disk temperature of $T_{\rm app} \simeq 0.8$ keV and an apparent radius of $R_{\rm app} \simeq$40–80 km (assuming a distance of 11 kpc and an inclination angle of 66$^\circ$; see Fender et al. 1999b). The power law has a photon index $\Gamma \simeq 2$ and requires an exponential cutoff at about 70 keV. The results of this spectral fit are characteristic of observations during which both the X-ray and radio flux were low. The spectrum in the middle panel is from an observation which exhibited radio plateau conditions when the X-ray emission was also faint ($2.6 \times10^{-8}$ [erg cm$^{-2}$ s$^{-1}$]{}). Despite the similarity in X-ray flux to the left panel, the shape of the X-ray spectrum is quite different. The thermal component of the model appears hot ($T_{\rm app} > 3$ keV) with a small apparent inner radius ($R_{\rm app} < 5$ km). This quasi-thermal component is not consistent with simple models of reflection from an accretion disk [e.g. @mz95]. It is possible that this broad thermal bump represents emission from a relatively small area of the disk which has been heated by, for example, a magnetic flare (di Matteo, Celotti, & Fabian 1999) or a spiral shock wave [@tp99], or that the disk develops an optically thick atmosphere which Comptonizes the emergent spectrum, so that the spectral hardening and decreased apparent radius are illusory effects of increased scattering in a “puffed up” disk (Merloni, Fabian, & Ross 2000). The photon index of the power law is $\Gamma \simeq 2.7$, and a cut-off at about 80 keV is required for observations with longer integration times and hence better statistics at energies above 100 keV. We note that @rao00 have presented a different model for spectra of [GRS 1915$+$105]{} such as these, consisting of a multi-temperature disk, a thermal-Compton spectrum [@st80], and a power law. For the purposes of this paper, we prefer to use our simpler model with fewer free parameters. The right panel is from an observation when the radio emission was faint (6 mJy) and the X-ray emission was bright ($8.4\times10^{-8}$ [erg cm$^{-2}$ s$^{-1}$]{}). The spectrum is similar to the radio bright, X-ray faint observation in the center panel, except that the relative contribution of the power law flux increases dramatically. The thermal component is hot ($T_{\rm app} > 2.5$ keV) and has a small apparent inner radius ($R_{\rm app} < 10$ km), while the power law index is steep ($\Gamma \simeq$ 3.2). The shape of the energy spectrum on 1997 September 17 is characteristic of all X-ray spectra from the hard state of [GRS 1915$+$105]{} when the total flux is greater than $\simeq 4\times10^{-8}$ [erg cm$^{-2}$ s$^{-1}$]{}. The count rate from [GRS 1915$+$105]{} tends to vary on time scales of hours when the source is bright in X-rays, so we only integrate spectra for individual [[*RXTE*]{}]{} orbits. With these exposure times, we can measure the power law only up to about 100 keV. However, when we combine spectra with similar photon indices, we find that the data are consistent with no cut off in the power law below 200 keV. The changes in the X-ray spectrum clearly do not track the radio flux from [GRS 1915$+$105]{} in a monotonic manner. The power law component steepens dramatically as the radio emission increases in the X-ray faint state (left and center panels of Figure \[spec\]), but is steepest when the X-ray emission is bright and the radio emission is faint (right panel). At low X-ray fluxes ($< 4\times10^{-8}$ [erg cm$^{-2}$ s$^{-1}$]{}), the thermal component is hot and has a small apparent inner radius during radio plateau observations, while it is cool and has a large apparent inner radius when the radio emission is faint [see also @mmr99]. ### Power Density Spectra and Cross Spectra The power spectra (here displayed as the frequency times the power density), Fourier phase lags, and coherence functions of the three observations from Figure \[spec\] are displayed in Figure \[pdsetc\]. In the top panel, two lines are plotted for each power spectrum: the black line indicates the 2–4.3 keV (soft) energy band, while the the grey line indicates the 11.5–60 keV (hard) band. The PDS in all three examples of Figure \[pdsetc\] have relatively flat spectra with a low frequency break, a QPO between 0.5–10 Hz, and a dramatic break at frequencies higher than the QPO. In the middle panel of Figure \[pdsetc\], we plot the phase lags between the 11.5–60 keV and the 2–4.3 keV bands, and in the bottom panel we plot the coherence function for the same two energy bands. To simplify the comparison between the power spectra, the phase lags, and the coherence functions, we have indicated the frequencies of the QPOs in our fits to the 2–4.3 keV band with vertical black short-dashed lines, and the continuum break frequencies with vertical black long-dashed lines. We have also indicated the break frequency in the 11.5–60 keV band with grey long-dashed lines (see below). For the following discussion we define three frequency regions related for the features of the power spectrum: region 1 represents frequencies lower than the low frequency continuum break, region 2 frequencies between the low frequency break and the QPO frequency (less its width), and region 3 frequencies higher than the QPO frequency (or its harmonic if present). The first point of interest is that the QPO frequency is not strictly correlated with the X-ray flux from [GRS 1915$+$105]{}. From the spectra in Figure \[spec\], it is clear that the X-ray flux increases from left to right, while in Figure \[pdsetc\] the QPO frequency is lowest in the center panel (radio plateau conditions). In the following sections, we will demonstrate that the QPO frequency is always lowest during radio plateau conditions. The second and most obvious difference between the timing properties of these three observations is in the Fourier phase lags. In region 2 (between the low frequency break and the QPO), the phase lags are negative when the radio emission is faint (left and right panels), and positive during radio plateau emission [center panels; compare @rei00]. This behavior is in sharp contrast to sources such as [GX 339$-$4]{} and [Cyg X$-$1]{}, for which the phase lags are always positive. Only the phase lags during the radio plateau state resemble those of [GX 339$-$4]{} [@nwd99] and [Cyg X$-$1]{} [@cui97]. The third notable result concerns the coherence function in region 1 (i.e. frequencies lower than the first break). The coherence function between the 11.5–60 keV and the 2–4.3 keV bands is lowest ($\simeq 0$) during the radio plateau conditions (center panel), somewhat higher ($\simeq 0.4$) when both the radio and X-ray emission are faint (left panel), and nearly unity when the radio emission is faint and the X-ray emission is bright (right panel). Only the latter case is reminiscent of [Cyg X$-$1]{} in its hard state [@cui97; @now99]. Finally, a comparison of the amplitude and strength of the power spectra as a function of energy reveals further differences in each of the three observations. While the radio emission is faint, (right and left panels) the root mean power at low frequencies is greater in the 11.5–60 keV energy band (gray curve) than in the 2–4.3 keV band (black curve). However, during radio plateau conditions (center panel) the root mean power is nearly equal in all energy bands. The shape of the power spectrum also changes— the low frequency break is strongest in the low energy band in the radio plateau observation (center panel), while the break is most apparent in the high energy band in the radio faint, X-ray faint observation (left panel). When the X-ray emission is bright, the low frequency break is evident in both energy bands (right panel). There are several other interesting details in the timing properties presented in Figure \[pdsetc\], although they are not directly relevant to the question of how the timing properties are related to the radio emission. The harmonic of the QPO tends to be strongest ($\simeq 10$%) when count rate is lowest (left and center panels), but is not apparent during X-ray bright observations (right panel). A weak feature at approximately half of frequency of the strongest QPO peak is also evident during observations when both the X-ray and radio flux are faint (right panel), with an RMS amplitude of about $\simeq$1% and $Q$ values of 1–10. Finally, a weak QPO is evident between 0.01–0.04 Hz during a stretch of radio faint observations from MJD 50450–40560 (left panel), with an RMS amplitude of about 0.1% in the 2-4.3 keV energy band. This low frequency QPO disappears at higher energies. Its centroid frequency decreases as the frequency of the 0.5–10 Hz QPO decreases (not shown). Since the 0.01–0.04 Hz QPO is present only in the low energy band, the coherence function at the QPO frequency is near zero. ### Relation of the Timing Properties to the Energy Spectrum The spectral fits which we have made to [GRS 1915$+$105]{} indicate that two spectral components are present (Figure \[spec\]). Most models of the emission from X-ray binaries invoke inverse-Compton scattering to produce hard photons from soft seed photons, and consequently the variability in the hard and soft energy bands should be intimately related. However, before finalizing any conclusions regarding the phase lag and coherence measures and how these change with photon energy, we need to examine how these timing properties depend upon the amount of the flux which is contributed by each of the spectral components in the relevant energy bands. In Figure \[vsspec\], we examine the dependence of the timing properties on photon energy for the same three representative observations as in Figures \[spec\] and \[pdsetc\] (diamonds connected by solid lines). In all of the panels, the fraction of the count rate contributed by the disk component is displayed with triangles connected with dashed lines, using the scale displayed on the right axis. Here it can be seen that the thermal component contributes no more than half of the flux in all energy bands for the observations which we analyze in this paper. Examining next the top panels, we find that the presence of the thermal component does not greatly affect the RMS power in a given energy band. In the second panels, the slopes of the phase lags show no obvious breaks; we might expect changes in the slopes if the soft and hard spectral components separately affected the phase lags. Finally, while the fraction of the flux in the thermal component increases in a similar manner during the X-ray faint, radio plateau (center panel) and the X-ray faint, radio bright observation (left panel), the coherence function at low frequencies plummets dramatically with increasing energy in the plateau observation (center panel), but is near unity when the X-ray flux is bright (right panel). We conclude that the changes in the timing properties are not related to the fraction of the emission which can be modeled with a thermal component. Time Evolution of X-ray and Radio Properties -------------------------------------------- The evolution of the parameters of the power spectrum, phase lags, and coherence function can be viewed as a function of time during the long periods in which [GRS 1915$+$105]{} remains in the hard-steady state. The results demonstrate that the three observations selected for Figures \[spec\]–\[vsspec\] are representative of the groups which we have defined. In Figure \[lhev\] we show two time intervals containing the hard-steady states of [GRS 1915$+$105]{}. In the first the radio emission is faint (MJD 50400–50600, left panels), and the second coincides with strong radio plateau emission (MJD 50725–50755, right panels). In the top three panels from both time intervals we plot: a) the intensity of [GRS 1915$+$105]{} as viewed by the ASM, b) the radio flux measured with the Ryle telescope at 15.2 GHz, and c) the radio spectral index between 2.25 and 8.3 GHz ($\alpha$) measured with the GBI. Notice that before the period of radio-plateau emission (right panels) the radio spectral index increases smoothly and becomes positive (more optically thick), and then it decreases symmetrically afterwards. Optically thin radio flares that are much more prominent at 2.25 and 8.3 GHz (not shown) occur both before and after this interval. In panel d), we plot the frequency of the 0.5-10 Hz QPO. @tcg99 and @cts97 separately established that the QPO frequency tracks the flux and the count rate in the PCA during hard-steady states similar to those displayed. However, it is clear that QPO frequency is systematically higher during radio faint emission ($> 2$ Hz) than during radio plateau emission ($< 2$ Hz), despite the fact that the ASM count rate is lower during the faint radio emission. In panel e) of Figure \[lhev\] we plot the average phase lags in region 2 between the 2–4.3 keV band and 1) the 7.8–11.5 keV band (diamonds connected by the solid line), and 2) the 11.5–60 keV band (triangles connected by the dashed line). During radio-faint hard-steady conditions, the phase lags are consistently negative and have a value of about $-0.2$ radians. During radio plateau emission, the phase lags are positive and increase strongly with energy. Only during radio plateau conditions are the X-ray phase lags positive, as might be expected from a simple model which seeks to explain the timing properties using Comptonization (see Section 3). In panel f) of Figure \[lhev\] we plot the average coherence function in region 1 between the 2–4.3 keV band and 1) the 7.8–11.5 keV bands (diamonds connected by the solid line), and 2) the 11.5–60 keV bands (triangles connected by the dashed line) as a function of time. The coherence over this frequency range drops dramatically when the radio emission is strong and optically thick, as shown in the representative observation of the radio plateau state in the center panel of Figure \[pdsetc\]. In the bottom panel we plot the total of the RMS noise at frequencies lower than the QPO (region 1 and region 2 combined). The absolute values of the RMS low frequency noise should not be compared directly from observation to observation, because as the QPO frequency varies, so does the upper limit to the frequencies which we consider. However, we have plotted the RMS low frequency noise for two energy bands: the 2–4.3 keV band (diamonds connected by a solid line) and the 11.5–60 keV band (triangles connected by dashed lines). It is clear that there is much more low frequency power in the 11.5–60 keV band than in the 2–4.3 keV band during radio-faint hard X-ray conditions, while during radio-plateau conditions, the amount of power in both energies is comparable. This can also be seen by comparing the radio plateau and X-ray-faint, radio-faint observations in Figure \[pdsetc\]. Timing Properties and the Radio States -------------------------------------- We now examine directly the relationship between the radio and the X-ray properties of [GRS 1915$+$105]{}. In Figures \[ratfl\] and \[qpocnt\], the various symbols denote the emission states in Figure \[rvx\]. In Figure \[ratfl\] we plot as a function of the radio flux density at 15.2 GHz: a) the QPO frequency, b) the phase lag between the 2–4.3 keV and 11.5–60 keV bands in region 2, c) the coherence for the same energy bands in region 1, and d) the ratio of the low-frequency (regions 1 and 2) power in the 11.5–60 keV to 2–4.3 keV energy bands. At first glance, there appear to be continuous distributions in the radio and timing properties in Figure \[ratfl\], but in fact the populations of radio faint observations and radio plateau observations are well-separated, with the radio steep observations falling in between the two. Moreover, the $x$-axes of the panels are plotted with a logarithmic scale, which emphasizes the large range in the radio flux for the various hard-steady observations. Four conclusions can be drawn from Figure \[ratfl\] regarding the changes in the timing properties as a function of the radio flux. The QPO frequency tends to decrease with increasing radio flux, and almost all of the observations with radio plateau conditions (triangles) have QPO frequencies less than 2 Hz. Second, during radio plateau hard-steady emission the hard photons [*lag*]{} the soft, while during radio faint ($\times$’s) conditions the hard photons [*lead*]{} the soft. Third, the coherence at low frequencies is lowest during radio plateau emission. Finally, the ratio of low frequency power in 11.5–60 keV photons to the power in 2–4.3 keV photons is nearly unity only when the radio flux is largest, during radio plateau conditions. Using our large dataset which incorporates a variety of radio conditions, we next examine correlations which other authors have reported to be well-defined. Two studies [@cts97; @tcg99] have examined the frequency of the 0.5–10 Hz QPO as a function of the PCA count rate (MJD 50275–50333) and the 3–20 keV X-ray flux (MJD 50363–50563), respectively. Both have found significant correlations among these values. In the top panels of Figure \[qpocnt\] we plot the QPO frequency as a function of total X-ray flux (defined in Section 2.3; left panel), thermal flux (middle panel), and power law flux (right panel). We confirm that the frequency of the 0.5–10 Hz QPO is tightly correlated with the X-ray flux from [GRS 1915$+$105]{}, but we also find that the correlations are not single-valued. In the left panel there exists a significant population of points from radio faint observations that have QPO frequencies between 2–5 Hz, for which the X-ray fluxes are low compared to those expected from the general trend. A plot of the average phase lags between 2–4.3 and 11.5–60 keV photons in region 2 as a function of the total, thermal, and power law fluxes (bottom panels of Figure \[qpocnt\]) reveals that there is no correlation between the X-ray flux and the sign or magnitude of the phase lags. The positive phase lags therefore appear to occur independently of any given X-ray flux level, and are better predicted by the presence of optically thick radio emission. Characteristics of the Faint Radio Emission ------------------------------------------- We can use the archival VLA observations to make some conclusions about the nature of the faint radio emission. The multi-frequency radio observations of [GRS 1915$+$105]{} are listed in Table \[vla\]. The first two observations occur within a day of radio faint, X-ray bright observations. In both cases, the VLA observed significant flaring, while Ryle observations, which were more nearly simultaneous with the [[*RXTE*]{}]{} observations, indicate that the radio emission was near 5 mJy within 12 hours of the PCA observations. It appears that the radio emission is unsteady within several hours of the X-ray bright hard-steady state, and we can not rule out that the X-ray emission is also unsteady on these time scales. Since the VLA observations show clear evidence for flaring while the Ryle observations do not, we can not infer the radio spectrum of the emission during the X-ray bright, hard-steady state. The remainder of the observations in Table \[vla\] are X-ray faint. There is clear evidence in the GBI data for $\simeq$30 mJy optically thin flares on 1996 December 28, 1997 January 12, and 1998 October 08. The flares decay on time scales of several hours, and no [[*RXTE*]{}]{} observations were taken within 0.5 day of these flares. Since GBI flux measurements below 20 mJy are uncertain, there remain two observations during which the VLA observations demonstrate that the radio emission was faint and steady. On 1998 September 14, 15.2 GHz Ryle and 5.0 GHz VLA observations separated by 6 hours provide a spectral index $\alpha = -0.5\pm0.1$, which suggests that the emission is optically thin. However, it is certainly possible that the radio flux varied by a few mJy during those several hours. Dual frequency VLA observations were made at 5.0 and 15.2 GHz on 1998 September 29, which indicate $\alpha = 0.12\pm0.01$. This suggests that the emission is optically thick, similar to a weak version of the radio plateau state. However, with only a single secure measurement of $\alpha$ during the radio faint state, more observations are certainly warranted. Discussion ========== We begin the discussion with a summary of our three hard-steady conditions: - [**Radio Faint**]{} (e.g. MJD 50450–50550; MJD 50708) The radio emission has a mean flux density of about 5 mJy at 15.2 GHz. When the X-ray flux is faint ($< 4 \times 10^{-8}$ [erg cm$^{-2}$ s$^{-1}$]{}), the X-ray spectrum is best described as the sum of a cool ($T_{\rm app}<$2 keV, $R_{\rm app} \sim 60$ km) multi-temperature blackbody and a power law of index $\Gamma \simeq 2.1$ with a cut-off near 70 keV. When the X-ray emission is brighter ($> 4 \times 10^{-8}$ [erg cm$^{-2}$ s$^{-1}$]{}), the X-ray spectrum is best described with a hot ($>$2 keV) multi-temperature black body and a steep power law ($\Gamma \simeq 3.2$) without a cut off below 200 keV (see Section 2.2). The QPO frequency is greater than 2 Hz, phase lags are negative (or, soft, indicating that hard photons precede soft photons) at intermediate continuum frequencies, the coherence at low frequencies in the continuum is $\simeq$ 0.5–1.0 (11.5–60 keV compared to 2–4.3 keV), and there is much more low frequency power in the high energy band than the low energy band. - [**Radio Plateau**]{} (e.g. MJD 50730–50748) The radio emission is bright ($\sim$100 mJy at 15 GHz) and optically thick ($\alpha > -0.2$). The X-ray spectrum is best described by a hot multi-temperature black body ($T_{\rm app} > 3$ keV and $R_{\rm app} < 5$ km) and a power law ($\Gamma \simeq 2.5$) which could be cut-off at $E > 80$ keV. The QPO frequency reaches its lowest values ($<2$ Hz), phase lags are positive (hard photons lag soft photons), the coherence is low (particularly at low frequencies), and there is about equal broad band RMS power at low frequencies in all energy bands. - [**Radio Steep**]{} (e.g MJD 50727-50730) These observations represent the transition into and out of radio plateau conditions. The radio flux is bright, while the radio spectrum is optically thin ($\alpha < -0.2$). The X-ray flux is often higher than during radio plateau conditions, but the spectrum is still best described by a hot multi-temperature black body ($T_{\rm app} > 3$ keV and $R_{\rm app} < 10$ km) and a steep power law ($\Gamma \simeq 2.5$) with a cut-off above 80 keV. The timing properties tend to be similar to radio faint conditions (although there are exceptions): the QPO frequency is between 2–5 Hz, the phase lags are negative, the coherence is around 0.5, and there is more broad band power at low frequencies in the highest energy band. Previous observations have established the nature of the bright radio emission from [GRS 1915$+$105]{}. The optically thick emission during plateau conditions has been resolved as a compact jet of relativistic electrons [@dmr00], similar to ones which may be present in [Cyg X$-$1]{} [@sti98] and [GX 339$-$4]{} [@cor00]. Our observations of the faint radio emission provide some evidence that it is similar to a weak radio plateau state. Radio flares with steep spectra originate from material which has been ejected from the central source [@fen99b; @dmr00], and so is most likely decoupled from the instantaneous conditions of the X-ray emitting regions. In the following discussion we will focus on comparing the radio faint and radio plateau conditions, during which the radio and X-ray properties of [GRS 1915$+$105]{} are closely related. The timing properties of [GRS 1915$+$105]{}— the QPO frequency, phase lags, coherence function, and broad band noise— appear to be highly dependent upon the radio conditions (Figure \[ratfl\]). However, it is not immediately obvious whether the changes in the timing properties are part of the mechanism which produces a compact jet, or whether they result from changes in the accretion flow induced when the compact radio jet is strong. Since the location and origin of the Comptonizing electrons which generate hard X-rays is unknown, the possible connection between these and the jet is doubly uncertain. Nevertheless, we may explore interpretations of these results using simple models for the accretion flow under consideration in the literature. In Section 4.1, we consider the extent to which Comptonization of a soft input signal by a static corona at the base of a compact radio jet is consistent with the phase lags and X-ray spectrum of [GRS 1915$+$105]{}. In Sections 4.2–4.3, we examine two generic models which incorporate more elements of the accretion flow. In both models, cool ($T \sim 1$ keV) thermal emission originates from an optically thick accretion disk, while the hot power law component originates from inverse Compton scattering. The first basic model places the Comptonizing electrons in a spherical corona within the inner radius of the accretion disk, while the second geometry assumes that the relativistic electrons are part of a planar corona sustained by magnetic flares above the disk. Generic Comptonization ---------------------- Hard phase lags are generic properties of the low-hard X-ray state of black hole candidates such as [Cyg X$-$1]{} [@now99] and [GX 339$-$4]{} [@nwd99]. Surprisingly, the phase lags in [GRS 1915$+$105]{} change from negative to positive as the strength of the radio emission increases (Figure \[ratfl\]b). Since @fen99a have suggested that the Comptonizing corona in the hard state of black hole candidates represents the base of a compact radio jet, we feel it is important to explore the possibility that the negative phase lags during radio faint observations of [GRS 1915$+$105]{} are intrinsic to the region close to the black hole, while the hard phase lags generated during the radio plateau state are due to the Comptonization of the intrinsic input signal in a large corona at the base of the compact radio jet. The hard phase lags in [GRS 1915$+$105]{} may constrain the structure of the Comptonizing region, because the phase is roughly constant as a function of frequency (Figure \[pdsetc\]). This is contrary to the constant time lags that would be expected from a homogeneous corona. Instead, @kht97 and Böttcher & Liang (1999) have demonstrated that constant phase lags can be produced in an isotropically illuminated, spherical Compton cloud with a uniform temperature and a density which decreases with radius as $R^{-1}$, since photons have an equal probability of scattering per decade of radius in such a cloud (see Nowak et al. 1999b for further discussion of these models). High-frequency variability originates from photons which have scattered from small radii, since the signal is washed out if it scatters with long time delays. On the other hand, the time lags of low-frequency signals are dominated by photons which have scattered over large radii. This produces time lags which decrease as a function of frequency, or phase lags which are nearly constant as a function of frequency. Several authors have pointed out that the phase lags which are observed during the hard states of black hole candidates imply large scales for inverse Compton scattering, which raises the question of how energy is transported to support a large, hot corona [@pfa99; @now99b]. In the plateau state of [GRS 1915$+$105]{} a phase lag of 0.5 radians at 0.3 Hz (middle panel of Figure \[pdsetc\]) would imply a light travel time of 0.3 s, or a distance of $R\tau \simeq 8\times10^9$ cm (where $\tau$ is the optical depth of the corona). For a 10 [M$_{\odot}$]{} black hole, this represents thousands of Schwarschild radii. However, this distance is much smaller than the size of the compact jet observed by Dhawan et al. (2000), 10 AU $= 5\times10^{14}$ cm, which is consistent with the assumption that the Comptonizing electrons reside at the base of the jet. If the Compton cloud is the base of a compact jet, this reduces the difficulty in explaining how energy is transported to support a very large corona, as the jet may carry up to 10$^{39}$ erg s$^{-1}$ of power to much larger radii [@fb99]. It is possible that the radio faint states are associated with a smaller Compton corona located near the inner accretion disk. In this case, negative phase lags could be caused by any number of effects intrinsic to the accretion disk, such as waves similar to those of @now99 which propagate outward in the disk rather than inward, or a modified form the of the magnetic flares of @pfa99. Even if a scattering corona intercepts a significant fraction of the photons from the disk, negative phase lags could be observed so long as 1) the scattering time in the corona is much smaller than the intrinsic time lags [@mil95] and 2) the differences between the energies of the seed photons and the observed signals are not too large (which should be the case in [GRS 1915$+$105]{}, since the apparent disk temperature is $> 0.8$ keV; Nowak & Vaughan 1996). Moreover, the radius of the Compton corona can change without greatly affecting the slope and cut-off energy of the power law. The electron temperature ($kT_e$) of the corona is constrained to be nearly constant by the cut-off in the power law at 70–80 keV measured during X-ray faint observations (left and center panels of Figure \[spec\]). If the temperature of the corona is fixed, a steeper power law (higher $\Gamma$) implies a lower optical depth [see @st80]. Since the power law is steeper during radio plateau observations ($\Gamma \simeq 2.5$) than during the radio faint, X-ray faint observations ($\Gamma \simeq 2.1$), we can suggest that the optical depth of the corona decreases and its radius increases as the compact radio jet becomes more luminous. However, the X-ray bright, radio faint observations would then have a small, optically thin corona, since the phase lags are negative and the power law is steep ($\Gamma \simeq 3.2$) during these observations (Figures \[spec\] and \[pdsetc\]). The electron acceleration mechanism could be different in the X-ray bright hard state, as is hinted at by the absence of an observable cutoff in the power law below 200 keV. As others have noted, estimates of the Lorentz factors for electrons emitting synchrotron emission [$\sim 400$; @fb99] suggest energies that are much larger than that of the Comptonizing electrons. One may consider the acceleration mechanism for the jet as an undetermined structure that draws material from the Compton corona near the inner disk, as in @fen99a. In the following two sections, we therefore briefly speculate on the implications which the simplest geometries for the Compton corona described in the literature have on the jet production mechanism. A Spherical Corona within the Inner Accretion Disk -------------------------------------------------- Many authors have suggested that a spherical corona is formed within the inner radius of an accretion disk, either as an advection-dominated accretion flow (ADAF; Esin, McClintock, & Narayan 1997), or by a shock in the accretion flow [@ct95]. The compact jet responsible for the radio emission could form from a wind emanating from an ADAF [@bb99] or a post-shock region [@cha99]. The distinctive feature of the spherical corona model is that the radius of the inner accretion disk is variable, which leads naturally to a scenario in which the X-ray timing and spectral properties, and the strength of the radio emission, are set by the inner radius of the accretion disk. The X-ray timing property that is perhaps most relevant to discuss in the spherical corona model is the 0.5–10 Hz QPO. The QPO may represent a characteristic time scale in the accretion flow such as an acoustic [e.g. @ct94; @act95], Keplerian [@rod00], or free-fall (Molteni, Sponholz, & Chakrabarti 1996) time scale. Since the QPO peak is narrow ($Q > 3$), it must originate in a localized region, as all of the time scales increase as the radius increases. Among the possibilities for the origin of such a QPO are oscillations in a shock in the accretion flow [@cm00] or a spiral density wave in the inner disk [@rod00]. Since the frequency of the QPO decreases during episodes in which the inner radius of the disk is thought to increase [@bel97b; @mst99], it seems plausible that the frequency of the 0.5–10 Hz QPO in [GRS 1915$+$105]{} tracks the inner radius of the accretion disk. Figure \[ratfl\]a then implies that the radio emission is strongest when the inner disk is farthest from the source. Nobili et al. (2000) have developed a model for the phase lags in [GRS 1915$+$105]{} by assuming a spherical corona with two temperature regions: a warm ($T \simeq 1.5$ keV) outer corona and a hot ($T \simeq 15$ keV) inner corona. The corona is isotropically illuminated by soft seed photons. The sign and magnitude of the phase lags are correlated with the radius of the inner accretion disk, because as the disk moves inward it compresses the corona, increasing its optical depth. Large, positive phase lags occur when the inner disk radius is large, because the hot corona has an optical depth $\tau_{\rm H} \sim 1$, while the warm corona is optically thin. When the inner radius is a factor of $\sim 3$ smaller, the optical depth of the hot corona becomes extremely high ($\tau_{\rm H} > 100$), and the soft seed photons from the disk thermalize to the temperature of the optically thick hot corona. The thermalized photons are then Compton down-scattered in the cool corona, which produces soft lags. This model does not attempt to explain the phase lags as a function of Fourier frequency, but it does predict that if the QPO frequency tracks the radius of the inner disk, negative phase lags are expected when the QPO frequency is high— exactly as Reig et al (2000) observed in [GRS 1915$+$105]{}. However, a corona with an optical depth of $\tau > 100$ would produce a thermal spectrum, while the energy spectrum of [GRS 1915$+$105]{} between 25–200 keV can not be described by an extremely optically thick spectrum (e.g. a spherical Compton cloud with optical depth $\tau > 5$; Sunyaev & Titarchuk 1980) when the phase lags are negative. Further work needs to be done to reconcile the negative phase lags with the spectral shape under this type of model. The interpretation of the coherence function and the continuum power as a function of energy is less certain. When the QPO frequency is low, the coherence at low frequencies is nearest to 0 (Figure \[ratfl\]c) and the ratio of the low frequency continuum power at high to low energies is near 1 (Figure \[ratfl\]d). A coherence less than 1 can occur if the variability is the superposition of separate linear signals with different properties [@vn97]. When the inner disk is farthest from the source (low QPO frequency) there may be more area available for multiple inputs to contribute to the variability. The radius of the inner disk should also determine the energy spectrum under a spherical corona model. We have already discussed the model of @nob00. A different set of assumptions are used by, for example @emn97 and @ct95, who take into account the thermal balance between the soft photons and the Comptonizing electrons. Soft spectra are formed when the inner radius of the disk is close to the last stable orbit, so that the disk produces copious amounts of seed photons which cool the corona. Hard states are formed when the inner radius of disk is far from the last stable orbit and most of the corona is photon-starved. However, if we consider the faint X-ray observations, the power law is steeper ($\Gamma \simeq 2.5$) when the QPO frequency is lower and the disk is inferred to be farther from the black hole (radio plateau observations; Figure \[spec\] and \[pdsetc\]). This is opposite the sense expected, because the Compton corona should be cooled less efficiently (which implies a lower $\Gamma$) when the inner disk is large. If the QPO frequency tracks the transition radius between the disk and the corona, there must an additional factor beside the radius of the inner accretion disk which affects the observed X-ray properties of the corona in [GRS 1915$+$105]{} under this model. A Planar Corona Sustained by Magnetic Flares -------------------------------------------- In a planar corona model, one can assume that magnetic flares powered by the differential rotation of the accretion disk produce a population of relativistic electrons above the disk (di Matteo et al. 2000; compare Dove et al. 1997 for difficulties with static models). The compact jet could represent a collimated wind which forms according to the magnetic-centrifugal mechanism of @bp82, and is strongest when the flares provide electrons enough energy to escape the potential of the black hole [@rom98]. In the magnetic flare model of @dim99, the inner radius of the accretion disk in general does not vary. A natural starting point for discussing the planar corona models are the phase lags observed in [GRS 1915$+$105]{}. The time lags which would be implied by the scale height of the flares in the model of @dim99 are too short ($10^7$ cm $= 10^{-3}$ s) to explain the phase lags observed in black hole candidates. However, @pfa99 have developed a model for [Cyg X$-$1]{} in which the spectral evolution of the flares which accelerate Comptonizing electrons produce the observed power density spectra and phase lags. Their model produces only hard (or zero) lags, because the electrons are assumed to be in thermal equilibrium during the flare, so the spectrum of the flares is set by the ratio of the heating in the flares to the amount of soft seed flux impinging on the flares. The flares become harder because the majority of the soft flux is produced by feedback in response to the flare, and the feedback is taken to decrease as a function of time as the flares move away from the disk. However, if the electrons accelerated in the magnetic flare have a non-thermal energy distribution which evolves as a function of time, it is conceivable that either hard or soft phase lags could be produced. The hard phase lags may represent the most efficient acceleration of the electrons, which could in turn explain why the radio plateau emission is associated with the largest hard phase lags. On the other hand, the negative phase lags could represent spectral evolution in which the soft energy band peaks later, and relativistic electrons are accelerated less efficiently. In the case of a magnetically sustained planar corona with a fixed inner disk radius, there is some difficulty in explaining what sets the frequency of the QPO and why this frequency varies. Since the power in the X-ray variability in [GRS 1915$+$105]{} drops sharply at frequencies higher than the 0.5–10 Hz QPO (Figure \[pdsetc\]), we can speculate this QPO represents a resonance at the minimum time scale at which magnetic flares can form. The radio emission, then, would occur most favorably when only relatively slow flares form (low QPO frequency). The coherence function and the ratio of the high to low energy broad-band power would also be set by the evolution of the magnetic flares. Since the coherence at low frequencies is lowest when the QPO frequency is low and the phase lags are positive, the long, hard flares may undergo significant non-linear spectral evolution. The ratio of broadband power in the high and low energy bands could be set by the shape of individual flares in the respective energy bands. Mathematically, this could be described in a manner similar to the shot noise models of @shi88. The model of @dim99 predicts the energy spectrum from magnetic flares, by assuming that a fixed fraction of the accretion power is released into the magnetic structures, and that the electrons which are accelerated by the flares are cooled inverse-Compton scattering of both cyclo-synchrotron and thermal (disk) seed photons. Soft spectral states are caused by flares which occur close to the disk and are flooded by seed photons, while hard states are caused by photon-starved flares which occur high above the disk. Although @dim99 do not attempt to explain the phase lags observed in black hole candidates, we might expect hard phase lags to correspond with flares that produce the hardest (lowest $\Gamma$) power law, since the flares would be highest above the disk and cooled least efficiently. However, when the phase lags are positive in [GRS 1915$+$105]{} we find a steeper power law ($\Gamma \simeq 2.5$ in the plateau state) than when the phase lags are negative ($\Gamma \simeq 2.1$ during radio faint, X-ray faint observations). Clearly, simple assumptions about the evolution of magnetic flares would have to be relaxed in order to explain the X-ray timing and spectral properties of [GRS 1915$+$105]{}. Conclusions =========== We have found that although radio emission is always associated with the hard-steady state of [GRS 1915$+$105]{} (between 1996-1999), the 15.2 GHz radio flux density and the 2–200 keV X-ray flux are not correlated. Instead, we find that the hard states can be described using three basic regimes of radio emission: a) radio plateau conditions in which bright, optically thick emission originates from a compact jet, b) radio steep conditions in which bright, optically thin emission originates from material most likely decoupled from the X-ray emitting region, and c) faint radio conditions for which our current study is unable to constrain the spectrum of the radio emission. As the radio flux increases, 1) the frequency of a ubiquitous 0.5–10 Hz QPO decreases, 2) the Fourier phase lags between hard (11.5–60 keV) and soft (2–4.3 keV) photons in the frequency range of 0.01–10 Hz change sign from negative to positive, 3) the coherence at low frequencies decreases, and 4) the relative amount of low frequency power in hard photons compared to soft photons decreases. We have attempted to understand these results qualitatively using simple models collected from the literature, but we find that our attempts to understand the X-ray timing properties of the hard state of [GRS 1915$+$105]{} inevitably lead to contradictions with the observed energy spectrum. These observational results therefore provide impetus for developing more sophisticated models of the accretion flow and the compact jet in [GRS 1915$+$105]{}. Radio astronomy at the Naval Research Laboratory is supported by the Office of Naval Research. The Green Bank Interferometer is a facility of the National Science Foundation, operated by the National Radio Astronomy Observatory in support of the NASA High Energy Astrophysics programs. The NRAO VLA is a facility of the National Science Foundation, operated under Cooperative Agreement by Associated Universities, Inc. This work was also supported in part by NASA contract NAS 5-30612. Robert Hjellming passed away in July 2000, and his contributions to astrophysics will be sorely missed. Abramowicz, M. A., Chen, X., & Taam, R. E. 1995, , 452, 379 Bandyopadhyay, R., Martini, P., Gerard, E., Charles, P. A., Wagner, R. M., Shrader, C., Shahbaz, T., & Mirabel, I. F. 1998, , 295, 623 Belloni, T., Klein-Wolt, M., Méndez, M., van der Klis, M., & van Paradijs, J., 2000 A&A, 335, 271 Belloni, T., Méndez, M., King, A. R., van der Klis, M., & van Paradijs, J. 1997, , 488, L109 Blandford, R. D. & Begelman, M. C. 1999, , 303, L1 Blandford, R. D. & Payne, D. G. 1982, , 199, 883 Böttcher, M. & Liang, E. P. 1998, , 506, 281 Brocksopp, C., Fender, R. P., Larionov, V., Lyuty, V. M., Tarasov, A. E., Pooley, G. G., Paciesas, W. S., & Roche, P. 1999, , 309, 1063 Castro-Tirado, A. J., Brandt, S., & Lund, N. 1992, , 5590 Chakrabarti, S. K. 1999, , 351, 185 Chakrabarti, S. K., & Manickam, S. G. 2000, , 531, L41 Chakrabarti, S. K. & Titarchuk, L. G. 1995, , 455, 623 Chen, X., Swank, J. H., & Taam, R. E. 1997, , 477, L41 Chen, X. & Taam, R. A. 1994, , 431, 732 Corbel, S., Fender, R. P., Tzioumis, A. K., Nowak, M., McIntyre, V., Durouchoux, P., & Sood, R. 2000, , 359, 251 Cui, W., Zhang, S. N., Focke, W., & Swank, J. H. 1997, , 484, 383 Dhawan, V, Mirabel, I. F., & Rodríguez, L. F. 2000, , 543, 373 di Matteo, T., Celotti, A., & Fabian, A. C. 1999, , 304, 809 Dove, J. B., Wilms, J., Maisack, M., & Begelman, M. C. 1997, 487, 759 Eikenberry, S. S., Matthews, K., Morgan, E. H., Remillard, R. A., & Nelson, R. W. 1998, , 494, L61 Esin, A. A., McClintock, J. E., & Narayan, R. 1997, , 489, 865 Falcke, H. & Biermann, P. L., , 342, 49 Fender, R. P. 20001, , 322, 31 Fender, R. P. et al. 1999a, , 519, L165 Fender, R. P., Garrington, S. T., McKay, D. J., Muxlow, T. W. B., Pooley, G. G., Spencer, R. E., Stirling, A. M., & Waltman, E. B. 1999b, , 304, 865 Foster, R. S., Waltman, E. B., Tavani, M., Harmon, B. A., Zhang, S. N., Paciesas, W. S., & Ghigo, F. D. 1996, , 467, L81 Greiner, J., Morgan, E. H., & Remillard, R. A. 1996, , 473, L107 Harmon, B. A., Deal, K. J., Paciesas, W. S., Zhang, S. N., Robinson, C. R., Gerard, E., Rodríguez, L. F., & Mirabel, I. F. 1997, , 477, L85 Hjellming, R. M., Gibson, D. M., & Owen, F. N. 1975, Nature, 256, 111 Jahoda, K., Swank, J. H., Giles, A. B., Stark, M. J., Strohmayer, T., Zhang, W., & Morgan, E. H. 1996, in Proc. SPIE 2808, EUV, X-Ray, and Gamma-Ray Instrumentation for Astronomy VII, ed. O. H. Siegmund & M. A. Gummin (Bellingham: SPIE), 59 Kazanas, D., Hua, X.-M., & Titarchuk, L. 1997, , 480, 735 Levine, A. M., Bradt, H., Cui, W., Jernigan, J. G., Morgan, E. H., Remillard, R., Shirey, R. E., & Smith, D. A. 1996, , 469, L33 Magdziarz, P. & Zdziarski, A. A. 1995, , 273, 837 Markwardt, C. B., Swank, J. H., Taam, R. E. 1999, , 513, L37 McCollough, M. L. et al. 1999, , 517, 951 Merloni, A., Fabian, A. C., & Ross, R. R. 2000, , 313, 193 Miller, M. C. 1995, , 441, 770 282, L17 Mirabel, I. F., Dhawan, V., Chaty, S., Rodríguez, L. F., Marti, J., Robinson, C. R., Swank, J., Geballe, T. 1998, , 300, L9 Mirabel, I. F. & Rodríguez, L. F. 1994, , 371, 46 Mirabel, I. F. & Rodríguez, L. F. 1999, , 37, 409 Molteni, D, Sponholz, H., & Chakrabarti, S. K. 1996, , 457, 805 Morgan, E. H., Remillard, R. A., & Greiner, J. 1997, , 482, 993 Muno, M. P., Morgan, E. H., & Remillard, R. A. 1999, , 527, 321 Nobili, L., Turolla, R., Zampieri, L., & Belloni, T. 2000, , 538, L137 Nowak, M. A. & Vaughan, B. A. 1996, , 280, 227 Nowak, M. A., Vaughan, B. A., Wilms, J., Dove, J. B., & Begelman, M. C. 1999a, , 510, 874 Nowak, M. A., Wilms, J., Vaughan, B. A., Dove, J. B., & Begelman, M. C. 1999b, , 515, 726 Nowak, M. A., Wilms, J., & Dove, J. B. 1999c, , 517, 355 Orosz, J. A. & Bailyn, C. D. 1997, , 477, 876 Orosz, J. A., et al. 2001, submitted, astro-ph/0103045 Pooley, G. G. & Fender, R. P. 1997 , 292, 925 Poutanen, J. & Fabian, A. C. 1999 , 306, L31 Rao, A. R., Naik, S., Vadawale, S. V., & Chakrabarti, S. K. 2000a, , 360, L25 Rao, A. R., Yadav, J. S., & Paul, B. 2000b, , 544, 443 Reig, P., Belloni, T., van der Klis, M., Méndez, M., Kylafis, N. D., & Ford, E. C. 2000, , 541, 883 Rodríguez, L. F, Gerard, E., Mirabel, I. F., Gómez, Y., & Velázquez, A. 1995, , 101, 173 Rodriguez, J. Varnière, P., Tagger, M., & Durouchoux, P. 2000, , submitted Romanova, M. M., Ustyugova, G. V., Koldoba, A. V., Chechetkin, V. M., & Lovelace, R. V. E. 1998, , 500, 703 Rothschild et al. 1998, , 496, 538 Shahbaz T., Groot P., Phillips S. N., Casares J., Charles P. A., & van Paradijs J. 2000, , 314, 747 Shibazaki, N., Elsner, R. F., Bussard, R. W., Ebisuzaki, T., & Wiesskopf, M. C. 1988, , 331, 247 Stirling, A., Spencer, R., & Garret, M. 1998, New Astronomy Reviews, 42, 657 Sunyaev, R. A. & Titarchuk, L. G. 1980, , 86, 121 Tagger, M. & Pellat, R. 1999, A&A, 349, 1003 Taylor, G. B., Carilli, C. L., & Perley, R. A. eds. 1999, Summer School on Synthesis Imaging, published as: ASP conf series, 1999, vol 180, ISBN 1-58381-005-6. Trudolyubov, S., Churazov, E., & Gilfanov, M. 1999, AstL, 25, 718 Vaughan, B. A. & Nowak, M. A. 1997, , 474, L43 Waltman, E. B., Ghigo, F. D., Johnston, K. J., Foster, R. S., Fiedler, R. L., & Spencer, J. H. 1995, , 110, 290 [llccccc]{} 10408-01-22-00 & 1996 Jul 11 02:09 & 4.15(1) & 0.063 & & & 54(2)\ 10408-01-22-01 & 1996 Jul 11 05:16 & 3.93(1) & 0.070 & & & 54(2)\ 10408-01-22-02 & 1996 Jul 11 08:38 & 3.78(1) & 0.073 & & & 54(2)\ 10408-01-23-00 & 1996 Jul 14 11:45 & 4.17(1) & 0.063 & & & 39.3(6)\ 10408-01-24-00 & 1996 Jul 16 04:04 & 3.62(1) & 0.075 & & & 66(1)\ 10408-01-25-00 & 1996 Jul 19 11:45 & 3.20(1) & 0.101 & & & 105.9(7)\ 10258-01-01-00 & 1996 Jul 23 10:33 & 3.54(1) & 0.118 & & & 65.8(7)\ 10408-01-27-00 & 1996 Jul 26 13:55 & 3.12(1) & 0.114 & & &\ 10258-01-02-00 & 1996 Jul 29 18:28 & 3.00(1) & 0.111 & & & 91.1(6)\ 10408-01-28-00 & 1996 Aug 03 12:43 & 3.32(1) & 0.105 & & & 103(2)\ 10258-01-03-00 & 1996 Aug 06 07:12 & 3.12(1) & 0.096 & & & 90.3(4)\ 10408-01-29-00 & 1996 Aug 10 08:52 & 3.59(1) & 0.090 & & & 89.4(6)\ 10258-01-04-00 & 1996 Aug 14 12:14 & 3.66(1) & 0.075 & & & 83.3(5)\ 10408-01-30-00 & 1996 Aug 18 07:26 & 5.01(1) & 0.058 & & & 40.8(6)\ 10258-01-05-00 & 1996 Aug 20 07:40 & 7.46(1) & 0.055 & & & 42.7(5)\ 10408-01-31-00 & 1996 Aug 25 04:47 & 5.15(1) & 0.065 & & & 61.1(5)\ 10258-01-06-00 & 1996 Aug 29 11:31 & 8.66(1) & 0.058 & & & 35.5(8)\ 10408-01-32-00 & 1996 Aug 31 07:55 & 8.03(1) & 0.058 & & & 34.1(4)\ 10408-01-33-00 & 1996 Sep 07 18:00 & 6.71(1) & 0.063 & & & 67.6(5)\ 10408-01-42-00 & 1996 Oct 23 03:36 & 6.37(1) & 0.079 & & & 16.6(4)\ 10408-01-43-00 & 1996 Oct 23 12:00 & 5.85(1) & 0.079 & & & 16.6(4)\ 10408-01-45-00 & 1996 Oct 29 12:00 & 4.21(1) & 0.096 & & & 6.7(6)\ 20402-01-01-00 & 1996 Nov 07 05:45 & 5.98(1) & 0.078 & & & 31.8(8)\ 20402-01-02-01 & 1996 Nov 14 02:09 & 4.76(1) & 0.090 & & & 16.5(5)\ 20402-01-02-02 & 1996 Nov 14 15:07 & 5.76(1) & 0.077 & & & 4.3(9)\ 20402-01-04-00 & 1996 Nov 28 03:21 & 4.95(1) & 0.086 & & & 21(1)\ 20402-01-05-00 & 1996 Dec 04 23:31 & 3.33(1) & 0.116 & & & 9.2(6)\ 20402-01-06-00 & 1996 Dec 11 18:43 & 3.23(1) & 0.119 & 10(4) & 7(6) & 5.3(7)\ 20402-01-07-00 & 1996 Dec 19 15:50 & 3.04(1) & 0.115 & 24(4) & 11(6) & 9.5(6)\ 20402-01-08-00 & 1996 Dec 24 22:04 & 2.96(1) & 0.102 & 16(4) & 10(6) & 5.5(5)\ 20402-01-08-01 & 1996 Dec 25 02:52 & 2.94(1) & 0.113 & 15(4) & 8(6) & 5.7(7)\ 20402-01-09-00 & 1996 Dec 31 06:43 & 2.64(1) & 0.124 & 16(4) & 5(6) & 5.5(7)\ 20402-01-10-00 & 1997 Jan 07 23:45 & 2.44(1) & 0.125 & 5(4) & 5(6) & 5.8(8)\ 20402-01-11-00 & 1997 Jan 14 01:26 & 2.29(1) & 0.124 & 12(4) & 7(6) & 5(1)\ 20402-01-12-00 & 1997 Jan 23 01:40 & 2.20(1) & 0.125 & 9(4) & 7(6) &\ 20402-01-13-00 & 1997 Jan 29 20:52 & 2.20(1) & 0.110 & 7(4) & 7(6) &\ 20402-01-14-00 & 1997 Feb 01 21:07 & 2.18(1) & 0.111 & 10(4) & 5(6) & 6.1(8)\ 20402-01-15-00 & 1997 Feb 09 18:43 & 2.20(1) & 0.144 & & & 5.8(7)\ 20402-01-16-00 & 1997 Feb 22 21:07 & 2.08(1) & 0.131 & 9(4) & 7(6) &\ 20402-01-17-00 & 1997 Feb 27 03:36 & 2.15(1) & 0.131 & 13(4) & 7(6) & 3.7(8)\ 20402-01-18-00 & 1997 Mar 05 21:21 & 2.13(1) & 0.120 & 10(4) & 8(6) &\ 20402-01-19-00 & 1997 Mar 10 01:12 & 1.98(1) & 0.146 & 13(4) & 6(6) &\ 20402-01-20-00 & 1997 Mar 17 22:04 & 1.98(1) & 0.121 & 9(4) & 5(6) & 4.6(8)\ 20402-01-21-00 & 1997 Mar 26 20:09 & 2.01(1) & 0.114 & 10(4) & 9(6) &\ 20402-01-21-01 & 1997 Mar 27 21:36 & 1.98(1) & 0.110 & 16(4) & 6(6) & 4.5(7)\ 20402-01-25-00 & 1997 Apr 19 10:47 & 2.16(1) & 0.133 & 6(4) & 7(6) &\ 20402-01-24-00 & 1997 Apr 23 03:07 & 2.33(1) & 0.132 & 9(4) & 7(6) &\ 20402-01-26-01 & 1997 Apr 30 22:33 & 2.48(1) & 0.100 & 8(4) & 10(6) & 2.0(7)\ 20402-01-26-02 & 1997 May 01 10:47 & 2.55(1) & 0.103 & 8(4) & 9(6) & 2.0(7)\ 20402-01-27-01 & 1997 May 08 16:04 & 2.77(1) & 0.097 & 15(4) & 14(6) & 22(1)\ 20402-01-27-03 & 1997 May 15 05:16 & 2.43(1) & 0.113 & 22(4) & 46(6) &\ 20187-02-02-00 & 1997 May 15 11:31 & 2.98(1) & 0.097 & 27(4) & 30(6) & 22.7(8)\ 20402-01-29-00 & 1997 May 21 11:45 & 2.93(1) & 0.107 & 13(4) & 11(6) & 30.9(8)\ 20186-03-02-05 & 1997 Sep 17 14:09 & 8.45(1) & 0.057 & 11(4) & 11(6) & 6.1(3)\ 20186-03-02-06 & 1997 Sep 18 03:07 & 5.70(1) & 0.062 & 10(4) & 10(6) & 11.2(3)\ 20402-01-47-01& 1997 Sep 19 00:00 & 6.14(1) & 0.055 & 15(4) & 14(6) & 19.3(6)\ 20187-02-03-00 & 1997 Oct 03 20:38 & 7.82(1) & 0.063 & 280(7)& 108(7)& 101(1)\ 20187-02-04-00 & 1997 Oct 05 08:23 & 7.37(1) & 0.057 & 32(4) & 13(6) & 21.4(6)\ 20187-02-05-00 & 1997 Oct 06 11:45 & 5.01(1) & 0.062 & 32(4) & 16(6) &\ 20187-02-06-00 & 1997 Oct 07 10:05 & 5.84(1) & 0.060 & 25(4) & 12(6) &\ 20402-01-49-00 & 1997 Oct 08 07:55 & 3.63(1) & 0.080 & 33(4) & 22(6) & 38.8(5)\ 20402-01-49-01 & 1997 Oct 09 09:21 & 3.43(1) & 0.082 & 44(4) & 35(6) & 36.2(5)\ 20402-01-50-00 & 1997 Oct 14 13:12 & 2.61(1) & 0.124 & 50(4) & 65(6) & 80.8(5)\ 20402-01-50-01 & 1997 Oct 16 09:36 & 2.64(1) & 0.117 & 50(4) & 74(6) & 83.5(6)\ 20402-01-51-00 & 1997 Oct 22 06:57 & 2.66(1) & 0.113 & 45(4) & 65(6) & 70.3(7)\ 20402-01-52-00 & 1997 Oct 25 06:28 & 2.73(1) & 0.108 & 46(4) & 41(6) & 63.6(5)\ 30703-01-14-00 & 1998 Apr 06 21:07 & 3.35(1) & 0.092 & 78(4) & 91(6) & 111.6(9)\ 30402-01-09-00 & 1998 Apr 09 21:07 & 3.54(1) & 0.087 & 87(4) & 85(6) & 105(1)\ 30402-01-09-01 & 1998 Apr 10 00:00 & 3.58(1) & 0.086 & 87(4) & 85(6) & 105(1)\ 30402-01-10-00 & 1998 Apr 11 09:21 & 3.46(1) & 0.089 & 100(4) & 122(7) &\ 30402-01-11-00 & 1998 Apr 20 06:14 & 5.45(1) & 0.052 & 188(5) & 80(6) & 49.3(6)\ 30703-01-15-00 & 1998 Apr 22 21:21 & 3.34(1) & 0.094 & 132(4) & 100(6) & 150(1)\ 30703-01-16-00 & 1998 Apr 28 16:19 & 3.10(1) & 0.101 & 82(4) & 111(7) & 102.5(6)\ 30703-01-17-00 & 1998 May 05 21:36 & 2.99(1) & 0.112 & 145(5) & 116(7) & 124(1)\ 30402-01-12-00 & 1998 May 11 22:33 & 3.18(1) & 0.093 & 70(4) & 87(6) & 87.3(4)\ 30402-01-12-01 & 1998 May 12 00:28 & 3.15(1) & 0.098 & 70(4) & 87(6) & 87.3(4)\ 30402-01-12-02 & 1998 May 12 02:09 & 3.17(1) & 0.099 & 70(4) & 87(6) & 87.3(4)\ 30402-01-12-03 & 1998 May 12 03:50 & 3.17(1) & 0.101 & 70(4) & 87(6) & 87.3(4)\ 30703-01-19-00 & 1998 May 19 14:23 & 2.79(1) & 0.112 & 99(4) & 95(6) & 99(1)\ 30703-01-20-00 & 1998 May 24 19:26 & 2.96(1) & 0.135 & 49(4) & 53(6) & 76.0(9)\ 30703-01-21-00 & 1998 May 31 19:26 & 3.17(1) & 0.098 & 52(4) & 46(6) & 66.3(8)\ 30402-01-14-00 & 1998 Jun 11 08:09 & 3.05(1) & 0.096 & 104(4) & 53(6) & 59.4(6)\ 30703-01-18-00 & 1998 Jun 16 18:14 & 3.79(1) & 0.087 & 91(4) & 43(6) & 36.2(5)\ 30703-01-22-00 & 1998 Jun 27 14:38 & 2.93(1) & 0.105 & 46(4) & 27(6) & 27.5(5)\ 30703-01-22-01 & 1998 Jun 27 16:33 & 2.94(1) & 0.106 & 46(4) & 27(6) & 27.5(5)\ 30182-01-01-01 & 1998 Jul 08 04:04 & 2.79(1) & 0.125 & 87(4) & 33(6) & 35.8(9)\ 30182-01-01-00 & 1998 Jul 08 05:02 & 2.80(1) & 0.116 & 87(4) & 33(6) & 35.8(9)\ 30182-01-02-01 & 1998 Jul 9 04:04 & 3.92(1) & 0.109 & 200(6) & 97(6) & 22.3(9)\ 30182-01-02-00 & 1998 Jul 09 05:02 & 4.16(1) & 0.098 & 200(6) & 97(6) & 22.3(9)\ 30182-01-03-01 & 1998 Jul 10 04:05 & 7.84(1) & 0.074 & 430(7) & 137(9) & 152(1)\ 30182-01-03-00 & 1998 Jul 10 10:05 & 7.10(1) & 0.079 & 350(7) & 139(9) & 90(1)\ 30182-01-04-00 & 1998 Jul 11 05:45 & 4.77(1) & 0.090 & 143(5) & 47(6) & 37.1(4)\ 30182-01-04-02 & 1998 Jul 12 04:04 & 3.46(1) & 0.102 & 93(4) & 38(6) & 23.9(5)\ 30182-01-04-01 & 1998 Jul 12 05:02 & 3.15(1) & 0.103 & 93(4) & 38(6) & 23.9(5)\ 30703-01-25-00 & 1998 Jul 23 18:00 & 3.56(1) & 0.103 & 28(4) & 12(6) & 7.0(7)\ 30402-01-16-00 & 1998 Aug 28 16:47 & 3.02(1) & 0.118 & 16(4) & 11(6) & 5.1(9)\ 30703-01-31-00 & 1998 Aug 31 18:28 & 3.63(1) & 0.106 & 40(4) & 28(6) & 10.3(5)\ 30402-01-17-00 & 1998 Sep 11 15:07 & 3.35(1) & 0.112 & 17(4) & 10(6) &\ 30703-01-33-00 & 1998 Sep 15 22:47 & 3.27(1) & 0.113 & & & 5(1)\ 30703-01-34-00 & 1998 Sep 22 00:14 & 2.27(1) & 0.146 & 13(4) & 15(6) & 10.8(7)\ 30703-01-35-00 & 1998 Sep 25 19:40 & 2.79(1) & 0.136 & 12(4) & 9(6) & 7.4(7)\ 30703-01-36-00 & 1998 Oct 03 17:02 & 2.94(1) & 0.125 & 28(4) & 10(6) & 7.7(7)\ 30402-01-18-00 & 1998 Oct 10 08:23 & 2.56(1) & 0.114 & 15(4) & 9(6) & 8.1(8)\ 30703-01-24-01 & 1998 Dec 18 16:47 & 2.92(1) & 0.129 & 12(4) & 13(6) & 6.6(6)\ 30703-01-24-02 & 1998 Dec 18 19:26 & 2.77(1) & 0.129 & 12(4) & 13(6) & 6.6(6)\ 30703-01-24-00 & 1998 Dec 19 03:21 & 2.81(1) & 0.135 & 12(4) & 13(6) & 6.6(6)\ 30703-01-41-00 & 1998 Dec 26 16:04 & 2.83(1) & 0.139 & 16(4) & 9(6) & 14(1)\ 40703-01-01-00 & 1999 Jan 01 01:26 & 2.86(1) & 0.136 & 37(4) & 16(6) & 15.2(7)\ 40703-01-02-00 & 1999 Jan 08 01:40 & 3.85(1) & 0.111 & & & 9.9(6)\ 40703-01-03-00 & 1999 Jan 15 23:45 & 2.80(1) & 0.139 & 21(4) & 30(6) & 18.8(6)\ 40403-01-01-00 & 1999 Jan 24 22:04 & 2.98(1) & 0.129 & 17(4) & 14(6) & 15.2(8)\ 40703-01-04-02 & 1999 Feb 06 08:09 & 2.88(1) & 0.100 & 12(4) & 14(6) &\ 40703-01-05-00 & 1999 Feb 12 01:12 & 3.71(1) & 0.101 & 68(4) & 65(6) & 47.3(8)\ [llcccc]{} 1996 Oct 29 & 01:15 – 02:56 & VLA & 8.4 & 30 – 70 & 0.26 $\times$ 0.23\ 1996 Oct 29 & 16:37 – 19:56 & Ryle & 15.5 & 6.7 & 30\ \[10pt\] 1996 Nov 14 & 23:37 – 01:21 & VLA & 8.4 & 18 – 8 & 0.27 $\times$ 0.24\ 1996 Nov 14 & 13:34 – 18:38 & Ryle & 15.5 & 4.3 & 30\ \[10pt\] 1996 Dec 28 & 20:50 – 23:00 & VLA & 8.4 & 4(1) & 0.32 $\times$ 0.26\ 1996 Dec 28 & 20:50 – 23:00 & GBI & 2.25 & 17(4) & 11\ 1996 Dec 28 & 20:50 – 23:00 & GBI & 8.3 & 14(6) & 3\ 1996 Dec 28 & 09:30 – 12:30 & Ryle & 15.2 & 6.4(4) & 30\ 1997 Jan 12 & 21:50 – 22:30 & VLA & 4.8 & 35 – 26 & 0.63 $\times$ 0.36\ 1997 Jan 12 & 21:50 – 22:30 & VLA & 8.4 & 30 – 19 & 0.41 $\times$ 0.29\ 1997 Jan 12 & 12:45 – 20:57 & GBI & 2.25 & 44(4) & 11\ 1997 Jan 12 & 12:45 – 20:57 & GBI & 8.3 & 21(6) & 3\ 1997 Jan 13 & 16:50 – 18:35 & VLA & 8.4 & 4.5(1) & 0.27 $\times$ 0.24\ 1997 Jan 13 & 16:50 – 18:35 & GBI & 2.25 & 14(4) & 11\ 1997 Jan 13 & 16:50 – 18:35 & GBI & 8.3 & 4(6) & 3\ 1997 Apr 19 & 10:40 – 15:20 & VLA & 1.4 & 3.9(1) & 4.5 $\times$ 3.5\ 1997 Apr 19 & 10:40 – 15:20 & GBI & 2.25 & 6(4) & 11\ 1997 Apr 19 & 10:40 – 15:20 & GBI & 8.3 & 6(6) & 3\ 1998 Sep 14 & 04:20 – 05:30 & VLA & 8.4 & 8.2(1) & 0.74 $\times$ 0.70\ 1998 Sep 13 & 23:07 – 04:26 & GBI & 2.25 & 18(4) & 11\ 1998 Sep 13 & 23:07 – 04:26 & GBI & 8.3 & 11(6) & 3\ 1998 Sep 13 & 21:39 – 22:39 & Ryle & 15.2 & 6.0(6) & 30\ 1998 Sep 29 & 04:00 – 06:30 & VLA & 5.0 & 12.2(1) & 2.26 $\times$ 1.2\ 1998 Sep 29 & 04:00 – 06:30 & VLA & 15.2 & 13.9(1) & 0.40 $\times$ 0.38\ 1998 Sep 28 & 21:36 – 03:36 & GBI & 2.25 & 20(4) & 11\ 1998 Sep 28 & 21:36 – 03:36 & GBI & 8.3 & 12(6) & 3\ 1998 Sep 28 & 19:07 – 21:30 & Ryle & 15.2 & 10.7(4) & 30\ 1998 Oct 08 & 03:30 – 05:30 & VLA & 5.0 & 31.7(1) & 1.6 $\times$ 1.3\ 1998 Oct 08 & 03:30 – 05:30 & VLA & 15.2 & 18.5(1) & 0.42 $\times$ 0.41\ 1998 Oct 07 & 21:22 – 03:07 & GBI & 2.25 & 60(4) & 11\ 1998 Oct 07 & 21:22 – 03:07 & GBI & 8.3 & 23(6) & 3\ 1998 Oct 07 & 19:05 – 21:19 & Ryle & 15.2 & 21.3(4) & 30\

ArXiv